Open sourcestormed the castle

Hey, everyone.

Welcome to ThursdAI. What's up, Wolfram? Let's add you in. Hey, man. How are you doing? Welcome back. Thank you. Thank you. That's so great.

How have you been?

I just came back from vacation, and I'm excited to dive

in to talk about this week's AI news, 'cause I was kinda, like, out. I tried to disconnect, and, and finally I'm back. So folks who are listening and tuning in, welcome to ThursdAI. Thank you so much for coming. it was an exciting, exciting week for sure, a lot of open source today, Wolfram. A lot of open source this week, and- Yeah, but open source- … and I'm excited to talk about- … Wolfram Ravenwolf: now on the mind of a lot of people because of some, certain incident. So I think there's renewed interest even more than before.

Yeah, I think let's start there, right?

So last week we told you about- Claude Fable. The Mythos level model that Anthropic finally decided to release after saying that Mythos is too dangerous to release, I have a picture, Wolfram, of me and my wife just Fabling all the way to our vacation. Just like we didn't even go to sleep during the flight, seven-hour flight. We're just like, we're literally just like both of us worked with Fable. and Fable was amazing. And the day after that, Fable was taken away from us, just like that. The US government decided to tell Anthropic that it's too dangerous and, someone, and I think it was somebody from Amazon, showed the US government like a jailbreak. Yeah … and, and that caused the US government to actually, tell Anthropic to, to stop releasing Fable for us.

Do not make Fable or Mythos available to anyone who is not of

US, nationality, including even people working in Anthropic, even developers of the model themselves were, to be prevented from accessing their own model if they are not, na- US Americans.

Just absolutely a crazy… I think this is the first time that

the US government bans an AI thing.

The legislation attempts we have seen in the recent

times that they want a month to basically test the model before it is released, so this happened of course before this was implemented. And yeah, the question is will we see even more advanced AI? I mean, that is the question on the mind of many people. Or if we see it, can it be removed at any time? What will you do if you depend on this?

so last week on the show we talked about… LDJ, you wa- you

wanna comment on this super quick before we, before we jump in, in the TLDR? Because I would love also to hear, like, from you on, on this topic. Just also make sure that you hear.

Yeah, I mean, just so many exciting things happened this week that I feel like

almost, Fable getting banned feels like it was, like, over a month ago . Like… But yeah, it's really exciting and… Fable b- getting banned, not necessarily exciting. I'm talking about the other things. But, Fable being banned, that aspect is unfortunate. I think it comes back to this broader question we could talk about later, which is will AI eventually become at a level of advancement where it's not allowed to regular people? And I think if it comes down to, hey, should it be disallowed to regular people or require some type of ID to access, I think most people would obviously pick the, the latter. Of course, they're both maybe things we wouldn't prefer, though, so we could talk about that later.

I expected to come to the show today, well, well rested after

vacation, but I also expected to come here and talk about Fable just exclusively. Hey Peter, welcome. and, just exclusively talk about how cool is Fable. And I think… so we finished the show last week exactly I think all of us got super excited and g- went to play with Fable, all of you listeners as well. And then my whole timeline was like Fable, Fable, Fable. I don't know if you g- you guys had this experience as well. It was just like Fable pill timeline 100%. Everybody was just like, "Holy crap, this model is something else." And then yoink. so the rest of the week was like open source and stuff. but then my whole timeline was, "What the fuck? The US government decided to ban, you know, Anthropic." there's… I actually don't know how deep to go into this because, like, Thursd AI is not a political show. We try to kind of stay away, but now AI is so involved with politics, we, we, essentially can't. US government has a bunch of statements about the security, and we have a bunch of statements. At least I can tell you that when we talked about Mythos at first, Anthropic, you know, used some very s- scary words to s- to say how good Mythos is. including mentioning that this is a weapons grade kinda like model and it's too dangerous to release. And then, and then Anthropic is also, "Hey, you know, US government should regulate." And then when, when they release a model, the US government regulates, Anthropic is like, "Oh, no." You know, it's, it's, yeah.

And although Anthropic has not been on the best

of standings of the US government because of the incident before where they tried to prevent the government from doing certain things with their models, and, yeah, it was a big clash. So basically, maybe is that a kind of revenge act, revenge act from the governments to-

I, I know, from Anthropic though, as, as far as reporting goes, at

earlier today somebody said from Anthropic that they expect that, this model will be back at some point soon so we're gonna wait maybe today, hopefully today. we'll just say the Fable was, was incredible. Peter Gustov from Marine AI, welcome to the show. Thoughts on the, the banning. Like what, what do you feel? Did you feel the-- they took away a piece of, you know… They, they let us have it for a little bit and then, and then the government took it away. What, what do you feel about that?

No, I still have some tests saved from the time when I was testing

it and, yeah, I come back to it and it's like, "Man, that feels good. That was something else." That was a big move. yeah, you test other models now, it's like, "Oh yeah, that's a big, big gap." yeah, it's very unfortunate. Yeah, it's-- But you know, it's trying to reflect, I think it's easy for us to kind of sit in our bubble and kind of say, "Well, it's clearly ridiculous, you know, what is this even a jailbreak?" You know, and then like, this is true, right? That's not really a jailbreak. You know, you can study the code, at least from reporting it or something like you can study the code and then can kind of patch vulnerabilities and therefore you can discover vulnerabilities. But then on the other hand, if you're just gonna talk to like a normal person who has no idea about this, and then you say, you keep saying, "Well, it's like a nuclear weapon and it's like gonna kill everyone, take e- everyone's jobs." And then they say, "Okay, can you fix it then?" I say, "Well, no, no I can't." So that's, that is like, well, then maybe you shouldn't have it. So i- yeah, it's like I, I know the logic steps are not quite exact there, but like at a very high level you can kind of also see the reaction of like people outside the vowel bubble. so yeah, I think it's just coming back to that, like we need to be careful with our words, right? What kind of imagery we use, because I, I think that's not accurate.

