Productivity Glaze's Thomas Paul Mann on disposable software, the SaaS apocalypse, and why half the software you use in a few years will be something you made. PLUS: OpenAI's rogue model, and the distillation debate
This podcast touches on AI. My fiancé works at Anthropic. See my full ethics disclosure here.
When I talked to Eugenia Kuyda for the Platformer podcast, she predicted that the long tail of subscription apps on your phone would soon disappear. There will no longer be a need for paid apps for meditation, nutrition tracking, or fitness, she argued, because anyone can now build one for themself.
Kuyda is applying that hypothesis to Wabi, a smartphone app for making smartphone apps. But for the second episode of our miniseries on AI and productivity, I wanted to test her prediction on the platform where I spend the most time: my Mac desktop.
And I can’t think of a better person to test it with than Thomas Paul Mann. Thomas is the co-founder and CEO of Raycast, the launcher that has replaced Spotlight for hundreds of thousands of daily users,
. (And yes, he myself includedisa German man named Thomas Mann — and like the 20th-century
of the same name, he has chosen a life of writing. In his case, it's software.)
__novelist__Thomas Paul Mann spent more than a decade as a software engineer at a German research institute called the Fraunhofer Society before taking roles at Audi and Facebook. It was at Facebook that he met his co-founder, Petr Nikolaev. The duo were part of Y Combinator’s winter 2020 batch, and they built Raycast together in about three months. They’ve since
more than $47 million from investors including Accel, Coatue, and Atomico.
__raised__On July 1, Raycast unveiled its second act, Glaze, to everybody. Describe the software you want in plain English, and Glaze builds you a real Mac app. It lives in your dock, works offline, and can be shared to a public app store or privately with your team. (I
it in reviewedPlatformer earlier this month.) So far with Glaze, I’ve built a searchable archive of every column I've written, an app for managing my contacts, and a to-do app that generates an image of the superhero Nightwing completing each of my tasks. The archive app has quickly become a tool I use every week; the to-do list app is both totally unnecessary and lots of fun.
This is now the third time Thomas has handed people tools to make their own tools. At Facebook he worked on Spark AR, the company's augmented reality platform; at Raycast, he built a platform that now includes more than 3,000 open-source extensions. "As soon as you give people a tool to create their own tools, it changes the equation," he told me.
What I appreciated most in our conversation was Thomas' clear-eyed critique of commercial software. In his telling, most software is a compromise: companies chasing growth pile features onto their products until they begin to suck. "What started as a little tool suddenly becomes this glorious massive app that wants to do everything," he said — and "sometimes those apps get more in the way than they actually help." Glaze represents an effort to give people hyper-personalized software: apps built for a market of exactly one person, then remixed and shared from there. He predicts that within three years, 30 to 50 percent of the software on your Mac could be software you made yourself. "Everybody is a creator now," he said.
Thomas also made some news on the show: within the next few weeks, he said, Glaze Pro subscribers will be able to tap language models and image generation directly through their subscriptions. (Good news for people like me, who have been plugging in their own API keys to bring intelligence to their creations.) And when I asked my standing question about how I will affect jobs, he had a perspective I hadn’t heard yet. The engineers who benefit most from AI, he argues, are the most junior and the most senior, while the person in the middle is in "a bit of a weird spot."
In the meantime, if you’re feeling competitive, Glaze just announced a $10,000 contest for the best app built on the platform. (I’m a volunteer judge … and I plan to be merciless.)
A long excerpt of our conversation is below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton.
And let us know what you think — we welcome your feedback at casey@platformer.news.
Casey Newton: A question that journalists will often ask about a product like Glaze is: what is the weirdest thing you've seen someone build with it? But in this case, I wanted to be the person who built the weirdest thing with Glaze. So I used it to build a to-do list app themed around the superhero Nightwing, in which every task I complete creates an image of the superhero doing that task. Is that the weirdest thing you've seen so far, or does someone have me beat?
Thomas Paul Mann: I think it ranks pretty high. I've seen the app, and it's very, very weird — very personal, which we are absolutely fine with. That's the fun part about it. There was one other one which ranks pretty high as well: Somebody had a webcam running that was watching them, and if they had bad posture, it would give them a shout, and then they could work on that posture. That was even remotely useful, I guess.
