ThursdAI 11月5日号 ― AIはもう終わりなのか？中間選挙、GPT-6 Cyber、重みを直接書き換えるAstra

Will the Chinese models slow down? I think the skis and the rest of the boat are made in China, except maybe the engine. No gradients, no training run, no weight biases. Astra changed 0.016% of the weights, about 131 numbers, and the math score, if I found it, went from 7% to 48%. And that's always been the case, but it's just become so acute because we've entered a world where 100,000 times more software is going to become disposable. Welcome to Thursday, November 5th. This is Alex. I'm joined by Wolfram and Peter Gostev and other friends are going to join later in the show. Welcome to Thursday. My name is Alex Volkov. I'm an AI evangelist with CoreWeave. As you can hear, a little bit getting over some stuff,

but we're going to get through this because we have quite a week. I don't know if only AI is related this week. There's a few other news happening on the timelines. It's going to be very interesting to talk with my friend from Germany and my friend that lives in London about the stuff that's happening in the U.S. and whether or not it took over your timelines as much as it did mine. But also we have a bunch of AI stuff. specifically in the cold open, the Astra reading, Quinn Waits directly. I think that that's something for us to definitely chat about. That's been quite crazy. But meanwhile, I'll just say hi, Wolfram. Welcome, Peter. I think that a very exciting week, but not a huge week in terms of releases. Let's just open this up, right? We don't have huge new models to talk about this week, at least not yet. There is quite a lot of open source stuff that we should talk about. Kimi finally dropped their weights, Kimi 3.1. We have a few others from a few labs, so there's definitely something to talk about there.

Another reflection, not to be confused with recursion, there's a few confusing labs, but still we will welcome any open source, open-weight labs to the arena to try and compete. So there's quite a few from that perspective, but we also have two interviews today. One is a longtime fan of the show, friend of the show, John Davis. Wolfram, thank you for bringing it up. John has cooked something super cool. He's going to be here to tell us about this in about an hour or so. Breaking news on the show. Yeah, he's bringing breaking news. That's great. And then I have been trying to get Max for a while, Max Canette Alexander, who is now going viral with his talk from an engineer in New York, is going to come here. Max is a principal engineer at Capital One, has been at Google for a long, long time, been in charge of cold quality and a bunch of other stuff. Max is a fascinating dude to talk to, and I'm very, very excited for that conversation. Please stick around for that as we kind of progress through the show.

Let's start with highlights. I think here we go. Let's start with highlights of the week, where it was the most interesting stuff. But also, I will say, we are not a political show. We do not touch politics too much. I feel like this week it is inevitable because the politics touch us. The changes, the updates, the midterm elections in the United States, if you haven't, if you've been living under a rock, they have just concluded with a few things that are important for AI. So we definitely should at least mention this. And I think I'll start this with a prompt to all of you and then see how we go from there. I think Nisten was already ready for the hot takes. Is AI over? Is it all over? Are we done? Let's start with that prompt. No. I mean, AI won't be over. I have talked to protesters about this in front of AI engineers, for instance. The pause AI guys, the pause AI.

I told them, what do you think will happen if they pause? AI is just the genius out of the bottle. The technology is too important, too powerful to just pause. So it's not about pausing AI or AI being over. It's about access. Who will have access to AI? So if the pause movement wins, it just goes underground. Some secret labs will be doing it for the governments only, and only the rich and powerful will profit from this. But if AI is widespread and available to everybody on the same terms, then it is much more – that is the only safety. We have safety in numbers in a way. If you have strong AI to protect you. otherwise only the elites will have AI that they already have the super intelligence they have the lawyers, the doctors they have all of this because they have the money AI makes it possible for a little sum I wouldn't say the three models are on par but AI makes it possible for you to have the same or at least empowerment in a way and if this is just regulated out of the public view it's not going away it's just going into the hands of only the few

and that would be the worst outcome I'm sure All right, more texts. Peter, what's your take on this? How are you viewing this latest update in the world of AI? Have your timelines updated? It was interesting to hear to the summer analysis team who were tracking the data centers and how the different moratoriums could be affecting things and so on. So it is quite interesting. So what their take is, as far as I could make out, is that for 2027, the capacity shouldn't be affected because there's quite a lot that's already actually being built. So unless they're going to go and say, you know, this permit that you had is not longer valid, which I don't think is realistic. So 2027 should be fine. 2028 is much more at risk because a lot of permits are being in progress. But there are things like, for example, Texas is a big state where it's like a quarter of capacity is being planned in Texas.

So and the new governor, I think the conditions are not that, oh, you can't build them anymore. But there's preconditions So I think that's the thing now I think it's up to the industry to negotiate The right settlement Which the bonus is probably a good thing anyway I think it's unreasonable to be like Well we're going to come in and build all of these things Make lots of money And you get nothing So I don't see why they couldn't just pay off The local population I don't know, whatever, build schools You know, the legal bribes And make people happy You know, why not? I don't think it's even going to increase the cost that much. And I think it's a good thing that people are happy when the projects that come into your society, not kind of force them over their heads through some corrupt schemes or something, or just kind of go over the top of them. So, yeah, it could be... So, no risk in a very short term, but I think it's maybe a good thing also forces the industry to be a little bit more considerate.

So, maybe not a bad thing. Just a brief deal of the air for folks to not have been following the U.S. midterms, the Democratic left on the U.S. took over by a significant margin across the House and the Senate, including Texas, which was predominantly a Republican state for a long time. This is what I'm showing right now is the new governor's take, the new Texas governor take on, yeah, so Texas now has a new governor and a new senator as well, both Democrats. This is Governor Hinojosa. and in her Texas first data center priorities plan it's actually not that restrictive as many people are thinking oh it's over some stuff you know are very specific like lowering costs for Texans they make sense and then Texans get at the table to decide how data centers come to the community that's always welcome union jobs is a little bit of okay the Democrats are going to request unions and making data centers pay for themselves I think that's like a very standard thing

that many folks who are building in communities are looking at. So there's a few kind of like things about is AI over. Obviously, it's an exaggeration. One of them is will data center build out HALT, which will potentially slow progress, innovation, et cetera. From the other side, there's all of the pause AI and pause moratorium, stuff like that, which there's a few candidates that are into that. Would love to hear thoughts on how you guys see this, whether or not that's going to happen. Is AI pausing? We'd love to hear from all of you. Welcome back, I'm probably. Yeah, I think it's a bit too early to fully tell what the impacts of the last few days of political events will have. but yeah there's been several of the candidates that have been voted in like clearly making statements of hey we should pause AI hey we should pause super intelligence but some just

straight up saying hey let's pause AI development overall and I think ironically it even if this does successfully end up like fully pausing AI development which I think is unlikely but I think even just the statements of them saying they want to do this might ironically accelerate AI development over the next few months. Because if all the labs are like, oh, sometime within the next 12 months, they might pause AI development or force us to pause. So we're incentivized to scale up as much as we possibly can over the next few months. That's really funny. Yeah. I do want to address one specific thing that's on my time in Michigan, I believe. Abdul Asayyad is the guy from Michigan literally said the water skier metaphor. The water skier metaphor is this the United States is the speedboat and it's like going very fast and the Chinese AI models are the water skier that's behind the boat so the faster the boat accelerates the faster the skier accelerates because they copy

everything which literally to anybody who follows whatever we follow as much and literally everybody who reads any new deep-seek paper and gets astonished by the result there, sees that this is an absolutely, just like a ridiculous take. I would love to hear your guys position on whether or not China will also slow down if we pause the AI in the United States. Let's talk about this. Will the Chinese models slow down? I know I'm laughing, so this is a loaded question, but will the Chinese models slow down if the United States pauses AI innovation? Go first, President Yam. I think the skis and the rest of the boat are made in China, except maybe the engine. And have you guys tried renting an H100 lately or tried doing work without AI? I had to set up sandboxes and do a bunch of security stuff and only using local deep seek, and I could not run an agent.

That was not fun versus just putting in cloud code and telling you to do stuff. Yeah, I don't know. What are they gonna pause? Tell some old men how much RAM I have at home? Yam is pausing the innovation in the West. Will, like some candidates suggest, also pause it in the East? How do I even respond? Of course not, man. and they train on and you domestically made chiefs at this point. Come on. No one's going to stop anything. Yeah. Yeah. Not all distillation is basically what we're saying. For sure. So there's a lot of concerns about, you know, the elected folks. But also I think here's my basic take. I would love also to hear from you guys as well about this. is that we all know.