Yeah.

So hopefully they're just gonna kind of get, gain a

bit more trust between each other. 'Cause I think probably what, what Anthropic is thinking is like, "We are like the safest lab. How can we not be safe enough?" Like but maybe not, not safe enough, you know, if, if it is a nuclear weapon.

The jailbreak itself that was reportedly identified by Amazon

and shared with the US government from Amazon, that's the reporting that I saw at least, you guys let me know if you saw something else. Was about the model identifying some bugs in some code and like, being asked to fix it, and it's not a generic jailbreak. And so Anthropic's response was, "Hey, this was not a generic jailbreak. It's not like somebody broke the whole model for us. And then the, the thing that I also saw is that US government's like, "Hey, you can only release this when you've patched all jailbreaks." And Anthropic's like, "It's impossible to patch all jailbreak. What are, what, what are you talking about? This is, you know, this is not how jailbreak works." Big open source week, folks. We had two big model releases We had GLM 5.2 and, and we have also Moonshot release, Kimi K2.7 Code. both of them performing very, very well. And we almost got a fat cat. Did you guys see this? the meme of Le Chaton Fat, or Le Gros Chatons. I

tested it already on Wolfpage, and it

was at the number one spot,

Yeah.

No- Talk about that … no Le Chaton Fat for us. but it was a really funny meme. So I think, I think worth jumping into the TLDR, and then we're gonna actually dive into open source first. Peter, you have some testing that you did, oh, and then we, yeah, we'll, we'll talk about a bunch of stuff. We should also mention, we'll mention this in a second. Cursor finally, accept or I guess SpaceX went IPO last week as well. Folks, it's, it's an insane week. SpaceX went to the biggest IPO in the world. The reason why we're mentioning this on the show, not because we love space. I personally love space, is because it's not just SpaceX anymore, it's SpaceX AI 'cause, you know, they, they grabbed, you know, space, they grabbed X AI and then X as well. and then SpaceX also acquired, Cursor for $60 billion, making the four co-founders of Cursor all billionaires. and, you know, now I think Composer 2.5 is Groq Composer 2.5. I don't know if you guys saw this. They already rebranded. It's available in, in Groq Build as well. So, you know, big, big moves. definitely Space- SpaceX AI is now joining OpenAI, Anthropic, a- and those folks up there. and also Noam Shazeer news. So a lot of news. I think worth going to TLDR, and then we'll cover open source. Let's do that.

All righty, folks, let's do a TLDR of our show. TLDR is a se- segment on ThursdAI where we talk about everything that happened really quick, at least everything that's very interesting to us, so that you will not miss even one thing. And then we'll dive into the most important things and cover them at length. So today on the show with you, Alex Volkov, AI Evangelist with Weights & Biases and CoreWeave. and then, our co-hosts for today is Wolfram Recht, LDJ and, and Peter Gostev. Peter, Gustav from Arena, and we have guests today. So Dexter Horthy is gonna join us, today to talk about, Human Layer, a relaunched Human Layer, and also the role of AI engineers in this world of, of SLAP. we also have, two folks from Weights & Biases joining us, Adrian Swanberg and Chris Van Velt, as we are launching something very cool from Weights & Biases that you all will definitely enjoy, I believe. So we'll talk about this as well. right. So in big company, in LLM and API news, I think the big news from yesterday from OpenAI is Noam Shazeer, one of the co-creators of the Transformers paper, one of the co-authors of the Transformers paper, the co-founder of Character AI, you know, previously in DeepMind twice, I believe. Noam Shazeer is joining OpenAI. I think this is big news. I don't know if there's a lot of details there to report, but Noam Shazeer joining Noam Brown and becoming the Noam-less, the Noam, the Noam-most frontier lab in the world. Noam Shazeer is joining OpenAI. I think it's huge news for everybody here. very interesting that they have been able to convince him to work on something that he wasn't able to work on in Google. So this is like a weird standoff, right? Like, nobody gets access to Fable-5. As far as I understand, inside Entropic, people don't get access to Mythos as well because of that order. it's a very interesting order. I think it's a huge implication that the US government is finally kinda like, regulating AI in a very direct way, and some of it is very murky. SpaceX, like we said, acquires Cursor or Anysphere, I guess, and the Cursor product for sixty billion dollars in all-stock deal. Obviously, SpaceX went to the largest IPO in the world last week as well. And so the, the genius math there is that, s- you know, SpaceX made so much money from the public market that this deal is barely a blimp on the radar. For everybody else, sixty billion dollars is a huge amount of money. For SpaceX, given the, what, two trillion bump val-valuation, v-v-v-valuation, this is almost, almost nothing. And still, SpaceX now has Cursor, and SpaceX, obviously is the owner of xAI. I think those are the major, major news from Big Labs. Folks, any other major thing we should mention in Big Labs that, that, that is worth mentioning this, this week?

