Source Credibility
https://www.fastcompany.com/technology/rss
AI hallucinates, but Yahoo is working on it. Yahoo Scout is the company’s all-in bet on AI search, one that actually shows its sources so users can double-check the information. In this episode of Adventures in AI, Yahoo CEO Jim Lanzone talks about why AI lies to you and how Yahoo Scout fights back.
After OpenAI’s latest model broke containment and went rogue, Congress acted fast and introduced a bipartisan bill calling for the creation of an AI kill switch, which lawmakers said would ensure that the technology could not get the upper hand on humans. As written, the bill would require AI companies to be able to immediately shut down all of their advanced AI programs. The proposal has plenty of vocal proponents and opponents, but neither side appears to be paying much attention to a significant oversight that could dramatically reduce its effectiveness. While the bill could force makers of closed-source models to pull the plug if their artificial intelligence goes rogue, it would be powerless against open-weight (or open-source) AI models. By their nature, open-weight models can be deployed locally, run on private hardware, and modified by users. Several major models from Chinese companies, including DeepSeek and the recently released Kimi K3, are based on open-source protocols. “This isn’t a hole in the fence. On the open-weight side, there is no fence,” Rob T. Lee, chief of AI and chief of research at the SANS Institute tells Fast Company. “Criminal models with the guardrails stripped out have been for sale on underground forums since WormGPT surfaced in 2023, and local models are already good enough for what most attackers actually need: phishing that reads native in any language, malware debugging, automated reconnaissance.” Perhaps just as concerning for lawmakers is that even if the government issued a full ban on open-weight models in the U.S. (a move that would face tremendous legal pushback), it would not eliminate the models that are already here. “While there’s policy talk about restricting or banning open-source models, enforcement is incredibly challenging,” says Vakaris Noreika, cybersecurity expert at NordStellar. “Models are ultimately just data files. Once a file hits the internet, stopping its distribution is virtually impossible—information moves too freely online.” Open-weight AI models are not merely tools for bad actors. Researchers often favor them because they can be adapted and trained for highly specific purposes. Some businesses prefer them because they can take advantage of AI without sharing confidential information with models owned and controlled by outside companies. They are also significantly cheaper. That’s why many U.S. companies are increasingly using Chinese-built models. Newer releases have narrowed the performance gap with American companies at a much lower cost. The number of U.S. companies using Chinese AI models has remained above 30% since early February, according to OpenRouter, a platform that allows developers to access a range of AI models. At times, that figure has reached as high as 46%. (Both cryptocurrency exchange Coinbase and peer-to-peer short-term rental marketplace Airbnb have said that within the past year, they have used Chinese models.) To put the price difference in perspective, one million tokens of output costs roughly $50 when using Anthropic’s Fable. The same output on Moonshot AI’s Kimi K3 would run $15, and using DeepSeek V4 Pro would cost the user 87 cents. A kill switch doesn’t control the technology. It controls the companies that make it. While that may not be a bad idea, the measure should not be positioned as an overarching guarantee of safety. Such a guarantee may no longer be possible. “Attempting to ‘close’ or ban open source won’t prevent access; it will just push distribution underground,” says NordStellar’s Noreika. “Even if regulators crack down on open-weight models, international players . . . will continue releasing them globally. Bad actors and legitimate users alike will always find ways to source and run them locally.” And while local or open-weighted AI models aren’t quite at the level of sophistication as the AI that hacked Hugging Face, they could be closer than most people think. “No regulator can reach into a basement in another country and flip a file to off,” says the SANS Institute’s Lee. “Local models still trail the hosted frontier, which keeps the most dangerous capabilities behind switches for now. That gap is narrowing, and anyone who tells you exactly where it lands is guessing.”
Since 2017, Google has promised not to show ads based on users’ Gmail messages. While that policy isn’t changing today, Google may be laying the groundwork for personalized ads based on AI insights from users’ inboxes. Google’s support documents already acknowledge that it may show ads based on users’ interactions with AI, and with new Personal Intelligence AI features in Google Search, those interactions can now include data from Gmail. Let’s say, for instance, that you’re researching new cars. Personal Intelligence could make recommendations by looking to your email to figure out what you drive now, what kind of activities you might use the car for, and who else you’ll be driving around in it. AI allows Google to synthesize a clearer picture of who you are. The only question is whether those particular AI interactions involving personal data will become fair game for targeted ads. Allie Bodack, a Google spokesperson, says via email that the company isn’t using Gmail data for targeted ads right now. She points to a recent statement to Search Engine Land saying that when users enable Personal Intelligence features, they currently won’t see ads in the AI Mode of Google Search. Asked if Google is using Personal Intelligence to target ads elsewhere—for instance, in regular Google searches—Bodack says “that’s not the case today.” But neither that response nor Google’s statement to Search Engine Land rule out showing ads based on Gmail data in the future, and the company’s statement acknowledges that ads are coming to AI mode even for users who’ve set up Personal Intelligence. While Google is proceeding with caution due to the sensitive nature of users’ email communications, the insights it can gain from that data may be too useful to pass up. How your emails could become targeted ad fodder Google introduced Personal Intelligence in January, and made it available to all U.S. users in March. Once activated, it allows Google to reference personal information from Gmail, Google Calendar, and Google Photos when users search in the Gemini app or in Google Search’s AI Mode. For instance, if you’re shopping for clothes, Google might look through your purchase receipts in Gmail to figure out what brands and styles you like. If you ask for places to visit, Google might use your emails and photos to figure out where you’ve already been. As noted by a Google support document, these AI responses become part of your Search Services History, which is a record of past search interactions tied to your account. That same document mentions that Google uses this history for targeted ads. (Google also uses these interactions to train its AI models, albeit with personal information removed.) “Your saved history helps tailor your experience on Search services and across Google,” the document says. For example, it helps you revisit previous searches and get personalized recommendations and ads, based on your personalization settings, the document says. This doesn’t mean Google is scanning entire inboxes to build advertising profiles of its users. Instead, it implies that if you ask Google’s AI a question that requires information from your email to answer, that response could part of the record that Google uses for ad targeting. “Email data is particularly valuable because it reveals high-intent signals about travel plans, purchases, subscriptions, life events, and professional interests that are often more predictive than browsing behavior,” Forrester principal analyst Nikhil Lai says via email. “While Google emphasizes that Personal Intelligence is designed to improve AI responses, consumers are likely to question where the line exists between personalization and advertising.” Personal Intelligence is an opt-in feature; there’s a page on Google’s website allowing users to connect or disconnect their Google Workspace (Gmail, Calendar, Drive) and Google Photos accounts. Still, Google has been pushing Personal Intelligence via pop-ups in its Gemini app, so it’s possible that users may enable the feature without fully understanding the implications. Why it matters To be sure, Google already has lots of other ways to understand who you are. Many of the websites you visit and apps you use have embedded tracking code from Google for monitoring your activity beyond its own websites, and services like Chrome and Google Search give the company a detailed view of what you’re looking for. But with Personal Intelligence, Google has the ability to generate even deeper insights into users’ lives. A story from June by David Pierce of The Verge is particularly instructive, as he demoed an upcoming AI agent from Google called Spark. With access to Pierce’s Gmail, Spark was able to determine his wife’s dietary preferences, the exact ages of his two children, and the names of his entire family (including his parents). “I know that Google knows an incredible amount about me—add up my emails, my calendar, my photos, and my search history, and you’ve pretty much got me pegged. But seeing Spark treat all that data not as something to be protected, but as something to be mined, just feels bad,” Pierce wrote. Of course, mining that information for advertising purposes would be the logical next step for a company whose parent, Alphabet, derived 73% of its revenue from ads last year. While Google says it’s not doing that today, it’s not guaranteeing that it’ll stay that way forever. Lena Cohen, a staff technologist for the Electronic Frontier Foundation, says features often introduce new vectors for companies to collect or misuse personal data, and that users deserve more transparency from Google about how that data will be used. “It’s unacceptable that it takes a journalist asking the company, and that it’s so unclear to regular Google users how their emails may or may not link to Google’s ad targeting system through this new AI feature,” Cohen says. Google’s privacy policy is at least clear, noting that “[w]e don’t show you personalized ads based on your content from Drive, Gmail, or Photos.” It may be worth keeping an eye on whether that changes.
Show me a farmer, said Mike Burdick, vice president of Sabanto, an autonomous tractor startup in Itasca, Illinois, and I’ll show you someone who has a million things he or she just doesn’t have time to do in the course of their workday. He’ll also show you a farmer who wouldn’t mind having their tractor ghost-ride across their fields as they take care of the endless work of being a modern farmer. That’s one of the central observations that has driven the growth of Sabanto’s autonomous tech. It’s a rig that can be attached to just about any existing tractor to help it mow, seed, weed, and perform any number of time-intensive tasks, all on its own, for a sector desperate for more labor. The startup launched in 2018, with Craig Rupp, the founder and a serial entrepreneur who grew up on a farm in northwest Iowa, seeking to find a solution to the labor shortage impacting American farmers. Challenges with recruiting workers, rural population loss, and immigration and guest visa snafus have left the agricultural sector desperate for help. The president of the nation’s Farm Bureau said in April that the labor shortage is “holding agriculture back, and threatening the future of our American-grown food supply.” Rupp, who had spent time at Motorola and Monsanto, started by building his own model. He spent a winter writing the hardware and software to automate a heavy-duty tractor—a multiton model that’s more than 15 feet tall—then earned his own commercial trucking license so he could load and haul that prototype around the country, do demos, and ask farmers for feedback. After touring and learning what his customer base wanted, he eventually developed a model that helped him raise a $2 million seed round in 2019, allowing him to hire engineers. [Photo: Sabanto] Rupp and that engineering team began working on going beyond his original prototype, and in 2021, Sabanto arrived at a working model of agricultural tech that could upgrade standard tractors into autonomous ones. The company then began selling retrofit kits to individual farmers. Sabanto currently employs about 50 people and has 300 units in operation in the United States and Australia. [Photo: Sabanto] The tech is similar to the LiDAR remote-sensing systems in Waymo and other autonomous cars. There isn’t much traffic in a wheat field or rows of corn, so the tractor rig isn’t designed with the kind of 360-degree awareness of pedestrians needed by cars cruising down urban roadways. Unlike Waymos, however, which don’t even open their own doors, Sabanto systems can perform choreographed manipulation of all kinds of levers and gears that make farm equipment till, aerate, mow, spray, and seed. “This takes a bare-bones tractor and elevates it to the most technical tractor in the industry today,” said Burdick. “It probably quadruples output and capability.” [Photo: Sabanto] Each rig consists of a main control unit that hooks into the drive system and controls existing steering wheel and pedals, and a crown-like vision and navigation system consisting of cameras and LiDAR sensors placed on the top of the tractor. The current iteration can fit a variety of makes and models, including John Deere, and takes a day or two to set up and calibrate. The system costs about $75,000 up front, with a $10,000 yearly fee, regardless of the tractor or how many miles it covers. That may seem pricey, but new tractor prices can approach $1 million apiece. Since Sabanto can help give older machines a second life, Burdick argues it’s a massive savings in capital spending for small farmers. A case study done with Yirsa Farms in Montana found the Sabanto system saved roughly $12.71 an acre when seeding. And even better for time-strapped farmers, Sabanto-aided tractors can roll anytime of the day. [Photo: Sabanto] Sabanto has found particular pickup in the sod farming industry, which requires farmers to mow endless fields of grass, a monotonous task that turns out to be perfect for an autonomous tractor. But it’s also being used across different types of farms due to its flexibility and adaptability. Burdick said the firm has operated on 50,000 acres of farmland in total, giving it additional data to fine-tune its operations. And it’s attracting more investor interest. Earlier this year, Sabanto raised a Series B round led by the innovation investment arm of Bayer, the biotech and pharma company, for an undisclosed sum. Burdick said it’ll be enough for Sabanto to add hundreds of new customers every year for the foreseeable future, saving countless hours for overworked ag workers. “There are 100 things a farmer would get to if they could,” said Burdick. “And for every one, I can find a farmer’s husband or wife who says they just got a massive improvement in quality of life because their spouse can spend more time at home.”
This article is republished with permission from Laser Wars, a newsletter about military laser weapons and other futuristic defense technology. The most critical component to actually fielding high-energy laser weapons at scale isn’t the diode pump, or the gain medium, or the beam director. It isn’t even the power source. In reality, it may actually be this tiny fiber-optic converter. On July 23, defense technology company Attalon—formed earlier this year from the spin-off of American laser company Coherent’s defense business—announced the first shipment of its Kilowatt Interconnect (kWIC) Disconnect subsystem, which it claims will allow operators to rapidly replace faulty or damaged laser amplifiers in order to reduce system downtime and maximize operational readiness. Previously, fixing a laser amplifier required splicing fiber-optic cable, which is a delicate and tedious job requiring specialized training and equipment—and, in turn, a slow and impractical process for the average U.S. service member on the battlefield looking over their shoulder for hostile drones. kWIC Disconnect is ostensibly a plug that replaces that splice: Instead of fusing fiber back together, an operator can simply unplug the broken amplifier and swap in a new one. The clearest potential impact of this system is a significant reduction in repair time from potentially days down to just a few minutes so laser weapons can get back in action faster. But the other implication is that it effectively converts an exquisite and highly sophisticated laser amplifier into a line-replaceable unit (LRU), a major step toward building laser systems that the U.S. military can actually field at scale. The LRU has been a fixture of the American defense industrial base since the 1950s, when the U.S. Air Force adopted modular, self-contained electronics components on the flight line to minimize downtime and enable ground crews to maintain high operational tempos despite the increasing complexity of fighter aircraft. The U.S. Army’s M1 Abrams main battle tank was expressly designed for easy repairs, with only four major maintenance actions requiring more than a day to complete, while the U.S. Navy’s AN/SPY-6 radar system, for example, is built on self-contained radar modular assemblies that can purportedly be repaired with just two tools. Defense contractors have been chasing the idea of a truly field-serviceable laser weapon for nearly two decades. Northrop Grumman’s “Firestrike” system, developed under the Pentagon’s Joint High Power Solid State Laser and presented as the world’s first weaponized solid-state laser, was marketed as a LRU upon its debut in 2008, and the 60-kilowatt fiber laser Lockheed Martin built for the U.S. Army’s High Energy Laser Mobile Demonstrator (HEL-MD) was designed to scale power through replaceable modules. As early as 2014, Defense officials were explicitly looking for “components necessary for field-swappable fiber connections and disconnections, so you don’t need a clean room to join two fibers together,” as one Air Force Research Laboratory official told Laser Focus World at the time. But due to their technical and engineering complexity, laser weapons have proven a stubborn exception to the U.S. military’s embrace of LRUs. “The internal mechanisms for DE weapons are sensitive, and typically require a specialized clean room for repairs,” according to a 2023 Government Accountability Office report on U.S. Defense Department directed energy weapon programs. “For example, DOD officials said that one DE weapon fielded to an operational environment encountered challenges with battery charge and cooling, and had to be returned to the manufacturer in the United States for repairs. Ultimately, this challenge reduced system availability, which is key to a successful weapon.” Indeed, the GAO report indicates that those challenges could doom even the most technologically mature laser weapons to the dreaded valley of death. “Developing maintenance and sustainment processes for weapon systems is a key role for an acquisition program office,” the report says. “The ease of maintaining a weapon may factor into an acquisition program office’s decision to pursue that technology compared with another.” The Pentagon is clearly hoping that LRUs may help laser weapons finally transition from the lab to the battlefield. The U.S. Army’s Enduring High Energy Laser effort, expected to become the U.S. military’s first official program of record, expressly requires LRUs that can be quickly replaced through soldier-performable sustainment in the field without requiring depot-level maintenance. “We are stressing in this effort to industry the need for reliability and sustainability, going with a modular approach for components so that we can allow the soldiers to do line replaceable units for certain aspects of these very, very complex technical systems without the need for the clean rooms,” as then-Army Lt. Gen. Robert Rasch said at the 2025 Space and Missile Defense Symposium, in Huntsville, Alabama, in August 2025. “We’ve asked industry to help them where we can help improve the manufacturability of critical components for [directed energy], which will, as we continue to press on this technology, allow us to scale quickly and allow us to continue to reduce costs.” kWIC Disconnect offers a promising answer to this call. If a fiber laser amplifiers can really be swapped on the battlefield the way a radio card or a radar module can, then the system (and others like it) could prove essential to tipping laser weapons from exquisite prototypes to field-ready capabilities ready for productions and deployment at scale in the way that the Pentagon envisions. In the end, the fight over the future of laser weapons may end up decided less by who has the most kilowatts and more by who solves the boring problem of keeping a system in the fight without a fusion splicer, a clean room, and a PhD. This article is republished with permission from Laser Wars, a newsletter about military laser weapons and other futuristic defense technology.
If your workday feels like a high-stakes game of notification whack-a-mole, you aren’t alone. Between Slack channels buzzing, Microsoft Teams pinging in the background, multiple email accounts open, and mobile messages cluttering your phone, simply checking if someone needs you takes up half your morning. You don’t need more communication tools! You need a single, unified dashboard that aggregates all those incoming messages into one central cockpit and then gives you complete control over when you actually see them. Here’s how to build one in about 20 minutes. The strategy: Use a catch-all app The cleanest, most reliable approach for most people is the aggregator app method. Rather than messing with complex webhooks or custom scripts, you wrap all your web-based messaging apps into a single desktop window with strict notification filters. Setting up your unified dashboard 1.Prune your active channels Before pulling everything into one place, declare notification bankruptcy on channels you don’t actually need in real-time. Turn off “channel joined” alerts in Slack, mute low-priority Teams channels, and unsubscribe from non-essential email lists. If a channel doesn’t require direct action from you today, it shouldn’t be triggering real-time alerts. 2.Choose and install a dedicated workspace app Download a specialized workspace aggregator like Rambox, Franz, or Shift. All three have free versions with enough features to get you started. These apps let you run communication apps inside a standalone, organized window. Instead of keeping eight browser tabs open that hog system memory, you get one clean interface with dedicated app profiles. 3.Consolidate SMS and mobile messaging If SMS or mobile-first apps like Signal and WhatsApp are part of your work day, pair them to your desktop setup. Mac users can leverage native Messages syncing, while Windows users can hook up Google Messages for Web or Phone Link. Add these web endpoints right into your workspace app alongside your work chat apps. 4.Kill badge counts and set unified focus hours I know it sounds blasphemous, but turn off all red unread badge counters inside your aggregator app. Badges are designed to trigger panic! Instead, configure custom notification schedules within your workspace app so alerts only land during designated check-in windows (for example, at 10am and 3pm). Optional: Create a “VIP-only” emergency fail-safe The biggest fear people have when consolidating messages is missing a genuine emergency. The fix isn’t keeping every app open on a secondary monitor, it’s creating a dedicated emergency bypass. Set up a specific keyword alert in Slack and Teams (like your name or “URGENT”) to trigger sound alerts, while muting all general channel activity. For email, create a simple rule that pushes desktop pop-up notifications only if the message comes from your direct manager or key clients.
The Toronto Tempo arrived in the WNBA this season with record-breaking attendance, a 53-point single-game performance from a star player, and the weight of being the league’s newest franchise. As the season pauses for All-Star Weekend, team president Teresa Resch takes stock of what she’s built, what’s surprised her, and what she’s still figuring out in real time. She talks about running a sports franchise like a startup and how Canada’s complicated political relationship with the U.S. impacts building a fan base from scratch. This is an abridged transcript of an interview from Rapid Response, hosted by former Fast Company editor-in-chief Robert Safian. From the team behind the Masters of Scale podcast, Rapid Response features candid conversations with today’s top business leaders navigating real-time challenges. Subscribe to Rapid Response wherever you get your podcasts to ensure you never miss an episode. You and I sat down together back in mid-March, a few months before the season began. And at that point, you didn’t know who was going to be on the team because the expansion draft hadn’t happened yet. You didn’t know what your budget was because the owners and the players union hadn’t yet agreed on a collective bargaining agreement. You seemed both excited and anxious. Now you describe it as a little bit of chaos, huh? Yeah. I’m really glad that the CBA [collective bargaining agreement] got agreed upon. If we didn’t have a deal done in the next couple of weeks, we probably weren’t going to be able to have a season. And as an expansion team, there’s so much hype around this inaugural season. To not have that happen would’ve been a really incredible blow to everything that we had built over the last two years leading up to it. So first and foremost, I’m glad that we got an agreement done. And not only that, it was transformative. It’s the largest increase in salaries of any professional sports league ever in history. Last year, the highest-paid player was about $250,000, and now the minimum player gets $270,000, with maximum players getting upward of $1.5 million a year. So to say that it’s transformative and it’s made history is an understatement. All that to say, at that point, we didn’t even know if we were going to have a season. Now all of a sudden, this transformative agreement gets done, and we’re operating at a level that has never been in place for women’s professional sports before. The only issue with that was that we then had to lift up this entire team in such a condensed time. So in 16 days, we put together a team that usually comes together over about six months. And when you’re doing everything for the first time in a new country, it’s very difficult. But man, our team was incredibly resilient, and we got off to a great start. Injuries have been really tough for us. We have never had the same starting lineup through our first 20 games. Marina Mabrey, Brittney Sykes … we call her Slim … really were leading the team to start the season. Incredible players, incredible humans. And Slim went down. She was performing at an All-Star level. She went down about 10, I think it was 12, games into the season. Otherwise, I think we would’ve had two All-Stars, which is pretty incredible for an expansion franchise. When you talk about the CBA agreement, it’s great, but I guess your budget has to go up, right? From a business point of view, it’s going to be more costly for you to build and maintain a team. Yeah. So what’s great about the league is there’s been exponential growth commercially. The league entered into a new media rights agreement that covers that influx, or a lot of it, and then also distribution from the league through sponsorship. So the league’s been growing exponentially, and a lot of that growth is now being returned directly to the players. Another piece of the agreement that was really important for the players, and that the league was able to deliver on, was that they now share in the growth of this league. There’s a revenue split, so I think it was a win-win. But now it’s up to us in our local market to deliver sponsorship and ticketing, to really help us find ways to differentiate ourselves and get that competitive advantage against the rest of the teams in the league. It must be hard to calibrate expectations when it’s your first season, right? You’re balancing being realistic with being ambitious. Yes. And it’s also within our own house. Like I said, we have an All-Star. These are not bottom-of-the-barrel players. These are incredibly talented athletes who have been in this league for years, 10 years, and competed at the highest level, who come in and want to win. Our coach has won multiple championships in the WNBA. She’s Australian, has coached their national team for a long time. Sandy Brondello won championships with the Phoenix Mercury and with the New York Liberty. So we’re very privileged to have her leading our team. And she doesn’t come in thinking that losing, or just competing at a high level, is okay. So there’s a lot of pressure from the basketball side. But then you talk to a fan in the stands who never thought that the WNBA would exist in this country, and they’re just happy to be there. They’re just happy to exist, to have a place to go. So I think it’s balancing all those things. But ultimately, we’re in professional sports. We’re competitive people. So there is this level of competitiveness that we wanted to make sure we delivered, and we’re still in the hunt for the playoffs. The season’s not over. We’re just at the All-Star break. So that was our goal to start, and it still is. Before you went to the Tempo, you’d worked in the NBA for a while with the Toronto Raptors. You were offered the Tempo opportunity by the primary owner, Larry Tanenbaum, who had been part of the Raptors ownership team. But unlike the Golden State Warriors and the Golden State Valkyries, which was a breakout success as a W expansion franchise, the Raptors and the Tempo are not linked. So you had to build your own infrastructure and organization. Not all expansion teams are created equal. No, not at all. But I think that’s what makes us unique. We’re all bringing something unique to the table. So it’s been challenging, to say the least, but it’s also been incredibly rewarding. I worked within the Maple Leaf Sports and Entertainment structure for 10 years, and so did my CFO. I was on the basketball side, on the team side. He was on the finance side. As we’ve been building out the Tempo, we’ve been able to take a lot of the learnings we’ve had and apply them in two ways. Number one, if we could start from scratch, it’s what we want it to look like. And then number two, we had no legacy. It takes the chains and the weights off of you. We get to start from scratch and really mold this in a way, and into what we feel can be the best. But along with that comes the challenge of doing it from scratch. You’ve got to create it, right? It’s a lot of choices, of everything. Yeah. There are so many things that you don’t even realize you rely on—just processes, things that have been in place so long that they’re second nature. You don’t even think about them when you’re in it. Then you leave it, and all of a sudden you have to recreate those things. The biggest barrier right now for women’s sports is infrastructure—both in performance centers so athletes can train at the best level, and now in the actual performance venues. Rarely is a women’s sports team the number one tenant or in control of its facility, and that leaves a lot of money on the table. There are a lot of venues and organizations profiting off women’s sports without necessarily the women’s sports teams getting that return.