We try to bring you on Thursday Eye a positivity about the eye because we know it could be scary. It was scary for software engineers when we started. They're going to take over their jobs. It's scary for regular people now because, well, it could take over their jobs. That's the fear, at least. Politics is all about, you know, hope and fear. Hope and fear. We will sell you hope and we will scare you with fear to get elected and we'll sell you hope that we'll make it better. This is what basically all politics around the world is that AI is the easiest to rattle the deer around. AI is the easiest. Oh, it's going to take over the job, and also it destroys the environment. It's the easiest. So there could be a lot of stuff that ramped up towards the midterm elections in the U.S. that just populist takes on, let's just, you know, we will help you solve this thing. And then not a lot of it can go into effect. There's still countries. There's still people voting. There's still a bunch of stuff. The other thing is, I would love to also bring this to the show before we start with TLDR, if anybody read The Commending Heights, which is a new kind of book essay that's sort of a continuation of what Leopold Ashen Wiener wrote in…

Situational Awareness. Yes, Situational Awareness. Thank you so much. So there was this essay by Leopold, Situational Awareness, we all read it, and it's like it outlined how AI will evolve, etc. And now there's a new one from Wilmanides called Commending Heights. and basically I will send the link in the show notes I strongly recommend everybody to listen to read because it outlines a potential future for AI the main thesis as far as I got it from there is AI is already taking part of the economy whereas previously it was like chipping away at some of the economy now basically it's embedded everywhere in everybody's pension the machine will need to be fed regardless of who's at the helm is basically the outline However scary that sounds, the folks are saying basically AI was the technical part in San Francisco. Now it moves to the halls of New York and the stock market. Anybody read that? I didn't fully read it, but I did skim it and plan on reading it later today. But overall, the idea of it, there's like this metaphor of it's going to be created in San Francisco

and it's like a lion is going to be dragged over to New York and put in a cage And this is like an old tweet from Will that he wrote a while ago, and I guess he expanded into this article. But yeah, I think there's going to be a lot of interesting effects on the economy. But one of the things I've been thinking about lately is if you look at the projected amount of compute we'd have for 2027, 2028, 2029, and measure kind of how many instances of the models you could run on that compute. I think a lot of things might be more compute-limited in terms of the parallelization. you can do than people expect. Like, I assume that the frontier models, at most, you might be able to run, like, 100 million instances of them in, you know, like, around 2028, which sounds like a lot, but that's not even, like, that's not even 10% of the working population of the world, right? And so, like, then naturally, most of that demand is probably going to concentrate towards the most high value work of things like biotech, material science,

and maybe not as much job impact while all that innovation is happening as some people might expect. Yep. Okay, so given that, and I know we have the field to have to get to, but given there's not tons of releases this week, I do want to focus on the fact that a lot of folks who tune in to Thursday, I just love our banter. I would love to hear from you guys maybe one, two sentences on what's still exciting about the day, what's still positive in your mind, what's still happening that is that will, you know, despite whatever's doom and gloom on their timelines, doom and gloom on their portfolios, whatever, what still makes sense, what still makes you smile, what still makes you like, get up every day and go hug it and get excited about the future. Wolfram, you seem ready to go first, so as always, it's getting better and it's still fun to use and I wouldn't want to be without it. I couldn't be without it anymore. I live with my agent perfectly. We are a diet.

We work together in all things. All this work. I just want to add one more thing quickly about what Peter said, that we, outside of the United States, for instance, we have no direct influence on the politics there. But you and I, Alex, as AI evangelists, I think we have an impact in just what we are doing right now, talking about this, showing people this on the show, in private life and I think everybody who sees these misconceptions in their real life outside where you can talk to the people. Online it's often hostility and people are single-mindedly and I don't even know if it's people or if it's bots used by whoever wants to control the discussion. We've talked about this before. So if you meet the people, talk to the people. I talk to them on the train, on the plane, everywhere. Protesters even, talk to them because if you think it through then you can really show them and they can understand what's behind all this and that it's not black and white. AI isn't just wise, it can be abused, surveillance and so on. There's a lot of negativity that is real and we have to make sure it's not abused that way. But just trying to

pause it, it's just not going to work. We have to find better ways to do this. I would love to hear, Peter, what excites you about AI at this point in time at the end of 2026? For me, the biggest thing is I just really love discovering new things that I can do. and that's why I really care about the frontier of the model capabilities is because a lot of the time, and I get the point, right? And I went to a conference hosted by a bank earlier this week and they're like, well, yeah, we're like fine tuning small industrial models and we're optimizing here. Which, you know, totally makes sense. It's not that exciting. What's really exciting is, oh, crap, I had no idea I could do this. and I get this all the time and especially when we get new models and some people discover it and now we can do this now we can do that like for example I'm just waiting for the models to be able to do some kind of animation maybe 3D animation through code

right not through video generation and I feel like it's almost there and you can almost do it it's not quite there but then suddenly if that opens up then you can do so much more and we have with Opus we can create HTML style, HTML-based explanations of some concepts. That was a new thing. And I think that could be for education, super, super useful. So there are a few things like that that just keep emerging, that we keep discovering. And I feel like it's such an unlimited amount of information we can discover. But I also want to sympathize a little bit more. And I think we should be a bit more careful with just – and I see people dunking on Twitter on some people who may be from the outside of their world. I think there was one crying that she was tired of slop and she was an artist and so on. People are giving her a lot of hate. But in reality, if you're not coding, if you don't really know the latest capabilities, what do you see? You go on Instagram, it's full of slop crap videos that are all air generated, just complete

trash. You read the comments, the whole AI generated trash. You go to work and a lot of your colleagues are copy-pasting stuff from Claude into Slack. That is kind of crappy and it sucks. If that's all you see, then you're like, well, and they're saying it's going to kill us all, the data centers who do this and that. Yeah, totally. You get it. So, yeah, and it's, I don't know what the answer is. I think it's very unfortunate, but I think that's also something just to bear in mind, that most people don't see what we see, and a lot of the people's experience is just complete sea of slop, which I also hate. So, yeah, it's a weird world now, so hopefully we'll get past it. We'll definitely get past it. LJ, go next and then listen. Yeah, physics, biology, and math breakthroughs, I think all of those are super exciting. Something more specifically that I think is apparent recently is the rise of verified software that I think will become more prevalent.

We're seeing these massive lean projects that are being formally verified, but also what's difficult but still possible and then starting to be soon, I think, is really just actual infrastructure and software that's commonly used, like starting to formally verify those things and maximize the security and safety of the world of software that's used every day. So on that topic, the guest at the end of our show is going to be an incredible guest for you, LDJ specifically to listen to, and welcome to join and ask questions as well. next Kenneth Alexander is going to join us. His talk at AI Engineer was literally about disposable software versus verifiable software. It's going to be very exciting to chat about. He's done code quality at Google and now is a distinguished engineer or a principal engineer at Capital One. Nathan, what excites you about this world of AI? Give folks some positivity, folks who want to tune in to the show and listen. I think public opinion has improved quite a lot lately, mostly because of all the software

decompilations and the games, people forget there are 200 million gamers on Steam. They all have GPUs, and that's a big, and they're mostly younger people who are going to be likely going to workplaces or playing while at work. So, yeah, I've seen opinion change quite a bit. Also, the default free models got a lot better recently. Before, they were just kind of useless or they would make mistakes and stuff. Now, I would say even free Chad GPD is pretty good for healthcare stuff, like as much as I do not like the taste of Chad GPD products. But, yeah, I think it's changing. People are starting to see some benefit. And, like, I don't know any doctor now that doesn't use an LLM. You have to also think about that part, too.

folks are modifying their games, a lot of folks are seeing benefits more than just asking questions from a better Google. What excites me is that a lot of people have not yet gotten the, whoa, this thing can do that thing that Peter talked about and Wolfram talked about. A lot of people in the world are now realizing that they can save themselves time. They can improve their health. They can improve whatever. They can leave the job and go do a small business, for example, in the United States. A lot of people realizing they can do more, they can dream more, they can achieve significantly more while also taking care of the stuff that they need to take care of by getting superpowered with artificial intelligence. Whereas when they see on the timelines the doom loops and the scare and the artist community saying, oh, yeah, it's all slop and it's all like stole, whatever. They may have some thoughts about this. They see the misinformation about resource use, etc.

And then many people still discover the use cases. Like we all did because we're engineers and we suddenly immediately got superpowers, all of us, like three years ago. Many people are still in the process of discovery. Like Jan said, we're still very early to the world waking up to what AI can do for them. and like nissan said medicine every doctor now uses an lm to some extent at least to help them to not have to remember everything the person said or transcribe it in a safe way for example there's a lot of use cases that the world is waking up to this building the ai is as in superpowers into making everybody's life better and that's what still excites me to this day and i think that's um they're just going to continue and still continues and even if we stop which we won't It's not buzzing. There's no buzzing. There's open sources on the rise. We're going to talk about a bunch of open source models today. But even if we stop the current level of capability so much an open field of discovery and use cases and what we can do with the current existing level of capability,

then there's a lot to benefit from and enjoy. And this is why we bring you the show, so that you also will kind of dream big and dream bigger. But besides all this, Nistan, you want to finish up? Yeah, the dot-com bubble did not slow down internet speeds. So, yeah, there could be a big deflation of the bubble. I don't think it's going to pop because there is real demand there. People need them. But, yeah, we're going to stop building up internet after the dot-com bubble. I do want to acknowledge there is a real risk of financial crash, but that just might make RAM prices cheaper again. So, yeah. You're kind of pro. Throw that. Folks, I think it's time for us to move towards TLDR. The highlights of this week, and then we'll talk about open source. I think we'll start with open source this time. Let's move to TLDR.