I was super excited to see, the Codex announcement from

OpenAI, that Codex is now fully available, not feature restricted in the EEA, so the EU, UK, and Switzerland. Because we didn't have the full memory feature, computer use, browser use. It's all available now. So while other things have been restricted, OpenAI has been opening up

I think I have it, lower down there, but yeah, Codex and, and

computer use also is now available in the EU with no restrictions. We have, Peter, you're in London. Wolfram, you're staying in Germany. You guys are finally getting, getting access to some of these tools. that's incredible. let's move on to open source, and I think we had a very big week in open source, folks. GLM 5.2 drops as the strongest open source coding model with solid one million, context in, tokens and context. this is-- I don't actually know the order of these, folks, but you tell me w-what are up first. Moonshot AI also, announces Kimi K2.7. Kimi K2.7 Code, one trillion MOE coding model with 30% fewer reasoning tokens and big benchmark jumps. And then GLM 5.2 comes, and GLM 5.2 takes number one on Design Arena open source that is now, I think, beating Fable in making websites. This is absolutely crazy. it's not like I can run GLM 5.2 on my home computer because it's a one trillion parameter model. But, I don't think that we've ever had open source matching like frontier labs in design, and I'm really interested in testing this out. So, Peter, you tested some of these models out, at least Kimi K2.7, but also GLM 5.2. so we'll definitely chat about this. I'm very excited to see how they perform. So we got Fable yanked away, but open source delivers. This is the type of week that we have, and I'm very excited to open the show with open source. we also did not get very big fat model from Mistral, but everybody kept talking about this. but, the from Mistral. In tools and agentic engineering, Claude Design gets a major, major update. Design system imports now. You can edit it in Canvas and the Claude Code sync bidirectional. It's really cool. We-- Last time we talked to you about Claude Co-- Claude Design, we showed off and we weren't able to use this. Now, hopefully, we'll be able to use this a little bit. HumanLayer launches an agentic IDE to fight AI code slop. HumanLayer's founder, Dexter Horsey, is a friend of the show. He's gonna show up here to talk about AI engineers doing a lot of slop. Dexter is one of the more famous kind of speakers from The AI engineer, scene, and we're both gonna talk about the AI engineer world, world fair. I believe Wolfram and I and old Dexter are all speakers there. and then we're gonna chat with Dexter about this. OpenRouter also launches Fusion API. Wolfram, I think you dropped this one. I saw this at a glimpse, but you wanna give this, like, a two-sentence description of what we… wha-what's important here?

So it's basically a way to achieve fable

intelligence at half the price. And what they are doing is they are, through the API, you just send one call, and it goes to multiple models, and then a, a judge model basically combines the best answers. And something like the consortium we have, seen before by Caparthy or how we have been doing it manually in the beginning, where you had asked multiple AI and make an answer out of the best parts of each response. So this is now on the API level, so you don't have to build it yourself. Just use it like a model, and bam, you get great results.

That is

super- I think it's super

important Shout out to OpenRouter for this combined effort.

European users- Yeah … in EA and UK and Switzerland are getting access to Claude Computer, Claude's computer use Chrome extensions, memory, and Chronicle. Chronicle-

Mm-hmm

… Alex Volkov: which I've never used since they announced it, and I've never used Chronicle since then. Chronicle is the feature in Codex where it takes a screenshot of your screen every five seconds or so, so you have full context of what you worked on. yeah.

Remember the, the, recall feature they had?

Yeah. And that was retracted because of privacy issues and so on. That's true. And now they have this as well.

Yeah.

I'm gonna try it again. I tried it at the beginning. It was killing my laptop,

By the way, mine is turned on and I, I don't know if I like noticed

any performance issues, but the, yeah.

Anyway.

thing in this week's Buzz we're gonna get to, the corner where I

talk about Weights & Biases and CoreWeave. Weights & Biases launches Hivemind, folks. We're gonna have the folks from Hivemind come here. It's a unified dashboard to track your spend and ROI across all of your AI coding agents. we support Codex and Cloud Code and a bunch of other stuff. it's really cool. I wanted something like this for the longest time, and honestly, I had access to Hivemind for the longest time, and I didn't, wa-wasn't able to talk about this 'cause it was in development. it is really, really cool, to see like a unified dashboard of all your, all your agents. and you know, obviously, Weights & Biases tracks your AI spend. so unlike Weave, which is a developer tool that you can track agents in production, this is for your coders in your company, so you can work as a team and see how many tokens together you spend. We also, launched Kimi K 2.75 code live on wandb Inference at very high speeds as well. I think we're, we're, we're tracking to be the number one on artificial analysis, I believe. Yep. Made it … with 289 tokens per second. Wolfram, go ahead.

So basically, we, with 289 tokens per second, we were in the

first place or are in the first place. And, they also have a quadrant where they show speed versus price, and we are in the most effective, quadrant as the only provider actually.

In vision video, Google DeepMind finally launches Omni.

So we talked to you about Omni at Google I/O, but Omni finally got launched, as an API to be able to get, tested, and now it's on different arenas as well. Peter, is it also on Arena AI to, for folks to test as well, Omni?

it scored really, really well.

Yeah. it was like, you know, GPT image two level in terms of the, of the- Of the

jump between the previous one, right?

It was, yeah, very impressive.