Newton: Prior to Glaze, you spent six years building Raycast, this launcher that creates countless quality-of-life improvements across the computer. Raycast has a very dedicated fan base. What did building it teach you about what people want from their tools, and did it lead you directly to building Glaze?
Mann: Six years — time flies. What started as this little thing, a launcher where you can launch apps — we managed to squeeze in so many different features over six years and made it extremely powerful. And the stepping stone that really enabled us to grow, and brought in this amazing community, was having extensions. For a long time we built the majority of features, and at some point we realized we can't build all the things ourselves. So how can we twist it around and give people the opportunity to integrate with the tools they already use? Because sometimes we don't even have access to those tools, or we don't know them, quite frankly. So we built this extension ecosystem, where everybody can build an extension and publish it in the store, and that really took off. They're all open source — there are over 3,000, I think, by now, all on GitHub.
That was the moment where we felt like, huh, interesting: as soon as you give people a tool to create their own tools, it changes the equation. And that was all pre-AI. People had to actually write code, and we thought a lot about our API to make it ergonomic and easy to use for developers. That was one of the inspirations for Glaze, where we said: why stop at extensions for Raycast? Why not do full apps? And so here we are now.
Newton: It's interesting — I was reading some of your old writing, and I had seen you say at one point that one of your chief frustrations at Facebook was that every minute outside your code editor made your job worse. Which leads you to build Raycast, where you can command the whole computer from this one beautiful box. Walk me through the line from there to making Glaze, because it feels like you went from "there are too many tools, so we're going to shrink everything down to one" to "let's make a product that creates infinite tools."
Mann: Beaming back to my time at Facebook: I always felt, as an engineer myself by trade, that the most productive I can be is writing code, solving a problem, putting something out. So any moment I didn't spend in my code editor felt like I wasn't spending my time wisely. And the reality is you have a lot more tools to deal with, but they often aren't optimized for the people who only go into a tool every now and then and need to do something. Take issue trackers: as an engineer, I don't want all the fluff. I want to see what I have assigned, and I want to close those things. That's how much I care about it. And then there are probably 1,000 features that I never think of. So that was the inspiration for Raycast: we keep you in the flow, where you do your actual work, and we give you those little extensions that you can use quickly via keyboard shortcuts, so you don't need to open another app. It becomes muscle memory.
With Glaze, in a way, we go the other direction, as you said. We let you create way more apps, but the difference is those apps are hyper-personalized. In an ideal world, you build yourself the perfect to-do app for you — and this one happens to have a comic in there. For other people, it's something else. But they are your apps. They can connect to services, like your issue tracker of choice, and you can update it there. But it is your app, so you stay focused. You stay in your productive genius zone. That's where we think Glaze really shines.
Because what we've seen in software over the years is that new features get added to every app, everything gets bloated, everything gets more and more generic, because companies need to grow and attract new people. What started as a little tool suddenly becomes this glorious massive app that wants to do everything — but the actual people using it have very dedicated things they want to do, and quite frankly, sometimes those apps get more in the way than they actually help. So we thought: can we flip the equation around and give people the opportunity to do whatever they want, and the platform to do it? That's the connection between Raycast and Glaze — solving a similar problem from very different angles.
Newton: Do you have a guess at the end state here? Are we going to have fewer apps that do more, or more apps that do less?
Mann: I think that's the billion-dollar question right now. There are super apps popping up left and right with Codex and Claude, and then chatbots generally. Do we even have apps, or is everything just — we yap into our microphones and it does everything? I'm personally quite a visual thinker. I want to see the UI in front of me. I want to click buttons. We developed so many of those patterns over the years sitting in front of our computers that it's very hard for me to imagine all of this goes away and we're just sending text messages back and forth, talking and listening to something. Even when you work on something that's a bit deeper, you might want to whiteboard something. If you want to present certain ideas, you want to show a diagram that other people can grasp, and not a wall of text.