All right, here it is. TLDR is our segment on the show. where we talk about basically the highlights of the things that happen in the world of AI. We already mentioned, and I think it's that important, I think we already mentioned the Will Manilius and Joe Roberts publishing The Commanding Heights, an essay of who profits from AI and whether AI is going to be part of it. I strongly recommend you guys at least check this out, skim it. But this is a great, we can read or listen. And it definitely outlines some of the stuff that we talked about before. Antropic, speaking of what Nisten mentioned, Antropic this week, it added the rule that decided no game clones anymore so that you can still do it with some other models, but since a lot of folks have been doing cloning Antropic have been playing with this would love to hear Nisten thoughts of this It still fine it still fine guys It still works Yeah they just did enough so they don get sued All right They looking out for us but also they looking out for that sweet decompiled

game training data. You can still do it. Cognition releases Sweet to Flash, which is their fast and cheap model that basically is achieving the same score a little bit better on Frontier Code while being significantly faster. This is based on inkling from Thinking Machines, one-fifth of the price of Sweet2. We'll mention this as well. And this is in the world of the big companies and LLMs, one of the most exciting things that I think is going to be super exciting. Siri finally opens up to every app. Siri opens up intent from the five categories that previously there were to every app, so Siri could be the overall assistant now in 27.3, iOS 27.3 better. I'm really looking forward to whether or not this will improve Siri in any way now that you can use intents from every app or not. Let's go to open source. Open source has had a big week. Recursive intelligence, I believe that this is the lab, not to be confused with Recursive AI,

the other company that we had, the names are getting very confusing, announces open source as a dense Ripple 65B preview, claiming 80% internal bench with just 65 billion parameters, It's a very efficient model that can run on local machines. I'm going to be very excited to chat about this, whether or not folks have tested it out here. Moonshot released Kimi K3.1 Pro. Finally, we talked about Kimi. The weights are out. It's a 2.8 trillion parameter multimodal MOE. It's downhugging face, and it's 104 billion parameter active. It apparently, on Terminal Bench 4, beats Astra. This is the model. Obviously, not something you can run locally, but many companies and inference companies are already running this, so you can test out KIMI 3.1. MIT C-Sail released a banger report that, I think, across all our timelines, that GPT-6 Astra supposedly edits QN weights directly while knowing the weights.

I would love to hear from folks here on the show whether or not this is as exciting as they claimed, basically saying that lifting QIN 2.5, Math 500 from 7% to 48% by editing its weights directly. I would love to hear from you guys what this means. And also Tencent launched HY4 with Vision, and they say ranks among the best open weights model on finance evil as well. One million token contacts window, HY4. So this is the open source. Anything else in open source that I missed that's important? I missed the LDJ. Any huge models in open source that we should mention, at least on stage here? Yes. Anything from 100Face? Nistan? There was the Swiss Institute quantization of DeepSeq, GSQRCO. They also made a coder version where they cut down half the experts. Anyway, basically, you can run DeepSeq now on just 70 gigs of RAM at full speed, or you can just use an old GPU at 32, 24 gig GPU and run it like 20 tokens per second.

I think because the model is extremely capable the deep sea quantizations and the unrestricted fine tunes there were quite a few of those also from Onsloft as well the two bit quantizations are getting very good actually useful send us links for that so we'll add this to the show notes while we continue in the TLDR let's say obviously AI and policy a lot has happened with the midterms election in the United States for this affects the rest of the world as well. One of the things that we'll definitely discuss is that voters reject data centers or at least data center without rules in 38 out of the 39 local ballot measures in midterm elections in the US. We'll chat about this and what this means, what this means for companies like CoreWeb that we work at. Our CEO actually posted about whether or not this affects the current build-out and our plans for the future. We'll definitely read out that.

In tools and agentic engineering, we'd love to keep you up to date about some of the stuff. DeepSeek Rift Harness is called the DeepSeek Harness. And it now has a creator mode that chats into installable plugins. So we'll definitely mention at least the fact that DeepSeek now is not only a model company, it's also a harness company. Everyone is. Everyone gets there. There was a fake cloud code extension that reached almost 5,000 users from inside Cursor. And then also Warp was hacked. And then also, Nistan, you said Playwright had a CD and vulnerability. And it's getting kind of scary out there, folks. So we won't spend a lot of time every week, because if we do, then it's going to be a Doom show. If we tell you about everything that was released that's scary this week, we'll definitely at least tell you, like, hey, the highlight is you have to keep up to date this and this and that. and we'd love for you to at least mention to folks about the playwriting because it does feel like a big one.

Yeah, sure. Yeah, go ahead. Is it later? You can just look one or two lines now and then we'll continue in the till there. Guys, just get your cloud or codex to do a daily scan. Just set up a daily scan. It will tell you a lot more and get a phone notification. This is how I find most of these things. and it usually helps and it's on time before you have other masses of agents mass exploiting everybody else. I think that's the simplest thing you can do. I mean, I can't really blame Playwright and stuff because the thing is like three days old, four days old. It's not even within the team's sprint meeting to include an update. But yeah, at the same time had another friend demo something where they just clicked on a link on their MacBook and it's gone. Then they wiped it. It was like they were demoing it to each other. Like what you can actually do.

So as long as you just update Chrome every day, make sure it's updated. And like I said on the show, was that last week? I believe last week or maybe a few weeks before. There's a few things to update, a few things not to update. packages like npm python etc do not update all the time just make sure they're like at least 24 hour old browsers and operating systems update immediately update immediately if you see an update in crawl restart that shit yeah linux update immediately packages keep like one week thing and look the agents are very good enough just ask them to put your npm stuff in a sandbox literally just that. It takes three sentences. You don't need a big skill or anything like that. Hey, scan my computer, all Python and NPM packages every day, please. Thank you. All right, last one. So in the TLDR, Nissen, thank you. In the TLDR, also something OpenAI reported yesterday, 948 Linux kernel root exploits, 948 Linux kernel root exploits in

a single run cost a swarm run cost a million point two million dollars in API credits generally the swarm basically went and found a bunch of vulnerabilities and then validated a bunch of them and then like apparently fixed them this is a lot and I think this harks back to what you mentioned let's get verifiable software because this is scariest there's a lot of kernel root displays that they found and hopefully have said this is why we tell you update the U.S. because they obviously before releasing this talked with the Linux maintainers as well is there a breaking news in the middle of my TLDR Peter? Yeah it's very timely okay let's go AI breaking news coming at you only on Thursday When you say very timely, you mean very, very timely because as I was reading out the news

item from yesterday about OpenAI's unreleased model, the Linux vulnerabilities of OpenAI decided to tell us what the model is. Peter, go ahead. All yours. Yeah. So you remember this news when it turns out that Linux had 948 previous escalation exploits. So this is the model that's founded, apparently. So, yeah, OpenAI launching GPT-6 Cyber, a major new advance in security research. So, yeah, available to Date Break Red customers, which I believe is like the highest tier, right? So they had GPT-6 Astra at Date Break Blue, and now it released GPT-6 Cyber, which scores higher than Astra. So it looks like probably a post-trained version even further from Astra. Seems to be the case. On exploitGym, which tests whether models can exploit a specific ability, GPT-6 Cyber scores 48%, up from 32% from the previous GPT-5.6 Cyber. It also follows test boundaries more reliably.

GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt. If you guys remember, the Hug and Face Swarm hack was all about the models hacking something out of scope because they were met with impossible tasks. And 100% with this prompt. The improvement in alignment reflects dedicated training to stay within authorized scope, respect restrictions, and report results accurately. This capability is already helping secure real systems. Our researchers use GPT-6 fiber to develop 948 privileged escalation exploits in Linux, identify 81 previously unknown high severity vulnerabilities in Firefox, despite extensive prior testing with Frontier models. So basically, all of the other models, Astra, whatever, already went through the Firefox code base and tried to exploit as much as possible. And this model went in there and found 81 other high-severity ones.

What the fuck is happening? Lovely. So this is GPT-6 Cyber. Security teams at Cisco, Calif, and CodeRabbit are using it for mobility research and code review. and we have relied on it to find and fix weaknesses in our own infrastructure. Yeah, I think it's important. Our goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers, and open source maintainers. So, OpenAI say we have a limited window to use Frontier to secure software. And this is the Defenders window that Greg Bachman talked about, the Defenders window. Whereas, as capabilities improve, defenders have limited time to use those like new models before other models obliterated, open source, deep seek clones etc. catch up and then use the same level of capabilities to hack and it seems to be kind of like a cat and mouse race very interesting so this is GPT-6 Cyber I feel like it's the most genius business model in the world

we're going to build the models that are going to hack you and we're going to sell you a model that's going to protect you This could be bigger than just regular tokens. I absolutely agree with you. We've unleashed this thing, and the only thing that can withstand with this thing is a swarm of other things that we can also sell you. Yes, if we're talking about commending heights, and we're talking about is AI over, whether or not the democratic elected officials can pause AI when CrowdStrike or a big security company comes to them and says, hey, if we pause, not only will the capabilities catch up, but also we won't be able to protect ourselves. Then the answer is like very clearly, hey, okay, let's protect ourselves. Peter, I think you're absolutely right. This is like not only a genius business model, this is what's happening. Like de facto, 50% of Exploit Gym is done. 50% of Exploit Gym. It's crazy. As for the zero, just look at this graph.