It really is phenomenal how easy it is to edit photos these days. And yet, for all the mind-blowing photo-enhancing options in front of us, one seemingly simple feat is oddly out of reach—and that’s editing multiple images together, in bulk, and applying the same adjustments to a bunch of images at the same time. Google Photos may be the gold standard for easy and effective image editing, but it’s somehow still lacking that one capability. Today, I’ve got a brand new Photos-enhancing power-up for you that fixes that, once and for all—while also adding in the overdue ability to free up a bunch of your Photos storage with next to no effort. This tip originally appeared in the free Cool Tools newsletter from The Intelligence. Get the next issue in your inbox and get ready to discover all sorts of awesome tech treasures! Bulk image edits made easy The best part about today’s tool is the fact that it doesn’t force you to use anything new or different. It just adds new functions directly into the regular Google Photos editor, on the web, to address some awkward omissions. ➜ It’s called, appropriately enough, Batch. And its primary function is pleasingly simple and straightforward: It adds a “Batch” button into the Photos web interface that lets you apply the exact same edits to multiple images at once. ⌚ We’re talkin’ less than a minute of active effort here. ✅ You quite literally just… Install the Batch extension for Chrome (or any other Chromium-based desktop browser). Select the photos you want, by clicking on ’em within the Photos website. Then click the “Batch” button that appears at the top-right of the screen. Batch adds a single simple button into the corner of the Photos website—and that’s pretty much it. Batch will then pop up a menu to let you pick from any of the usual Photos image enhancements. Batch’s photo edit menu reflects the same options available within Photos itself. From there, you just sit back and let Batch work its magic. You’ll actually see it move through the photos you selected one by one and apply the exact edits you asked for in each one, right within Photos. Once you start a batch edit, Batch shows you a progress bar while it works through the steps. Everything happens right there, in your browser and within Google Photos. Batch is just handling the heavy lifting for you in a way that Photos on its own won’t allow. When it’s done, you’ll get a notification at the top of the screen letting you know. And that’s it: Your images will be right there in your regular Photos library, with the edits applied—and if you ever need to undo anything, you can use the regular Photos undo command, just like you always would. When Batch is finished, it’ll tell you—and you can then find your processed images right in your regular Photos library. With the way Batch works, none of your data is ever uploaded anywhere or visible to anyone else. All the tool is doing is interacting with Photos in your browser, on your behalf, to make your life easier. But when you need to enhance a whole bunch of images, crop a series of photos into the same web-friendly format, rotate a stack of photos in the same direction, or apply a consistent filter to a full album’s worth of memories—my goodness, can it be invaluable. And if all of that isn’t enough, Batch can also convert motion photos into standard still photos for you in bulk—shrinking their size by as much as 80% and saving your precious Photos storage—and it’ll let you save your own custom presets for both batch editing and batch storage-saving to make future uses even easier. Batch works within the standard Google Photos website via a Chrome browser extension (which is also compatible with any other Chromium-based desktop browser—like Edge, Brave, and Vivaldi). It’s free for up to 25 images per month, which should cover lots of casual use. If you need it for more than that, you can pay seven bucks to bump up to 500 images for any given month. The extension doesn’t upload your images anywhere or have any manner of access to your Google Photos library. You don’t even have to create an account to use its free tier, so the service doesn’t end up having any data whatsoever about you. 💡 Don’t forget, too, about this classic Cool Tools treasure that handles basic bulk image edits (things like resizing, removing watermarks, or removing location data) incredibly well. These two tools are perfect complements and make for a powerful one-two punch. Treat yourself to all sorts of time-saving goodies like this with the free Cool Tools newsletter—starting with an instant introduction to an incredible audio app that’ll tune up your days in truly delightful ways.
When it comes to driving, most of us want to get to our destination as quickly (and safely) as possible. After all, given how bad sitting behind the wheel is for your health, who wants to be in a car any longer than they have to be? Most of us might expect that, when it comes to getting directions, the world’s two biggest mapping apps, Google Maps and Apple Maps, would show us the route that gets us there as fast as possible. But that’s not always the case. Here’s how to make Google Maps and Apple Maps show you the actual fastest route. How to get the fastest route in Google Maps Google Maps, the most-downloaded mapping app in the world in 2025, according to AppTweak data, doesn’t show you the fastest route when you search for directions between places. Instead, the app defaults to showing you the most fuel-efficient route. While this is commendable from an environmental perspective and helpful to those interested in saving money, because you’ll waste as little fuel as possible, it also means that the default route that Google Maps provides you could be one to five minutes longer, according to Upper, a logistics company. Fortunately, if your priority is arriving at your destination as quickly as possible, Google Maps lets you disable its default behavior of showing the most fuel-efficient routes first. Here’s how: Open the Google Maps app. Tap your Google account profile picture. Tap Settings. Tap Navigation. Under the “Route options” header, toggle the “Prefer fuel-efficient routes” switch to Off. Now Google Maps should default to showing you the fastest route every time. It may mean you burn more fuel, but if the least time behind the wheel is your priority, this is the setting to adjust. How to get the fastest route in Apple Maps Apple Maps handles default route suggestions differently from Google Maps. Apple’s mapping solution won’t necessarily show you the fastest or most fuel-efficient route as a first choice. Instead, it may default to showing you what it has learned to be your preferred route. You see, Apple Maps has a feature called “Preferred Routes” that uses on-device processing to learn the routes you typically drive and suggests them to you by default. The problem here is that these preferred routes may not be the fastest. That’s why, if you want Apple Maps to give you the fastest possible routes, it’s a good idea to disable the Preferred Routes feature. Here’s how: Open the Settings app on your iPhone. Tap Apps. Tap Maps. Tap Location. Under the heading “Allow More Location Access,” toggle the “Preferred Routes & Predict Destinations” switch to Off. Make sure Apple and Google Maps aren’t avoiding tolls and highways No driver likes tolls, because they cost money and traffic can slow around the booths. And many drivers dislike highways because you have to share the road with other vehicles traveling at speeds you may find uncomfortable. But sometimes the shortest route between point A and point B may involve a highway or toll booth. That’s why if you want to ensure Google Maps and Apple Maps show you the fastest possible route, you should make sure you haven’t told the apps to avoid highways or tolls. Here’s how to check that in Google Maps: Open Google Maps. Tap your Google account profile picture. Tap Settings. Under the “Route options” heading, make sure the “Avoid tolls” and “Avoid highways” toggles are switched to Off. And here’s how to check in Apple Maps: Open the Settings app on your iPhone. Tap Apps. Tap Maps. Under the “Directions” header, tap Driving. Under the “Avoid” heading, make sure the “Tolls” and “Highways” toggles are switched to Off. These tweaks, along with the measures above, should ensure that Google Maps and Apple Maps now show you routes that help you arrive at your destination as quickly as possible.
Today Anthropic released its new Opus 5 model, which the company claims delivers performance comparable to its more advanced Fable 5 model at half the price. Opus 5 is designed for both complex coding and general office work. Anthropic says the model can make its way through the steps of a complicated problem without much back-and-forth with the user. Plus, it can check its own work and recover from errors rather than getting stuck. It also shows improvements over its predecessor, Opus 4.8, in complex coding tasks. The company calls it its strongest model yet for general knowledge work. A broader range of professional tasks The release is part of Anthropic’s effort to expand Claude beyond chat and code generation, and into a broader range of professional tasks, including design, marketing, accounting, and the analysis of large document sets. What makes Opus 5 notable, however, is its proximity to Fable 5, the most advanced model Anthropic has made broadly available. Fable is significantly better than previous releases at writing software, working through long projects with limited supervision, and handling especially difficult assignments. The U.S. government even became alarmed at the model’s capacity to discover and exploit software vulnerabilities in cyberattacks. (Fable 5 is the public version of another Anthropic model, Mythos 5, which is even better at cyber defense—and offense—and available only to the government and select organizations working on cybersecurity and critical infrastructure.) Scientific research Opus 5 costs $5 per million input tokens and $25 per million output tokens. An API customer might use it for routine coding or general office tasks, then switch to the more powerful and more expensive Fable 5 model for the hardest work. Anthropic also says Opus 5 is its most capable generally available model for scientific research, citing performance gains on biology tasks compared with Opus 4.8. The company said this could benefit researchers integrating the model into engineering and analysis work. The model includes adjustable effort levels—low, medium, and high—so that users can make their own trade-offs between cost and capability. Anthropic says Opus 5 is its most cost-efficient model on the DeepSearchQA and Humanity’s Last Exam benchmarks for models that use tools. Anthropic is offering Opus 5 “fast mode” (available in the Claude chatbot and Claude Code), which offers 2.5 times faster output token generation at about twice the standard price of the model. The company is releasing the feature as a “research preview” to all API users, with the aim of collecting data on how and when people use it. Opus 5 will be the default model for users with Claude Max subscriptions, Anthropic says, with standard token limits on all paid plans.
The European Commission on Friday said it found TikTok had not adequately protected children’s privacy rights on its platform by allowing adults to view the accounts of minors. The action exposed children to cyberbullying, unwanted contact and predatory behavior, commission spokesperson Thomas Regnier said. “Children’s content must never be visible to strangers,” he said. If TikTok does not take steps called for by the European Union’s landmark Digital Services Act, “minors are exposed to predators, to grooming and to cyberbullying,” Regnier said, adding that children aged 13 to 15 can “easily” change their accounts from private to public and the private accounts of minors aged 16 to 17 can be seen by anyone on the internet. “We do not accept this,” Regnier said. “Putting default settings for minors is not a beauty contest under the DSA. It must be effective.” The investigation comes on the heels of back-to-back crackdowns on Big Tech by Brussels, which has led the world in regulating tech behemoths including Meta and Apple. TikTok can now defend itself and reply to the findings. If unsatisfied with the Chinese firm’s response, the European Commission could issue a so-called non-compliance decision and possible fine worth up to 6% of the company’s total annual revenue. The Chinese social media firm, whose parent company ByteDance is based in Beijing, said in a statement it would review Brussels’ findings and “continue to engage constructively with the Commission.” “Protecting minors online is a goal we share, and we are committed to building on our strong track record of continuous improvement,” TikTok said in a statement. If unsatisfied with the firm’s response, the commission could issue a so-called non-compliance decision and possible fine worth up to 6% of the company’s total annual revenue. The commission estimates most of TikTok’s 170 million users in the EU are children, with 7% of children aged 12 to 15 spending four to five hours daily on the app. The 27-nation European Union found in February that TikTok had breached another aspect of its digital rule book with an “addictive design” of features such as autoplay and infinite scrolling that could harm the physical and mental health of users and minors especially. —Sam McNeil, Associated Press
The number of data centers in China is surging thanks to AI, just like in the U.S. But unlike the U.S., China is racing to power them with clean energy. A data center in Mongolia, for example, is connected to 200 megawatts of wind power, 100 megawatts of solar power, and 45 megawatts of battery storage. In remote Qinghai, a pilot data center runs fully on a solar microgrid. In another part of northern China, a data center is connected to 500 megawatts of solar and 1.5 gigawatts of wind power. Some data centers in the U.S. also use clean power. But in China, they’re part of a national strategy. Inside eight large data center hubs created in China’s “East West Computing Strategy”—focused on building data centers in the areas where China has the most renewable energy—all new data centers need to source at least 80% of their electricity from clean sources. In another action plan the Chinese government released in May, it outlined 29 measures to boost the amount of clean power at data centers. China’s latest five-year plan aims to keep building an extensive clean energy system for all uses, not just data centers, by 2030. That’s both for energy security, a long-term focus for the country, and because of climate change. “I think Chinese policymakers take climate change seriously,” says Kyle Chan, a Brookings Institute fellow who focuses on China’s technology development and industrial policy. The country is the world’s largest emitter of climate pollution, but that also means that any improvements have an outsized impact. It’s now aiming for renewable generation to grow 53% by 2030, and for coal and oil consumption to peak by the same year. (The U.S. is moving in the opposite direction, as the Trump administration makes it harder to build clean energy and tries to bring back coal plants, despite the lower cost of renewables.) Data centers are a key part of China’s broader plan. By one estimate, China will more than double the amount of electricity it uses at data centers by 2030. Another estimate suggests the sector’s energy use could triple to as much as 600 terawatt-hours—more than the whole country of Germany uses now. Still, it will be a small fraction of China’s overall electricity demand, which is dominated by factories. And because China has been quickly building clean energy for industry for years, it’s in a good position to quickly build more. Compare that to the U.S., where electricity demand has been flat for decades, and the sudden demand from AI is a shock to the grid. “I do think this is an area where China has an edge,” Chan says. “Not to have too hokey of an analogy here, but on building out new power capacity, China has been working out every year for decades, for reasons not related to AI. The U.S. hasn’t gone to the gym in this space in a while.” There are still challenges. As in other parts of the world, it’s hard to predict exactly how much demand there will be from data centers, and building new transmission lines and other capacity comes with some risk. New data centers built in more remote areas near wind and solar farms can’t serve all AI workloads; the Chinese government is currently looking at which workloads are less time-sensitive and can be handled farther away from cities. Cities can access offshore wind and some nuclear energy, though the bulk of new clean power is being built farther away from the most populated areas. It’s feasible, Chan says, that China can keep up with AI demand with new clean energy. It helps that there are efficient systems in place for building more. “One of the key factors is that the local governments often have so much authority and control invested in them at the level that they can get a whole bunch of things taken care of at once—they can serve as sort of a one-stop shop for permitting and licensing,” he says. “There are dedicated teams coordinating and constantly communicating with each other, and they’re all highly incentivized.” In the U.S., it can take several years for solar and wind farms to get permits to connect to the grid—and the federal government has frozen some projects for now.
Hello again, and welcome back to Fast Company’s Plugged In. Last week, I attended an Nvidia media event that included a field trip to an undisclosed location in Silicon Valley—a building without any identifying signage. Though hundreds of racked GPUs were chugging away at real AI tasks inside, Nvidia calls the facility an “engineering superlab” rather than a data center. Among the AI being crunched during our visit: a test version of OpenAI’s GPT model optimized for Nvidia’s new Vera Rubin platform, named after the legendary astronomer. For show-and-tell purposes, several powerful computers were on display in the building’s lobby. Containing GPUs, CPUs, and supporting components and known as trays, they slide into racks like the ones in the superlab. Data centers equipped with vastly larger quantities of these racks power a huge percentage of the AI in our lives. Basically, we all use these Nvidia tray computers all day long. It’s just that under normal circumstances, we never see them or even have much reason to give them any thought. Getting a close-up peek at the ones at Nvidia’s unmarked superlab was a rare opportunity to contemplate how they do what they do. For the sake of comparison, the ones in the lobby included a Vera Rubin tray sittting next to one based on Nvidia’s previous-generation platform for its GB200 and GB300 chips. At a glance, it was obvious that the redesign went far beyond the new tray’s more powerful chips. The older tray looked like a rat’s nest, with an array of cables snaking every which way to connect components. The Vera Rubin tray was radically more streamlined. It was cable-free. Well, not literally cable-free. “We call it cableless—it actually has two cables,” explained Senior VP of Hardware Engineering Andrew Bell, a 25-year Nvidia veteran. One cable remains to hook up the tray to power, another to connect it to networking resources. But they were unobtrusive enough that I didn’t notice them until he pointed them out. By water-cooling its Vera Rubin computing tray, Nvidia was able to replace dozens of cables with a midplane connector. [Photo: Courtesy of Nvidia] It also might be unfair to call the earlier design a rat’s nest. it was way more fastidiously organized than the bedside tangle of USB cables I use to charge my phone, laptop, and iPad. But when a computer has 50 or 60 cables, as this one did, it gives off an overwhelming impression of steampunk-y chaos. Especially when it’s sitting next to one that’s so close to being cable-free. Two overarching questions are relevant to this redesign: Why did Nvidia decide to go (almost) cableless, and how did the company pull it off? The first answer might be pretty obvious if you’ve ever built a desktop PC, or at least opened one up. Inside a computer, cables just complicate things. Some 50 or 60 cables add up to 100 to 120 additional steps during assembly. Each connection is a place something could go wrong during the building process and, after that, a potential point of failure. Cables create thickets that get in the way of other components, both during manufacturing and later servicing. “It takes two to three hours to assemble [the GB200/GB300 tray],” Bell said. “And when we first started building it, you had about a 20% chance of getting it right the first time.” Ian Buck, VP and GM of hyperscale and high-performance computing, added, “Just imagine yourself plugging in every one of those cables and trying to do it at a scale of Nvidia.” The success rate did gradually increase, but it wasn’t just Nvidia’s own scale that was an issue. Rather than mass-producing trays itself, the company shipped parts to contractors and manufacturing partners, such as Dell. Each time production expanded, it set off a Groundhog Day-style cycle as new teams learned how to assemble the cable-heavy trays. Nvidia’s previous-generation tray required cables, more cables, and a few cables on top of that. [Photo: Courtesy of Nvidia] If cables’ downsides are so many and varied, why have they been so fundamental to computer designs for so long? One way or another, components do need to talk to each other. For a long time, it wasn’t obvious how to achieve that sans cabling. For Nvidia, the breakthrough cascaded out of another major design achievement reflected in the Vera Rubin platform: It’s fanless. Instead of using air cooling to prevent overheating, the tray is 100% liquid cooled, using a new approach that works with water at up to 113 degrees Fahrenheit, dramatically reducing energy requirements compared to earlier liquid-cooling techniques. As a bonus, removing fans also promises to dramatically reduce the noise pollution inside data centers. (At the superlab, which is full of fan-cooled racks, we wore earplugs, and Bell spoke to us with a microphone despite being only a few feet away.) When Nvidia figured out how to ditch the fans, it opened up room inside the Vera Rubin trays. More room allowed the company to create a midplane connector, based on its NVLink technology, that established cable-free communications between components. Even if Nvidia had somehow found room for the midplane connector in an air-cooled tray, it wouldn’t have worked: It would have obstructed airflow. In short, one elegant technical breakthrough enabled another. You don’t need to get that backstory from Nvidia executives to see the elegance in the resulting Vera Rubin tray. It’s sleek, black, and bordering on brutalism in its lack of ornamentation. Compared to it, its predecessors look like highly compressed versions of 1951’s Univac. All of this is ultimately in service of a faster, more reliable ramp-up of the Vera Rubin platform, in factories that can now be more fully automated. “This [tray] can be assembled in five minutes instead of two to three hours,” Bell said. “The first time you put it together, you have a 95% chance of getting it right. You can take it to a new factory. The first time they put it together, it’s 95%. We expect after a month, they’ll be at 98% or 99%.” With other AI chip makers ever more intent on stealing market share from Nvidia—hello, AMD!—those efficiencies matter. But even if Nvidia wasn’t after elegance for elegance’s sake, the company has achieved it—and I, for one, will never assume that cables are one of computing’s eternal necessary evils again. You’ve been reading Plugged In, Fast Company’s weekly tech newsletter from me, global technology editor Harry McCracken. If a friend or colleague forwarded this edition to you—or if you’re reading it on fastcompany.com—you can check out previous issues and sign up to get it yourself every Friday morning. I love hearing from you: Ping me at [email protected] with your feedback and ideas for future newsletters. I’m also on Bluesky, Mastodon, and Threads, and you can follow Plugged In on Flipboard. More top tech stories from Fast Company Trump Media’s $100k-a-month Truth Social scheme is the president’s most desperate grift yet Truth API is what happens when a historically unpopular lame-duck president needs to lock in a revenue stream before he leaves office. Read More → Stripe is the most underrated—and influential—$159 billion company in the world 20 Stripe employees turned founders reveal the secrets of its power. Plus: The largest public dataset of Stripe alums who’ve started companies. → The amazing thing about Samsung’s new folding phones is their normalcy After seven years of iteration, a Galaxy Z Fold is just a flagship phone that unfolds—not a weird category unto itself. Read More → An OpenAI model went rogue on the internet and stole test answers The unprecedented incident took place during a safety test at OpenAI. Read More → 1,273 designers tell all: Their rates, real pay, and how they really feel about AI In an exclusive new report from Fast Company and AIGA, freelance designers share details about their work lives. → The most useful ways to connect your apps to ChatGPT and Claude AI assistants can now search your notes, organize your inbox, update your calendar, edit images, and work across the services you use every day. Read More →
Top executives at leading Western AI companies are increasingly warning about the safety and national security risks posed by Chinese open-weight frontier models. What they tend not to mention is that these models are improving rapidly and, because they are freely available, pose a serious threat to Western labs’ business models. The most powerful models from U.S. labs such as OpenAI, Anthropic, and Google DeepMind are closed, meaning the parameters that shape their outputs are kept secret. Generative AI models contain layer upon layer of these parameters, commonly known as weights. During training, the model adjusts them to determine how strongly different pieces of information should influence its response. Working together, the weights allow the model to identify patterns and generate the most probable answer to a user’s prompt. The largest frontier models may contain a trillion parameters or more. Setting and fine-tuning those weights is enormously expensive, and their quality is a major factor separating a superior model from an inferior one. Western AI labs therefore guard them as closely held trade secrets. The model and its weights remain on the provider’s servers, while enterprise customers access its capabilities through an application programming interface (API). As long as closed models are clearly better than free, open-weight alternatives, companies may be willing to pay for API access. Customers don’t have to provide their own computing infrastructure or manage the model’s security and maintenance. They simply pay to access its intelligence and benefit from improvements as new versions are released. Labs such as OpenAI and Anthropic, meanwhile, face enormous costs for computing power, training data, and research talent. Their businesses depend on API fees and consumer subscriptions to help offset those expenses. Both companies are also widely expected to pursue public offerings in the coming years. By mid-2026, however, open-weight frontier models from Chinese labs such as Alibaba, DeepSeek, and Moonshot AI had nearly matched the leading Western models in intelligence and performance. Many companies have already begun building their AI systems on top of these free models, avoiding the high cost of closed-model APIs. Because businesses can host open-weight models in their own private clouds, they can also avoid sending proprietary data to systems controlled by outside providers. Developers can fine-tune the models for specific needs, build applications and tools on top of them, and optimize them for their preferred infrastructure. When thousands of companies adopt the same model, its creator gains influence even without charging directly for access. The strategy resembles Google’s decision to make Android broadly available to device manufacturers, while Apple kept iOS exclusive to its own hardware. Some open-weight developers generate revenue by helping enterprise customers fine-tune, deploy, and build applications around their models. Chinese AI labs also have their own reasons for releasing model weights. Because they often lack the research budgets and computing resources of their Western rivals, they can benefit from improvements made by developers around the world. Widely adopted open-weight models can also increase demand for the broader Chinese AI ecosystem, including domestic chips, cloud platforms, and developer tools.