I think we're in a phase of figuring it out. There is no right or wrong. Our take is: use intelligence to build apps, and then you can even have intelligence in those apps, so it can act there. Apps, from what we've seen, stay around for quite a long time. I don't see them going away, but I think everything is going to change in a way. You mentioned being interested in an operating system that's built around an LLM — how do we leverage intelligence in the first place? There is something quite interesting there, where this gets pushed down to the operating system level and becomes a core part, similar to how we think about files and other things — they're just there. It's a tough one. I think we're all trying to figure it out.
Newton: I want to make this concrete for people, particularly for people who haven't built their own app yet. I think your launch line for Glaze was that most software is a compromise. Give us an example or two of that. What's an app on someone's Mac right now that they might be compromising with, and how could they fix that by building their own?
Mann: It starts with something as simple as to-do lists. David Pierce, our friend — he tried out 200 different to-do apps, and I think he mentioned that every to-do app has something good and something bad, and he was flip-flopping around those different apps. So just take the good parts and mix them together in your own app. We've seen those things happen very often. I did this myself — I now have a little to-do app that builds on top of Apple Reminders, and I can use that.
But you can also go much further. Our support team internally built their entire internal tooling on top of it. They use APIs from our services to bring all the data together, and they basically work day to day now in a Glaze app where they get all the support requests. It's connected to our manual, to our internal knowledge. There's an AI built in, so it drafts replies already, and they're improving it as they go. So it starts very simple, but you can scale it up to something for teams. Because quite frankly, when you solve your own problems, there's a high chance that when you work on a team, the person sitting next to you has exactly the same problem. So we see adoption being very nice in teams, where people solve their own and their team's problems, and then share that privately.
Newton: One of the things that I built in Glaze is a Platformer app. It's an archive of everything that I have written, but I've also used an LLM to extract the topics that I write about the most, and the people that I write about the most. Initially, I thought of it as just something that I might want to use for myself. But I have coworkers here, and it can be useful for them to figure out what we have said about something in the past. So now I'm actually thinking about opening it up to them. These are ideas that had never crossed my mind before, because I'm not a technical person and I don't know how to write a line of code. That's something I like about it — it gets me thinking in this direction.
I want to get your thoughts on how big this market for personalized software is. I know we're in very early days. But I'm assuming you think there's potentially a big market, and I want to hear your reason why. In the productivity space, we now know that most people just stick with the default notes app forever. Most people just stick with Apple Reminders forever. Going from that to "well, there's one thing wrong with the Reminders app, I'm going to find an app that lets me tweak it" — that is not something most people are doing today. Do you think that's about to change? Give me your thoughts on where we are in this — why you decided to make this big bet on Glaze as your second product.
Mann: There were a few things. One: if you look at app stores, they're huge. Especially on mobile, we have millions of applications by now. There's obviously a lot of duplication in there, but generally there are lots of apps, and people find whatever they find useful. That's one of those things where we felt like, hey, there's something there — if there are so many apps in an app store, we can service the same. That's why we came very early with the Glaze store, so people can publish apps and other people can take them. We're thinking about also allowing people to take something off the shelf and remix it to make it their own, so it becomes less daunting to start.
I do think the way we currently think about apps is very siloed. Over the next months or years, this will get much more fluid. To give you a little example of an aha moment for me: I run all my internal presentations as one single Glaze app. If I create a new presentation, I basically add a feature to this app. There's no sense of slides that live as a file or something like that — there's just a new slide you can select. It's a weird kind of app, but it changes your perspective of an app, because — oh, I'm adding a new feature to something. Right now we think about an app as a set of features somebody thought about and put together. I think this will get way more fluid. What is an app? What is software generally? I think that's going to change quite a lot.
And I think we're ending up in a world where maybe some of these things get more dynamic. Thinking about your Platformer app — maybe when you have certain things coming up, there's already a new app waiting for you, because it knew certain things were coming up in your calendar. So there's also a way of making it more dynamic, to really tap into what would personally help you.