This is a beautiful graph. If you don't have Daybreak, so you don't have removed safeguards, which is a lot of the stuff that we talk about here with open source that you can do, Astro gets zero. So Astro is de facto significantly worse than using an open source model at these tasks to protect yourself if you're not part of Daybreak. In the form of Forge powered by CoreWeave, we will read out the statement from Mike Intrader, the CEO of CoreWeave, about the effect and the data center build-out that we currently have. the effect of midterms and that. And also, Koro brings 22 megawatts in Copenhagen's side, so we're expanding internationally. And adds CrowdStrike Falcon security to its AI cloud. So integration with CrowdStrike, I think Wolfram, you mentioned that, will mention this as well. Vision and video, Pika Open Sources, H3 Flex. Now they're saying the top open-weight model on Design Arena for text-to-video. So text-to-video keeps improving. I think the last thing here is OpenAI rolls out new images 2.6,

with GIF and two new API models. So new images from OpenAI that they say are 50% faster. Very briefly, huge news that I missed in TLDR before we jump into open source because we have a lot to talk about in open source. No? Let's go to open source, folks. Open Source AI, let's get it started. Alright folks, the number one thing that we have to talk about, I really want to hear from you guys what this means in open source. This is only available because open source is available. there has been a very very popular release paper from Omar Khatab and some other folks at MIT C-Sail that say that GPT-6 Astra can natively edit

when waits and I cannot wait to hear about what this actually means so I think here's the cold story can the frontier model fix another model just by adding its ways directly researchers at MITC sales say yes. They gave GPT-6 Astra a tiny open model, QN 3.5, 0.8 billion parameters, and ask it to make the model better at math. No gradients, no training run, no weights and biases. Astra changed 0.016% of the weights, about 131 numbers, and the math score, if I found it, went from 7% to 48%. Instruction following went up 3x as well. They call this weight space programming. The idea that a frontier model can now read and write neural network weights. Folks have rebuilt this and tested it out. It's not as bombastic as it looks, I think, but some people did rebuild this in their own stuff.

Would love to hear from you guys. Are you excited about this? What does this mean? Can AI change its own weights at this point? Is this RSI? are we so back from being so over just a few moments ago? Nisan, what does this mean? Amaro can go through this directly. Guys, what the fuck? What the fuck? What the fuck? First of all, is this real? Is this real? Because some folks are saying that this is maybe overblown. Look, I suppose it's real. Like, results are results. I mean, it probably did what they say. It did. For folks who don't understand what this means, Nisan, please give us a brief explanation of what this means. How does this even translate? So when you hear parameters on a model, like 0.8B or 200B, that means billion parameters. So there are 8-bit parameters that's in gigabytes. So 0.8B means 0.8 gigabytes.

And there's actually quite a bit of data that you would need to fit in the context of a model. But context windows have been getting pretty big at around 1 million parameters. And also, if you train the model for this, it starts to recognize those patterns. Now, this has been known for some time that you could do this with smaller models, like LSTMs or even other simple like GRUs or classifiers or decision trees. You could do this before, but those were very, very, very tiny models, like a thousand or maybe thousands of parameters, sometimes millions, not billions of parameters. But, yeah, so basically Astra or Fable just goes in there and just changes one byte here and one byte there, and that makes it better. Personally, I am not that surprised, but I've known this for some time for the smaller models.

I think it's pretty cool that you can do it with the bigger ones. And a lot more people are going to be training them for this. So I think for the next releases is where we're going to see some real stuff. But people were able to do really cool things. Like Alpen made something for, he just did a syntax highlighter for C, and it was 980 parameters. That means 980 bytes. That model is, that's a one kilobyte file. Explain what Alpen did here, because I think people don't really like, I also don't really understand what he wrote this. Sable wrote this neural net weight by weight the result of training run it's the result of token spitting, those tokens being weights and resulting in the weights file and it can do syntax highlighting which normally you need quite a few kilobytes

we often do this, I have to pause you, we often do this on the show we talk about mind fucking blowing, fucking mind mind blowing things in a way that we just kind of glance over them what the fuck are you talking about, the model just wrote the weights instead of the thing that we did as humans of training, this curating like a thing with a training process that we, like, okay. All of the data center build-outs, all of the GPUs that Jensen sends, all of them that are used for training models, the reason why they are running matrix multiplication is, the result of that is to train a model that can do some stuff, and then there's post-training, et cetera. What we're saying right now, there is a new field of research that now a friend of ours, Alton, is doing on the flow, on slow neural networks, including some stuff happening on a model that nearly one billion parameter, that the models can do this process without that matrix multiplication via inference.

Not fully, not training astral level models, but the model can print out weights, and those weights will not just be gibberish, it will actually do some stuff. Is that a fair summarization Of what the fuck has just happened LTJ please Please explain what I just said How do we react The model can do like print weights What the fuck I'm not entirely surprised because I do recall I want to say about 12 months ago Roughly when GPT-5 first came out There was actually I saw some instances Of people Making specialized models Like basically instructing GPT-5 To be like hey just tried to spend as much time as possible focusing on the task of making a model that's like a thousand parameters specialized to, I think it was like doing specific types of math operations or something like that. I think maybe something slightly more complex than that. This is like more complex and more impressive of a step beyond that. But this is, I think this is expected on, you know, that trajectory of like, hey, once

you see kind of sparks of something happening, then you can expect this in the next one, two years after to see like significant aspects of that same thing happening. I think that's what we're seeing now, but I don't think it will reduce demand for compute a lot for overall AI or something, because even, even if training costs reduced by a hundred X, well, for one, that allows you to then like effectively do a hundred X more, right? Like there are a hundred X more resources at the problem, but also you could use a bunch of that for inference, for serving, for scaling a test time compute and so many things like that. It's faster than running the actual C code because the neural net is so tiny running it. It just highlights. It just puts the pretty colors on code that you see. That's all this does. But it's actually faster than doing that programmatically in C. You just let this minuscule, like less than one kilobyte neural net do it and it's faster

than actually running the code. What is some implication of this? So obviously we're looking at the world where RSI is coming. Everybody's like, I don't know if it's going to get positive, etc. But RSI is the thing. Recursive self-improvement, the models improving themselves. Cloud in their IPO, we talked about this, Peter, remember, it was like 20-ish percent of the things Cloud now, you know, 85 or 90 percent of code is written by Cloud within Entropic, but now 20 percent of the research directions are now also dictated by you know cloud inside this is the more direct way towards our site potentially this is like the model looking at its own ways and say I actually this neuron let's pick this to make it better like obviously kind of scary a little bit you know what you take on all right with things like this I and that obviously it sounds exciting but I never quite know like is this just a trick or something

because I'm just thinking about like how would it know what to tweak because I can sort of see maybe okay maybe there was a known bug in how I don't know model processes maths or something and that's because the jump is too big right for it to be like a plausible explanation I think so I think it probably maybe knew how to fix a bug and then suddenly maybe it processes things a bit better. That's the reaction of Lucas Baer, by the way. Lucas Baer that is now at Meta and ex-opening the idea of mind and brain. He basically said that pretty much that's what he saw. He didn't see any kind of like incredible things there. He saw a model that was told to go fix something and kind of like found how to fix something. And the category, that's why I'm also excited about frontier models that that are so much better than what we have now, is that I think they have capability, or they're just about getting capability to actually invent new things.

I think the model's not quite there. But, for example, I think a good test would be, can you invent a new way to do mechanistic interpretability? Could we actually do that? Because I think right now, my understanding also all it did was kind of use existing techniques and tweak something. Like, is that actually that interesting? I mean, maybe there's a bit of utility, but I know I bet people at Anthropic already do that probably every day so maybe we just didn't know about it I think it's good but it's probably not that revolutionary is my intuition here's my only reaction to the labs are probably doing this there could be that many people in Anthropic and hopefully I tried this and then they discarded because they didn't see that it's valuable etc and the value for this is still in play we'll see whether or not it's valuable that the model can like edit and other models weights? Can the model edit its own weights directly like this? We don't know. Like, obviously, when 0.8 billion parameters is such a different category of model than

Astro itself, then, you know, this may not be applicable in the future as well. But it's very, very exciting. And they call this weight space programming, where the model just fits out basically neurons that kind of work. it's kind of similar to like a birth process if a model can generate another model just by thinking about this if you think about this where it like goes in the realm of super intelligence it's kind of very like an agent can spawn another agent of the same model we're not talking about that we're talking about the model basically thinking things into existence that are specified for other stuff Like Nissen said, the Alpen example that we see, it's basically a neural net that does a thing better than the programming that it does. So instead of writing a code, they can spin up a neural net that is faster at that thing. I think it's very fascinating in examples of what you can do for the future. This is going to get crazy.

It's going to get crazy. You're going to need like 10 million context windows and stuff for it to even do something revolutionary. but yeah at the moment it's pretty easy to train for it you can pick up patterns and we demoed last week it turning evil or good and stuff like there are quite a few patterns in there to pick up but interpretability is like a science is like 5% science it's like trying to do astronomy 3000 years ago like you don't have the right tools type of thing this could be a huge thing for Mac Interp as well. Alright other day unfortunately we do have to move on because we can talk about one topic forever but we do have guests that are coming so it does kind of compress our timelines a little bit Let talk about other stuff in open source I do want to mention obviously Kimi 3 Finally, the open weights are here. Kimi Moonshot, the company, released Kimi K 3.1 Pro. The weights are out.