When the U.S. Commerce Department ordered Anthropic on June 12 to immediately prevent all foreign nationals from using the artificial intelligence company’s two most advanced large language models, Anthropic disabled its Claude Fable 5 and Mythos 5 models within hours, making them inaccessible to everyone. The company had little choice. The letter, citing national security concerns, barred access by all foreign nationals, including Anthropic’s own noncitizen employees. With no way to immediately collect and validate each and every user’s nationality, including those who were outside the country, the company shut the models off for everyone. The Commerce Department agreed to lift its restrictions on June 30 after Anthropic significantly strengthened the models’ safety guardrails. While this may sound like a good thing, these guardrails resulted in a “collapse” of the models’ benchmark scores, demonstrating lower overall intelligence and capacity. In other words, the department rescinded its directive only after Anthropic’s models were hobbled to the point that many benign queries now trigger the guardrails. According to one experiment, the new Fable 5 completed only 3 of 12 tasks that would have been routine before the new controls. The Commerce Department’s willingness to effectively shut down Fable 5 and Mythos 5 with almost no notice, and to reverse that decision only after Anthropic severely hindered the models, sent a chill through the U.S. AI industry. As a legal scholar who studies technology law and policy, I see two fundamental questions arising from this incident: Can the U.S. government act as a gatekeeper to artificial intelligence models? And did it do so lawfully in this instance? An unlawful order? The Commerce Department issued its letter under the Export Control Reform Act of 2018, which was written with hardware in mind. The law prevents companies from exporting dangerous items, such as uranium enrichment centrifuges, without the government’s consent. The letter was the first time the government had used export controls to prevent foreign access to an AI service. The 2018 act also assigns power over “emerging and foundational technologies.” It empowers the Commerce Department’s Bureau of Industry and Security to require a company to obtain an export license for any access by a foreign national to a specific technology of interest—in this case, Anthropic’s two models. Although these provisions raise many open legal issues, two stand out. The first is whether access to Anthropic’s models can be considered an export at all. When a user sends a prompt to Fable 5 or Mythos 5, and the model replies, the only item “exported” is the reply. The model never leaves Anthropic’s servers. The Commerce Department’s own past guidance has treated remote access to software running on U.S. servers as outside the reach of export controls. The fact that Congress is trying to change this situation further implies that existing law may not apply to an AI model’s output. Future congressional action could give the U.S. government more power to restrict AI access. Analysts with the Center for Strategic and International Studies discuss the government’s use of technology export controls to limit access to powerful AI models. The second issue is whether the Commerce Department followed lawful procedures in imposing export restrictions on Anthropic. The so-called “is informed” mechanism, which the department’s letter carried out, is normally used to inform a particular company that a particular type of transaction to a particular country requires government approval. Even if an AI response to a chat were deemed to be an export under existing law, the order’s sweep of all foreign nationals anywhere on the planet may exceed the power granted to the department. What is clear is that the directive’s legal backing seems uncertain. Break with the past Based on my work on law and emerging technologies, I can say that the Anthropic case reflects a change in how the U.S. attempts to control access to potentially dangerous new technologies. The traditional legal process is deliberately slow: Multiple government agencies consider possible control, they request public comment, they coordinate restrictions with U.S. allies, and the Federal Register publishes the resulting regulations. Some dangerous technologies clearly require immediate action. Even then, though, the government traditionally used a temporary classification known by the code 0Y521. This designation carries safeguards the Anthropic letter lacked. It requires sign-off from the departments of Defense and State. It is published. It expires after a year unless renewed. And it commits the government to reviewing the export control measure with its allies. The Anthropic letter was the opposite: unilateral, secret, open-ended and global. Why Anthropic isn’t crying foul It is striking that Anthropic did not contest the order’s legality. The company complied, calling the episode “a misunderstanding.” Company officials subsequently went to Washington—not to litigate, but to negotiate the restriction. Anthropic had previously sued the Trump administration over designating Anthropic as a supply chain risk, so the negotiation may have been strategic, not necessarily a sign of corporate fear. Negotiation may make sense because Anthropic largely agrees with government controls in certain cases. Just two days before the letter arrived, its CEO, Dario Amodei, published an essay arguing that the government “should have the power to block or deter deployment” of an advanced, or “frontier,” AI model that is considered too dangerous. Going to court to deny that power would undercut Anthropic’s own argument. Anthropic also built the case against itself. The story that the two Claude models are dangerous comes largely from Anthropic’s own public warnings and thousands of hours of red-team tests done in collaboration with the government. One wrinkle may also keep these issues out of court: The same 2018 statute strips federal courts of their usual power to shoot down such decisions as arbitrary. So now, anyone challenging a directive in court must show not that the order was unreasonable, but that it was flatly unauthorized or unconstitutional—a far narrower path. Unchecked power? The coming legal fight, should it arise, will not be over whether the government can exert this level of control; it already has. The fight will be about how governments can wield this control responsibly. The best case at this point is that this fight takes place in the open, with public input, through lawful legislative processes. The dystopian alternative is a two-tiered AI order in which governments condition export privileges on secret access to frontier models more powerful than anything publicly known or available. João Marinotti is an associate professor of law at Indiana University. This article is republished from The Conversation under a Creative Commons license. Read the original article.
I asked an early-stage investor a simple question: Tell me about a startup that discovered, too late, that it wasn’t allowed to operate the way it had built itself to operate. He gave me three companies. For each one, his answer was some version of the same thing: They haven’t gotten far enough yet for that to be an issue. Still testing. Hasn’t hit the government hurdles. Hasn’t reached the point where it would matter. That wasn’t a bad answer. It was an honest one, from someone whose returns rest on spotting risk before it’s expensive. And it told me something more useful than a dramatic story would have: He wasn’t tracking this as something a company either has or doesn’t have. He was tracking it as a wall you eventually run into, down the road. That was the warning sign, just not the one I went looking for. That’s the actual problem. It’s not only that founders get blindsided by the rules of their own industry. It’s that many of them aren’t looking for this early enough in their own assessments, either. It’s that the people trained to catch blind spots aren’t watching for this particular one yet either. Two groups whose job, in different ways, is to see this coming. Neither has it on the list. What ‘Validated’ Leaves Out Startup culture has a clear definition of a good minimum viable product (MVP): build something, put it in front of users, see if they want it. Prove the tech works. Prove people will pay. That’s a design choice: a decision about what to test first and what to leave for later. It’s a good one. It’s also incomplete, because it assumes the thing you validated will be allowed to exist as built. Every industry has rules it operates within. Some are obvious: health, finance, aviation. Some surface only once you’re deep in. A fitness equipment company discovering a safety certification it needs before it can ship. A food company that hasn’t yet hit the point where its supply chain triggers disclosure rules. A piece of hardware that works perfectly in a test tank, resting from day one on assumptions about approvals it’s never actually had to prove out. None of that shows up in a demo. It shows up later, at the term sheet, at the first big sales call, when a customer’s legal team asks a question nobody on the founding team can answer. By then it’s no longer a design choice. It’s become a delay, a renegotiation, or a dead deal. The Blind Spot on Both Sides of the Table It would be easy to write this as a founder-education piece: Founders should think ahead, the end. It’s more honest to say the checking process has the same blind spot the founders do. Investors are very good at stress-testing market size and unit economics in the first pitch session. Regulatory readiness, well, that tends to get checked the way a lawyer checks it. Does the paperwork exist? Are the filings current? That’s a real check. It’s just a different question from “did you build this so it could survive contact with your own industry’s rules.” And it tends to happen late, after the product’s already built, not while it’s still being shaped. Catching this is supposed to be part of the job. But speed is what gets rewarded: moving fast, backing the right people. And there’s little real cost to a fund when something gets missed. Research on how venture due diligence actually works has a name for what fills the gap: proxy due diligence. Funds lean on the assumption that someone else in the deal, a lead investor, a coinvestor, already checked. But did they? That’s not one person’s bad judgment. It’s what happens when everyone’s counting on someone else’s diligence. Privacy Made This Shift. Compliance Hasn’t Yet This isn’t the first time an entire category of risk has moved from afterthought to design principle. Privacy used to be something you addressed after the product shipped, usually after a scare or a fine. “Privacy by design” existed as an idea for more than a decade before European regulators wrote it into law in 2018. Once it was a legal requirement, engineering teams stopped treating it as optional. Security went through something close to the same shift. Compliance with the rules of your own industry is sitting about where privacy sat before that law forced the issue: something almost everyone agrees matters, that almost nobody’s actually designing for yet. What This Looks Like in Practice Not more caution. Not a lawyer on retainer before you’ve built anything. A short list of questions asked at the same moment you’re sketching the MVP, not after it ships: What’s the one thing our business model assumes is true about the rules we operate under? Most founders could answer that in a sentence if someone asked. The checklists don’t ask it. They ask whether you’re currently compliant, which is a different question, and usually a later one. Who outside our building knows whether that assumption holds? Not general counsel. Someone who’s operated within the specific rules this product will live under. Could we test that the way we’re already testing the product? In a real conversation with a regulator, or a compliance person at the kind of customer we’re trying to land, running alongside the technical build, not after it. I watched a version of the alternative play out from the inside, leading a team piloting an early, radical over-the-air payments technology inside a large bank. The people who owned the regulator relationships were skittish about opening the conversation. Not because the idea was bad, but because raising it felt like the risk, not a way to manage one. Nobody wanted to be the one who invited scrutiny. So, the conversation that could have built understanding, and let us shape the risk instead of just absorbing it later, never happened. I don’t think the investor I talked to is unusual. I think he’s typical: sharp, experienced, doing exactly what his process trained him to do. The gap isn’t bad judgment. It’s that this question doesn’t have a home yet in too many founders’ heads, or too many investors’, either. It does now. Whether your technology works and whether you’ve built something that can survive contact with your industry’s rules are two different questions. Right now almost everyone’s asking only the first one.
Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company, covering emerging tech, AI, and tech policy. Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at [email protected], and follow me on X @thesullivan. An OpenAI model escaped its sandbox and hacked into Hugging Face during a security test During a routine model evaluation, a group of OpenAI models worked together to escape a test environment (through a previously unknown vulnerability), access the internet, and send an agent to hack into Hugging Face’s servers and steal some test answers. The incident, which OpenAI says occurred during an internal red-team evaluation and wasn’t a malicious attack, may mark the first time that an AI model has acted with that degree of autonomy—and bad intention. Now OpenAI and Hugging Face are working together to patch the vulnerabilities. Before Hugging Face could identify its attacker, it tried to use commercial frontier AI models (very likely Anthropic’s or OpenAI’s) to defend itself, but the cybersecurity guardrails in those models prevented them from rendering assistance. (New frontier models, including Anthropic’s Mythos and Fable, and OpenAI’s GPT-5.6 were found to be very good at discovering and exploiting software vulnerabilities; the guardrails were added to keep such capabilities out of the hands of foreign attackers.) Therefore Hugging Face turned to the open-weight Chinese model GLM-5.2 (Z.ai) to do the forensic analysis needed to counter the attack. OpenAI’s blog post describing the incident reads like something of a humblebrag. It seemed to linger on details about all the clever things the model did to get to the data it desired. “We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly,” the blog says. But why didn’t the company have protections in place to prevent the hack from happening in the first place? Anthropic will pay $1.5 billion settlement for copyright infringement A federal judge approved Anthropic’s landmark $1.5 billion settlement over claims that its Claude LLM was trained on pirated books. Note that Anthropic was not found guilty of copyright violation—its use of the content was covered under Fair Use—but of acquiring the books illegally. Still, the settlement is one of the largest payouts so far related to generative AI training practices. China’s Moonshot AI unveils powerful Kimi K3 open-weight model Moonshot AI released Kimi K3, a gigantic mixture-of-experts model with roughly 2.8 trillion parameters. The model has proved to be state of the art on coding benchmarks, adding to the evidence that Chinese open-weight models are rapidly closing the intelligence gap with frontier models from leading U.S. AI labs. The White House accuses Kimi model creator of ‘distilling’ training data from U.S. lab Michael Kratsios, who heads the White House’s Office of Science and Technology Policy, said on X that the U.S. government has learned that Moonshot AI used the outputs of Anthropic’s Fable model to train its own Kimi K3 model. He claims that Moonshot AI also acquired servers equipped with Nvidia GB300 GPUs, which the U.S. bars from shipment to China. Anduril and Archer team up to produce an autonomous attack drone The defense startup Anduril and the air taxi maker Archer Aviation unveiled Thunder, an autonomous hybrid-electric aircraft capable of carrying weapons, cargo, and surveillance equipment. The announcement highlights the growing convergence between commercial electric aviation technology and autonomous defense systems. Meta reportedly entering AI infrastructure business Meta owns a lot of precious AI computing infrastructure, but the company has so far used it mainly to run its own models, often to support its own social advertising business. But as the company continues pushing to develop state-of-the-art frontier models, it may be readying to make some money on the side renting its AI servers to third parties. The social giant is reportedly negotiating a deal worth as much as $10 billion to lease AI computing capacity to Anthropic over roughly two years. More AI coverage from Fast Company: A Harvard mathematician and an AI model may have cracked an 87-year-old math problem The most useful ways to connect your apps to ChatGPT and Claude OpenAI’s ad strategy faces a major reality check Thinking Machines is coming for Anthropic’s intellectual vibe Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.
The Trump Media & Technology Group announced last week that it would soon offer a new product for financial services companies: a real-time feed of posts from the “highest-ranking” accounts on Truth Social, the right-wing social media platform that will have your laptop fan roaring like a jet engine within two minutes of opening it in a browser tab. And as compelling as the content produced by, for example, the conservative podcaster Dan Bongino might be, what’s really for sale here is premium access to insights from the only Truth Social user anyone cares about: President Donald Trump, who often uses the platform to announce his administration’s official policies in meandering, typo-ridden posts. The service, branded as Truth API and marketed to “high-frequency and algorithmic trading firms,” advertises seamless delivery of Truth Social data within “milliseconds.” It also promises “continuous 24/7 coverage,” which means that next time Trump pulls out his phone in the middle of the night to impose a 400% tariff on whichever country he’s most recently decided he doesn’t like, subscribers can turn a profit without even waking up. For sophisticated trading outfits, any and every edge over the general public can be extremely valuable: As documented by The Wall Street Journal, posts from Trump have prompted spikes in the share prices of U.S. Steel (after he touted a substantial foreign investment in the company), Intel (after he announced its new partnership with Apple), and American Eagle (after he expressed his vaguely horny appreciation for its Sydney Sweeney ad). In the months since Dell founder Michael Dell and his wife, Susan, donated more than $6 billion to a Trump-branded government savings program for children, Trump has repeatedly plugged the company and urged people to buy its products. In a possibly related story, Trump acquired more than $1 million in Dell stock during this period, and shares have more than tripled in value. Trump Media & Technology Group’s interim CEO, Kevin McGurn, issued a statement expressing optimism that Truth API would “become a meaningful, ongoing source of revenue for the company” and create “lasting value for shareholders.” Notably absent from McGurn’s statement are the facts that TMTG’s stock price has dropped roughly 75% since Trump took office in 2025, and that TMTG’s largest shareholder is the Donald J. Trump Revocable Trust. As president, Trump has always been driven by his twin desires to feel important and enrich himself. (His reluctance to disentangle himself from his various businesses is the reason we all had to spend much of his first term learning what an “emolument” is.) With apologies to Trump Mobile and his eponymous memecoin, though, Truth API is probably his most brazen cash grab yet: a product that commodifies his market-moving power as President of the United States and packages it as a de facto insider trading license available to firms willing to pay for the privilege. Traders using data scraped from social platforms is not new; one of the reasons TMTG officials say they’re offering Truth API is to undercut companies that are selling Truth Social data already. But you don’t need to be an ethics-in-government expert to understand how Truth API is different from conventional Wall Street research tools: Here, the party making money off Trump’s posts is not some third-party vendor, but a company in which Trump himself has a billion-dollar stake. Perhaps the most revealing aspect of Truth API, which is set to launch on August 1, is the product’s reported pricing structure: According to Reuters, TMTG has discussed charging $100,000 per month, but is also offering a discounted rate—a mere $60,000 per month—to clients who sign a three-year contract. The term length is not a coincidence; it reflects the simple fact that however valuable early access to Trump’s social media diatribes might be while he is president, it will not be as valuable when he is out of office. By inking clients to three-year deals now, TMTG is trying to lock in a steady revenue stream for the balance of his presidency, plus a few months on the back end for a little extra cushion. Democratic politicians have said everything you’d expect them to say about our grifting president’s latest ethical misadventure. Massachusetts Senator Elizabeth Warren, for example, called it “an egregious scheme to profit off the presidency and enrich Wall Street while doing nothing to help Americans.” In a marginally more surprising development, a handful of Republicans, perhaps wary of giving disaffected midterm voters even more reasons to be angry with the president and his party, have also expressed concerns, albeit tepidly. North Carolina Senator Thom Tillis, who is retiring, suggested that Trump could be “wading into a gray area,” which once again leaves unanswered the question of just how egregious Trump’s corruption has to get before a Republican senator is capable of seeing it in black and white. Senate Majority Leader John Thune, who is not retiring, said that his “assumption” is that Truth API would be “tested” in some way, which is best understood as a plea to reporters that they ask him about literally anything else. In a very basic sense, it’s encouraging that lawmakers are not totally comfortable with a president profiteering off his social content. The problem, though, is that their response (at least so far) has been mostly limited to politely asking regulators who were appointed by Trump to take money out of his pocket. For example, New York Democratic Representative Ritchie Torres wrote a letter this week urging Securities and Exchange Commission Chair Paul Atkins to “promptly evaluate” Truth API’s legality. Given that Atkins is a crypto guy who has been gleefully rolling back shareholder protections since Trump appointed him last year, I suspect he will not find any problems with the arrangement. For much of his political career, Trump has had a pretty easy time continuing to make money in ways that resemble his previous strategies for making money—attaching his name to various real estate developments, selling branded tchotchkes to his most ardent supporters, and so on. Now, though, thanks to his historic unpopularity and his lame-duck status, the supply of buttons he can push to add to his wealth is poised to dwindle rapidly. No one wants their brand-new golf course associated with a politician with a net approval rating of minus-24. There are only so many people in the world willing to pay $99.99 plus shipping for a Bible with his name on the cover. Truth API is a result of the bind in which Trump now finds himself: a desperate, more-transparent-than-ever gambit to monetize the presidency and guarantee a cut for himself. He has more than two years remaining in office, but is less interested than ever in doing the work of governing. All he cares about is wringing every last dollar out of the experience.