So Google's the first any-to-any model.

but specifically I think folks are testing it in the, you know, fr- from the video generation aspect, and I think it's scoring incredibly well. So shout out to Omni and we're gonna take a look at a few more. And xAI, relaunches or launches finally Grok Imagine Video 1.5. they call it 2X faster generation. Honestly, xAI Grok video has been more and more incredible. They've been working really hard. it's kinda hard to track xAI Grok's Imagine, because they keep releasing this without like actual version numbers and it's like, "Oh, we updated Grok Imagine, we updated Grok Imagine," so it's annoying. But finally they gave us a version pinpoint, 1.5. It's native audio, and they claimed number one leadership position also, at different places, at Design Arena, Image to Video Leadboard and other places as well. So, we're gonna take a look at Grok Imagine Video, and I think it's really, really cool. Folks, yesterday, Midjourney decided, Midjourney, the company that creates Midjourney image generation model, decided to say, "Hey, we have an announcement, a hardware announcement. Let's, let's talk about this." And, Midjourney announced Midjourney Medical. Midjourney Medical, the first hardware device that they announced as part of it is the full body ultrasonic scanner. It captures 80-- 800 terabytes of data per scan in under 60 seconds. and the reason why we're talking about this here is because the only way to handle all this data is AI. and also Midjourney is an AI company. It's self-funded. There's, there's no VCs there, so they can do crazy stuff like this. it is definitely part of sci-fi is here. Like we'll tell you all about this. It's really, really cool. I think this is all on the TLDR. Wolf, anything that I missed? I think, I had a, a few more things, but I think that's there's plenty of us- plenty of topics here to talk about.

That's the most important stuff, for sure.

Most important stuff.

All righty. all right, folks, let's dive into open source. I think we're very excited to talk about the two open source models that we got this week, after Fable was yanked away. Let's cover open source.

open source AI. Let's get it started. Folks, today, we're talking about two open source drops. Actually, three. Let's start with Le Chaton Fat. I… Guys, I think I missed the meme. What-- C-can somebody please describe to me what is happening with this fat cat from Mistral? Like, what's going on? Somebody says 30 trillion parameters, MWE, what, what? please, somebody tell me what's going on. LDJ, I see your hand raised, and then we'll go to Peter.

Yeah.

So I believe the origin of this is actually, I think it was, Arthur Mensch or one of the, o- one of the founders of, of Mistral, they posted, a loss curve, a chart of loss also just being a way of measuring error essentially- Mm-hmm as the training improves and it goes through, happens over time. What they showed is like in the top right corner or like the corner of this picture with this loss chart, it said- Mm-hmm … like the name of the training run and it said something like "Le Chaton Fat" and then some, some other things. And so people are like, "Holy shit," like, "What is Le Chaton Fat?" And then people started making a bunch of jokes of it from there.

Ah, interesting.

and so people kinda decided to treat this as a release w- with benchmarks and, and details, but it's all bullshit, right? There's no, there's no G- Le Chaton Fat, Le Gros Chaton, whatever. Peter, w- what's your take on this Mistral thingy?

Yeah.

It's like I, I posted this in internal Slack, and then our CEO said like, "Oh yeah, post it." 'Cause I was like, "Is this, is this too dumb?" but yeah, this was the GPT-2 image, by the way.

Oh, nice.

That's it. Could you, for folks who are just listening, could you describe what we're looking at?

Yeah.

So it's, it's our normal Agent Arena chart, and Le Chaton Fat is a bar fable on top and goes outside of the axis, and the caption says, "We have requested an urgent authorization from the French government to extend access to display the result on the Agent Arena." and I feel like everything was getting like enormous amount of engagement, which is I think just people minds. I think e- everything, the, the whole I think dynamic with Anthropic has been just completely insane last several months, and if I look back at my tweets that got like the most engagement, like pretty much all of them is some stupid Anthropic related thing.

Yeah.

So I think I get up to that as well,

There's definitely a vacuum between the announcement of

Fable getting inked away and some of the open source stuff that people wanted to, to talk about things. So we're not getting a fat cat from Anthropic. They did, sorry, from Mistral. There's no Le Chaton…

Yeah, I posted the source in the side chat-

Oh, nice.

stream yard here.

Let's take a look.

Oh, Guillaume Lampel, the CTO, I, I believe, or chief scientist of, Mistral, posted this. Let's take a look. Oh, he posted a WNB curve. Oh, that's interesting.

Like, so this is an actual post, like, not, not a, not a troll.

Not a meme. And then people took this and you see in the corner it says BLC_Le Chaton Fat FPA. Mm-hmm. And so they just named that the training run, and then people started making a bunch of speculations which turned into jokes from there.

Yeah.

So, you know, it's obvious from this chart that Mistral is using Weights & Biases. Just a shout-out, to the Weights & Biases team. this blew up. This is 1.7 thousand likes on a training run. He posted this with nothing. This is just a loss curve, a beautiful loss curve. w- he posted it with no commentary, and then people, people just got, like, super, super excited, about this. All right. let's add Yam Peleg to the stage here, and then let's move on to talk about actual open source models. I think that-- let's talk about GLM 5.2, I think. GLM 5.2. Peter, I think when you said, "Hey, I want to come and talk about this, I can't stay for long," this is, this is the model that you wanted to cover the most of it, right?

Yeah, I can talk about both, and I actually really curious to hear what

everyone else thinks, 'cause, this was one of those releases that is quite confusing. because if you look at the benchmarks, normally the way it goes, right, a company that is like caching up, right, not, not Anthropic, for example, they would release a model. They'll have a bunch of benchmarks with it, which are super cherry pick. Maybe it's like sometimes it crosses the line into like, abuse or not. but normally it's like, you know, a few good benchmarks, and then other benchmarks come out and they all kinda, "Ah, actually it's not as good as you think," kinda thing. That's normal. That's what I normally see. And here it's like almost every benchmark is like, "Oh, actually it is really good." And, and you know, we had some where it's even, like above Fable, right, at, at certain points, which is like, okay. and then I was, I was trying to use it over the last couple of days, and I had some, I think inference is still in ensure. I don't know what, whether you guys have a, whether you have good, good access. But I was trying to use it, and I don't know, I wasn't getting like particularly impressive results. Like, I'm sure, like I haven't done a lot of side by side with like, 5.0 or 5.1. I'm sure it is better. but like, I don't know. It's like I'm not like that impressed. So in the sense that it's amazing to have such a great model that is like not massive and is open source like that, that's kind of as given. But I wouldn't go as far as saying that it's anywhere near in the category of Fable. so I don't know. What, what do you guys think? Like I- I'm, I'm kind of a bit confused, like how come all the benchmarks are so good?