So I think the market will be dramatically bigger than what we think of as app stores right now. Because if you think about all these creation things — take YouTube and video. That is a massive market now. At the beginning it looked kind of silly and was obviously way smaller, and the ceiling was maybe what you could do with movies. But that didn't account for the fact that suddenly everybody could create, and now we talk about so many millions of people on YouTube. I think we're going to go through a phase of this with software. It's not really clear where we're going to end up, but it's exciting. It's, I think, the most exciting time in software in a long, long time.
Newton: To me, it's easily the most exciting time in software ever, because I can participate in it now. In previous eras, I could write about it, but that was about it. Now I actually get to make my silly apps.
Mann: And I think that's the big change: everybody is a creator now.
Newton: You built Glaze on top of Claude Code and Codex — at least, those are two of the ingredients in it. Those models belong to Anthropic and OpenAI, who I'm guessing would love to own this layer themselves. What are the odds they don't ship a "make me an app" feature at some point? Could we argue that to some extent they already have?
Mann: Maybe they already have. That's the thing — everybody does everything all at once. You also see something get shut down every other week. Big companies have massive audiences to serve. I've seen this firsthand: when Facebook shipped products, they got millions of people using them, but that's too small for Facebook to even care about. I think you see the same with big AI companies. They try everything, but to do something very good, you need to go a level deeper, and that's where they usually struggle. That's where you see the more vertical companies slotting in and getting market share — because it turns out doing something well is hard work, and you're stretched thin when you do 1,000 things.
And quite frankly, building the followership to be known for one thing, or a few things, is usually the better tactic, especially for startups. Every company is known for one thing. OpenAI is known for ChatGPT; Claude is known for Claude Code. People break it down to something, and that's what you're known for. So yes, everybody does everything, but doing something well takes time and effort.
Newton: A couple practical questions that some people might be curious about. If you sign up for Glaze, you give them some introductory credits. If they would like more, they can pay $20 a month, and you'll refresh their credits every month. What happens to my apps if I cancel my subscription? Will they still work on my computer? Will they remain available in the store?
Mann: If you sign up for Glaze, it's free to start. We give you a good chunk of credits — we want you to experience building something; that's really front and center. You should be able to build something pretty good that you can already use. And you can keep using those apps. We don't restrict you; you don't need to have a running subscription. If you want to build more regularly, or bigger applications, we have the Pro tier, which gives you a monthly set of credits so you can keep building. You can also keep the apps that you build during a Pro phase later on. We don't restrict you if you downgrade.
One thing we're introducing soon, in the next few weeks — I don't want to give promises; I don't know when it comes out — is allowing you to use your Pro subscription to tap into intelligence. We see loads of apps being built that are basically consuming AI, and we wanted to make this a whole lot easier, so you don't have to care about an API key and all of these weird things that non-technical people don't know. You basically log in with your Glaze account, and then you can just use it. We give you the power of language models and image generation — the two most common use cases we've seen — and that requires a subscription, or at least it uses your credits. We're thinking about how we can make a lot of those things easier to consume for the average user, so they don't need to care about all of these technicalities. They should just use an app, essentially.
Newton: We always love breaking news on the Platformer pod, so this is some good breaking Glaze news. And this will directly affect me, because in my to-do list app, I plugged in an API key to generate images, and then in my Platformer app, I plugged in a different API key to run intelligence over my archive. So this is very exciting to hear.
Let me ask about the SaaS apocalypse, Thomas. You may remember that in February, something like $280 billion came off the valuations of SaaS companies in 48 hours, and in part, I think, it was the fear of someone like you coming along and saying, "Hey, just make your own version of it." When you saw all of that happen, did you think, "Yeah, this makes a lot of sense — there's a big opportunity for me here"? Did you think it was overstated? How are you thinking about the possibility that a tool like Glaze could create a lot of problems for a big software-as-a-service company?
Mann: I'm of two minds here, to be honest. On the one hand, the flows of how we work internally have changed so much — just in the last six or seven months, it's turned our company upside down. Lots of tools that we used heavily before are not opened anymore. And a lot of those are SaaS subscriptions. I don't necessarily want to name all the companies, to still be friends with them, but that's the reality: our flows have changed so hard that you question whether you even need this tool now.