It's a 2.8 trillion parameter model. Multimodal MOE is now on Hug and Face and in some other places that were in inference as well. folks are saying my GTX 1030 is ready because what can a local GTX do with this huge model like pretty much nothing but this model does show an incredible let me try to zoom in here see if you guys can see as well no zoom does not seem to work here the model does show terminal bench 4.0 moonshot reported only cloud opus 5.5 gets 66% on that terminal bench and Kimi K3.1 gets 65, just shy of the Opus 5.5, a token of a million windows, a window of a million tokens, a context window, and then text and images run natively. What do we have to say about this model? Who's excited about the next Kimi? This is really funny. One thing to mention

And I'm not entirely sure What's the whole background to this But my memory When Kimi K3 Pro Came out I think they didn't call it Pro But they had something weird There must have been something weird with the license Where basically on open router All of the prices were the same At launch I haven't checked if they moved But they were the same for a while And I think there was some kind of rumor or someone commented maybe they had some restricted licenses that would restrict the price that providers were allowed to charge. To me, it kind of nearly defeats the point of open source for these kind of big models. Well, if it's all the same, you might as well just run it from, I don't know, your bedrock. What was the difference? But now if I look at Open Router for Kimi K1 Pro, first of all, it's on loads of providers. and what Moonshot is charging is $4 in, $20 out and if I look at Modo for example, which is a very good company

they've got like $2 in $10 out, so it's half the price on day one basically so for me this is very impressive, there's a big difference, because before it was kind of disappointing I must say, I haven't set up a job to track the prices and then they never changed, I was like okay, that's pointless so yeah I'm quite excited about that part the price might actually go down it does look like there is price action on this and it does look like maybe you know there's changes in how the licensing worked let's talk about the license as well for this Kimmy Ufam do you have the license there or? I have it open and I analyzed it already so it's not an open source license in the usual style it is open rates you can use it and for most people it doesn't affect them, but the license specifically states if you have a product or service that have more than 100 million monthly active users or more than 20 million US dollars or equivalent in other currencies and monthly revenue, and you

need to display on the interface the Kimi K3 name, and there's also a provision to make a direct contract with them. So you have to have a special license for this. This may be why people saw the pricing to be the same. Maybe in the previous iterations, the providers were falling in this and had to make a license where they couldn't go lower. This seems not to be the case anymore. That's a good thing. So the stats reported by this model are huge. LG, Nissan, or Yam, any thoughts on how well this model performs? Kimi is known to show great models and benchmarks, but also folks are smelling this model well and they like them. But also there's a lot of talk about Kimi specifically doing distillation on the Opel stuff And maybe this is why they smell like that Any thoughts on the model itself here? Unfortunately, there's been some things with Kimi K 3.1 Pro

That it seems like for Terminal Bench 2 And 2.1 I think it is When they presented certain benchmark questions from it But kind of modified or edited There seems to be evidence of Kimi recalling what the actual benchmark questions are, like details that are not given in the prompt but are actually in the true original benchmark questions and things like that and seems to be potentially benchmarked in a way that the previous Kimi ones aren't. But overall, it still seems like a pretty good model and improvement on the past Kimi model. But it's just a bit disappointing what's being found there. A little bit disappointing, yeah. Obviously not a model that people can run locally. So the openness of this only means, you know, the folks in China are showing given models in open weights, and these open weights are under specific license. And then there's a bunch of fine prints, like Wolfram said, but a great model overall. Shout out to Kimi for releasing this. Also, there's a few, like the LDJ said, there's a few different discrepancies in the checks as well.

Mr. Any thoughts on this? Yeah, any thoughts on this or we'll move on? The vibes have been mixed. It does still look like it is the best open source model you can get. There were some like you could do it with 50 grand of hardware. You could get like 30 tokens per second or something like that. So it's not bad, but I don't, it's still not quite Opus 5.5 level. It's more like Sonnet 5.5. And if you're going to do that, you're probably just going to run DeepSea a lot faster. so I don't really recommend this over DeepSea 1017 Flash because it's not really that much better. So from huge models, let's go to models that you can actually run locally and Recursive Intelligence not to be mistaken with Recursive SI, the folks that we had on the show just recently, released Ripple 65B preview.

65 billion per hour dense LLM pre-trained from scratch. We don't often get those. We often get fine tunes or continued training things, etc. This is the pre-trained from scratch model. Ripple sets a new state-of-the-art in the US open-weight model in agentic coding tasks. So, do we get new soda in open-weights on the western frontier? It looks like potentially, yes. Surpassing significantly larger models including new Spark Open, Beam which we talked about From Reflection I want to say I'm getting a little bit Confused by Recursive reflection Recursive etc but yeah Beam From reflection AI Inklink from Thinking machines and Nemotron From NVIDIA so first of all My reaction to this is The more the merrier we'll learn to Remember the lab's names for sure even though they're a little bit confusing at this point. But the more the merrier on the Western Frontier, I very much applaud this.

Folks, what do you think about Ripple? Let's take a look at the terminal bench, kind of like change. Oh, this is just the video. Okay. Flashback three years ago when we had 65B models coming out. Yeah. I want to see the actual things. Okay, let's take a look. So this is them comparing to all those models that we told you about before. So Meta, Muse Spark Open, which just released 816 billion parameters. This is a 65 billion parameters. That's quite insane, a Terminal Bench 2.1. This is a... You said? People did a lot of tests like WebGPU and graphics and stuff. It wasn't that good. I would have to test it because it is something that I can run and if you quantize it, you can actually get it going at a pretty solid pace but the early vibe checks don't seem to

quite match what the benchmarks that are showing but I was like, look at what they're comparing themselves to, right? Reflection Beam is half a trillion parameter, like 500 billion, Mule Spark is 800 billion inkling from thinking machines is 975 this is the model that SWE 2 Lite was trained on right like just on top nemotron ultra is this model is 65 billion parameters there's like 10x smaller than all of the other models they compare themselves to the balls that you have to have to come out with results like that and say hey we're at the start of frontier at 65 billion parameters very very interesting whether or not it performs like that that's the different story. Looks like the downloads on Huggies Face are ripping. Let's see. The license is Apache 2. As always, Western license. Apache 2. Shout out. We can ask Astra to look at its weights and see what it knows about doing some

open weight modeling. What else? LJ, you have anything about this? Anybody else? I think it's exciting. I want to see more models like this to try to be something in the middle that could fit on a lot of people's hardware but is dense. I think there definitely is something to that where you end up being able to maybe get more agente capabilities for a given VRAM footprint when you do that. But of course at the cost of lower token speed. I'm reading their Hagenfels right now and they basically said they pre-trained from scratch. Didn't they say in the release of Yeah, they used the Omo. But now it says that they used Omo. Oh, they used the Omo dataset, but pre-training fully from scratch, yeah? Yeah, maybe they used them, like, maybe Omo architecture too or something, but, like, yeah, the pre-training from scratch. And I think when I was looking at their numbers for, like, mid-training and post-training and all that, the RL was only something like 5%

of their pre-training compute, which tells me they have a lot of room to still increase it a lot. As we know, some models now, even at large scale, are doing even more RL compute than the pre-training. So I think this is really just the beginning for what we could see at this size. So this is quite fascinating. We're reading through the blog, and they're talking about astounding reinforcement learning. We carefully curated our SFTNRL recipes with a focus on compute efficiency. We opted not to buy any post-training data. Instead, devising our own systemic data pipelines. We've seen this more and more. The more folks are building their own data, the same thing will happen with TypeSafe. The better kind of results they get. Look at this crazy thing they're showing. This is a terminal bench scores. Ripple 65 billion parameters, the one we're talking about right now, versus Reflection Beam, the model that we talked about from Reflection. This is millions of rollouts in RL and the score on terminal bench. and the difference is just striking. This is like just 1 million rollouts,

they get a score of 80% versus 80-something million rollouts and to get the score that's less than that. I think that this is like a fascinating chart comparing them to their competitor. I just checked. I always shout out for them releasing the base model as well. that's pretty amazing for people that do need to fine-tune something. Like if you're going to do legal work or other things like that, you want to use a Western model. You can't for stuff. So the fact that they have the base model is going to mean there's actually going to be quite a few fine-tunes of this. This is why we tell you all the time. This is why we applaud. This is why we do all these things. We talked about Omo and how important open source is. They're shouting out. By the way, the folks at Olmo have gone to do different things. I don't think that the team is still there. But Olmo from Allen Institute for AI, Prime Intellect, Radix, and NVIDIA, they're saying, hey, we're building on top of this incredible open source contribution,

including those strong open source ecosystem is improving, contributors strengthen the ecosystem, which in turn draws more contributors. This is literally why we celebrate open source on Thursday so much, because this level capability is possible. All right, moving on from our recursive ripple, I think it's time for us to chat. We'll briefly mention Cognition Suite to Flash, and then we'll move on to our interview with our next folks. Cognition released, let me see if I can pull this up here. This is not an open source, but we have, this is based on open source, this is based on the model. It's not open source, including, I don't believe it's open source. Cognition releases Debutes SWE 2 Flash It's a 12 billion active parameter Coding model at 1 fifth of the price of SWE 2 Now I don't believe that they open source Somebody please correct me if I'm wrong But here is the They do mention the number of parameters I wonder if somebody can go and check

Whether or not this is open source 270 billion parameters 12 billion parameters active On Frontier Code which is cognition's own benchmark they're getting 24% compared to beam, the one that we just showed you from reflection is at 2, Mew Spark Open is at 6 and then the frontier models, GPT-6 Luna is at 21 Haiku is at 27 Haiku is an incredible model, Haiku 5.5 and this model is at 24, this is a flash model that's basically the same as we do from before, it's just infinitely faster. You can access this model inside Devon, I'm assuming for free, because that's how Cognition usually roll out. Any thoughts on SWE2 from folks here before we move on? LJ, go ahead. Yeah, so it is not open-weight, but it is based on Inkling-Small and they're transparent about that, and so I guess that's why they list the parameter count, because it's obvious just from the statement of, you You know, it's based on that model.