I think that, we've seen similar things from kind of GLM before.

some of this was claimed to be benchmark, benchmark. Some people are like excited because it's open source and like it's cheap to, to access. as far as I saw, I didn't have time to… Like I just came back from vacation yesterday, so I didn't have time to test it myself. I tried to test it this morning via Devin Desktop, so like it didn't complete yet, so I need to like finish that. but as far as I saw, people are getting very excited about, design stuff with GLM specifically. LDJ, what, what's, what's your thoughts on GLM 5.2 specifically?

Yeah.

So my thoughts on it, especially with the last capability you said there, the, the design aspects of it. I feel like from the perspective of the fact that capabilities are getting to a point now where even if you do specialize in a certain area, that could still be genuinely valuable. I feel like GLM is going in that direction when it comes to front end design and these types of things. I think it's definitely not at the level of Fable or even 5.5 when it comes to maybe more math related things or more, let's say, experimental design for actual research. But when it comes to something specifically like designing a website and actually having good taste there, it does seem like in some cases even competing almost with Fable.

Design Arena also goes for the vote, which one you prefer

type, kind of like voting in ELO score. And so supposedly, if they're doing their job correctly in, you know, collecting those, you know, votes, folks have been selecting GLM 5.2 results specifically for design tasks higher than Fable and by quite a margin, 13, 60 ELO score, 1350 for Claude Fable, which we know is great. Last week we showed you Fable. I did the whole website design. Peter, you did your designs with the 3D stuff. Like Fable was incredible and, I think b-between the excitement of Fable and actually the deliverance of Fable, there's a one-to-one correlation. Like everybody who got excited w-- like, yeah, Fable's that good. Tell me, tell me what, what, what are your thoughts on, on, on,

I, I don't know what design arena do specifically, but for us is

that we have kind of tentatively kept it on the leaderboards for now, even though we kind of, we… If it doesn't come back, we, we're gonna remove it. But one nuance there is that when you have these kind of side-by-side things, so it just kind of, for example, on agent arena, we, we have, like, the overall set of models and, and then, we measure different signals to compare to each other. But the point is, if our model is actually not available, then you can't really compare, Mm-hmm … to, to it, so it's not like an absolute benchmark. so one nuance could be is that actually if you were to put them side by side, maybe it, it, that kind of, that these numbers wouldn't hold. So this might be just one thing to co- to bear in mind. So maybe one thing to expect is that if there are numbers like that, if Fable comes back, it might just shoot up again. So, but we'll see.

Yam, any, any comments from you on this?

Have you been able to test out GLM 5.0?

I usually keep around, one of the, the cheaper tier of models,

just for, you know, k- kind of, un- unlimited, unlimited quota kind of model. Like, so you can add up and just offload very, you know, large, parallel agentic things to it. And for, for quite a while, it was, 5.1, so it was really good for this use case. I'm just happy that, now we have 5.2. But, it's gr- fu- It, I just, I was just kind of surprised that the benchmarks, weren't better, if, if it makes sense. I mean, it is better than 5.1, but I was expecting I, I was expecting a little bit more. I mean, yeah, it, it is a confusing release to, to say the least.

So let's talk about- Yeah … the actual, the actual kinda

like facts of the release, right? So GLM 5- 5.2 from ZAI is, you know, they, they ment- mentioned it specifically, "Strongest open source model for long horizon tasks. on Frontier SWE measures whether agent can complete full engineering projects. GLM 5.2 scores 74.4 trailing Opus 4.8 by just 1%." but again, the benchmarks don't tell the full story, but we'll mention them as well. Terminal Bench 2.1, it's getting 81%. Wolfram, we haven't tested this on Wolf Bench yet, like we're waiting for this model to go up on inference, but, Terminal Bench jump should mean something great. And also, Peter, I wanna bring you back 'cause like we did talk about the Agentic Arena once you guys launched it, and I believe the GLM 5.1 was the, the top, like open source one.

Yeah

I remembering this

Yeah, please … yeah, 'cause, in there I can, I can show you.

Yeah, so it did do very well, so it's number 10-

Mm-hmm

on the- Number

10 overall and number one in open source, right?

Like from open source model- Yeah models, this is like the, the, the leading model in Agentic Arena.

Mm-hmm.

So it's only like slightly better than 5.1, for, for now. So we'll, we'll need to see, you know, if, how, how that shakes out. But I, I would say, like I've seen some other benchmarks and it's like so close and so on. Like I, I feel like ours is a good reflection of where I would, I would place them. so yeah, it's… Yeah, and I, I've been trying a few things, but I'll, I'll just show you while, while I have the screen.

Yeah.

so I've got like, "Find an interesting dataset, research it,

like, a- and show to me, to kind of educate me." Like- I know I don't have the side by side, but like Fable is so much better than this. It's like that doesn't feel that engaging. I,

I wonder if it's fair though, Peter, to compare this to Fable.

like, I wonder if it's fair to com- to compare any open source model to any, even the last frontier model that's not like Fable style. or, you know, h- how do you compare this to like other open source? Like this looks decent, although very AI- Yeah like native design. Yeah, no- This thing where, where like rounded corners and a little bit of color, all AI models do this. Specifically the Opus models do this, this specific thing.

Yeah.