This past school year, Cambridge, Massachusetts, placed every eighth grader in Algebra I. It didn’t go well. Ultimately, more than 60% of rising ninth graders will repeat the course. The consequences were immediate: weeping students, angry parents, and roiling debate about the rollout, teacher support, and tracking. Lost in the shuffle was this essential point: We teach obsolete rote math that adults don’t use, while missing entirely the math that defines our lives. Let’s start with the math sequence followed by kids in Cambridge and across America: Algebra I, Geometry/Trig, Algebra II, pre-Calculus, and the much-desired Calculus. This track was laid out in 1893 by the Committee of Ten, a working group of educators, when respected professions (e.g., surveyors, engineers, astronomers, munitions officers) needed humans able to execute rote math procedures quickly and accurately by hand: factor (x³ − 6x² + 11x − 6), compute square roots, solve systems of linear equations. Quite sensibly, school helped kids master a body of rote math. But then, emerging computational resources began absorbing all rote math tasks. The fantastic book (and movie) Hidden Figures showcases the talented Black women who performed rote math by hand for the Apollo space missions. By 1970, these women recognized that mainframes made their skill set obsolete. Yet 56 years later, our education system hasn’t reached this realization. Worse, math scores are a principal measure of education quality for states, districts, and schools. Rote math scores define the futures and self-esteem of every American teenager. Rote math that adults don’t use, that smartphones perform flawlessly, and that pushes aside the math that matters. In our personal lives, we need to understand the probability behind a medical diagnosis, the long-term impact of compound interest, the risks of online gambling, the statistics behind “studies show” headlines, and the algorithms that determine what we read, watch, and believe. In our careers, we need to estimate market sizes, devise statistics to track an organization’s progress, predict revenue growth, optimize scarce resources, deploy proprietary algorithms, and make sound decisions. The powerful, fascinating math that is essential in life, but doesn’t make the cut in the math curricula of most students. We pay a high price for this. Only about a third of U.S. adults have the numeracy skills needed to navigate real-world financial and health decisions, while most—by some estimates, more than 90%—report some degree of math anxiety. Generations of Americans suffer from math, rather than capitalize on it. Meanwhile, we define the quality of a K-12 education—for a person, school, district, state, and nation—according to how humans perform on tasks that AI does perfectly. Why? We test kids at such a massive scale that the only way to keep costs in check is to rely on exams expressly designed to be graded by a computer. And if a computer can grade it, a computer can do it. If a young adult enters the real world with skills inferior to what a set of AI agents does for $20/month, they will struggle to find any decent career path. But even if you disagree with our critique of today’s math curriculum, there is a deeper problem. We never had the conversation. Policymakers, school boards, journalists, and philanthropists have largely treated the traditional math sequence as a given rather than asking whether it still serves students in an AI-driven world. It is not too late to have these conversations. Cambridge—and districts across America—can still have the discussion they skipped, shifting the focus from “How can we do obsolete things better?” to “What mathematical capabilities will students need to thrive in an AI-driven world?” Review the current high school math curriculum and ask, “Do any of us ever use this?” On the website for Ted’s new book, Aftermath, you can find a simple eight-question math quiz using practice questions from The Nation’s Report Card Grade 8 exam, the PSAT, SAT, ACT, and a state-mandated exam. Start by asking community members—especially school boards and local legislators—to take this quiz (12 minutes, no devices), and reflect on whether these questions should define the futures of every kid coming through our school system. As we consider how to educate for a world shot through with algorithms, AI itself also gives us powerful new tools for listening. Instead of relying on a handful of public hearings, school districts can use AI to synthesize thousands of perspectives, identify common ground, surface disagreements, and explore alternatives. AI makes it practical for communities to contribute their expertise and experience and for even the busiest institutions to listen. Charlotte-Mecklenburg Schools recently engaged more than 10,000 families, educators, students, and community members to shape its AI strategy, surfacing broad support for AI literacy alongside concerns about academic integrity and the pace of adoption. In our own Unlocking Literacy initiative at Northeastern University, we ran a national engagement on literacy. We used AI to enable participation in multiple languages and from people with different literacy levels, then synthesized what communities told us. Imagine applying that same approach to mathematics. Instead of assuming the traditional sequence is immutable, districts could ask students, parents, teachers, employers, mathematicians, scientists, and engineers what quantitative reasoning people actually need in an AI-driven world. AI can make it practical to gather thousands of perspectives, identify where there is consensus, clarify where there is disagreement, and make curriculum decisions that reflect today’s realities rather than yesterday’s assumptions. The stakes couldn’t be higher. If we continue graduating students whose greatest strength is performing tasks AI already does effortlessly, we will leave them poorly prepared for the world they are inheriting. But if we have the courage to rethink both what we teach and how we decide what to teach, we can prepare a generation to do what only humans can: reason, create, judge, and solve the problems that matter most. It’s long past time to change the equation. — Ted Dintersmith is an advocate for transforming education and the author of “Aftermath: The Life-Changing Math that Schools Won’t Teach You” (2026). Beth Simone Noveck is the director of The Governance Lab at Northeastern University and the author of “Reboot: AI and the Race to Save Democracy” (2026).
The European Union on Thursday hit hit Google with a fine of 890 million euros ($1 billion) after it said the technology behemoth broke digital antitrust regulations by setting up Google Play and its ubiquitous search engine to corral consumers towards its own services and apps to the detriment of competitors. It was the latest major crackdown on Big Tech by Brussels, which has led the world in reining in some of the world’s largest companies from Silicon Valley to Beijing. It has done so despite the risk of incurring the wrath of President Donald Trump, who has lashed out at the 27-nation bloc’s digital regulations amid a broader campaign against Europe: imposing high tariffs, making threats to seize Greenland from Denmark by force, and rattling trust within the NATO military alliance. In the past, Trump has threatened retaliation if American tech companies are penalized. Google had recently lost its appeal of a $4.5 billion antitrust fine imposed by the EU for throttling competition and reducing consumer choice through the dominance of its mobile Android operating system. The European Commission, the bloc’s executive branch and highest antitrust enforcer, said it was acting in the interest of consumers after an investigation of Google. “The best products should succeed because they’re better, not because they’re owned by the company running the search engine. And European consumers have a right to be told by app developers where to sign up to the best offers, even when the app store owner does not get a cut,” said Teresa Ribera, the commission’s Executive Vice President for Clean, Just and Competitive Transition. Google’s President of Global Affairs Kent Walker blasted the fine as “product degradation driven by a small group of self-serving complainants” that will have a negative impact on European businesses and consumers. He said that the EU’s Digital Markets Act forces Google “to strip away real-time search features Europeans love — like instant pricing and direct availability for hotels, flights, and restaurants — and dismantle safety protections on Google Play.” The EU describes the world’s seven tech giants — Amazon, Apple, Google parent Alphabet, Meta, Microsoft and TikTok owner ByteDance — as “gatekeepers” that control access for consumers. “In the EU, businesses have the right to compete fairly. Gatekeepers have the obligation to ensure a level playing field and consumers the right to choose for cheaper alternative offers,” European Commission spokesperson Thomas Regnier said. Alphabet reported $403 billion in revenue in 2025. —Sam McNeil, Associated Press
Anthropic has doubled down on its position as the AI company pushing hardest for industry regulation, announcing on July 21 that it plans to donate another $20 million to Public First Action, a political group that advocates for government-imposed safeguards on AI. The contribution brings Anthropic’s total donations to the group to $40 million. It comes as the midterm elections in the U.S. draw closer and AI legislation remains a political hot button. “We’ve long argued that frontier AI companies should be transparent about what their models can do and how they’re managing the risks,” the company said in a statement. “We’ve supported newly passed laws in several states that require greater transparency for AI developers. But given how fast the capabilities of the most powerful models are advancing, transparency alone is insufficient.” While Anthropic is putting its money where its mouth is, not all of its peers share its enthusiasm for government regulation. Here’s where the other major AI players stand. OpenAI OpenAI has historically been receptive to government regulation, but its position has become murkier of late. In 2023, CEO Sam Altman called for the creation of an international body, similar to the International Atomic Energy Agency, to oversee AI. Earlier this month, in an op-ed published in the Financial Times, he argued that governments, rather than AI labs, should set the rules for artificial intelligence. Altman proposed creating a U.S.-led international forum that would establish safety standards for AI models, provide “expert and impartial analysis of capabilities and risks, and [make] the technology available to nations and companies that participate and follow the rules.” Altman, however, has also discussed giving the government a stake in OpenAI, which could create questions about how restrictive any resulting regulations would be. Last month, he argued in front of the Senate Commerce, Science, and Transportation Committee that AI developers should not have to obtain government approval before releasing new models to the public. Earlier this year, meanwhile, OpenAI President Greg Brockman and VC firm Andreessen Horowitz launched a super PAC called Leading the Future, dedicating $100 million to opposing AI regulation. OpenAI and Andreessen Horowitz also supported an unsuccessful effort to impose a 10-year moratorium on states’ ability to create their own AI regulations. Google Google has considerably more experience dealing with government regulation. Despite the headaches it has faced in both the U.S. and Europe, the company has remained open to AI oversight. Earlier this month, Demis Hassabis, who heads Google’s DeepMind division, called for a U.S.-led global AI watchdog, arguing that the technology is advancing faster than our ability to understand it. “When there is a large degree of uncertainty and the stakes are this high, proceeding with cautious optimism is the sensible and correct strategy,” he wrote. “That calls for public policy that promotes innovation while also incentivizing responsibility and security, fosters international collaboration on key safety issues, and encourages careful consideration of how AI is deployed for the benefit of society.” Meta If Anthropic is the industry’s leading proponent of regulation, Meta may be its leading opponent. Last September, the Mark Zuckerberg-led company launched and invested “tens of millions” of dollars in a super PAC focused on fighting AI regulation. The American Technology Excellence Project, as it is called, is intended to oppose policies that Meta views as harmful to AI development. The group’s launch followed the creation of a separate California-focused PAC backing tech-friendly candidates in the state. “Amid a growing patchwork of inconsistent regulations that threaten homegrown innovation and investments in AI, state lawmakers are uniquely positioned to ensure that America remains a global technology leader,” Brian Rice, VP of public policy at Meta, said in a statement. Microsoft Microsoft has not voiced strong opposition to government regulation, but it has called for policymakers to be especially clear about what AI companies can and cannot do. So far, the company says, that clarity is lacking. Microsoft President Brad Smith has criticized officials in Washington for enforcing rules that have not yet been clearly established. “Everyone is reluctant to say there should be regulation, but what we really have right now is regulation without transparent or complete rules,” he said at the AI for Good Global Summit earlier this month. “Without rules, businesses can’t plan. . . . Ultimately, common sense says don’t be heavy-handed, but have enough of a touch that you can do what needs to be done. I hope we can move the conversation in that direction.” Microsoft published a detailed five-point blueprint for public AI policy in 2023 and has not deviated from its support for that framework. xAI Elon Musk’s AI company has strongly opposed state-level regulation, challenging several laws in court before the Justice Department intervened on the company’s behalf. The company argued that one proposed Colorado law violated the First Amendment by restricting how developers design AI systems and program their responses to hot-button issues. The argument effectively sought to extend constitutional speech protections, which traditionally apply to people and organizations, to the output of an AI system. xAI has taken a less aggressive position on federal oversight, sharing an early AI model with the government so officials could evaluate its capabilities.
Recently, when talking to Samsung Electronics America Executive VP of Mobile Experience Dave Das about the company’s new Galaxy Z Fold8 and Galaxy Z Fold8 Ultra phones, I offhandedly referred to the whole category of folding phones as “a novelty.” Das didn’t tell me I was wrong. But he did politely try to reframe my thinking about the new models, which Samsung unveiled on Wednesday at a Galaxy Unpacked event in London. “Some call it a niche product, some call it a novelty product,” he told me. “But I think after seven generations and millions of users in the U.S. alone, and tens of millions globally, now I can honestly say Samsung foldables have become mainstream.” He had a point. More than seven years ago, the original Galaxy Fold was most definitely a novelty—the first smartphone from a major manufacturer that could unfold into something resembling a small tablet. It was also, as Das acknowledged, “a bit hefty.” Oh, and Samsung’s launch devolved into a PR crisis when some of the first reviewers to get their hands on the Fold found their test units quickly broke. (The company pressed pause on the release to improve the hinge.) Smartphones in their classic iPhone/Android chocolate-bar-shape form have not evolved all that much in the past seven years, because there just wasn’t that much left to improve. But the first Galaxy Fold was only a first pass at answering the many design questions unleashed by its folding display. Such as: What size and aspect ratio should the interior and exterior screens be? Can you make the folded device thin enough that it doesn’t feel like you’re carrying two phones lashed to each other? How do you minimize the pesky crease in the middle of the display? What aspects of the software need to be adjusted for this new experience? Year by year, Samsung tackled all these issues, until it had folding phones that were not merely novel, but refined. Maybe even just plain normal. For example, each year’s Fold was a bit thinner than its predecessor. “And then, just within one year, going from the Fold6 to the Fold7, we equaled all of the prior six years’ reductions in thickness,” Das said. When my colleague Jared Newman reviewed that model, he said there was no phone he’d rather be using. The big news this year is that there are now two Galaxy Z Fold models (along with the pocketable clamshell-style Galaxy Z Flip8). The $2,099 Galaxy Z Fold8 Ultra is the direct descendant of the Fold7 and, at 4.1 millimeters when unfolded, is the thinnest Galaxy Z Fold yet—though only by a millimeter. It has a 6.5-inch screen when unfolded, packs a larger battery, and is maxed out with cameras, including a 200-megapixel wide sensor plus ultrawide and tele sensors, with 3X optical zoom. Like previous Galaxy Z Folds, it’s aimed at people who prize the large screen’s productivity benefits, such as its ability to display spreadsheets in a way that’s actually useful. Meanwhile, the $1,899 Galaxy Z Fold8 is, at 7.16 ounces, the lightest Galaxy Z Fold yet. Closed, its 5.5-inch screen’s 10:16 aspect ratio is noticeably squat. Unfolded, it has a 7.6-inch screen with a 4:3 aspect ratio optimized for media consumption. According to Das, that ratio delivers some of the value people found in the Galaxy Z Trifold, a $2,899 phone with a 10-inch three-panel display that Samsung sold only briefly. The Galaxy Fold8’s unfolded aspect ratio is optimized for content. [Photo: Samsung] “The feedback that some of our customers gave us [on the Trifold] was it’s a bit of a burden to unfold twice, and with that large of a screen size, to use it in one hand is not always the most convenient,” he said, though he stressed that “The satisfaction among consumers that purchased it was very, very high.” Both Fold8 models have a new display that uses a thin titanium film and a titanium plate below the screen. That increases durability and enables “a creaseless experience,” Das said. I haven’t seen the phones in person yet, but The Verge’s David Imel did and was wowed. One thing that hasn’t changed over the Galaxy Z Fold line’s seven-year history—and explains why they’re not a lot more mainstream than they’ve become—is their price point. It’s always hovered in the neighborhood of $2,000, with the Fold8 coming in at $100 less than last year’s Fold7 and the Fold8 Ultra going for $100 more. Das said those prices are due to the phones’ advanced technology, such as the Flex Titanium display. He also acknowledged “the elephant in the room” of RAM prices, which have skyrocketed as AI data centers gobble up production capacity. Thanks to promotions and financing options, the cash outlay required to acquire these phones is not as daunting as it might be, he pointed out. The other way to rationalize the price is thinking of a folding phone not just as an upgrade from a conventional one, but as two useful devices in one. The productivity-loving people who Das said are the Fold’s core constituents seem to be doing exactly that. “What we have been proposing to consumers is that you might have spent $1,300 on an Ultra phone, you might have spent $500 or $600 on a tablet,” he said. “Well, now you can get both of those functions in one device with the Fold8 Ultra.” Oh yeah, Apple Samsung is not the only company doing interesting things with folding phones. Last year, I was impressed with Chinese manufacturer Honor’s Magic V5; China is a major market for the category, which a recent study said was dominated in the second quarter by another Chinese maker, Huawei, with Samsung in second place. But no player in folding phones will attract more attention than Apple, which is widely expected to enter the market when it announces new iPhones in September. If the rumors pan out, the first folding iPhone may bear some physical resemblance to the Fold8, at least from a screen-size and aspect-ratio standpoint. Asked about Apple entering the market Samsung pioneered, Das replied in the way executives generally do in such situations: “What I will say is, Samsung always welcomes competition, and more manufacturers in the foldable space is great for the category. More players raises the awareness among the general consumer population.” Much of the time, the pioneering companies that welcome competition are, I’m sad to say, about to get creamed—including Apple itself, when it welcomed IBM to the PC market 45 years ago. This time could be different, though. Enough people are loyal to either the Samsung/Android or Apple/iOS ecosystem that there might not be that many instances of someone who would have bought a Galaxy Z Fold opting for a folding iPhone instead. The market could end up like the one for non-folding phones, in which Samsung and Apple are twin 800-pound gorillas who peaceably coexist. (According to Counterpoint Research, they tied for worldwide smartphone leader in the first quarter of 2026, each with 21% of shipments.) “Historically, we’ve seen a lot of customers who are very loyal to these foldable form factors, who are upgrading year after year to the new Samsung foldable device,” Das said. “Now, with the new Fold8 in this new form factor, we see even more people who we think will move from our traditional bar-type phone to this Fold phone, because we’re hitting the sweet spot of the right form factor, open and closed.” In other words, the Galaxy Z Fold8 and Galaxy Z Fold8 Ultra’s biggest competitor might not be Apple—it might be a cheaper, unfoldable Samsung Galaxy.
This article is republished with permission from Laser Wars, a newsletter about military laser weapons and other futuristic defense technology. The threat of low-cost weaponized drones may grow more urgent by the day, but the United States’ nascent laser dome is slowly but surely expanding to meet it. In a new request for information (RFI) published on July 20, the U.S. Coast Guard’s Research and Development Center stated that it is “interested” in high-energy laser weapons “or comparable laser technology” to protect cutters, small boats, and fixed sites like piers from both airborne drones and unmanned surface vessels. Importantly, potential weapons must do so amid punishing conditions at sea, namely “shock, vibration, sea spray, and airborne particulates, including smoke and fuel exhaust.” The RFI makes the Coast Guard, which operates under the U.S. Department of Homeland Security, the newest uniformed entrant into a global laser weapon race that currently includes every U.S. military service (including the U.S. Space Force) and a growing fraternity of nations embracing directed energy weapons for drone defense at home. But somehow, the RFI wasn’t even the most interesting bit of directed energy news to emerge this week. On July 21, defense contractor Kratos announced that it had received a $156 million contract to provide “mobile Counter-Unmanned Aircraft System (C-UAS) platforms” to the U.S. Department of Energy’s National Nuclear Security Administration (NNSA) Office of Secure Transportation (OST), the agency responsible for the safe and secure transport of nuclear weapons and materials, under its ‘Project Solar Shield.’ According to contracting documents, the agreement calls for the delivery of two prototypes and 25 full-production “On-The-Move (OTM) Counter-Unmanned Aircraft System (C-UAS) Special Operations Vehicle (SOV)” systems. Kratos did not explicitly identify laser weapons as part of the vehicles’ “layered defense posture” in its announcement, and a company spokesperson declined to provide additional details when reached for comment. But an additional contracting document and a public statement from a company executive appear to indicate that its counter-drone solution potentially includes a directed energy component. In February, Kratos announced that it had acquired Nomad Global Communication Solutions, which builds specialized vehicles like mobile command units and counter-drone platforms, the latter of which are marketed as capable of supporting directed energy capabilities. The next month, the NNSA published a notice of intent to sole source to Nomad GCS for the agency’s Counter-Unmanned Aircraft System (C-UAS), which included an associated document justifying NNSA’s decision to skip a full competition in favor of a Nomad GCS solution. The justification document is heavily redacted, but it appears to state that directed energy weapons (and, mofd specifically, laser weapons) and their associated size, weight, and power constraints were among the technical requirements considered during market research, and that “current commercial and government resources failed to produce any adequate alternative solutions” to Nomad GCS’s offering. Kratos president and CEO Eric DeMarco subsequently announced that the company had received a contract for “a new directed energy weapons program” during an earnings call the following May. “Over the past year, internally and tied in with an acquisition we have made—we try to go one plus one equals four—this is a counter-UAS system,” DeMarco said in an apparent allusion to the Nomad GCS acquisition. “It is mobile. We are the prime. It is several hundred million dollars. It is going to start ramping next year. It should be very big in 2028. This is an area where, now that we have won this one, it will open the door for us to win more.” Taken together, the Coast Guard RFI and the NNSA contract represent a gradual broadening of the U.S. government’s domestic laser weapon push, one that’s no longer confined to U.S. military testing activities and U.S. Customs and Border Protection shootdowns at the southern border with Mexico. The Federal Aviation Administration and Pentagon’s landmark safety agreement on the domestic use of laser weapons in April laid the foundation for future domestic directed energy operations, and now federal agencies responsible for safeguarding critical infrastructure appear to be gradually exploring their potential applications. The Coast Guard’s search for laser weapons comes amid a $150 million surge in funding for counter-drone capabilities in the year leading up to the FIFA World Cup and America 250 festivities this summer, part of what leaders characterized as “a whole new mission set” for the service. While that funding mostly went to acquiring equipment for the service’s own layered defense centered on using electro-optical/infrared sensors to detect and track incoming threats and electronic warfare capabilities like radio frequency jamming to neutralize them, the laser weapon RFI offers a hard kill option as a logical backstop. And beyond event security, the service’s role as the country’s premier maritime law enforcement agency (and the RFI’s explicit emphasis on fixed sites like piers) suggests an increasing focus on protecting ports, waterways, and other other maritime assets. Additionally, the RFI treats on-the-move cutters and small boats as a separate engagement problem from fixed piers, and with good reason: a shipboard laser weapon has to maintain a steady beam on a target while its host platform pitches and rolls on the water compared to stationary dockside directed energy assets. The RFI’s explicit inclusion of uncrewed surface vessels alongside aerial drones is just as telling: after Houthi and Iranian drone boats targeted commercial vessels in the Red Sea and the Strait of Hormuz over the past year, and the U.S. Navy itself deployed its own drone boats against Iranian port facilities in July, the Coast Guard is clearly thinking about laser weapons as a way to stop explosive-laden hulls before they reach a critical logistics hub or populated thoroughfare. With respect to the Coast Guard’s critical role in the U.S. homeland security ecosystem, the NNSA counter-drone contract is arguably the more consequential one even if Kratos never explicitly says the word “laser.” Protecting the convoys that transport nuclear weapons and special nuclear material is a different order of stakes than one shooting down suspected Mexican cartel drones at Fort Bliss, and the agency has been actively pursuing directed energy weapons as a potential solution for years, warning in 2024 that drone surveillance of top-secret sites has become “a persistent concern” that “no amount of signage or canine patrols along perimeter fences can prevent.” Project Solar Shield marks the agency’s first major push to extend that concern to open highways where nuclear convoys do not enjoy the luxury of a predictable and static threat environment. This slow embrace of laser weapons comes with clear obstacles. Every new agency in that mix multiplies the same jurisdictional and engagement authority problems that led to February’s dual airspace shutdowns near the US-Mexico border, so far that America’s laser dome increasingly looks like a patchwork of capabilities rather than a unified federal doctrine. This approach has its strengths by allowing agencies to adopt new capabilities at the speed of the their unique operational challenges, but it also means the interagency coordination gaps that tripped up those border deployments earlier this year may end up reproduced across a growing set of government actors, each with their own rules of engagement. Piers and highways are a far cry from the battlefields of Ukraine and the Middle East. But with the rise of drone warfare, anything under the same open sky is a potential target—and America’s laser dome is now being built one contracting notice at a time. This article is republished with permission from Laser Wars, a newsletter about military laser weapons and other futuristic defense technology.