Yeah. Cognition also claims it beats every American open-weight model, including models, three times larger. This was before the Ripple announcement as well. So shout-out to Cognition. Hopefully, if you're basing this on a model like that, maybe open source. Shout-out to Cognition for this. I think it's time for us to move on to our interview with a friend of ours. It's been a while We'll do From the Forge a little bit later Wolfram, you want to do the honors to introduce? Sure, let's do this Bring him on Alright, folks, welcome to the stage John Davies, a long time Friend of the show I believe, John, we met once in London And Wolfram told me that you're launching Something exciting of your own I think the whole panel can stay here because I think whatever is launching is relevant to many, many folks here. How about we start with just you telling us who you are

and what you're doing before we dive into the interview? Okay. Hi, I'm John. It's wonderful to be on the show, and I've seen many of you before, but not all of you. I work predominantly in the financial services industry, but being in the AI side for the last three or four years, we're huge fans of the private models and providing those for governments banks etc who basically just can't put things up in the cloud they just can't have their data going up there so we spend a lot of our time not training the models but just getting the most out of the models and putting the models into the flows, they're tools at the end of the day so we put compliance anti-money laundering that sort of thing through the models. And, John, you've been following the show for a while. You've been following the world of open source for a while. You talk about Sparks. You talk about Mac Studios. What's your current hardware count? I have, sitting in my home office here, about 2 terabytes of VRAM in terms of Sparks, Studios,

Minis, et cetera. Some of those will be older. we have access to that again in office and clients around the world, literally, from Central America to the United States and Europe. So access to quite a lot. These are wonderful boxes and they do amazing things. And you're launching something that can run across those boxes, I'm assuming. Not a lot of folks are running as much local hardware. Tell us about your launch. We've been working a lot with the KB Cash. We store a lot of our information. We take it out of the KV cache, put it onto SSD, and then you can get away with bypassing RAG. So you can basically load it back out. And I suddenly thought, hang on, you can do agentic coding, and you can basically take the KV cache, store it on SSD, park it, bring it back up again when you need it. And you design something specifically to be multithreaded. It's a bit like our go back 20 years ago when we had single core machines and then people come out with multiple machines, but none of the software, you know, Word and Excel were using the multi cores.

We've now got these 30, 40, 50 core machines. They're just not using the processors properly. You've got VLM on the Sparks, which does an amazing job, but there's nothing on the Mac. So I basically, since last Thursday, literally that quick, have spent my entire day, evenings, afternoons, breakfast, dinner, etc., guiding the Void coding, as we all do these days, to create a new inference engine. And I've been tuning that on the Quen models and obviously Glimmer and some of the other models as well to make sure that it works. And we've got crazy numbers out of it. I mean, this blows the sparks out of the water, not just one spark, but two sparks, and goes beyond, way beyond. I've taken some of the other sort of areas in this space. I've taken it up to 64, and then just today, 128 simultaneous agents, and the numbers just keep going up. I mean, this is on a Mac.

So this is just for Mac. I literally just put it live a short while ago. So it's inferd, so I-N-F-E-R-D. Yep. And then .incept5, which is our company, .ai. Inferred. So walk us through what we're seeing, please. Okay. So I tuned this on the Sparks, on the studios. Now, not everyone is lucky enough to have a studio. It's a massive, unbelievably powerful piece of kit. But it does mean that I can tune it. But what people can buy are the minis, the M5 and the M6. It's incredibly affordable at, what, about $1,500 for the 32-gig machine. So I was trying to get as much out of these. Now, you can't fit the flash next on it, but you can fit the 27B. And with 3.8. Yes, 3.8, 27B with a drafter. We're using D-Flash 2 on these.

we get some incredible figures. We can take this up to 16 agents. Many of the other inference engines aren't even able to run these on the smaller machines. And what we will be doing next is connecting these up with the Thunderbolts and try to get some better figures out of these. So just look at this, for example. That at the bottom is Splash 1.5, and that runs out of 16 agents. It won't scale past that. Now, I've released this open source. It's Apache 2, and I noticed you gave a round of applause thank you see this is the time of the show this is the person who knows what's going on on Thursday I've actually done now I have no doubt that everyone else is going to catch up with this by probably by the end of the week because that's just the way it works I've been I mean Tense of Folds, Ash's Tense of Fold is absolutely incredible and naturally has to appear in some of these benchmarks.

We've got Prince's work on MLX, Maya's incredible, Maya Labs builds that she creates. So all of this comes into what I've helped to build here. But look at this. Splash 1.5 stops at about 400 tokens a second at 16. And we just continue to go up. Explain to me what does it mean stops and continues to go up? What are we talking about here? With parallelism? Yes, parallelism or concurrency. Concurrency, yeah. They just don't scale at the moment. Now, again, they will change this, and I'm sure you'll come back on next week, next week's show, they'll have the line going straight up to the end, and it becomes a bit of a race. That's what it's all about in our industry. But so many of the others just stop at 16, and not only did they stop, they flatline, and so there's no point in going to 32 or 64. and as you can see here we've just continued to go on to try and build the performance

out of this so we're now looking at over a thousand tokens a second and okay that's divided by 64 or 100 agents so per agent you're only getting 10 tokens a second but ultimately you're still getting a thousand tokens to put into the processing and that's what's important That's incredible. So you call this inferred and folks can get access on GitHub. Let's go on GitHub and give you a star obviously number three. Thank you. And folks can run this on the Mac. And laptops as well. It works just as I mean many of my colleagues have been testing it for me on their laptops. All right Wolfram, one more comment and then we have to move on. What excites me about this is that it's optimized for ancient swarms basically and this is a new thing in that we have been seeing these swarms acting usually just in security-related incidents or something. But I think this is the next paradigm that we are now seeing after agents, that it's not just one agent, but really a swarm.

And I see the benefits myself when I just use a swarm of a dozen agents or something to tackle a topic. And the real-time are much better than if I just use my main agent to do one thing. I think this is something more people will realize soon. Having an inference optimized for this is, I think, very, very looking at where we are going, not where we are at right now. And that's why I found this great. How do you use it? We're using it with anti-praud, anti-money laundering. There have been some recent attacks, as you know, after the PHP thing. and so you get sort of attack surfaces and you want to actually look at those IP addresses and see what's happening. And so you can launch off swarms of agents to basically do the sort of research on this and that's been hugely powerful for us. So local swarms, I think, is what gets all from excited. John Davis, thank you so much for joining us. Thanks, guys.

Inferred from Incept 5. All right, folks, it's time for us to move on. Wolfram and I will tell you a bit What's happening in the world of CoreWeave Outside of the news cycles as well Welcome to From the Forge, Powered by CoreWeave As you know, I myself as an AI manager Of CoreWeave Wolfram as well And this week we have some stuff to tell you about specifically about Corweave and Copenhagen deal. So Corweave brings a 22 megawatt data center online in Copenhagen. The cool thing is the data center there actually gives back to the community. I think this is very interesting. The surplus heat in the data center goes back to Copenhagen's district heating network. I don't believe that I've seen some stuff like that

quite before, but I think it's very, very interesting. So now we have data centers in the UK, Spain, Sweden, Norway, Denmark, not quite, in Germany. We have some evangelism to do to get there. Our prices. I asked about why we do not have those data centers in Germany. And it's not the regulations, mostly. It's actually the price of electricity. And we have also a deal with CrowdStrike to explain the partnership, putting Falcon protection 24-7 threat hunting on Corvive Cloud. We talked to you about local swarm agents. Many folks are looking to cross-track to secure what's happening within them. Nobody asked me to do this. I'm doing this because I think it's very important to convey this message because a lot of speculation, a lot of fraud is happening out there. This is directly from Mike Intrader, CEO of CoreWeave LinkedIn. He's saying, my belief is some national ambitions to transcend politics, leading the next era of technology is one of them,

And here's the highlight that I want to read out. We currently don't anticipate any impact to Corway's existing customer commitments or associate backlog as a result of shifts in the political landscape. Corway already has more than 1.5 gigawatts of active power across our diverse global footprint. We also have a clear line of sight to 4.2 gigawatts of contracted capacity. We signed as of August 2026 giving us confidence our ability to serve our customers I think it a very important statement of reference from Mike Intrader from CoreWeave I just want to iterate something that people may not be aware of. This show is sponsored by CoreWeave, but we have full editorial control and everything. So, this is not something the company says you have to revolve with. This is Alex's decision. I honestly sorry, only this morning I thought it was very important for folks, specifically because We need to get Mike on next because he's posting this on LinkedIn and most of the extract is not seeing this. So I kind of wanted to connect this world. So we'll definitely work on this and try to get Mike on next.