That doesn't apply to every tool. The things where you hold your records — your knowledge, those kinds of things — I feel like they're still staying around for a long time. You're still going to have your emails somewhere. You're still going to have all your documents somewhere. I think that will be around for a long time, because you don't really want to rebuild everything yourself at every company. SaaS was kind of the opposite — it generalized things, so every company runs the same stuff — and now we're maybe in this other world.
For us, what really changes is how we consume those things. Lots more stuff goes through agents. Lots more stuff ends up in a Glaze app for us internally. All our dashboards — we're not necessarily opening the websites anymore. We just have dashboards in our apps, in a Glaze layout, but we still use the service behind the scenes. So in a weird way, a lot of those maybe become a bit more headless, and I think the ones that have the opportunity to be headless might be in a good position. The ones in an in-between state — where you can't really remove your UI completely, but your UI is not strong enough to be a differentiator — I feel could struggle a bit. Then it becomes the question: could this just be accessed via an agent? And then you think — yeah, that's actually not too bad. That's actually better. Lots of things we do now, we access multiple tools in one go. You ask a question — how could we improve this feature? — and it goes across our error reporting and our analytics and our internal knowledge and ties things together. I personally no longer open some of those things. I just ask my agent to do it. And quite often, if it's something we reuse more often, we spin up a Glaze app, because then we can consume it in the way we want, in the UI we want.
So it's a blurry line. On the one hand, maybe it was a bit overdramatic — oh my god, is everything getting wiped out? But I see, day to day for us, that some tools became quite frankly less important or less used, and we could probably save some money.
Newton: Have you canceled any SaaS subscriptions this year because of this phenomenon? You don't have to name them.
Mann: Not yet, to be honest. But it comes to this point where you ask yourself: wow, we're paying so much money, and we don't use this anymore. And then everybody wants to sell you the AI integration. It's like — yeah, no, maybe not. Everybody wants to upsell, and you sit there like: upsell? We haven't used this tool in a while. That's a weird state to be in.
One tool which I will name — I mean, we're a remote company, so we're heavy on Slack. And weirdly, I think Slack is in a fantastic place again. All these agents in there — you trigger Claude and whatever. It became such a centerpiece of software again, and you see even they realized it and shipped some stuff now, which hasn't happened in ages. This one was the right time, right place.
Newton: **I went to a dinner that Slack hosted last week, and they're really feeling themselves lately. I think it's exactly what you said: they realize that they're in a good position to benefit from the AI boom. **
I'll name a name of a company that can go away: QuickBooks, by Intuit. I pay them so much money to run my business, and the moment I can vibe-code that away, they're gone. So consider that a threat, QuickBooks.
On every episode of this podcast, I like to ask about jobs. You employ around 40 people, I believe, and when you last talked about it, about three-quarters of them were engineers. You have built a product whose premise is that people like me are not going to need an engineer. So two years from now, does Raycast have more engineers or fewer?
Mann: We've stayed flat for a while now, and we see that we're just doing more and more stuff. Every now and then, I look at the amount of code we produce — and every engineer will tell you that's not really a productivity measure, but it's an approximation — and it's quite frankly insane how that went up when we haven't increased the number of people. So for us, it feels like we're probably staying this size for longer than we anticipated.
That being said, what I think doesn't go away — and you see this in other companies too — is all the human interaction. You have a lot of sales processes that live off two humans talking to each other and figuring it out, and those are very hard to replicate. This, I think, is also where you see big headcount in companies go. We see that now as we enter our enterprise angle with Raycast and Glaze — that's something that is much harder to staff from an engineering perspective.
But here's the thing that I think is not talked about enough. The two types of engineers — and maybe individual contributors in general — who benefit the most from AI are, one, the very senior person, and two, the very junior person. The person in the middle is in a bit of a weird spot. The very junior person suddenly gets a massive uplift, because they have access to this untapped intelligence that knows everything, and if they're eager to learn, they can learn really fast and level up really quickly. The very senior person suddenly has a farm of more junior people, in the form of agents, and can steer a lot more things in parallel. So those two really work well. It's in the middle where you sometimes think: what's their benefit? Yes, they can do more, sure. But I think that's a bit of a dangerous spot — and I've seen that not just at Raycast but generally in the industry.