That's a tell.

But yeah.

Yeah.

And, and that's the thing, like I think we-- I just want to calibrate this a bit, like that it's not, I don't want to say that this is like a bad model or something. Like this is clearly an amazing model. Yeah. Like the fact that it's open source, like, so it's, it's kind of trying to keep the things in my head, like the fact that it's open source and we can like take it and run it, like that's amazing. That's- Open source. It's

MIT licensed.

It's 753 billion parameters MOE with MIT license. So it's incredible. But yeah.

Yeah.

So that, that's all like, I'm, I'm not arguing against that part, but I'm definitely arguing against that when we kind of see in the benchmarks of people saying, "Oh, it's like, it's, like almost there at the level of Fable." Like I, I think it's nowhere near. Like it, it's the-

Yeah.

let's just hold our horses a little bit.

Like it's good model. It's maybe better than what you'd expect, but that's not nowhere near that level yet. And like it would be kind of crazy if it was, wouldn't it be? Like if it's like- Could be … a 700 odd billion parameter model, same as Fable, like that, that sounds just implausible.

Fa- Fable is in the what, trillions?

We don't actually know. but yeah- Yeah … I think somebody mentioned like in the, in the trillions of parameters. so this model- That's like- Yeah, go ahead.

That's also the jagged prompt here where models can

be close in one pa- one area and they can be very far apart in another area and you can't, in the benchmarks, if a model is great in a benchmark, you can say the same model with the same score is as good as the other one. You can only say it as g- as it did as well as the other one in this specific test, this kind of benchmark. So saying this is on the same level as Fable would always need the disclaimer in this particular benchmark with these tests we tested. So that has to be kept in mind because every benchmark is only a, a small part of the model that we are actually seeing and testing for this specific use case.

I don't know why.

not because of the model. It's just dev and desktop, just like literally just crashes in the middle. I have no idea what's going on, but, you know, once, once I have something, I'll show you. what else do we have here? So f- terminal bench are good. any- anything else we should mention about this, folks? Should, should folks try this out? I mean, if you are focused on price and performance, I think this is the model is, like, very interesting for you. also, Peter, the thing why I don't wanna compare this to Fables, but w- we don't have access to Fable anymore. So, let's compare this to model we have access to. and API pricing is very cheap, so, you get, I think, for free in dev and desktop, actually. Yeah, I'm running this, and I think it's for free. and, it, you know, it's available pretty much everywhere, hopefully soon on, Weights & Biases Coreweave Inference as well. Let me just add this. Like, folks are saying that, like, they find output quality comparable to Opus 4.8, but, like, roughly eight times cheaper. some folks have t-tested for a few days, long horizon tasks, very good for agentic use. All right, folks, I think let's move on and talk about, Kimi K2.7 Code. I think it's the other big, big model release that we had from this week. and I think obviously Kimi is, the darling of the open source industry while we wait for DeepSeek, although Jan mentioned y- you do use DeepSeek. so Moonshot AI released Kimi 2.7 Code this week as well. Also a one trillion parameter MIT. So, so GLM is not quite one trillion, right? Like seven hundred eighty-six, seven hundred fifty-three billion parameter, million parameters. billion parameters. Sorry. Moonshot AI's, Kimi 2.7 obviously is a one trillion parameter model. it's an MoE model, thirty percent fewer reasoning tokens and jumps on benchmark. Let's take a look. they have their own Kimi Claw benchmark, which is very interesting, kind of like OpenClaw-ish, benchmark. Program bench is at, fifty-three percent, beating the previous Kimi at forty-eight, and Kimi Code. I don't see any major benchmarks here, though, so let's take a look if we can find them. But, folks, thoughts on Kimi 2.7. Wolfram, I think you mentioned that we already have this running on the WB inference or CW inference, I guess, at incredible speeds.

It's running.

It's fast. I did a test with it, but my initial test was with thinking off, and after that I realized it does not support thinking off. So my score, I didn't publish it because of that. I will rerun it with, thinking on, which is, basically the default and the only supported way right now.

Yeah.

so six x high speed mode coming soon. So this is not, like, yet even, like a speed, speed model. twenty-one percent plus on Kimi Code Bench V2, eleven percent on program bench. Peter, thoughts on Kimi 2.7 Code?

one important note is that, it's not multi-model,

so it doesn't have vision.

Yeah,

Which is, like, qui- quite a big part of coding right

now, which is kind of, I guess we need to make allowances for that. but on, on Kimi, maybe if I, if I just show- Yeah my screen again. Like-- And then we saw this on a few benchmarks. I was kind of slightly worried about this, our, our benchmark here that it, it showed that Kimi, 32.6 is actually slightly better than 2.7, so that's 2.6 and that's 2.7 Code.

This is on Agentic Arena, right?

And I was like, n- you don't normally see

that," so I was a bit worried. Is the-- Are we doing something wrong here? Sometimes there are weird bugs or something, right? I think- Actually, it's almost

Just, I saw the same that it did not beat, 2.6, but since

I can't really say anything about the real models, so I have to redo the test.

Yeah.

So w- there was one, benchmark I remember. I don't have a link for it, but I think the, Van- VengeBen- bench, I think, they saw the same, that 2.7 did worse. and I didn't look into it, so this is more kind of a speculation off the top of my head, but they did talk about kind of being faster. So I don't know, did, did it get smaller or something or did they, did the sparsity increase? I wonder if they did something to make it, like, a bit m- nicer to use, but actually maybe they had to compromise on other things. This is not informed view, but I, I just, I'm trying to explain, like, how could this be that, like, the next version i- is weaker? So it could be that maybe it got more narrowly trained on code and tasks, so maybe it'll do better on certain things but not others, But, yes, it, it wasn't like a big jump that we saw with the other models.