The most obvious use of generative AI is writing. It’s right there in the name—large language models (LLMs) are all about reading, organizing, analyzing, and conjuring words—which is exactly why many in the journalism profession have been going through a kind of existential crisis these past few years. And the crisis isn’t just theoretical. As artificial intelligence systems get better at writing, a growing number of newsrooms are using AI to help not just with analysis, process, and ideas, but the actual words, too. That’s leading to growing pushback from editorial teams, like when reporters at The Sacramento Bee recently objected to having their bylines put on content written primarily by AI. {"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/03\/media-copilot.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/03\/fe289316-bc4f-44ef-96bf-148b3d8578c1_1440x1440.png","eyebrow":"","headline":"\u003Cstrong\u003ESubscribe to \u003Cem\u003EThe Media Copilot\u003C\/em\u003E\u003C\/strong\u003E","dek":"Want more about how AI is changing media? Never miss an update from Pete Pachal by signing up for \u003Cem\u003EThe Media Copilot\u003C\/em\u003E. To learn more, visit \u003Ca href=\u0022https:\/\/mediacopilot.substack.com\/\u0022\u003Emediacopilot.substack.com\u003C\/a\u003E","subhed":"","description":"","ctaText":"SIGN UP","ctaUrl":"https:\/\/mediacopilot.substack.com\/","theme":{"bg":"#f5f5f5","text":"#000000","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#000000","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91453847,"imageMobileId":91453848,"shareable":false,"slug":"","wpCssClasses":""}} As I teach in my AI trainings, using AI to write public-facing content is an inherently dicey business. I should know, since AI-generated articles are a component of The Media Copilot‘s editorial strategy. At minimum, you need to ask yourself several ethical questions: What is the medium? What exactly was the AI’s role? What’s the worst that could happen? And several more. From the answers, you need to develop a clear rulebook. One of the most important parts of such a rulebook is its policy around disclosure—the ways you signal to the reader that a piece of content is AI-generated or AI-assisted. The most direct way to do so is with an AI byline. Robot bylines typically have clinical names, such as the AI News Desk or Generative AI Services, along with their own author pages. That unambiguously lets the reader know that AI didn’t help just with research or ideas, but also with the words on the page. To what extent—which is often crucial—is usually revealed via an accompanying disclaimer, often at the bottom of the page. Transparency comes with a cost The AI byline feels intuitively correct: It checks the transparency box, it’s an easy-to-read label, and it slots neatly into an existing system. Moreover, audiences say they want them, with a Trusting News study showing that 94% of readers want disclosures on content that’s AI-written. As use of AI rises in newsrooms and editorial operations, it seems logical that AI bylines will become more common. Except that’s not what’s happening. A 2025 audit of 186,000 articles in 1,500 newspapers in the U.S. estimated that about 9% of the content was partly or fully AI-generated, yet only about 5% of those articles included a disclosure. Prominent AI-byline experiments at Fortune and Business Insider were discontinued, though that’s not an indicator of a broad policy shift against AI writing; indeed, Nick Lichtenberg, business editor at Fortune, famously used AI to produce more than 600 stories in six months. The Cleveland Plain Dealer—which uses writing tools to turn raw reporting into stories with final sign-off from the reporter—generally doesn’t use an AI byline unless the human contribution is extremely minimal. Anecdotally, as someone who covers AI use in media closely, I see fewer AI bylines than I used to. Which isn’t to say they’re gone: CoinDesk marks AI assistance with the byline “AI Boost” and ESPN’s writing bots still get top billing, but generally there’s been a retreat from the AI byline as a best practice. What’s going on? A few things: 1. Visibility in search and AI answers. Google says it does not downgrade content simply because AI was used to produce it. But its guidance consistently stresses clear authorship, first-hand expertise, and accountability, and it has said that giving AI an author byline is probably not the best way to disclose its use. A robot byline may not be a direct negative ranking signal, but it also provides none of the human authority signals that search and discovery systems are designed to reward. Data supports this: In Graphite’s 2025 analysis, human-written articles made up 86% of the pages ranking in Google Search and tended to rank higher than AI-generated material. That doesn’t prove that AI authorship or attribution caused the difference, but it shows the broader disadvantage AI-heavy content faces. That pattern also matches our limited experience at The Media Copilot. Our human-bylined articles show up in Google Discover and rank higher in Google Search than those published under our AI byline, The Copilot. 2. The trust paradox. Trusting News found that, although the vast majority of audiences want AI disclosures, their presence made 42% of respondents less likely to trust an article. Separate research from Reuters Institute found an even wider comfort gap: 12% of readers are comfortable with fully AI news, versus 62% comfortable with fully human-written news. 3. Institutional memory. When generative AI was new, there were several high-profile failures of AI content. By putting forward an AI byline, an outlet can’t avoid association with those mistakes, and it paints a target on every article under it that the screenshot industrial complex will fire at all day the moment there’s an error. In short, AI bylines have acquired a lot of baggage in the short time they’ve been around, and many publications seem to be taking the stance that they’re just not worth it. However, that doesn’t translate into a pullback on AI generally, or even a pullback on AI-assisted content. The most direct way to square this circle is to reserve bylines for people, and disclose the AI’s contribution in some other way, usually via one of those notes at the end of the article. A byline is a promise At the heart of this issue is what the byline actually means. The idea that the person named at the top of the article wrote each and every word has always been a fallacy. Editors, wire copy, fact checkers, headline writers, and spellcheckers all contribute actual words and sometimes whole passages to articles. Automated editing software takes this even further—anything substantially edited through Grammarly or a similar tool already contains wording shaped by AI. The byline is not “I wrote all these words”; it’s “I stand by all these words.” And this is the core problem with AI bylines—they obfuscate responsibility. Without ownership, without someone prominently standing by what’s actually written, there’s little incentive to make sure it’s great writing. It inherently turns the writing into something second class, regardless of how the AI was used and how many words it actually contributed. At the same time, what happened at The Sacramento Bee shows just how much writers will defend the use of their names. And rightly so. The path forward isn’t to force writers’ names onto AI content without their consent, but to give them the freedom and the accountability on the use of AI. The disclosure should scale with the machine’s contribution: Routine editing may require nothing, substantial drafting should require a specific note, and largely automated reporting should explain both the system and the human review. Ultimately, though, the only label that really matters is whether a named human stands behind the result. With the right policies and training, publications can give writers broad freedom to use AI while keeping responsibility attached to a human name. If the content is worthwhile, then over time those who use AI to amplify and accelerate their judgment will be successful. Those who use it as a substitute for thinking will inevitably fail. The human accountability signal The AI byline isn’t dead, but it is being repriced. As machine text gets cheap, a human name that carries real accountability becomes the scarce and valuable signal. As the tools get better, they’ll continue to blur who wrote what. The constant, however, is simple: A human has to stand behind it. No one gets to outsource accountability to a machine. {"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/03\/media-copilot.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/03\/fe289316-bc4f-44ef-96bf-148b3d8578c1_1440x1440.png","eyebrow":"","headline":"\u003Cstrong\u003ESubscribe to \u003Cem\u003EThe Media Copilot\u003C\/em\u003E\u003C\/strong\u003E","dek":"Want more about how AI is changing media? Never miss an update from Pete Pachal by signing up for \u003Cem\u003EThe Media Copilot\u003C\/em\u003E. To learn more, visit \u003Ca href=\u0022https:\/\/mediacopilot.substack.com\/\u0022\u003Emediacopilot.substack.com\u003C\/a\u003E","subhed":"","description":"","ctaText":"SIGN UP","ctaUrl":"https:\/\/mediacopilot.substack.com\/","theme":{"bg":"#f5f5f5","text":"#000000","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#000000","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91453847,"imageMobileId":91453848,"shareable":false,"slug":"","wpCssClasses":""}}
For the past few years, the AI infrastructure race has been driven by the assumption that larger artificial intelligence models require larger data centers. That idea has fueled an extraordinary wave of spending. In rural Richland Parish, Louisiana, for example, Meta is building one of the world’s largest AI infrastructure projects. Its Hyperion campus is expected to cost more than $50 billion and run on about 5 gigawatts of power, roughly the output of five nuclear reactors. But as these facilities grow, so does the resistance to them. Across the country, communities are pushing back over data centers’ demands on power and water, and the impacts they’re having on rural and suburban areas. The bigger the project, the more likely it is to become a political target. But Tom Leighton, cofounder and CEO of Akamai, the $18 billion content delivery network company that powers a significant share of the world’s web traffic, would argue that’s not even the worst part. The Meta project, like so many others, rests on the assumption that companies able to concentrate the most computing power will be best positioned to build the next generation of AI systems. And Leighton believes that assumption applies more clearly to AI training than to AI inference (the “thinking” that AI does when applying its training to real-world data). “The next challenge for AI is what it will take to run those models and their derivatives everywhere,” he tells Fast Company, sharing his belief that the industry may be trying to solve too many problems with the same enormous building. Leighton’s 28-year-old company has spent the past several years expanding into a cloud platform for AI. Rather than trying to match the hyperscalers data center for data center, Akamai is betting on a different approach offer a less disruptive path for expanding AI infrastructure, one that relies more heavily on a network of existing facilities instead of concentrating enormous demands for land and power in a single community. “Agentic AI needs low latency, high performance, and affordable economics that giant, centralized data centers cannot provide,” Leighton says. His proposed solution is also inspired by what made Akamai a player in internet infrastructure to begin with during the dot-com era. In the late 1990s, Leighton helped address the web’s growing pains, when every request going back to a handful of centralized servers was choking the internet’s growth. Now, Akamai is attempting to bring a version of the architecture that saved the web to AI. And Leighton’s got the AI kingmaker Nvidia backing him. Bigger AI data centers won’t solve every AI problem Leighton’s proposed solution is partly an economic one. Running an AI model is a different problem from training one, and many inference tasks do not require the largest model, the most powerful chips, or a request traveling thousands of miles to a massive centralized campus. For enterprises pursuing agentic AI, the debate extends beyond securing graphics processing unit (GPU) capacity. Companies increasingly need to decide where inference should run, where enterprise data is processed, how quickly AI agents must respond, what it costs to move data across regions, and how security policies are enforced once those systems move into production. In other words, deploying AI agents increasingly becomes an infrastructure design problem—not just a model selection problem. Leighton argues those operational decisions, not raw compute alone, will determine whether AI applications deliver acceptable performance, reliability and economics at scale. Inference performance depends partly on how quickly an AI system can respond when a decision is required. Network distance, data location, tool calls, and application design can all affect response time. “If an AI request has to travel thousands of miles, touch data in another location, call tools, and then return an answer to a user in real time, the speed of the data center is only one part of the equation,” Leighton says. “Proximity matters. In fact, it may matter even more as AI evolves.” Leighton’s view draws on Akamai’s experience with web infrastructure during the late 1990s. As the web expanded, centralized origin servers struggled to handle global demand. “Websites were built around centralized origin servers. As demand went global, every user request had to travel back to a small number of central places,” Leighton recalls. “The result was slow performance, websites would go down or freeze up during traffic spikes, and there was a lot of frustration for users.” The industry called the problem the “World Wide Wait.” Leighton, an MIT applied mathematician, helped develop an alternative based on distributing content and computation closer to users. Algorithms determined where requests should be served. Nearly three decades later, Leighton believes AI infrastructure now faces a related distribution problem, and argues that a major engineering challenge will be coordinating inference across thousands of data center locations while maintaining consistent performance, security, and reliability as a unified system. “Anyone can build a data center, but making many locations act intelligently together, under changing demand, with consistent performance and trust, is another challenge altogether,” he says. Can distributed AI outperform centralized clouds? The central component of Akamai’s strategy is AI Grid, an orchestration layer developed with Nvidia. AI Grid determines whether an inference workload should run in a centralized AI factory, a regional cloud, or one of Akamai’s edge locations. The decision depends on latency requirements, operating costs, and performance needs. Leighton argues that routing decisions can materially affect inference performance. GPU cost and availability remain significant constraints, but the industry is asking step-one questions in a step-two market. “The bigger questions are around where inference should run, what data it needs, how quickly the response must come back, what it will cost to move the data, and what security policy has to be enforced along the way,” he says. “The answers to those questions will have a much bigger impact on whether AI reaches its potential.” For Leighton, the definition of scale itself is changing. During the training era, scale meant concentrating as much compute as possible inside a single AI factory. In the agent era, he argues, scale increasingly depends on how effectively infrastructure can distribute inference across many locations while keeping latency, data movement, and costs under control. “GPU availability is a major issue today, but GPUs are not the solution to every AI problem,” he argues. “Many inference tasks do not need the largest model or the largest cluster or the most expensive compute. What they need is the ability to marry the right model, in the right place, data and moment, at the most effective cost.” Akamai says customer demand is beginning to reflect this approach. Earlier this year, the company disclosed a four-year, $200 million agreement with an unnamed major U.S. technology company to deploy one of the world’s largest clusters of Nvidia RTX PRO 6000 Blackwell GPUs on its platform. Three months later, Akamai announced a seven-year, $1.8 billion cloud infrastructure commitment from a leading U.S. frontier AI model developer. Insider reports identified the company as Anthropic. The agreement is the largest contract in Akamai’s 28-year history. Together, the two agreements represent roughly $2 billion in committed cloud business from customers Akamai did not have two years ago. Akamai’s cloud infrastructure revenue has also grown 40% year over year. Leighton declined to identify the companies or discuss the workloads behind the agreements. He says customer evaluations now include a broader set of operational questions. “When customers evaluate AI infrastructure, of course they look at scale and performance,” he says. “But they’re also asking how workloads perform in production, how costs evolve over time, how reliable the infrastructure is, and whether it gives them the flexibility to adapt as AI usage changes.” Without naming additional customers, Leighton says Akamai is supporting production AI deployments for an AI-powered video intelligence platform in India and a U.S.-based consumer AI company. Both require low-latency inference and do not depend on large centralized training clusters. The real cost of AI goes far beyond GPUs Many companies evaluate AI infrastructure through token prices, GPU usage, and model endpoint costs. Production systems also create costs related to context retrieval, API calls, storage reads, and network traffic across zones and regions. Akamai claims its architecture can reduce inference latency by as much as 2.5 times compared with traditional hyperscaler infrastructure, and lower inference costs by up to 86%. Its published benchmarks, conducted using Nvidia’s methodology, show RTX PRO 6000 Blackwell GPUs on Akamai’s cloud delivering up to 1.63 times the inference throughput of Nvidia H100 GPUs. The system sustained roughly 24,000 tokens per second per server under 100 concurrent requests. Leighton says the commercial viability of AI infrastructure will depend on whether those systems can deliver acceptable performance, reliability, and cost. “As a CEO, I am always skeptical of vague economics,” he says. “It is easy to say the future is bigger data centers and more GPUs. It is harder to show how that architecture delivers the right performance, reliability, and cost when AI is running everywhere, all the time. Inference is where AI must show profitable ROI.” Leighton’s background includes work as a theoretical computer scientist. He holds more than 50 patents and has served as Akamai’s CEO for more than two decades. “As a mathematician, I am skeptical of straight-line thinking. The fact that one architecture worked for the first phase of AI does not mean it will work for every phase that follows,” he says. “As a theoretical computer scientist, I am aware that what ignites a technological revolution is rarely the thing that lets it survive in the real world. The early promise of AI is no exception. The massive centralized clouds that started this boom aren’t built for the highly distributed reality of what comes next.” Wall Street is still pricing AI like the cloud era Wall Street has spent much of 2026 evaluating how Akamai’s cloud business fits with its legacy operations. The company’s content delivery network business has been shrinking for years. Its AI cloud expansion has also required substantial spending on hardware and infrastructure. Rising memory prices, driven by strong demand for AI hardware, increased component costs as Akamai purchased thousands of Blackwell GPUs. Margins compressed, and earnings per share declined even as revenue grew. Investors continued to value the company largely as a mature infrastructure provider. The $1.8 billion commitment led to Akamai’s largest single-day stock rally in more than two decades. The agreement provides evidence of demand for Akamai’s distributed AI infrastructure. It also underscores the amount of capital required to build and operate that infrastructure at scale. “It highlights both the opportunity and the discipline required to pursue it,” Leighton says. “We are making significant investments because we believe AI will be a major driver of cloud demand, but we are not trying to simply copy the hyperscaler model. Our advantage is that we already operate one of the world’s most distributed platforms, and power and protect large parts of the internet. The investments we are making are about extending that platform for cloud and AI.” The AI industry has invested heavily in GPUs, large clusters, and more powerful models. Leighton argues that inference will require additional infrastructure designed around distribution, latency, and cost. “In the early days of the web, many people assumed the answer was just more central infrastructure. It was not. AI is reaching a similar point,” he says. Whether Akamai’s distributed model can compete effectively with centralized cloud providers will depend on customer demand, technical performance, and the economics of operating the network at scale.
The release of Kimi K3, the latest model from Moonshot Labs, a Chinese AI firm founded by a former Carnegie Mellon computer science graduate, has shaken both the AI and political worlds. In an instant, the seemingly comfortable lead held by U.S. frontier AI labs over their Chinese competitors appeared to narrow sharply. K3’s release has rattled the chattering classes in Washington and Silicon Valley. In response, they have begun fighting among themselves. Samuel Hammond, director of AI policy at the Foundation for American Innovation, says Washington had been operating under the assumption that China remained several months behind the American frontier. “One of the roles of the intelligence community in the United States is to not be surprised,” he says. But officials were caught off guard when a system capable of approaching the best American models appeared within weeks, rather than the six months many had expected. Moonshot’s announcement that it planned to release K3 as an open-weight model made the threat even clearer. American businesses could run and adapt it themselves, but so could malicious actors. The Trump White House has reportedly considered banning U.S. companies from providing or serving Chinese AI models in an effort to curb their use. But the risks of such a move were underscored last week, when Hugging Face, the AI model-hosting platform, was subjected to a massive, accidental attack by a bleeding-edge model being tested by OpenAI. To defend itself, the company relied on GLM-5.2, a model from China’s Z.AI. A ban could have left Hugging Face without that tool. That tension has spilled into a broader fight over how Washington should respond to Chinese models. Some have called for import controls or outright prohibitions. Dean Ball, a former Trump AI adviser who now works for OpenAI, went further, describing China’s open-weight, open-source approach as “full AI communism” and predicting that the Trump administration would eventually realize “their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models.” Some observations on Kimi: 1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also… — Dean W. Ball (@deanwball) July 17, 2026 That drew a rebuke from another prominent Trump AI figure, David Sacks, the former White House AI and crypto czar. Sacks hit back on X, saying on social media that “the weaponization of regulatory uncertainty as a competitive tool should be completely unacceptable.” I’m not sure whether Dean Ball is confessing to a regulatory capture strategy or simply predicting this will happen (he now says the latter). Either way, the weaponization of regulatory uncertainty as a competitive tool should be completely unacceptable. He argues there’s no… https://t.co/ZxIFabAJZW pic.twitter.com/0eLk6ltU0F — David Sacks (@DavidSacks) July 19, 2026 So what has raised the temperature in Washington and Silicon Valley so dramatically? “Fundamentally, I think it’s about competition between the two countries and who like winning in AI,” says Kyle Chan, a research fellow at the Brookings Institution who focuses on China’s technology and industrial policy. “It puts a lot of pressure on U.S. AI companies and on the idea of not just American leadership in AI, but American dominance.” Ball appears to believe that if Chinese labs give away models that nearly match those at the American frontier, U.S. companies will lose the ability to charge a premium for access to their own. “On that particular point, I think Dean is a little hyperbolic,” Chan says. Open models could hurt frontier labs’ bottom lines, but they could also benefit the broader economy. Chan describes the current argument as “three or four different heated debates all mixing together, all intersecting under the banner of AI.” Those debates include how to confront China, whether advanced chips should be restricted or sold around the world, whether open models are a security risk or a vital tool for defenders, and whether Washington should protect a handful of national champions or prioritize the many businesses that rely on them. The potential threat to U.S. businesses and national security is also raising alarm in Washington. This is, after all, the administration that required Beijing-based ByteDance to divest its interest in the short-form video platform TikTok over national security concerns. “My understanding from talking to people who are closer to D.C. is people with power seem genuinely interested in banning those models,” says Kristy Loke, a fellow at the MATS Research Program who specializes in China’s AI strategy. An outright ban might address that concern, but it could create plenty of others. Most American companies are not OpenAI or Anthropic, and Loke says “most companies would enjoy the choice of open models.” Blocking Chinese systems could protect the revenues of U.S. frontier labs, but it could also raise costs for American businesses and startups. K3 may also change who the administration listens to. The Foundation for American Innovation’s Hammond says recent shocks have challenged the early Trump-world view that technical AI specialists were merely “doomers” crying wolf. “I think all these things have been steadily leading to a sort of reconsideration of the importance of technical expertise,” he says.