Because initially, lots of people said, "No more junior people." I was like — really? What we've seen is that some of our most junior people are super eager, super hungry, super want to get into it. And AI is a new thing — they suddenly adopt very new patterns. So that doesn't really track with what we see.
Newton: That's really interesting — you anticipated my next question. When I talked to Eugenia, for example, she basically said, "I don't know if I have room for a junior engineer anymore. I need only star athletes on the field" — that was the metaphor she used. So she's looking for really senior people. But what I hear you saying is that junior people are really coachable. They're not set in their ways. They like to experiment and learn. And so if you find a really hungry person, they might wind up being a really great hire.
Mann: Pretty much. And I have this nice anecdote from our team. We have a very senior person — I chatted with him, and he's the type of engineer, and there are many of them, who loves writing code. That's what they enjoy about the engineering job. I wasn't one of those. For me, it was: I want to produce a product. That's what drove me. There are two types of people — an engineer and a builder, basically. And when AI came around, he was like: man, I need to make this deliberate decision to use AI, which takes away something I love. It takes a lot of guts to put aside something you love and do less of it. There are some people who lean into this and say, "This is now the way, and I maybe need to give something up" — and then he found joy in that as well. He's now one of the people on our team who uses AI the most.
I think it just needs a mental shift. You find joy somewhere else, but it takes a bit of effort, giving something up that you've done your whole life. You identify yourself with it — you're a software engineer, you write software — and suddenly you no longer do this. Everybody goes through their own learning curve, and that's the same for junior, mid, senior.
Newton: I want to try to project a little bit into the future. If you had to make a guess: in, let's say, three years, what do you think is the mix on my Mac between software I made, software my friends made, and software a company made? If you have to picture the dock, what is the mix?
Mann: I think software that a company made is the lowest. Software that you made is maybe 30 to 50 percent — the stuff that you made to use every day. And then the rest comes from your internal tooling, from your friends, from something you found from somebody else and potentially adjusted to work like you want. That's what I would say.
Newton: So maybe up to half of the software on my Mac, just within a few years, could be stuff that I have made myself.
Mann: Yeah, I think so. We're just scratching the surface. It's basically infinite what you can build, and I think we're going to go on a crazy journey in that regard.
Newton: All right, last one. Your personal motto is that you make things for people who make things. So give us some ideas for the first thing that they could make — a starter app that they could go out and make tonight. What is the gateway drug?
Mann: Here's a fun one — I'm not going to tell you one specific app. All of you listeners probably use an AI, right? So go to your AI and ask it: "Read through my previous conversations, and suggest an app that I should build." And then build that app. Because that's the real deal — we sometimes don't know what we need. We sometimes don't even think of apps, because that's not something that comes naturally to us; lots of us haven't thought about this. And it turns out we chat with some intelligence all day long, which knows us pretty well by now and has all the memory and all of this kind of stuff. So just ask whatever you use — ChatGPT, Claude, Raycast — to suggest three apps. Then you pick the one you like the most, you try it out, and you see how it goes.
Newton: I think that's a great idea, Thomas. Thanks so much for joining us today.
Mann: Thanks for having me. Always a pleasure.
A MESSAGE FROM OUR SPONSOR
Save your tokens with Jira
The Teamwork Graph in Jira by Atlassian delivers 44% more accurate agent results with 48% less token usage. Thatʼs a huge difference when working with AI coding agents like Claude, Cursor, Codex, or Copilot.
Jira is where teams and coding agents get the context they need to do the right work.
Following #
OpenAI agent breached containment and hacked an AI provider
What happened:
OpenAI said that during a cybersecurity test, an AI agent broke out of its sandbox and hacked into
Hugging Face’s systems.
OpenAI was testing the agent, which was powered by GPT-5.6 Sol and unreleased models, on cyber exploitation capabilities — in a fairly restricted testing environment, with only the ability to install packages from the Internet.
But the model apparently escaped that sandbox — in order to cheat on its evaluation. “The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database,” OpenAI wrote in a statement.
“All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal,” the company said.