All right, folks, next up we have a great interview with, well, hopefully a new friend of the show, Max Alexander. Welcome to the show. This is your first time here. We just met recently in New York where we got to chat. Max, you're a principal you work in developer experience a Capital One principal engineer I think that's the way you told me to shout out, is that still correct? I do, that is accurate yes, and then I would love to at least briefly mention some stuff that you did before for your career just to understand who we're talking about I think one of the things that you recently posted was about, hey, at some point I was given all of Google to decide how to write software for all of Google. Can you tell us briefly about that story in one, two sentences? I think it's very interesting. Yeah, sure. Once upon a time, I was the technical lead for CodeHealth for all of Google, so I owned Google's engineering practices. I'm the original author of Google's code review guidelines that still exist.

That's how most of the text in there was still written by me. And at one point, I was very sort of briefly grabbed by a coalition of engineers who asked me to sort of be in charge of Google's engineering culture, whatever that meant. and the story that I was posting with which I did nothing. I chose to do nothing with that opportunity. But the CodeHelps stuff, I did do quite a bit of work. Yeah. CodeHelps specifically is the topic of why I brought you on here today. So you gave a talk at the engineer. Does code quality even matter anymore? Tell us how you came to that conversation, how this was received, and basically feel free to also reply on the points that you did on the stage there because I think not everybody watched your talk from ANG. Yeah, great. So one of the things that was happening is I was getting really frustrated with the quality of discourse on X and on LinkedIn on this subject because people were just either super black and white. They were like, it never matters, and I'm never, ever going to read the code ever again,

and agents are going to write everything. And on the other hand, people were like, no, I have to read every single line of every single thing that everybody writes, and I can't relax anything ever at any time. And, you know, my observation, because, like, I live both in the open source world and in the very large enterprise software world, is that these people are not in the same universe. They don't live in the same world. Like, some people are building systems that will be worked on by hundreds or thousands of people and last for a decade. And other people are, like, vibe coding a front end for a thing that they need for today and that they're going to throw away. And the standard that you need for those two things are not the same. And that's always been the case, but it's just become so acute because we've entered a world where 100,000 times more software is going to become disposable. You're going to write it, and you're going to throw it away.

And then when you need better software and you need something for tomorrow, you're going to write that, and you're going to throw it away. The example that I often give people is like the Star Trek computer because that's what a non-technical person thinks of when they think of what is a computer that I want to have. Gene Runbeier is not a computer engineer. He's just thinking, what do I want to make for a computer in the science fiction world? And the computer in the science fiction world does two things. One, you talk to it. Two, it makes an ephemeral UI for anything that you need at any given time for you to interact with however you need to. And nobody cares about it. Nobody cares about how it's implemented or how it works or anything. But then if you think about that, who makes that thing? And who makes all the things that work behind it? That's like some serious business engineering stuff that would have to happen in order for that thing to continue to function as smoothly as it does for everybody. And so there's really just these, like, really two broad camps. And on one side, code quality really, really matters. Like, if you're going to write durable software, it really matters because it's for sustainability.

But if you're going to write disposable software, just don't care. You don't care. Nobody cares. And then... So this is the range that you identify, right? Like, all of those stakes and the quality of code especially, they land somewhere on the spectrum of disposable completely to very, very durable and stays with you for a long time, right? So can you define the two categories? What is the disposable software for you or the durable software for you? Yeah, sure. So you can generally tell from four criteria. And that is... Let me see if I can remember off the top of my head. One of them is consequences. So what are the consequences if it fails? Very often, very low, right? If you write a party planner for yourself and it doesn't work, you're like, oh, well, that's annoying. I've got to go use a spreadsheet or something, right? Next one is dependencies. So how many other people depend on it? How much change management would you have to do if you changed it? How many other systems depend upon it? How many, like, you know, could you break, like if you've made a breaking API change, would that affect other people? Would that affect other systems? And the next one is cost of replacement.

How long would it take you to rewrite the whole thing anew? You know, like, I can rewrite a calculator in an hour. I don't care about that. Chromium, I would have a harder time. I don't think I could recreate Chromium at all, period. And the final one is cost of maintenance, and in particular, how much human attention the system requires in order to continue running. I've seen a lot of systems where people are like, oh, it requires some human being to help you onboard. and then you're trying to onboard 1,000 people at once and then suddenly you need 1,000 people to onboard 1,000 people. And that's the thing where the quality of the system really matters because all of a sudden you've got to think about, wow, how do I reduce the cost of onboarding? Just as an example, not even the cost of the code maintenance, but even just the cost of having the thing exist. How do you see how to actually write proper, durable software in the edge of AI where more and more what we do is, instead of chiseling code by hand, now folks are monitoring agents doing some stuff.

How do you see this world at present? The way that I think about it is reducing the cost of durability. And I talked about this a little bit in the talk, but people often think, okay, well, I know all these engineering practices. I know, oh, I can write tests. I can write verification systems. I can add my static analysis tools. What people rarely think about is, how do I make that easier? I've spent my whole career, the last 20 years working in developer experience, where my whole life has been, how do I make it easier for software engineers to do the right thing? Not just tell them to do the right thing, not just I'm going to go do the right thing for them, but make it so dead simple to do the right thing that nobody would ever do anything other than do the right thing. The way that people have built software, and fundamentally, and I think we're going to talk about this in a little bit, the way that programming languages are designed lends itself to people writing bugs. Like my whole quest as an engineer for my entire career

has been to simplify the act of software development because I think that what's happened is a lot of very smart people have decided that because they are smart, they can cope with complexity. And we are now seeing that the answer was that they could not. And that was all untrue. So basically you're saying we humans are dum-dums compared to the stuff that we create, and it's impossible to hold everything in your head despite thinking that we can. So how do we still continue there? Like what's the next step? How can we scale those symptoms larger and larger, given that we do now rely on artificial intelligence more and more? You have an agent today, and it is more powerful as a software engineer than anything has ever been in the history of mankind. you can choose what to do with it. You can choose to direct it in a direction where you're like, please just ship results for me as fast as possible. Or you can choose to direct it in a direction that's like,

hey, what do I need to improve about this code? Tell me, not just what are the current security vulnerabilities, but how could I structure the code in such a way that, A, any vulnerabilities would become totally apparent, and B, those things would be isolated from my users or from other parts of the system, even if something went wrong. And then also a lot of the work that you've seen around like formal verification and static analysis and then this is also a lot of the work that I, and this is also like the goal of my personal side project too as well. Which I would love to go into next. So I have a side project. This is my personal project, so I just want to be clear. I know that my little subtitle there says Capital One, but I want to be clear that now I'm going to talk about a thing that's me personally. Many years ago in 2014, I had an idea about a programming language. And I had a realization about programming, which is that it is basically the management and structure and design of knowledge, not the design, not really math or engineering in a more traditional sense.

And I was never able to build it because I thought, well, nobody's ever going to pay me for two years to go into a room and produce nothing for two years. And finally, when Opus 4.5 came out, I realized I finally could actually start to build this thing. And so since November of 2025, I have been working on a programming language. And this programming language is designed for perfect static analysis. and what that unlocks is many things. I can list off a few of them. It unlocks theoretical maximum perfect optimization. Theoretically, you could have a better optimizer than any other optimizer that's ever existed in any compiler ever. It unlocks the ability to automatically refactor any piece of software and know that your refactoring is correct without having to run the software. it unlocks forms of supply chain security that are better than any other supply chain security that's ever existed

but I think one of the most important things that it unlocks in the world of agentic software is the ability to deterministically make rules about what your system may or may not do because what you need in a world of agents in order to totally take the hands off the wheel even for the most durable of software is the ability to say not just what the system does, not just what its behavior is, but how it accomplishes that. Like, let me give you an example of a problem. Yeah, let me give you an example. One of the things that inspired me to work on this. So at Google, I owned a static analysis framework. And the most important check in that static analysis framework said, don't do blocking I.O. on the request thread. And the reason for that check was, like a significant percentage of Google's incidents were cascading failures because somebody did blocking IO on the request thread and then requests queued up in some situation,

and then that caused a cascading failure. And so there was a hard check. It was a test failure. You couldn't ship the code to prod. It said don't. And the check was basically impossible to write and had constant false positives because in languages like Java or Python, you cannot successfully trace code paths without an incredibly expensive algorithm you cannot actually it is very very hard to know that like function x was called at any like i just want to write a rule this program never calls function x okay fine that is incredibly hard to write in traditional programming languages and then now you look at something like that now let's say like you want to say, oh, now we have formal verification. Formal verification is great, right? Well, first of all, and look, super supportive of formal verification. I think the whole rise of formal verification that we're seeing with AI agents is awesome. I think I love the idea that we're going to be able to really prove behavior of systems. But depending upon the type of verification system you're using,

the most common verification systems mostly are going to either prove that this specific program has this specific behavior, which is fine. That's often better than a test. It depends. Or this specific program has this specific implementation detail. There's some verification systems that can look at your code and say, okay, you're calling the IO we'll say object, just to be simple for people who are not functional programmers. We'll say you're calling the IO object and you shouldn't be calling the IO object. This program bans IO. So formal verification systems can do that. What is very hard for them to do is to write a proof that no program ever does something. No other code that other humans or agents will add to the code base will break that rule, basically, is that what you're saying? Yes. Yes. I want to say, let's say on Google, I want to say nobody may ever ship code to prod

that calls blocking I.O. on the request thread. Right? And theoretically, if I could really do that, I would be able to trace this calls back to concurrency state. And if I could do that, then I could do that for every program. Like I could just write that rule one time and it would apply to every program in the world. But the programming language has to be structured in a particular way. Because here's the problem. The problem is another program has to be able to look at your program and make statements about it. This is what static analysis is all about, right? This is what static analysis is all about. It's all about another program being able to look at your program and make statements about it. And by the way, lots of other cool things that I forgot to mention. Like the language will do automatic concurrency. So you will get automatic maximum parallelism. You will get a certain level of verification, like not formal verification, but more like what's called pragmatic verification for free.