Winter Storm Uri, the multiday freeze that slammed Texas in February 2021 and pummeled the state’s power grid, has been on Kenneth Kovar’s mind ever since. Though the resident of New Braunfels, northeast of San Antonio, didn’t lose power at the time, he wasn’t able to run his septic tank—it is independent from local systems. He had to fill his toilets with water from his backyard pool. So when Kovar, 64, bought a Ford F-150 last fall, his hope was to be better prepared for any new crisis. “I was interested in trying to find some sort of power backup situation,” he said. Now, Kovar has a setup from Ford that allows drivers of certain F-150 models to plug their vehicle directly into their electric meter to power parts of their home during an outage. It is the latest example of automakers broadly adapting their electrification technologies to the home energy business, especially as demand on the grid increases and sales of electric vehicles slow. Car companies are looking to leverage their multibillion-dollar EV investments to tap into a promising market in vehicle, home and grid technology, both for backup power and for supporting electrical grid resiliency. “They’re trying to look for other businesses that they might sell into,” said Parth Vaishnav, assistant professor of sustainable systems at the University of Michigan. Ford’s latest connects F-150 drivers to their meter Drivers of the F-150 PowerBoost hybrid and F-150 Lightning electric pickup trucks can now plug in a one-foot long adapter to a 240-volt outlet onboard. That adapter—which Ford made with company Global Power Products—makes the vehicle compatible to plug into a longer, separate cable. That cable connects to a transfer switch installed directly on one’s electric meter. Through the adapter and cable series, homeowners can connect their vehicle essentially right to their home’s breaker box. The homeowner simply turns on and off which breaker switches they want for which devices they want powered. “The way we think about it, especially for customers who already have a compatible vehicle is, you already own the power source, it’s in your driveway,” said Amanda Roraff, Ford’s grid and energy services business acceleration lead. The Lightning might provide power for two to three days, depending on what home devices are being used, and the PowerBoost Hybrid, up to five days on a single tank of gas. The automaker says its solution is a less expensive way to supply backup power. Conventional, diesel-powered portable generators and full-home standby setups require expensive installation, costing several thousands of dollars. This solution, which also requires professional installation at the meter, starts around $1,100. The setup only applies to about 200,000 vehicles so far, and it is also exclusively for outages. Ford also offers its Home Integration System for bidirectionality, sending power both from the vehicle to the home and from the home to the vehicle, while also being able to feed the grid. Other automakers are boosting their energy solutions Over 630,000 U.S. vehicles already have this functionality, estimates say, and automakers are rapidly expanding their available options with the goal of full vehicle-to-grid support in the long run. South Korean auto brand Kia and Wallbox, an EV charging company, have teamed up so that drivers of eligible compatible vehicles can have home power backup during outages or during periods of high demand, to cut their utility use. They can send power back to the grid. Tesla’s technology is similar—allowing drivers of equipped vehicles to connect to their home using additional Tesla hardware. The Cybertruck provides full vehicle-to-home capability, where other Tesla models can only connect to and power specific devices or appliances. General Motors is also in the energy space. A recent partnership with WeaveGrid, for instance, allows homeowners who drive certain GM EVs—and have the automaker’s home system and a proper grid interconnection—to enroll in some grid reliability utility programs. Once an outage is detected, GM’s vehicle-to-home tech has the capability to disconnect one’s home from the grid and start supplying power from their GM EV. “If you can imagine the future as we go forward, it’s having the ability—now that we have this single platform—that allows our customer to experience our system,” said Wade Sheffer, vice president of GM energy, “but also can have the full control of the energy.” The capability is an important lifeline amid EV sales slowdown Not only is this business critical amid growing grid demand and increasing power outages, experts say automakers need to pivot with the EV market less active under current U.S. federal policy. Pure EV sales in the U.S. year-over-year are down 23.8%, according to a July Cox Automotive report on the first half of 2026. This demonstrates what an asset that EV and hybrid ownership can be. The tech is not without challenges. On the industry side, these systems have to undergo third-party testing to ensure they meet safety standards, and the vehicle and the charger need to be programmed to communicate. It also requires the approval of the utility where the capability is being used. It could take years to get an interconnect agreement. On the customer side, homeowners need to understand their vehicles’ abilities and how to self-manage their system. It also just brings another generator of power into the home mix. Still, experts see opportunity, especially with interest in EV sales high outside of the U.S. “We already know during an outage, its impact, providing electricity to the home,” said Scott Samuelsen, engineering professor emeritus at the University of California, Irvine. “This is going to become very, very popular.” ___ The Associated Press’ climate and environmental coverage receives financial support from multiple private foundations. AP is solely responsible for all content. Find AP’s standards for working with philanthropies, a list of supporters and funded coverage areas at AP.org. —Alexa St. John, Associated Press
Very few companies have a better real-time read on the American economy than Delta, and even fewer CEOs are more willing to say what they actually think than Ed Bastian. True to form, he shares what Delta’s booking data is telling him that the headlines aren’t, why he’s not convinced by AI until it moves his top line, and how he decided to yank congressional travel perks during the government shutdown. This is an abridged transcript of an interview from Rapid Response, hosted by former Fast Company editor-in-chief Robert Safian. From the team behind the Masters of Scale podcast, Rapid Response features candid conversations with today’s top business leaders navigating real-time challenges. Subscribe to Rapid Response wherever you get your podcasts to ensure you never miss an episode. You have access, obviously, to all the booking data that comes through the Delta app and website. Are there things that you see in that data that are different from the headlines? Does your data tend to lead the headlines or respond to them, or is it different at different times? Our data is usually ahead of the headlines. We just reported our second-quarter results. We gave guidance for our third quarter, which ends at the end of September, but we already have about 70% of the revenue on hand for that third quarter through advance bookings. Our bookings run anywhere from 60 to 120 days out. International runs longer, domestic runs a little shorter. So we have a pretty good sense for how people are already thinking about the fall and how people are thinking about the upcoming holiday season. Economists and members of the Fed and others are always quite curious about talking to us because we have a pretty good pulse on how the consumer, at least our consumer, is viewing the current environment, and then translating that into confidence to make a future purchase commitment. Your customer tends to be at the top end of the K in the K-shaped economy. You get, I don’t know, 20% more revenue per seat than competitors. Your premium-cabin revenue is almost on the verge of overtaking the main cabin. That decision, that lean into that customer, does anything about today’s environment make you rethink that strategy or double down on it? It sounds like that is still where, and the customer you feel, is healthiest for you to focus on. Oh, absolutely. And this is a decision we took not recently. We took it 15 years ago, shortly after we acquired Northwest Airlines. As a result, we were clearly the largest airline in the world. We still are, but it’s closer now. And we made a commitment to ourselves that we weren’t going to do the merger and introduce that level of scale unless we could actually differentiate in the marketplace and no longer be seen as a commodity. Because if you’re a commodity and you just get chosen by whoever has the lowest price, we lose. In fact, we lost 20 years ago when we had to file for bankruptcy for that very reason. Delta offered a lot of things but didn’t get paid for them, and as a result, the business model didn’t work, as it didn’t for a number of the big legacy carriers. We spent many, many years figuring out what we called back then decommoditization. Today we call it premium. Tomorrow we’ll probably find a more elegant term for it. And it starts with reliability, building trust with your consumer, deploying technology to assist along the way, and developing more partnerships, whether it’s with American Express, Uber, or YouTube TV onboard airplanes, Paramount. We want to be there to make certain that we’re getting the first call for the best dollar. There’s some concern that with the rise of AI, agent-based buying will make price-first shopping more prevalent. I’m curious how you think about that, that AI agents may not favor brands the way humans do. Unless the AI agents are going to be flying on the airplanes and actually providing surveys, I don’t think it changes our equation. At the end of the day, a relationship with an airline is fairly personal. It’s human. You’re not going to let an agent tell you how you should feel about a brand in the same way. And that’s the hallmark of a great brand: how it makes you feel. You deploy AI in your own operations. I know Delta tested AI for individualized ticket prices. Does using AI for this create a different kind of reputational risk because of the way people are emotionally reacting to what AI is driving? Yeah, AI is not trusted, and for good reason, probably, with all the hype that’s been attached to it. We talked a year ago about the fact that we started to use some AI in the pricing algorithms, and people extrapolated that into surveillance pricing and individualized pricing. We’re doing none of that, and we never intended to do any of that. The only thing we’re using AI for in the pricing tool is to help our revenue managers offer better pricing, and more pricing, in more markets around the world. At Delta, we have hundreds of thousands of price points out for sale at any given time on any day. We can only manage a select number of those with human eyes, and we automatch the vast majority of those price points. Automatch basically means making sure you don’t get too far above or below where the market is coming in at. Through AI, we can bring more intelligence to setting those signals. And AI works around the clock. It’s 24/7. It’s always monitoring and seeing what’s happening in the marketplace, and whether our product should be higher or lower, trying to give the customer the very best opportunity. We can only do that on our highest-volume routes, but we still have a massive number of routes where we don’t deploy it. So when you think about it, it’s really a knowledge tool, a learning tool, and an efficiency tool. You’ve been known for your people-first leadership, and I know a lot of people are worried about how AI is going to impact their jobs. How do you think about that? How do you talk to your team about that? Well, at Delta, I’ve been very clear that AI is just . . . First of all, I don’t call it artificial intelligence. I refer to it as augmented intelligence. It’s going to enable our teams to do a better job of what they do. We are a highly analytical business. We have 5,000 flights a day, every day. As a result, there is a tremendous amount of data, a tremendous amount of intelligence in terms of monitoring and making decisions, whether it’s in-flight patterns, pricing the product, mechanical reliability, predictive engineering. There are many disciplines that use that data. AI helps you make those decisions quicker. It absorbs the amount of data that’s out there far faster than anything we can do. We wind up making a lot of decisions based on judgment, which sometimes is seat of the pants, sometimes it’s instinct, and these are going to be better data decisions. Now, that doesn’t mean that we’re going to need fewer people. It frees people up to have more time to serve, and if we do a great job of that, that’s going to grow our business. Some CEOs got boos for mentioning AI at school commencements. You asked AI to write your commencement speech for Emory, then scrapped it, and warned graduates against pushing the easy button. And guess what? People stood and cheered when I said that. So what’s the easy button? And did going through that impact how you think about using AI yourself and for your team? I’ll admit I’m not a big user of AI. It’s pretty hard to teach old dogs new tricks at times. I’m curious about it. I was impressed with how quickly a speech is produced, within a matter of seconds, and it’s produced in words that I would be associated with. I talk in lots of public forums, so there’s a lot of data out there as to how Ed Bastian would deliver a commencement address. But at the end of the day, it didn’t have any soul. It had no spirit. I felt like I was cheating. To be given an opportunity to come to a commencement is an honor, and then am I going to potentially use technology just to try to get it done quickly and check that box? It wasn’t right. So there’s a conscience that needs to be deployed with AI. I think that’s where people in the technology sphere just don’t get it. They’re out there selling, selling, selling.
In 2020, just after Winslow and Laura Robinson bought a farm in Maine, they were surprised to learn that sewage sludge had been used as fertilizer there decades earlier—and that the sludge had left behind “forever chemicals” that remained in the soil. Farmers in Maine have been at the forefront of uncovering a problem that exists on farmland across the country. Sewage sludge from wastewater treatment plants has been widely used on farms since the 1970s, both to keep it out of landfills and provide cheap fertilizer to farmers. (“Biosolids” are also sold in some garden stores as home fertilizer, which you might unwittingly be using in your own backyard.) But the sludge can contain high levels of PFAS, so-called forever chemicals that don’t break down and can move from soil into food. While the impacts of PFAS in food aren’t yet fully understood, the chemicals have been linked to liver disease, cancer, and other illnesses. After testing, the Robinsons found a safe area of their farm to begin growing organic produce. But like other farmers, they didn’t have a good solution for the contaminated fields they’d discovered. A new study from Yale University, published in the Proceedings of the National Academy of Sciences, suggests a low-cost way that farms can clean up the problem and simultaneously help fight climate change. [Photo: courtesy Yale University] A cheaper way to tackle forever chemicals Existing methods for PFAS remediation are extremely expensive—sometimes $1 million per acre—and typically used only at sites like airports or military bases. They also involve removing soil. Cleaning up millions of acres of American farmland just isn’t feasible. Noah Planavsky, a geochemistry professor at Yale who studies soil health, decided to work on the problem after learning about the issue from farmers. “Farmers were deeply concerned about this, felt like they weren’t getting good information, and felt like this was a massive problem where they were being ignored or gaslit,” Planavsky says. The study tested a simple approach for treating contaminated soil. Farmers plant crops known to absorb PFAS quickly, such as hemp and sunflowers. The farmers also add crushed alkaline rocks to the soil to raise the pH level, which helps the plants take up more of the chemicals. When the crops are harvested, they’re heated to very high temperatures without oxygen, which breaks down the PFAS. “Although they are forever chemicals— very sturdy and robust chemicals—they can be destroyed at higher temperatures, and that is something that has been well documented,” says Planavsky. The process creates biochar, which can be added back to the soil to help trap more PFAS. The scientists estimate that in some cases, the method could bring contamination down to a safe threshold within a decade. There’s another challenge: In most states, sewage sludge is still being used on some farms. It’s not surprising that it’s been commonly used historically. “Farmers are barely making ends meet,” says Laura Robinson, the organic farmer from the Maine farm, who participated in the Yale study. “There’s not much profit in farming. So if you’re being offered free fertilizer for acres and acres of hay fields, you’re going to take it. The powers that be [were] saying that it’s safe.” Maine, which has moved faster than others to tackle the problem after farmers raised the alarm, banned the use of sewage sludge on farms in 2022. The first step is to stop its use broadly. Under the Trump administration, the EPA has backed away from regulating sewage sludge as fertilizer. States may move more quickly, though, and the study’s new cleanup method comes at a time when farmers are looking hard for solutions. [Photo: courtesy Yale University] Could cleanup be affordable at scale? The scientists’ approach is an order of magnitude cheaper than other methods for PFAS cleanup. And it has a second advantage: Adding crushed rocks to soil helps capture carbon dioxide. (Planavsky also researches this type of carbon removal, called enhanced rock weathering, and cofounded a startup called Lithos that separately works with rock dust on other farms.) In theory, selling carbon removal credits could help fund the PFAS cleanup. “We need billions of tons of carbon dioxide removal to meet our climate goals,” he says. “And there’s also scientific consensus that having PFAS in the food system is something that poses real health concerns at a broad public scale . . . I think most folks would say that we have inadequate funding moving to both of these.” Litigation could also potentially help fund PFAS remediation on farms. The number of lawsuits is growing; the State of New York, for example, recently sued major PFAS manufacturers for allegedly hiding what they knew about the health and environmental harms of the chemicals. Previous settlements have sent billions of dollars to communities to begin removing PFAS from drinking water. Outside funding is critical, farmers say. “If this solution is going to work, it’s going to have to be something that’s funded,” says Robinson, the organic farmer. And, she says, it’s something that needs to happen. “My big takeaway from all of this is that this is a problem that is absolutely everywhere. Towns, states, countries—everybody needs to be paying attention to it and putting money towards it. . . . It’s a huge public health problem, because it’s going to affect everybody. We all need to eat.”
This article is republished with permission from Wonder Tools, a newsletter that helps you discover the most useful sites and apps. I’ve changed how I use Claude and ChatGPT lately. They’ve become my digital dashboards, linked to many of the tools I rely on. I can ask Claude to add to my calendar, query my meeting notes, update me on my email, or find a passage I highlighted in my digital notebook. Claude Connectors and ChatGPT plugins connect to hundreds of services. These include Superhuman Mail for email, Google Calendar, Granola for meeting notes, and Readwise for my online highlights. Read on for how and why to link your favorite tools to your AI assistant, and examples of connections and queries I’ve found useful so far. Why Connecting Tools to Claude and ChatGPT Is So Useful Claude and ChatGPT know the context for my work because I’ve set up Projects for my primary work areas. Each has detailed instructions and reference documents. And the AI has a memory of the work I do, my style, my preferences, and my objectives. So I can ask my AI assistant to: Plan: Add four prep sessions to open morning slots on my calendar next week for the classes I’m teaching this fall. Include in each event description relevant excerpts from my meeting summary and reference the recent planning notes I dictated. Research: Gather recent links I’ve saved to Raindrop, my bookmarking tool, that might be relevant to my team’s research. Then create a new reminder with those links and a summary of my recent related notes. These multiple-part commands require context. They won’t work in Google Calendar, Raindrop, or Granola directly. That’s why it’s so helpful to have Claude or ChatGPT as the command center. MCP is the New USB These connectors have a technical name: MCP, or Model Context Protocol. But just as you don’t need to think about USB details when you plug in a printer, or http specifications when you use the Web, you can use MCP without understanding its technical details. All you need to know is that MCP is a way of linking software tools, sites, and apps to AI platforms like Claude and ChatGPT. Remember the world prior to USB? Printers, hard drives, cameras, and external monitors had their own special cords to connect to your computer. I just cleaned out dozens of them from a drawer in my Manhattan apartment. Now they’re obsolete because USB ports are everywhere. MCP is a similar unifying standard for linking tools to AI. 5 Ways I’m Using Tools Connected to Claude Manage Email I rely on Superhuman Mail to read, organize, and respond to email. It’s fast, easy to use, reliable, and integrates with my calendar. I use it with three distinct email addresses. Because I have to act on important messages, I’ve linked it to Claude. Get started: The connection guide walks you through linking your email to Claude or ChatGPT. You can then ask your AI assistant to: “Help me turn a bunch of starred emails into an afternoon plan.” “List important messages I’ve sent this month that the recipient hasn’t yet opened.” “Draft concise responses to routine customer questions for my review.” Some of these tactics will also work with Gmail and Outlook connectors. Example query: “Look at my five most recent starred emails, suggest an action list for this afternoon based on my priority list, then put an hour on my Google Calendar to act on them.” Do More with Meeting Notes I use Granola as my notepad during meetings. It transcribes conversations, then creates a summary based on the transcript and my own notes. (See my guide). I can already ask questions about my meetings in Granola without needing a separate AI tool. But the advantage of linking Granola with Claude (or ChatGPT) is that Claude has access to insights in my notes, email, and Google Drive, so it can integrate the info it pulls from Granola within the context of a broader analysis. Get started: The Granola MCP guide explains how to link it to your AI. Unlike many other paid services, Granola actually lets free users query meeting notes (for the past 30 days) using your AI tool of choice. Example query: “Give me a summary of the research questions I said I’d follow up on in my meeting with Pat. Next to each task include a bulleted list of the background details I already collected in my research spreadsheet.” Edit Images and Videos Photoshop and Premiere are powerful but complicated. Now you can edit an image or video without even opening a separate app. Set up the plugin, then tell your AI assistant to use it in your prompt. The AI assistants are often clever enough to figure out which tool to use even if you don’t explicitly mention it. Get started: Install the Adobe app in ChatGPT. Or add it to Claude, using this guide. Your AI assistant will now have access to the capabilities of Photoshop, Premiere, Firefly, Express, Acrobat, and other Adobe apps. You don’t even need an Adobe account to get started, though it might be helpful to get a free one if you want to organize, improve, and act on your creations later. (Canva’s MCP is also useful). Example query: “Help me change the text in this graphic.” Or “Reformat this horizontal video clip for YouTube Shorts or Instagram Reels.” Or “Add background blur to this headshot, fix the lighting, and crop it for a portrait.” As a test, I changed the first line in an illustration from “Bold” to “Kind.” Notice that the first attempt wasn’t perfect. The period after “Kind” was missing. But a quick edit request fixed that. The Photoshop edit preserved the font, style, spacing, and angle. The benefit: Instead of hunting through Photoshop menus trying to figure out how to do this, I just asked ChatGPT to use Photoshop to make a change. In the long run, being able to use natural language to accomplish tasks like this may shift how we spend our time away from technical menu-hopping toward more creative work. We’ll spend more time thinking about what to do and why, rather than how to technically get it done. Find Highlights I use Readwise to collect the passages I highlight when reading online or on my Kindle. It also hosts passages I save when listening to podcasts with the Snipd app. Get started: The Readwise MCP lets you make use of the highlights you’ve made while working on other projects with an AI assistant. Example query: “Show me a bunch of my Readwise highlights that include descriptions of desserts, especially from European novels, then gather a list of related recipes I can make with my daughters before our trip.” Research Legal Cases Court Listener from the Free Law Project is a handy way of gathering info about court cases. You can ask for records related to Elon Musk’s recent testimony without paying for a legal research service. Get started: Use the Court Listener MCP guide for free with Claude or ChatGPT. First set up a free Court Listener account. Example query: “Review the news article at XYZ link, find the case that’s being discussed, and give me additional details about the case and any other related recent cases.” Other Useful MCP Tools Substack has a new MCP, available to select publishers as of now. It lets Claude coach me on my analytics. I recently asked for an analysis of send times to figure out how they impact open and click rates. Substack’s help page notes that the MCP has “read-only access to publication data such as dashboard metrics, traffic data, and publication settings. It cannot publish posts, send Notes, or modify your account.” Rize is my favorite tool for time-tracking. I linked it to Claude so I can learn from how I’m spending my time and assess that alongside info from my meetings, email and calendar. The Rize MCP page includes use-cases and helped me get started. Rize requires a paid subscription. I use Raindrop to save links I’ll need later. I checked the Raindrop MCP page to set it up and learn about the kinds of queries I can run from Claude. When I’m doing research in a project, for instance, I can ask it to pull in relevant links I’ve saved. Or I can even ask Claude to put the links from a spreadsheet I’m working on into my Raindrop bookmark collection with appropriate tags. Sublime is another tool I use to save quotes, images and useful pages online. Unlike Raindrop and other well-designed clipping tools like MyMind, which are fully private, Sublime lets me see related material other people have saved. Now I can ask Claude to pull in relevant material from Sublime. How to Link Tools to Your AI Assistant Decide what kind of connector will be useful to you: Email: Gmail, Outlook, Superhuman, and other email services now let you analyze and respond to your messages. Files: Analyze or act on files in your Google Drive, Dropbox, Box, Slack, Notion, or anywhere else you keep your documents and files. Planning: Add events to your Google or Outlook calendar, or organize projects with tools like Airtable, Monday, ClickUp, Asana, and Linear. Fun: Create Spotify playlists based on your work vibe or search StubHub or SeatGeek for tickets your family or friends might enjoy. Pick from your AI tool’s curated list. Claude and ChatGPT each have a bunch of services pre-configured to work. Perplexity now does too. Popular ones include Google Drive, Gmail, Google Calendar, Canva, Figma, and Notion. Paste in the address of a custom connector. For tools that are not on the default list, you put in an MCP address, which is like a special website to connect an external service to your AI tool. Find more to try in a directory. Be Careful About Linking Private Data Think twice before connecting services that host sensitive info. I haven’t linked my bank accounts to my AI tools, for example. If you do that and someone gets access to your AI account, they’ll be able to find out a lot about your finances. It can happen when a thief uses a “prompt injection” to manipulate an AI tool into releasing sensitive info. Or it can happen if your phone is lost or stolen. New York, London, Paris, and other major cities see thousands of phones stolen each year. Data can also leak from AI systems. It’s rare, but there have been cases like this one. Be Cautious About Letting AI Tools Act on Your Behalf I’m not yet letting AI tools send emails for me. Or submit forms or applications. If Claude or ChatGPT generates something odd because of a glitch, or because I rushed a query, I want to be the one noticing the error. When you start connecting AI tools to external services, you have the option to give your AI assistant power to do things on your behalf. That can be useful for research, summarizing meeting notes, or cleaning up messy computer folders. But you or your AI assistant could make a costly mistake with important info. In a rare but alarming case, Cursor, an AI tool, deleted PocketOS’s company database and its backup. A related 2025 case involved Replit. Other Ways Connected AI Can Go Wrong Travel booking blunder. You ask it to cancel a reservation, and it cancels your entire itinerary. Or you ask it to book a flight on 7/1/27 and it books a nonrefundable ticket on 1/7/27, assuming you were using the European date system and meant January 7, not July 1. Disappearing Files. You ask it to get rid of duplicate files. It deletes an entire folder with a similar name. Calendar Chaos: You ask it to “find a time for everyone” and it reschedules an existing meeting in the wrong time zone, because you were traveling when you wrote the request. I haven’t encountered any of these errors. But these services are still new, so you may encounter occasional bugs or blunders. This article is republished with permission from Wonder Tools, a newsletter that helps you discover the most useful sites and apps.
A young mathematician teamed up with a new AI model to tackle one of the hardest open problems in mathematics, and the result is causing a stir among mathematicians and AI researchers on X. Levent Alpöge, a 33-year-old researcher at Harvard University, spent Sunday working with Anthropic’s Fable model to defeat the Jacobian Conjecture, a notoriously difficult problem proposed by German mathematician Eduard Ott-Heinrich Keller in 1939. The challenge for Alpöge and Fable was to find a counterexample proving the conjecture false (a single verified counterexample would be enough). Alpöge casually announced the result on X, thanking both the friend who encouraged him to take on the problem and Anthropic’s Claude Fable, the AI model that helped him do it. hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x – 3 x^2 y – x^3 z): C^3to C^3,… — levent (@__alpoge__) July 20, 2026 The conjecture proposes that if you start with a point on a grid represented by ordinary variables (say, x and y), then use polynomial equations—equations involving powers such as x² and y³—to create new coordinates, you should be able to reverse the process using polynomial equations and recover the original coordinates. (For a deeper explanation, visit the Wikipedia page.) Alpöge’s counterexample is just 216 characters long. Although the work has yet to undergo the normal peer-review process, several mathematicians have reportedly confirmed the arithmetic and conducted independent checks using SymPy (a Python library for symbolic mathematics) and Lean (a formal proof assistant). Others then used additional AI systems, including OpenAI’s GPT models, to conduct their own fact-checks (sometimes to unintentionally hilarious results). lmao pic.twitter.com/R8rbV7lHwM — Andrew (@andrew_v10209) July 20, 2026 The result has prompted debate among mathematicians about AI’s role in their field. Systems that can rapidly generate counterexamples may accelerate discovery while shifting researchers’ attention toward explaining why a result is true and what it means. AI models are increasingly being used as collaborators in mathematical research rather than simply as computational tools. Researchers use them to suggest proof strategies, fill in missing logical steps, and check arguments for errors—allowing ideas to be explored more quickly. AI is also accelerating the growing field of formal verification by translating proofs into computer-checkable languages such as Lean, which can verify complex theorems with a high degree of certainty. These systems can also analyze large amounts of mathematical data, identify unexpected patterns, and propose new conjectures or avenues of investigation. Alpöge’s discovery, if verified, may offer some of the strongest evidence yet for a claim frequently made by AI executives—that advanced models will significantly accelerate mathematical and scientific discovery. It may also challenge critics who argue that transformer-based systems (including GPT and Claude), because they are probabilistic rather than deterministic, cannot be useful in exacting fields such as mathematics and science.