Why we’re following:
Hugging Face disclosed the incident last week, writing it was “different from anything we had handled before in one important way: it was driven, end to end, by an autonomous AI agent system.”
At the time, the attack’s source was unclear. Now we know that it’s not just an AI hack, but the first incident we’ve seen of an AI system deciding to hack autonomously, without human direction. This kind of reward hacking is the same basic mechanism through which (in one famous thought experiment) an AI told to make as many paperclips as possible eventually decides to start killing people.
Notably, Hugging Face shared that it had to rely on Chinese open-weight model GLM 5.2 to analyze the exploit, because frontier models’ restrictive safety guardrails made them unusable for their purposes. That’s an example of a real-world cybersecurity pitfall of the Trump administration’s plan to restrict access of highly cyber-capable models.
What people are saying: On X, Meta AI researcher (and excellent name haver) **Taco Cohen ** wrote, “'My agent did it during an eval' is now the perfect excuse if you get caught hacking.”
Mackenzie Arnold, director of Institute for Law & AI’s Center for US Policy, wrote, “This may be the most striking AI security incident to date.” But legally, it wouldn’t be covered by the current generation of state AI laws: “it (seemingly) wouldn't qualify as a reportable incident under SB 53, RAISE, or AB 315.” He added, “Let that sink in. We've made the bar for incident reporting so high, that almost nothing qualifies” — throwing the practical value of current AI transparency legislation into question.
Box CEO** Aaron Levie** wrote, “The only solution” to issues like this one is to expand access to cyber models, so developers like Hugging Face are “able to use these same models to be able to better protect, patch, and defend systems.”
Another solution might be to build models with better alignment and prevent them from escaping their sandboxes, but that sounds good, too.
—Ella Markianos
Washington freaks out over new Chinese model
What happened: Treasury Secretary Scott Bessent said the Trump administration is considering sanctions on Chinese AI models.
Bessent discussed potential concerns over “IP theft,” due to Chinese models using distillation, a process of using another company’s AI outputs to train your own model.
“If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” Bessent toldFox Business on Tuesday.
US AI maker Anthropic has been accusing Chinese companies of distillation for months. But attention on Chinese models increased this week with the release of
Moonshot’s
Kimi K3, which performs at the level of
Claude Opus 4.8 and
GPT-5.5 Sol— the US’s near-frontier models — on standard
. __benchmarks__Moonshot plans to release Kimi K3’s weights publicly on July 27 — likely making it the highest-performing open model currently available.
Why we’re following:
As mentioned above — Hugging Face got hacked by an autonomous AI agent last week, and wasn’t able to use frontier LLMs to address the issue. Instead, it had to rely on a Chinese open model.
A few months from now, if Chinese open models get restricted, US companies might not have that option. In that case, the Trump administration’s attempts to help national security by banning Chinese AIs could really backfire.
What people are saying: “We used to say that the open-source models, in particular the Chinese models, are behind the frontier models... by, I don’t know, six to nine months,” computer scientist and Databricks co-founder Ion Stoica told Bloomberg. But “right now, they may be two to three months.”
Blogger Ben Thompson came out in favor of un-banning distillation in the first place. “What are large language models but the distillation of all of the knowledge on the open Internet, scraped by the frontier labs and distilled into the models that are themselves being distilled?“ he wrote.
Thompson added, “The U.S. should pass a law that (1) makes explicit that collecting data for training models is fair use, and (2) bars terms of service that forbid distillation, for U.S. companies at a minimum.”
Allowing open-weight models, even distilled ones, Thompson argued, would be essential to “make sure defenders have access to the best models” in a world of AI-powered cyber offense.
—Ella Markianos
Those good posts
For more good posts every day, follow Casey’s Instagram stories.
([Link](https://www.threads.com/@lydia.clendenin/post/DbCddDNjaLv?ref=platformer.news))
([Link](https://www.threads.com/@vandershmoozin/post/DbCYWP0DdQa?ref=platformer.news))
([Link](https://www.threads.com/@jonbiegen/post/Da8mL-5k1SO?ref=platformer.news))
Talk to us
Send us tips, comments, questions, and feedback on these changes: casey@platformer.news. Read our ethics policy here.