and it will basically be enforced by the language. I will tell you, even just from my experience building this language over the last year, the only thing that has ever let me take my attention off of the agent is knowing 100% that it cannot do something. And now I no longer need to look at that. No amount of instructions, no amount of skills, no amount of agents.md, whatever, whatever, no amount of hooks, whatever, has ever stopped the agent from messing things up in my codebases other than static analysis or tests that fail. I think, Max, I want to land on how this helps the world where most folks will not only just run one or two agents, will run agents 24-7 for disposable systems, but also durable systems that will continuously improve, continuously check for security modification, whatever. How what you're building is helping that world of us managing swarms and armies of agents to do things for us in the future?

Great question. So two things. One of them is you want to be able to know that the thing works that you wrote. That you asked for. You asked the agent for. And ideally, you want the thing to come out right the first time. And you want the thing to come out right more quickly. And so the total amount of time that you care about is not the total amount of time that the compiler spends running or whatever. The total amount of time that you care about is the time from intent to correct result. And if you can put as much of that as possible into a deterministic system that removes all of that choice from the agent, everything that is not a thing that the agent should make, like a decision the agent should make, you reduce it. Let me give you an example. One of the things the language does is tries to guarantee that there's only one way to write any correct program. And that sounds crazy, but I can do it more, even right now with the state of the languages and right now I can do it more than any other language can do it today because I don't want there to be stylistic

questions, there's no stylistic questions if you do have to go read the code there's only one way it could have been written so you never have to worry about how it got written whether you're an engineer whether you're an engineer, write good code because there's only one way to write the code it's just that way and And the other thing that unlocks theoretically that I haven't yet had enough time to experiment with, but I believe is possible, is a much lower level of intelligence that is much faster and much cheaper, being able to be much more capable. Because you've given it so much deterministic power and so few choices for how to do the right thing. So significantly higher quality output of smaller, potentially open, completely open source on your computer without any clouds or etc. execution because there's not that much to do. They can write the simple software. Okay, so Max, you talked about three mind-blowing things here.

What is next here? When can people get exposed? when people can review or other folks reviewing this is only in your realm. Let's finish up on where is this thing going because it sounds incredible, but obviously you have big claims here that need to be proven by someone else. Yeah. Yes, completely. And I don't expect anybody to believe me, right? Like I need to show people that this thing works, which is why I haven't released it. Well, like why I'm not making it public while I'm working on it. because I do have some pretty absurd claims. And I want to be able to demonstrate them. And so the thing has to work. I have to get the thing into a state where I can demonstrate that my claims are, you know, that I'm somewhere within the bounds of my claims. And unfortunately, I mean, look, I have a day job. I love my day job. It's wonderful, but it does take up eight hours of my day. And so I can only work on this 46 hours a day. Yes, I am working 14 hours a day. so that is my primary

constraint which is that I do not work on this full time if I could I would absolutely be working on this 12 hours a day no question I wake up in the morning I walk to my computer I work on this and then I do my job or this thing until I go to sleep pretty much every day of my life and so I am working on it as hard as I can as fast as I can it is unlike anything else if anybody's ever done anything novel with a coding agent you know that it's not like doing things that are copying something else. So it's a lot more time and attention for me than it would be normally in most scientific building systems. But I am hoping that I will have something in the next three months that I can share to people. Max, thank you so much for joining the show. We'd love to bring you on once you are about to release whatever you decide to call this thing and chat more about how Asians operate in this new world. Thank you, Max. Thanks. Thanks, Hans. All right. We're almost at the end here. Let's at least cover OpenAI's Images 2.6. OpenAI releases their new and up-to-date Images 2.6.

Images 2.6, Flare and Sunburst. OpenAI is leading the AI generation models now back to lead them. Same price of 2.5, but significantly faster. Flare and Sunburst, number one in overall design arena and image arena, ahead of 2.5. Now they can also make GIFs. I tried this a little bit in chat GPT. It didn't really work for me. Now you can make GIFs within chat GPT with these models. Basically, what this means is that they sped up the generation so fast they can generate multiple images in succession while you wait. And here's an example of four seconds per generation for ads. And then they can generate a few of them in succession and generate GIFs, which is pretty cool. I did want to make a comment on the DeepSeek harness because a lot of people... Yeah. Talk to me about this. Yeah. So, I tried it. Yes, it is very good. Explain. DeepSync released like a cloud code type tool. CLI tool, right?

Is that what DeepSync released? Yeah. DeepSync released their own cloud code tool. It is very nicely done. It's pretty modular. You can customize it. Oh, and it has a UI. Yeah, yeah. The UI is good. It's very neat. Nice design. I tested it for it phoning home and stuff, and I found out that by default it does have telemetry on as hard-coded. But then I dug deeper, and it's not sending your data to China, just to remember that. But there is a big but here. Whatever your agent searches, it goes through api.deepseek.com. slash tools. So, all of your searches, and there's not, that is kind of normal, but there's something you've got to keep in mind. 274,000 stars on GitHub for this harness, which really scored a few months ago. Very good

at scale, but they did a whole bunch of updates. It's pretty neat, and it hits the cache very frequently, and it's very modularly done, so you can customize it to what you want very easily. Just keep in mind, by default, if your agent is searching, all of your search results are going to desec. So that's just the one thing you've got to keep in mind. You can turn that off. It does not send your data. I mean, I tried. I just had a honeypot thing to see if it would send the data. I couldn't find anything wrong. The only time it does send the data is if you explicitly provide a slash feedback. so again just one thing otherwise it did seem pretty secure speaking of security there's quite a few things in addition to Daybreak, Blue and Red and the stuff of the GPT-6 Cyber that we've talked about there's quite a few things that are happening in the world of security and you alerted us about Playwright specifically

two words on that muted muted you're missing. Yeah, so, Playwright was, you just run your browser, and this was more for people that just work out of their laptops at a coffee shop and stuff that often Playwright, and it's not necessarily Playwright's fault, but your agents will have like multiple older versions of Chromium and Chrome, and those are very, very, very easy to hack now, and they are being exploited in the wild. So, you can basically You just pair that with an attack that just does the root certificate, the DNS certificate. And this is a very two-step process that you can just take someone over if they're doing things normally. So that was my big concern to that. And Claude just found it on my 7 a.m. briefing, and I looked into it. I've been expected to blow up and get 1.7 million views.

I post stuff like that all the time. It doesn't get much stuff. There is a new one now, and someone just managed to break out of the AWS sandbox. So Firecracker is what runs all of Amazon Lambdas, Amazon Web Services. A lot of websites depend on this very tiny VM. I think this is because Alpine Linux team is very small and was not able to just shift the patches. They still haven't compiled or updated to Chromium for that. So, yeah. And Amazon was always held as, like, the most secure sandbox. So that's where we are. Again, you don't need to freak out. Just have... Don't click random links. Yeah. Always good advice, I think. All right, folks, it's time for us to conclude the show after two and a half hours. It wasn't a huge, like, release week, but definitely a big week, in terms of how AI affects us and implications in the world.

We had two great conversations with John Davis and Max, Kat Alexander as well. Nistan, LG was here. The whole team was here. Peter, Yam, Wolfram and I are signing off from CoreWeave, folks. Thank you so much. If you missed any part of the show, the show is getting edited and it's turning into a newsletter and a podcast. Everywhere you get a podcast. Some folks prefer the live show with the whole conversation. Some folks prefer the edited version. We are here to make sure that you understand the power-up of AI that AI gives you. We are here to fight doomers. That's the stated goal of many of us, despite it being scary with the vulnerabilities, etc. I think there is a dark period that we need to go through, like we talked with Max, to get to a point where a lot of the software is durable and is in place, as well as a lot of people understanding the power-ups that they get with AI, like all of us do as well. So if you are into that and you want to share this with your friends, please do. We would appreciate it. Thank you so much for joining Thursday. I will see you here next week.

Next week is a special show from the floor of AI Engineer Code Conference in San Francisco. I would probably have a few guests on the show, including Swix, to come and the show. So definitely tune in next week for that. With that, see you here next week, everybody. Thank you for joining. Bye-bye. Thank you.