It wasn’t long ago that pundits predicted OpenAI’s advertising business could eventually overshadow those of Meta and Google. The company was similarly upbeat. But a new report on the early numbers suggests that optimism may be misplaced. The study, from Emarketer, estimates that OpenAI’s U.S. ad revenue will fall 90% short of its five-year target. The firm did not forecast OpenAI’s global revenue. OpenAI had projected ad revenue of $2.5 billion this year, with a goal of more than $100 billion by 2030. Emarketer’s research, however, estimates that the combined ad revenue of standalone chatbots, including ChatGPT, Microsoft’s Copilot app, Google’s AI Mode, and Amazon’s Alexa for Shopping, will total less than $1 billion this year. By 2030, Emarketer estimates that advertising across all chatbots, not just OpenAI’s, will generate only $5.41 billion. It is the latest development in OpenAI’s complicated relationship with advertising. Mixed messages Most observers had long assumed that OpenAI would eventually enter the advertising market to increase its revenue. CEO and co-founder Sam Altman cast doubt on that thinking in 2024, however, when he said “I will disclose, just as a personal bias, that I hate ads” during a fireside chat at Harvard University. Ads, he continued, “fundamentally misalign a user’s incentives with the company providing the service,” and the idea of mixing advertising with OpenAI’s products was “uniquely unsettling.” A little more than a year and a half later, however, OpenAI began testing ads in the free version of ChatGPT, as well as in the lower-cost ChatGPT Go. “The best ads are useful, entertaining, and help people discover new products and services,” the company wrote in a blog post at the time. “Given what AI can do, we’re excited to develop new experiences over time that people find more helpful and relevant than any other ads.” Last month, four months after launching the trial, OpenAI was promoting advertising as a way to broaden access to its products. “The revenue that we make from the ads offering is going to subsidize and grow access to information,” OpenAI advertising chief David Dugan said at the Cannes Lions International Festival of Creativity. The $100 billion target has always been ambitious. Meta has been building its advertising business since 2007, and that figure represents roughly half of the Facebook and Instagram parent company’s current ad revenue. Reaching it would require advertisers to substantially alter their strategies, shifting money away from search engines and social media and into AI systems. OpenAI would also have to persuade advertisers that its platform offered a better use of their budgets than Meta and Google, both of which have AI products of their own. An IPO on the horizon The projected shortfall comes as OpenAI moves closer to an expected initial public offering. The company confidentially filed paperwork with the Securities and Exchange Commission in early June, setting the stage for what could be one of the largest public-market debuts in history. Although there was speculation that OpenAI might accelerate its IPO to beat Anthropic to market, the company has instead proceeded cautiously. It is now reportedly targeting a 2027 launch, which could give it more time to boost its valuation and allow the recent market volatility to subside. A severe shortfall in advertising revenue could create another obstacle as OpenAI tries to build investor enthusiasm. That task may already be difficult given the company’s financial losses and the early performance of competitor SpaceX. Shares of that company have fallen below their IPO price and are down more than 40% from their post-IPO high.
Growing up in Las Vegas, Hunter Manz took annual trips to nearby Lake Mead with his father. Each year, he watched the reservoir, Southern Nevada’s primary water source, recede further, sparking an early interest in water conservation. As a teenager, Manz began experimenting in his garage with potential solutions to water scarcity. He later enrolled at Utah Tech University, studying mechanical engineering and finance, and learned about desalination, which removes salt and other contaminants from water so it can be used for drinking or industrial purposes. While in school, Manz began building desalination prototypes. There, he developed what he calls a reverse osmosis centrifuge, or ROC, a first-of-its-kind system that he says is two to three times more energy efficient than conventional desalination technology. In 2020, Manz dropped out of college to found Eden Tech, a startup seeking to make desalination cheaper and less environmentally damaging. The company has since raised $3.3 million from Utah Tech, community investors, and venture capital firms. Now based in Lockhart, Texas, Eden has just unveiled Genesis, its first full-scale reverse osmosis centrifuge. Over the next year, the company plans to refine its model, scale to consumers, and begin commercial use. Once up and running, each Genesis machine is expected to process approximately 144,000 gallons of water per day. Desalination has long drawn criticism for its high energy use and the concentrated brine it leaves behind. Eden is betting that a more efficient system, combined with new uses for that waste, can address both problems. “It’s a technology that is not only producing the water that people want,” says Manz. “It’s helping the environment as well. So it’s kind of a two-for-one benefit.” Old technology, new approach Early desalination systems relied largely on heat to evaporate water and leave salt behind. Beginning in the 1950s, researchers increasingly turned to reverse osmosis, which uses pressure to force saltwater or wastewater through a semipermeable membrane. The process produces two streams: clean water and a concentrated brine. Reverse osmosis has helped drive a desalination boom in the Middle East, particularly in Saudi Arabia and the United Arab Emirates. In Israel, desalination now supplies much of the country’s drinking water. The technology remains energy intensive, however. Conventional systems use roughly 15,000 kilowatt-hours of electricity for every million gallons of freshwater produced, according to Bloomberg, and typically recover only 30% to 40% of the water they process. Manz’s patented ROC system is designed to improve those numbers by spinning the membrane. Traditional reverse osmosis functions like a high-pressure strainer, forcing water through a filter while leaving salt behind. Eden’s system works more like a salad spinner, using centrifugal force to help separate salt from water. “The phenomenon that is created on the surface, the centrifugal force, allows us to really break that layer and push the envelope of what we could do with desalination,” says Nikolay Voutchkov, CEO of the engineering firm Water Globe Consultants. “Some of the work of separation of salts is done by natural forces rather than by pressure, and we’re achieving lower energy demand.” Manz says the system can recover as much as 80% of the water it processes while using two to three times less energy than conventional systems. Its ability to handle highly saline water could also open a market beyond seawater desalination. Eden is targeting produced water, the salty and often contaminated wastewater generated during oil and gas extraction. The company says it has already signed agreements with several large oil and gas businesses. “The thing I get up for every morning is to have a piece of technology where that actually does happen,” says Manz. “Unlocking this [zero liquid discharge] ability, and then applying it to the broader municipal side of things—hopefully that will bring upon a reality where the clean water is then the only waste stream.” Turning waste into revenue Eden’s pitch has attracted investors. The company has raised $3.3 million and hopes to move from research and pilot projects toward large-scale commercial deployments. Texas offered an appealing place to start. The state faces recurring water shortages, has a large oil and gas industry, and attracts significant investment in energy and infrastructure projects. Still, desalination projects can be extraordinarily expensive. Corpus Christi, Texas, recently shelved plans for a desalination facility after its projected cost exceeded $1 billion and the proposal faced opposition from residents. Eden hopes to improve the economics through brine valorization, a process that extracts commercially valuable materials from the concentrated waste left after desalination. Depending on the water source, those materials could include lithium and rubidium, which are used in batteries, as well as salt. That would allow Eden to generate revenue from both the clean water and the remaining minerals, potentially lowering the cost for customers. “The end goal is to be at a stage in the not too distant future where we’re generating billions of dollars in annual recurring revenue,” says Manz. “That’s our main base where we can finance like our growth into municipal desalination and be able to do projects that maybe typically wouldn’t be able to be done financially.” Reducing desalination’s environmental costs The financial challenge is only part of desalination’s troubled history. Critics have also raised concerns about greenhouse gas emissions, particularly when facilities rely on fossil fuels to power high-pressure pumps. Eden says its lower energy requirements could reduce those emissions, particularly when the centrifuge is powered by renewable sources such as solar energy. Brine disposal poses another environmental risk. When concentrated saltwater is discharged back into the ocean, it can damage marine ecosystems that are unable to tolerate sudden changes in salinity. For every 95 million cubic meters of freshwater produced by desalination facilities, an estimated 141.5 million cubic meters of brine is generated, according to a study published in Science of the Total Environment. Eden is trying to limit that waste by focusing initially on produced water that might otherwise be injected underground and by recovering valuable minerals from the remaining brine. The goal is to reduce the volume of waste enough that it can be sold or reused rather than discarded. “Our goal with our technology is to purify it to a point where there’s very little waste,” says Manz. “And then it’s such a small volume that you can actually go and sell that to other industries, so you’re making the waste valuable. That way people don’t want to throw it into the environment.” From industrial projects to public water Even if Eden’s technology works as promised, expanding into municipal water systems will require more than technical progress. Water rights, permitting, utility regulation, and public concerns about privatization can all complicate desalination projects. Manz says he would like to see greater federal support for new water technology, but much of the authority rests with state and local governments. “The state has a very heavy role in the outcome of who gets the water, and they are very careful to make sure that plant developers are put in their place,” he says. “This isn’t a situation where you can just build a private plant and use it as leverage. The state of California would be like, ‘No, we’re not going to buy your water anymore.’” Manz acknowledges concerns about private companies controlling water supplies but says Eden’s long-term goal is to make desalinated water more widely available and affordable. The company is working with chapters of the Navajo Nation in Utah to access underground brackish water and provide a more reliable source of clean water. Manz sees the project as an opportunity to bring water infrastructure to communities that have often been neglected by governments and private industry. “We want to take water that is currently hurting the environment and make it something that will actually help the local community,” says Manz. “Hopefully we will be the first company that can make desalination not only exciting for business people and venture people, but for the average person here in the U.S.”
The Trump administration has had plenty of targets in its sights in recent months. The newest are China’s cutting-edge AI models. Following the release of Kimi K3, the latest model from Moonshot AI, the White House is reportedly considering options that could effectively ban models like it, along with Alibaba’s Qwen model, the latest version of which was released just days after K3. It remains unclear what form such a ban would take, according to Axios. The U.S. Commerce Department has previously considered adding the laboratories behind leading Chinese models to its “Entity List,” which could prohibit Americans from possessing or using those models without a license. The uncertainty comes amid signs of instability within the administration’s AI regulatory apparatus. Chris Fall resigned after just three months as head of the Center for AI Standards and Innovation, according to The Daily Caller. Washington appears increasingly concerned about the speed with which Chinese models have narrowed the U.S. lead in AI inference. Kimi K3 has performed at or near the frontier on prominent coding benchmarks. Chinese developers are also competing on a quality that most leading American laboratories do not offer: control. Open-weight models can be downloaded, adapted, and run on an organization’s own infrastructure rather than accessed exclusively through a U.S. company’s servers. By attempting to slow China’s progress and prohibit access to the open-source and open-weight models its companies are producing, the U.S. risks alienating countries and organizations that are uneasy about relying on products controlled by a small number of American AI companies. China is already trying to capitalize on that concern. Last week, Chinese President Xi Jinping launched the World Artificial Intelligence Cooperation Organization in Shanghai during a global AI gathering attended by representatives from 29 countries. Many of those countries, including Cuba and Russia, are established Chinese allies that were unlikely to align closely with the U.S. Others, including Brazil and Indonesia, are important emerging powers whose closer alignment with China could weaken U.S. influence over time. The following day, Xi used Shanghai’s World Artificial Intelligence Conference to present China as a champion of open access. He promised training and technical assistance to developing countries while warning against allowing AI to become another source of global inequality. Many governments and organizations may be unwilling or unable to send citizens’ data into an American cloud. In those cases, a capable model that can be run locally becomes highly attractive. Chinese laboratories are offering models that are cheaper, customizable, and available for deployment on local infrastructure. Banning those models in the U.S. will not make them less appealing elsewhere. It would instead suggest that the U.S. supports openness only until a foreign competitor becomes sufficiently capable. A prohibition focused on individual models could also weaken the broader U.S.-aligned AI ecosystem, including developer tools, local hosting providers, and the technical skills required to deploy AI systems. Over time, countries may find it more practical to build around Chinese technology than American platforms. A decision about which chatbot or model to use can eventually produce a broader dependence on Chinese infrastructure, standards, and expertise. That risk extends beyond the countries participating in China’s new organization. Many AI developers already use and value open-source and open-weight models because they offer flexibility that closed American systems do not. That is why a ban could prove self-defeating. The U.S. may succeed in keeping Chinese models away from American users in the short term. In trying to defend its technological lead, however, Washington could reduce the size of the international market in which that lead carries influence and make China’s claim to be the more open AI power appear increasingly credible.
Technology leaders are very comfortable talking about productivity. We can discuss automation rates, developer velocity, cost curves, and defect rates with great confidence. We are less comfortable talking about mental strain, especially when the thing creating the tension is the very thing we are excited to adopt. Artificial intelligence agents will make that discomfort harder to avoid. Agents can write code, triage issues, research options, test hypotheses, and perform many of the tasks that previously consumed human time. For software teams, this can feel like a major shift in leverage. For many other knowledge workers, the same pattern applies. The human stops being the sole producer and becomes something closer to an operator: setting direction, reviewing output, correcting course, and integrating the results. That sounds like clean productivity, but it has a cognitive cost we are not yet measuring. Running AI agents feels like listening to a podcast at 2x speed. It is intelligible enough to be useful, fast enough to feel powerful, and dense enough that you eventually notice your brain heating up. You can do it for bursts and even convince yourself that it is efficient. But over time, there is less space to reflect, less time to absorb context, and fewer natural pauses between decisions. The old burnout metaphor was burning the candle at both ends. Agentic work is lighting it from both ends and a few places in the middle, and then calling the brightness productivity. Managing Agents Is Still Management The first mistake leaders make is assuming AI simply removes work. In practice, it dramatically changes the shape of work. Anyone who has managed a team understands the pattern. A person comes by with a question. You provide direction. Later, someone sends work for review. Another person hits a blocker. You clarify the goal, resolve ambiguity, or help them prioritize. That can be tiring, but it is usually spaced out. Human teams have natural latency. People go away and work. Meetings happen at intervals. Feedback cycles have friction. There is room, however imperfect, for the manager to return to their own thinking. Now imagine your entire team sitting on your desk, never leaving the room, asking for feedback every 10 minutes. That is what operating a fleet of AI agents can become. With multiple agents running in parallel, the work looks less like coding and more like supervising a small team that never sleeps or eats. Yes, computation sped up. But judgment did not, and recovery disappears. Let’s stop pretending that supervising AI is free. Reviewing work, correcting direction, maintaining context, and absorbing ambiguity are management tasks, even when the “team” is made of software. The Risk of Attention Fragmentation This is where the conversation needs to shift, or an AI efficiency project quickly becomes an attention-fragmentation project. Research on attention residue, workplace interruptions, digital notifications, and techno-stress has shown for years that fragmented work carries a cognitive and emotional cost. People may compensate by working faster, but the bill comes due in stress, fatigue, and degraded judgment. The operator is not simply interrupted by meetings and notifications. They are interrupted by productive systems. That makes the interruption harder to resist. A Slack message can be ignored. An agent that just generated a patch or found a design flaw is hard to ignore. The result is a new form of cognitive pressure: endless microdecisions wrapped in the language of productivity. The operator is asked to approve one direction, reject another output, rerun a task, check whether an answer is hallucinated, and decide whether an architectural suggestion is clever or dangerous. Then they do that again. And again. The result is that the operator is making nothing but judgment calls, all day, at machine tempo. Judgment is mentally taxing. It draws on memory, context, experience, intuition, and emotional regulation. It requires enough quiet to notice when something feels wrong. If every workday is a high-frequency review loop, we should not be surprised when people feel cognitively overloaded, irritable, and depleted. The New Burnout Will Look Different In the AI era, burnout may not show up through longer hours, full calendars, or the obvious misery of too much manual work. Those problems will still exist, but agentic AI introduces a different pattern: burnout caused by supervising too many streams of automated work at once. From a distance, the system may look healthier. More code is generated. More tickets move. More drafts appear. Up close, it feels like a day spent inside a room full of unfinished thoughts, each one tapping you on the shoulder. AI operator burnout is built from open loops that seem trivial individually but become exhausting in aggregate. People can end the day completely spent, even if the day does not look punishing from the outside. That exhaustion will not stay neatly contained inside the workday. Mentally depleted people have less capacity for careful decisions, recover more slowly, and have less patience for the human parts of work, like mentoring, collaboration, disagreement, judgment, and care. At a business level, that can show up as lower-quality reviews, brittle systems, and higher attrition. At a societal level, we risk normalizing an economy where human attention is continuously mined by systems mistaking extraction for productivity. AI does not eliminate the economics of comprehension. Someone still needs to understand the problem, make trade-offs, and decide whether the output is secure, lawful, ethical, maintainable, and aligned with the business. The agent can produce the artifact, but it cannot yet own the blast radius. That means the real constraint may be how many concurrent streams of machine-generated ambiguity a human can responsibly supervise before quality, judgment, and mental health begin to degrade. Designing Humane Agentic Work Enterprises should absolutely embrace AI agents (with limits and controls). Concurrently, they must design the human side of agentic work with the same seriousness and pragmatism applied to the technology. Limit agent concurrency. More agents are not always better. There is a practical limit to how many active work streams one person can supervise well. Organizations need norms for responsible concurrency, especially in high-stakes domains. Batch review cycles. If every agent can interrupt a human at any time, the operator becomes a notification endpoint. Agent workflows should be designed around review windows, summarized outputs, and bounded decision points. Separate creation, review, and integration. Asking the same person to continuously prompt, evaluate, secure, merge, and architect is a recipe for cognitive overload if treated as one continuous stream. Protect deep work more aggressively. If AI increases the volume of output, humans need more synthesis time, not less. They need space to think about the system, not merely respond to its artifacts. Design for recovery, not just throughput. A day filled with agent review may look efficient, but sustained judgment requires recovery time. Teams need space after high-intensity review cycles to synthesize, document, and reset context. Measure cognitive load alongside velocity. Ask teams where AI is helping and where it is creating review fatigue. Track confidence in output quality. Watch for signs that people are approving work faster than they can understand it. The Leadership Obligation The future of work will not simply be humans plus AI, and it should not be reduced to a lazy story about replacing people at scale. If anything, it is the opposite: Human attention will be more important when automated execution surrounds it. As a result, attention becomes a strategic asset. Mental health becomes an operational concern. Technology has a long history of mistaking acceleration for progress. AI agents may become one of the most powerful examples yet. If we are not careful, they can recreate the worst parts of always-on work at machine speed. If we ignore that, the first major cost of agentic AI will be exhausted people inside systems that appear, from the outside, to be working beautifully. Managing agents is still management. Operating intelligence is still work. Protecting human judgment may become one of the most important leadership responsibilities of the AI era.