I think that the code, that they specifically called it the code

model that goes in that direction, that it may not be as strong in the other parts. And also that they re- reduced the reasoning, that could also have an effect like that. So in Wolfbench, on the second spot is, Gemini 3.5 Flash, which is super strong, but if you look at how many tokens it needed to get there, it actually cost more than if you were using GPT 5.5 because, it needed so many tokens

I think it's worth mentioning here, Wolfram, as you said, that we're

now offering this model on CW inference, and this is like very, very fresh, so you guys are welcome to go and try it. Kimi K two point seven code, is, the highest output, per artificial analysis, at least as of June eighteenth. we're serving this model at two hundred eighty-nine tokens per second, you know, very, very fast, with reasoning as well. Blackwell runs FP4 natively, and so our applied training team re-quantized it to NVFP4 to full GP300 and GP200, throughput. So accuracy held across benchmarks and li-Life Core Bench suite bench and Pro and Multilingual, and we added the speculative decoder on top. most AI models write the answer one word at a time, which is slow. speculative decoding is the technique that let the model draft a few words ahead of, ahead of it and check all of them in one shot with the bigger model. So we added the speculative decoding on top and, we're now showing speed versus price as the highest in a tier. So that's very, very exciting. folks, CW inference is available, and you can definitely check it out. Kimi has been kinda like the darling of the open source folks. Not-- I, I don't believe the license is quite open source. Wolfram, you, you wanna chat about the license of Kimi?

license.

Let's

take a look.

I

license modified MIT.

So I, I think, I believe that the, like, the license includes some provisions for, like, bigger companies required, to, to pay for.

Yeah, well, a hundred million monthly active

users or more than 20 million US dollars in monthly revenue, then it is required to display it properly.

but yeah, gen- generally MIT license, use it for anything.

But if you are a company and have these requirements, 100 million monthly active users or more than $20 million revenue, in monthly, you shall prominently display Kimi 2, K 2.7 code on the user interface of such product and inference. So you cannot just, like, use this model and claim it to be your own. I think, though, generally, as, you know, a- a- as a podcast that celebrates open source, I think we absolutely must celebrate the fact that, like, th- these models continue to push the frontier of what open source models can do, in, in complete open source. LDJ, I haven't heard from you about Kimi K2.7. what's your take, at least from your, zeitgeist about Kimi K2.7? Anything interesting that we also should mention?

Yeah, I didn't get the chance to go too deep into it, but, based

on the benchmarks, it's interesting. Compared to 2.6, like it's, in some of the benchmarks, over fifteen, twenty percent jump. yeah, I mean, if we don't get the, the Fable model accessible to us within the next few months, it's inevitable that I think like six months max before we probably have a Chinese model that's like clearly parity or better.

Yeah.

I think, o-one thing worth calling out is that Kimi K2.7, according to Moonshot folks, is built specifically for coding available through Kimi Code and the Kimi API. may not be just like a general agentic model, which, Peter, could be something that the, the, agentic arena is testing more favorably to. And so they're saying that, you know, for, for Kimi K2.6, and is more recommended for the general purpose tasks, non-coding tasks. so that's a very interesting call-out from the, the Moonshot folks as well. all right. I think that's all on open source, and we, we, we did quite a lot of it. let's move on to big companies. I think there's a lot to talk about Fable. Peter, I know you have to drop at some point, but like at least about Fable, let's stick with us if you can, because I do, do wanna cover that part as well. We don't have any transitions, but folks, last week, just what a day after or two days after, I believe it was on Saturday, when I saw the blog post. I actually disconnected, deleted Twitter, can you believe it, for a few days over the weekend and, and I got a notification from Reddit that the US government-- that Anthropic posted a post about the US government requesting a block of Fable and Mythos to every, foreign nationals, which caused Anthropic to close Fable access to everyone. and so I do wanna go through the actual kinda announcement because Anthropic did write some stuff there, and also cover where, where it's at currently. So let's see if I can find the actual post from Anthropic, which saw-- I believe this post now has ninety million views, which is insane. This, this was by far the biggest news of last week, b- in terms of, like, what we saw. The US government s- says Anthropic publicly on their Twitter, "The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable-05 and Mythos-05 by any foreign national, whether inside or outside of the United States, including foreign national Anthropic employees. The net effect of this is that we must abruptly disable Fable-05 and Mythos-05 to all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption. We believe this is a misunderstanding. We're working to restore access as soon as possible. Read our full statement." The full statement now includes the supposed jailbreak that somebody found and told the US government, that the, the supposed jailbreak is allowing the model to, to code, I think, and find vulnerabilities in the website. but I think the ba-- the, the matter is, access to all other Anthropic models will not be affected, so, like, we still have, you know, f- Opus 4.8 for example. but, the government believes it has become aware of method of bypassing the jailbreak in Fable-05. We reviewed the demonstration of this specific technique, and these vulnerabilities all appear relatively simple. We have found that other publicly available models are able to discover them as well without requiring a bypass. So basically, somebody said to the US government, "Hey, this supposed weapons-grade, model from Anthropic is jailbreakable," so the security and safety stuff they added are not enough.

I also heard that it was by an Amazon, person.

The Amazon, wasn't it a CEO or something from Amazon who did it? Somebody from- Which

are investors … yeah, requested the,

the product.