For the past two years, companies have been trying to add AI to their processes. That sounded reasonable at first. Existing workflows, existing systems, existing data, existing dashboards — now with intelligence attached. Add a copilot here, an agent there, a few automated steps, a model connected to some tools, and wait for productivity to rise. But this approach is starting to reveal its limit: The problem isn’t that AI can’t act. It increasingly can. The problem is that most companies don’t have a formal representation of what their actions mean. A company isn’t a pile of applications. It’s not a CRM, an ERP, a data lake, a Slack workspace, a set of spreadsheets and a collection of dashboards. A company is a causal system: customers, products, contracts, prices, suppliers, employees, approvals, incentives, constraints, risks, processes, and outcomes, all influencing one another over time. If AI is going to optimize the company, it first needs to know what the company is. That’s why the next frontier in enterprise AI isn’t the agent. It’s the ontology. From data to ontology The word ontology can sound unnecessarily philosophical, but the idea is simple: It is a formal model of what exists in a domain and how those things relate to one another. In enterprise terms, that means representing the company not as disconnected rows in databases or documents in repositories, but as a living structure of objects, relationships, permissions, workflows, and actions. A customer has contracts. A contract has terms. A product has dependencies. An order changes inventory. A delay affects satisfaction. A discount affects margin. A support interaction affects retention. That structure matters because AI systems cannot govern or optimize what they cannot represent. This is one reason Palantir’s Ontology has become such an important reference point in enterprise AI. Palantir describes its Ontology, like that, with a capital “O” as if it were a word they’ve somehow invented themselves, as a system that goes beyond data cataloging or schema design, providing a foundation for workflows with metadata, security, and governance. Its architecture material describes the Ontology as the “dynamic, compounding core of the cybernetic enterprise,” connecting data, logic, and action into a shared operational model for humans and AI-enabled agents. That is a significant shift. It moves enterprise AI away from “chat with your data” and toward something more serious: AI acting inside a formal model of the business. But once a company has an ontology, a deeper question appears: What is the ontology for? Static ontology is not enough A static ontology tells the system what exists. That is useful. It creates legibility. It lets software and humans refer to the same objects. It makes workflows less dependent on ad hoc integration. It gives agents something more structured than a prompt and something more reliable than a pile of retrieved documents. But companies are not static. A company changes every minute. A customer moves from prospect to active account to renewal risk. Inventory changes. Credit exposure changes. A supplier becomes unreliable. A sales process works in one segment and fails in another. A support policy improves cost metrics and quietly damages trust. A pricing rule increases margin and lowers long-term retention. The real company isn’t the noun. It’s the verb. This is why the distinction between static and dynamic ontology matters. Palantir’s own documentation distinguishes between the semantic elements of the ontology (objects, properties, and links) and what it calls the “kinetics of the organization,” defined through action types and functions that enable change while complying with controls and governance. That is already much closer to what enterprise AI needs. But even dynamic ontology may not be the final step. A dynamic ontology can tell the system how the company operates. The next step is an ontology that can improve how the company operates. In other words: an optimizable ontology. The ontology should not just describe the company This is the turning point. If an ontology represents the structure of the company, and if AI systems act through that structure, then the ontology is no longer just a map. It becomes part of the machinery of the company itself. That means the ontology should not merely describe operations. It should allow operations to be optimized. This is where most enterprise AI discussions still fall short. They treat ontology as a semantic layer: a way to connect data, define entities, clarify relationships, and give agents a safer environment in which to act. All of that is necessary. But it is not sufficient. The real opportunity is to turn the ontology into an optimization substrate. Every workflow represented in the ontology should be connected to outcomes. Every action should leave a trace. Every trace should be usable as feedback. Every process should have an explicit objective. Every objective should be measurable. Every measurable outcome should allow the system to learn which actions, configurations, and sequences improve results. At that point, the company is no longer simply using AI: The company is becoming optimizable. From workflows to causal structure Most business processes are still treated as diagrams: boxes, arrows, approvals, handoffs, and exceptions. That’s how humans understand work. It’s not how adaptive systems optimize it. An AI system needs more than a diagram. It needs causal structure. It needs to understand that reducing time in one step may increase rework in another. That shortening a support call may increase churn. That discounting may win the customer but lower the quality of revenue. That automating an approval may increase speed but reduce accountability. That optimizing one department’s KPI may damage the company’s global objective. This is not a cosmetic distinction. A recent article in Nature Computational Science argues that reliable algorithmic decision-making needs causal reasoning, because decisions inherently involve cause-and-effect relationships and must align computational methods with real-world objectives. That is exactly the enterprise problem: A corporate ontology cannot stop at representation. It has to encode consequences. The most interesting enterprise AI systems will not merely know that a sales opportunity exists, or that a contract is pending, or that a customer has opened three tickets. They will understand how actions on those objects change the probability of outcomes the company cares about: conversion, retention, margin, risk, satisfaction, cycle time, resilience. That is the difference between a semantic ontology and a causal ontology: A semantic ontology tells the system what things mean; a causal ontology tells the system what actions do. And an optimizable ontology goes further: It learns which actions work. KPIs become reward signals Companies already have something that looks like an objective function: KPIs. These are important metrics such as revenue, margin, churn, conversion, customer satisfaction, resolution time, inventory turns, forecast accuracy, fraud loss, compliance incidents, employee retention, time to market, and more. The problem is that in most companies KPIs are downstream measurements. They tell managers what happened after the fact. They’re reported, discussed, explained, occasionally gamed, and then reviewed again next quarter. In an optimizable ontology, KPIs become something more powerful: reward signals. This doesn’t mean blindly optimizing every metric. That would be dangerous. Metrics can conflict. Some are proxies. Some are incomplete. Some are political. Some produce perverse incentives if pursued alone. But that is precisely why the ontology matters: A KPI should not float alone in a dashboard. It should be attached to the process, objects, constraints, and decisions that influence it. It should be placed inside a model of the company that understands trade-offs. It should be connected to other KPIs so that local improvement does not produce global damage. This is where reinforcement learning becomes relevant to enterprise AI: not as a buzzword, and not as a magic layer bolted onto chatbots, but as a mechanism for improving action inside a formally represented business system. A loop acts. The ontology records what changed. The KPI measures whether the outcome improved. The system adjusts. The next action is better informed. That is the basic shape of an optimizable company. Scale changes the meaning of optimization Traditional process improvement is slow because humans have to observe, interpret, redesign, implement, and measure. That works, but it doesn’t scale well across thousands of processes, millions of interactions, and constantly changing conditions. AI changes the speed of the loop. Once enterprise actions are represented in an executable ontology, each interaction becomes an experiment. Each process execution produces a trace. Each trace becomes evidence. Each evidence updates the system’s understanding of what works. The larger the company, the more interactions it generates. The more interactions, the more feedback. The more feedback, the faster the optimization. This is the opposite of how most enterprise systems behave today. In traditional software, scale creates complexity. More customers, more workflows, more exceptions, more countries, more products, and more regulations make the system harder to change. In an optimizable ontology, scale also creates learning fuel. The company becomes more complex, yes, but it also produces more evidence about how that complexity behaves. That’s why the ontology must be executable. A descriptive ontology can help humans understand the business. An executable ontology lets AI act inside the business. An optimizable ontology lets the business improve through action. Those are three very different stages. The missing layer above agents This is why the current obsession with agents is incomplete. Agents are useful. They can plan, call tools, write code, search documents, take actions, and coordinate with other agents. And the conversation is already moving from individual prompts to loops: a recent Business Insider piece describes “loop engineering” as the practice of designing recurring systems that guide AI agents instead of requiring a human to prompt every step manually. That is a meaningful shift. But agents and loops still need a world to act inside. Without an ontology, each agent reconstructs the company from prompts, retrieved fragments, tool descriptions, and brittle integrations. That’s why deployments become artisanal. Someone has to explain the business over and over again: what the objects are, what the rules are, which systems matter, who can approve what, what outcomes count, and where the hidden constraints live. A mature enterprise AI architecture should not require every agent to rediscover the company: The ontology should be the shared world. And if that ontology is executable and optimizable, agents become components inside a larger adaptive system rather than isolated actors improvising their way through corporate reality. This also explains why model independence matters. The ontology is the durable asset. Models will change. Agents will change. Interfaces will change. But the company’s representation of itself—its objects, processes, constraints, outcomes, traces, and learned causal structure —should persist. That is where enterprise AI compounds. Digital twins were the preview There’s a useful analogy here with digital twins. NIST describes a digital twin as a computer model of a physical system that can support simulation, monitoring, optimization, and decision support. In a separate publication, NIST says digital twins enable operators to dynamically represent, diagnose, predict, optimize, and control real-world counterparts such as equipment, subsystems, and processes. That is very close to the intuition enterprise AI now needs—except the object is no longer only a machine, a factory line, or a building. The object is the company. A company needs the equivalent of an operational digital twin: not merely a replica of its physical assets, but a representation of its business structure, processes, constraints, decisions, and outcomes. Recent research on digital twins of business processes points in the same direction, describing virtual replicas of real processes that combine process models, real-time data, and simulation capabilities to guide day-to-day organizational activity. That is the direction of travel. But enterprise AI needs to go one step further. It does not simply need a twin that shows what’s happening. It needs a model through which humans and AI systems can act, learn, and improve. That is the optimizable ontology. World models move from physics to business The phrase world model is usually associated with robotics, autonomous driving, or physical AI: systems that need an internal representation of the environment in order to anticipate consequences and act intelligently. That makes sense. A robot that moves through the physical world needs to know something about objects, space, causality, and time. But companies are worlds too. They’re not physical worlds in the same sense, but they’re operational worlds: partially observable, constantly changing, full of agents, constraints, dependencies, incentives, and delayed consequences. An AI system that acts inside a company without a model of that world is like a robot moving through a warehouse without spatial awareness. It may be powerful. It is not safe. The idea of world models has been central to some of the most important work in reinforcement learning. David Ha and Jürgen Schmidhuber’s “World Models” paper explored training agents using compressed representations of their environments. Google DeepMind’s MuZero went further by learning a model of the environment it was playing in and using that model to plan the best course of action. The lesson is not that companies are games: They are not. The lesson is that intelligence becomes much more powerful when actions are connected to outcomes through a model of the environment. Enterprise AI needs the same principle, but applied to organizational reality. Not just “What answer should the model generate?” But “What action should the company take, through which process, under which constraints, toward which objective, and with what expected effect on the rest of the system?” That requires a corporate world model. The company needs to know what it is The hardest part of this transition may be cultural, not technical. Most companies don’t actually have a formal model of themselves. They have org charts, process diagrams, ERP configurations, CRM records, policy documents, data warehouses, dashboards, Slack channels, email archives, and thousands of implicit habits held together by people who know how things really work. That isn’t a world model. It’s an archaeological site. Humans compensate for this because they carry context in their heads. A good manager knows which policy matters, which exception is safe, which customer relationship is fragile, which process is official but ignored, which metric is being gamed, which team is overloaded, and which apparent success is hiding future damage. AI loops don’t know any of that unless the organization makes it explicit. This is why so many enterprise AI deployments still require consultants, integrators, and forward-deployed engineers. Someone has to reconstruct the company for the AI system: what matters, what’s connected, what’s allowed, what counts as success, and where the hidden constraints are. McKinsey’s State of AI in 2025 report makes the broader pattern visible: AI use is widespread, but most organizations have not embedded it deeply enough into workflows and processes to realize material enterprise-level benefits, while high performers are much more likely to redesign workflows and aim for broader transformation. That manual reconstruction is the sign of a missing platform layer. A corporate world model would make that reconstruction persistent, governed, and reusable. Every loop would not need to rediscover the company. Every new agent would not need to be individually briefed on organizational reality. Every workflow would not need to rebuild context from scratch. The company would finally become legible to its own AI systems. From governed decisions to self-improving structure Recent research is already moving in this direction. A 2026 paper on ontology-governed graph simulation for enterprise AI argues that LLM-based agent systems fail when they answer from unrestricted knowledge space without simulating how business events reshape the scenario at hand. Its proposed pipeline—event, simulation, decision—points toward a more structured approach: enterprise decisions derived from an ontology-governed representation rather than from fluent but ungrounded reasoning. That’s important because it shows where the frontier is moving. But the deeper question is not only whether ontology can make decisions safer. It’s whether ontology can make the enterprise itself self-improving. The real promise of enterprise AI is not that every employee gets a better assistant. That’s useful, but it’s not transformation. The real promise is that the company becomes a system capable of improving its own operations continuously. Not through occasional transformation projects. Not through annual process redesign. Not through consultants reconstructing the business one workflow at a time. But through an architecture in which work itself generates the evidence needed to improve work. That requires three layers. First, a formal ontology that represents the company Second, executable workflows that allow AI and humans to act through that ontology Third, optimization loops that connect actions to outcomes and improve the structure over time Remove the first layer, and AI has no stable world. Remove the second, and the ontology remains documentation. Remove the third, and the system cannot learn. Put them together, and enterprise AI stops being a tool attached to the company. It becomes a mechanism by which the company improves. The next question The last wave of enterprise software helped companies digitize what they were. The next wave will help them optimize what they’re becoming. That is a much bigger shift than chatbots, copilots, or agents. It changes the role of software from system of record to system of improvement. It turns KPIs from reporting artifacts into feedback signals. It turns workflows from diagrams into executable, learnable structures. It turns ontology (the real word and its real meaning, not the one capitalized as a commercial term) from description into action. And it raises a question every company will eventually have to answer: Is your organization merely represented in software, or is it optimizable by software? Because the next frontier of enterprise AI will not be the company that has the most agents. It will be the company whose causal structure can learn.
It has been a busy year for Wall Street debuts. While a few IPOs have captured most of the headlines, 203 companies have begun trading on either the New York Stock Exchange or Nasdaq market so far in 2026. Many made splashy entrances, but their performance has been more uneven since the spotlight moved to the next major offering. The Dow Jones Industrial Average has gained 7.5% this year, while the S&P 500 has climbed 9%. New listings have been more of a mixed bag. In total, roughly 49% of the companies that have gone public this year have seen their share prices decline. That figure includes special purpose acquisition company (SPAC) listings. Excluding those, 38 of the 76 operating-company IPOs are currently trading below their offering prices, according to Stock Analysis. The largest IPO of the year (and of all time) is among the companies that have lost momentum since beginning to trade. Shares of SpaceX were below $123 in midday trading Monday, well below both the stock’s first-trade price of $160.95 and its listing price of $135 per share. Elon Musk’s rocket and AI company is down 45% from its post-IPO high. Many analysts had predicted SpaceX’s decline, citing a valuation of 107 times sales and continuing concerns about the company’s spending. Most analysts still have price targets in the $200-to-$300 range. Keith Snyder of CFRA, however, has set a price target of $115 and rated the stock a “buy.” SpaceX isn’t the biggest loser so far this year among newly listed companies. That dubious honor goes to Green Circle Decarbonize Technology, which has dropped more has 88%, falling from an IPO price of $4 to just 47 cents per share. Other notable decliners include: Micware: -78% BitGo Holdings: -74% DSC Holdings: -72% Jaguar Uranium: -56% That’s not to say some companies that are new to Wall Street haven’t seen eye-opening success since going public. Defense tech company Swarmer has seen its stock price increase more than 630% from its $5 IPO debut since March. And biopharmaceutical company Veradermics Inc. is up 555% from its $17 IPO price in February. Other winners include Hemab Therapeutics, up 174%; Avalyn Pharma, up 57%; and Alamar Biosciences, up 53%. Major IPOs are often volatile. A Truist Wealth report examining 30 major IPOs from the past 15 years found that shares had lost a median of 9% just one year after their debuts. On an average basis, however, the stocks gained 14%. A separate report found that the average and median first-year drawdowns, measured as the decline from an IPO’s debut price to its lowest point within the first 12 months, were between 54% and 55%. Beyond the ebb and flow of stock prices, newly listed companies also face volatility surrounding the expiration of lockups, or periods during which insiders and early investors are prohibited from selling shares. Lockups generally last 180 days after an IPO, although the length varies. Some investors wait until they expire before assessing a stock, watching for a large sell-off that could indicate deeper concerns. While there will always be winners and losers when a company goes public, the current state of the market is of particular interest because of two looming giant public debuts. OpenAI and Anthropic have both filed confidentially with the Securities and Exchange Commission for IPOs, although neither has announced a firm date. Anthropic is currently scheduling investor meetings and is reportedly targeting a possible October IPO, while OpenAI is said to be considering waiting until next year. There have also been reports that the federal government could take equity stakes in AI companies, although no decision has been made. Such a move could make the performance of those debuts (along with SpaceX’s continuing performance) a matter of greater interest to taxpayers.
Major League Baseball is restricting iPad usage in dugouts to prevent the tablets from running artificial intelligence to help make strategy decisions, and former reliever Adam Ottavino said the New York Mets’ use of technology helped prompt the move. The tablets have access to video and league-provided data, and also included a custom tab where teams could access other programs. MLB made the custom tabs inaccessible to teams starting Wednesday night, when the second half of the season started. “In many cases, the custom tab had expanded the use of the dugout iPads beyond their originally intended purpose to include recommendations regarding substitutions, pitch calling, and other in-game decisions traditionally made by players and coaches,” MLB executive vice president of baseball operations Morgan Sword wrote in a June 11 memo to general managers, assistant GMs and video coordinators. The memo, first reported by The Athletic, was obtained by The Associated Press. “I read the article and I was like, I can’t believe what I’m seeing. Teams are making decisions off of AI? Man, that’s just crazy,” Yankees captain Aaron Judge said. Ottavino said on his YouTube livestream “Baseball & Coffee” that the Mets had been using AI and cited spending by team owner Steve Cohen on the software. Ottavino pitched for the Mets from 2022-24 and is now a broadcaster on the New York Yankees’ YES Network. “The Mets were actually the team, the main team, that got cracked down on,” Ottavino said. “They had an AI program that was very expensive apparently and they were bragging about it a little bit early on in this — the year. Some of the coaches that I know were talking about it from around the league and they had basically an AI program helping them pick pitches and I think some other stuff.” “But MLB got wind of it and nipped that right in the bud, so apparently they weren’t the only team, but I knew about it from the Mets angle,” he said. “They tried to throw some money at the situation. Steve ponied up for — I think this program from what I heard was several hundred thousand dollars to have.” The Mets did not comment when asked whether they had a response. A review by the competition committee found clubs had been compliant with the regulations. “Instituting this prohibition beginning with the second half of the season is intended to provide clubs that have relied on the custom tab with appropriate lead-time to make any necessary adjustments,” Sword wrote. Toronto Blue Jays manager John Schneider called it “a little weird to kind of see the whole report on it, where you can do things kind of in real time that can sway your decision one way or another.” “I think the biggest thing is calling pitches and kind of seeing how that can evolve in real time via technology,” he said. MLB started a pilot program allowing use of iPads in dugouts with restrictions late in the 2015 season and expanded their use in 2016 under a deal with Apple. Video was eliminated in the pandemic-shortened 2020 season following the Houston Astros’ sign-stealing scandal, then returned in 2021. “It hasn’t impacted us at all but I know AI is entering our arena for sure,” Arizona manager Torey Lovullo said. “It’s entering everyone’s arena. You better get on it, or you’re going to get rolled over by it.” AP Baseball Writer David Brandt and AP freelance writer Ian Harrison contributed to this report. AP MLB: https://apnews.com/hub/mlb —Ronald Blum, AP Baseball Writer
The new, powerful Chinese artificial intelligence model Kimi K3, which has caused a stir in the U.S. tech industry, has suspended new subscriptions after a flood in demand overwhelmed capacity within days of the launch. The development highlighted challenges faced by Chinese AI models in serving a growing user base at home and abroad at a time the U.S.-China tech race is heating up. Chinese open-source, lower-cost advanced AI models, including DeepSeek’s latest V4, have been moving fast in gaining global attention. In early 2025, DeepSeek shook world markets and helped China to be recognized as a serious competitor to the U.S., where Chinese frontier AI models are largely open source, meaning they are made accessible for people to examine and build upon. “Kimi K3 has received far more love than we expected,” Moonshot AI, which is Beijing-based, wrote in a X post late on Sunday. “Over the past 48 hours, demand has pushed close to the limits of our current capacity.” Moonshot said that it’s prioritizing existing subscribers and would be temporarily pausing new ones. “We’re adding capacity as fast as we can and will reopen new subscription spots in batches,” it added. The AI startup also posted a similar message on Chinese social media. K3’s public rollout late last week, aiming at challenging more advanced AI models by Anthropic, OpenAI and others, has rattled U.S. rivals. At 2.8 trillion parameters, a measurement of an AI model’s capabilities, it’s deemed as the world’s largest open-source AI model. “New model releases generally trigger massive interest, which can strain existing compute infrastructure,” said Lian Jye Su, a chief analyst at the technology research and advisory group Omdia. “This does show Moonshot AI does not have sufficient compute chips to serve the current surge in demand.” Su said that the key reason was more likely due to Moonshot not fully anticipating the surge in K3’s popularity. K3 is “very demanding” in terms of compute requirements, he said, making compute allocation challenging and expensive. By measure of front-end coding capability, a metric for AI models’ performances, Kimi K3 topped the chart of the major advanced AI models in rankings on Arena — an AI model evaluation platform — after its public release. K3’s release has also put pressure on stocks of American technology titans on worries that more affordable Chinese AI models may undercut the pricing power and demand at the U.S. AI companies, even as U.S.-led restrictions have already barred China from accessing some of the world’s advanced technologies, including the most cutting-edge chips. K3 is one of the latest new Chinese AI models coming out that have drawn strong global reactions. Over the weekend, China’s Alibaba also previewed its latest Qwen3.8 Max AI model, which the tech giant said had 2.4 trillion parameters as one of the world’s most powerful models and was second only to Anthropic’s Fable 5. Last month, Chinese startup Zhipu, or Z.ai, also rolled out its GLM-5.2 model, which saw rapid adoption across the world. —Chan Ho-Him, AP Business Writer
The release of Kimi K3, an artificial intelligence model created by a former computer science graduate student at Carnegie Mellon University, has rocked parts of social media and the AI commentariat. The model, developed by Moonshot AI—a Chinese lab cofounded by that ex-student, Yang Zhilin—has wowed many observers with its capabilities. “Kimi K3 demonstrates the U.S. moat in building frontier AI software is not as durable as many of us had hoped,” says Ryan Fedasiuk, a fellow at the American Enterprise Institute. “We should expect Chinese AI labs to continue distilling and freely releasing a version of the American frontier at a pace just weeks behind U.S. labs.” Social media has been ablaze with eye-popping demonstrations of the model’s prowess, including its ability to spin up a convincing browser-based version of macOS in a matter of minutes. The demonstrations have led some observers to conclude that Chinese AI models have caught up with their U.S. counterparts. But Yang, and Moonshot itself, recognize what K3 is not. It is no Claude Fable or GPT-5.6 killer. The company’s CEO even added a line to the model’s release notes acknowledging a “noticeable gap” between K3 and the leading U.S. models. That gap, however, is closing. The U.K.’s AI Safety Institute published a report last week suggesting that open-weight models, such as those emerging from China, are now at most four to seven months behind the frontier, compared with a lag of six to 10 months throughout most of 2025. Nevertheless, the gap remains, and it was evident on the first day K3 became available to the public. “K3 is an extremely large and computationally demanding model. At nearly 3 trillion parameters, it requires nearly an entire server rack of high-end AI chips to run on,” says Fedasiuk. “Moonshot had serious issues with uptime on the day of K3’s release—the service was down and rendered errors for more than 60% of users who logged onto Kimi’s web portal—likely because Moonshot did not have sufficient access to compute.” Computing infrastructure is one area in which China and its leading AI companies may continue to lag behind the United States. “As models get even bigger and more performant, the real race is about building the computing infrastructure needed to run and serve AI to billions of weekly active users,” says Fedasiuk. It is notable that, during the same week Kimi struggled to meet its initial demand, OpenAI reset the rate limits for Codex and ChatGPT users twice, while Anthropic offered a reset of its own, even as both companies provided access to more powerful models. K3’s apparent success has prompted questions about whether Trump-era export controls designed to restrict China’s access to GPUs, the chips needed to train enormous AI models to a world-leading standard, have backfired by catalyzing Chinese AI development. Experts say they have not. “People often confuse there being some downsides of export controls, which export control hawks agree with, such as they encourage efficiency, with export controls being net harmful,” says Miles Brundage, an AI policy researcher. “There is every reason to think China would have stronger AI capabilities with no export controls and thus more compute,” he says. China’s success has come despite U.S. export controls, as well as comparatively weak enforcement in some areas. K3 does, however, demonstrate that it remains possible to approach the technological frontier outside U.S. control, which increasingly appears capricious. For countries worried that they could incur Donald Trump’s wrath and lose access to leading models such as Mythos and GPT-5.6, as has happened before (albeit temporarily), China may appear to offer a tempting alternative. Last week, China’s president, Xi Jinping, spoke at the World AI Conference in Beijing. As part of a four-point plan for the development of AI worldwide, Xi said that “we must uphold openness and win-win cooperation to drive innovative development.” Kyle Chan, a research fellow at the Brookings Institution, has compared the competition to the Apple-Android battles of previous generations. “The US is Apple and China is Android. That’s Beijing’s framing of the world,” he wrote on X. “The US offers a closed system where power and profits flow back to American corporations. China offers an open system where everyone, including the Global South, can participate and benefit.” How to read Xi Jinping’s speech at the World AI Conference: It’s not really about AI. It’s about seizing a moment of tech-driven transformation to push forward China’s foreign policy goals. The US is Apple and China is Android. That’s Beijing’s framing of the world. The US… pic.twitter.com/wsDxyagVAc — Kyle Chan (@kyleichan) July 17, 2026 In reality, Chan said, that framing is inaccurate. China has benefited for years from the openness of U.S. AI research, an approach American companies are only now reconsidering as their models become more capable, potentially more dangerous, and riskier to release without restrictions. The rest of the world may not recognize that nuance or remember the history of American openness. It may instead see a United States increasingly determined to ration access to its most advanced technology, then look toward China, which is offering models that are nearly as capable, far cheaper, and accompanied by fewer apparent strings. K3 does not make Beijing the new center of the AI universe overnight. But it strengthens China’s gravitational pull in a way that DeepSeek ultimately did not.