Which are investors of, Anthropic as well,

Oh, the Amazon investors in Anthropic, yep.

folks, what do we think? This is the first time, as far as I have seen this, that besides when the Department of War and Anthropic went head-to-head, and then they threatened to use the Defense Production Act to Identify Anthropic as a supply chain risk. Besides that, this is the first time the US government actually declares something that blocks access to an AI, model, the frontier AI model. What the fuck? I think at some point we decided, like, we'll get there, but, like, what the fuck? Let's talk about this.

And especially if it is also possible to do the

same with GPT 5.5 or other models. I mean- Yeah … they have not been blocked. So what's going on here?

Yeah.

Peter, thoughts on the, the, the best model that we've tested so far getting yanked?

Yeah.

No, it's a, it's a definitely a sad day, and it's very confusing, right? There, there's like a lot of reporting and, that's… I think journalists doing a good job, like digging stuff out, so I'm quite, quite enjoying reading the covers. But yeah, it's kind of-- it's one of those moments where you, you're trying to think, "Is this it now? Is this like our life or is this just a blip?" And I guess, I don't know. Standing here, I have no idea. Like it could just go back tomorrow to, to normal releases and labs kind of don't pay like any special attention. but I think international competition, well, from China is gonna matter, right? Like, okay, US government can block things, but, but it's like six months or twelve months behind, whatever, right? Clearly, GLM models already, by the way, sanctioned by the US government, I believe, already. and despite that, they're doing a very good job.

the model.

No, no.

As in, as a GLM, as a entity. Mm. So the, the ZAI, Zihu AI, they're, I think they're sanctioned by the US government, like especially not, not like every other Chinese company. so, yeah. So I, I think it's kind of also not tenable position to kinda keep that. And I think also the fact that so much investment is going in and, I think US government's quite keen to also keep the investment flowing. So it's kind of, I can't quite imagine this position holding, but I guess it's different people within the government also have their own opinions, which I think is not unreasonable. But yeah, I think a shock to everyone how quickly that happened.

it happened very quickly.

Everybody was very excited to play with Fable, continue to play with Fable. some folks were holding out to the Quoras. Also, as a reminder, Anthropic only let us play with Fable in the max tier until like June 22nd. So everyone's like trying to like-- to, to token max, including me and my wife on the plane to our vacation, like both are just like token max Fable as much as possible. The amount of time my wife came to me and said, "Hey, when's Fable back? Like, what's going on? I want Fable. Nothing else works as well," is more than five times, I believe, at this point. I think we have some good news-ish from earlier today. So some reporter quotes an Anthropic official that said, "Anthropic is confident that in the coming days, Fable 5 will become available again," international managing director from Anthropic, says in Seoul. So there are discussions even over the weekend with the US government from Anthropic officials to how to bring this model back, because folks, you must remember Mythos-5 is being used by other companies via the Glasswing Project, Glasswing. and Mythos-5 level model, Fable is also Mythos-5, is also being used within Anthropic. And that model is very, very good at producing other models continuously. Like the frontier of AI is being deliberately, how should I say, held back by the US government because of this issue. And if this issue is just a jailbreak, that all model as are jailbreakable, all models are jailbreakable, but not for the specific purposes, then what are we doing here? Are we, like, pausing ourselves while China catches up? So, as-- Let's see. Counting on this panel, given that I'm a naturalized citizen, LDJ, assuming that you're a citizen as well, please raise your hand if you, if you are.

Yes, I am.

So me-- only me and you technically are allowed.

will this lead to us having to provide a, government ID before we get access to Fable? Like, like this, this could potentially have huge implications in how frontier AI continues, right? Wolfram, we keep talking about a, y-you being in the EU, and the EU restricting some access to the more, like, computer use stuff, et cetera. And EU kinda like is, like, in a stronger on regulation side. but this is, this is definitely, like, a big thing because of, Peter, of what you said. Of we don't know whether or not this is the new way of accessing frontier models, and we all are h- it's worse L- let me pause here and say, last week on the show, I asked all of you if it's better to have access to Fable-05 with the restriction, the Tropic added, you guys remember this? And we all said it's better to have access with the restrictions. Like it's all better to have a restriction with it falling back to Opus 4.8 sometimes or saying, you know, or being down when you ask it to, to do some frontier AI research. we all got access and then we all quickly learned, what is it like to get access to a very good frontier model, Nisten level frontier model, and then getting yanked. LDJ, I think you raised your hand, as, as a commenter. What?

Oh, I was just raising my hand to, confirm yes,

I heard that you-- that, that if they bring it back and it's

only still restricted to US citizens, then you'd be able to, to get it. Peter, thoughts on, on, o-o-on like the jagged frontier?

Yeah.

one thing that I found interesting is that, it's kinda hard to read between the lines or like from the signals, but it does look like, you know, there's G7 right now, right, in, in France, and it does look like Macron was, cozying up to Dario. And I wonder if like the Europeans are like starting to see some opportunities there. I don't know what, what the plan of action is there, but I think you mentioned, Alex, like, or Europeans like regulating more heavily, but like no one's banned it here. So it's kind of-- I, I wonder, I don't think there is anything they can particularly do. I think US can still block it, but I wonder if there is like more interest in, in having some more AI in, in Europe somehow, because I think a lot of the governments and people waking up is like, "Well, this could happen with any model." And it's like, what, what are you gonna do if US just says we're keeping it for ourselves? I think it's bad for the US 'cause they just spend money and they, they don't, don't have any alternatives. But if, it, it's like commercially bad for the US, but like completely disastrous for Europe. So yeah. I mean, we've only got basically one lab. There are a bunch of like product companies, startups in the UK which are doing very well, like Granola, Eleven Labs. They, they, I guess they, they are-- Eleven Labs have their own models. but yeah, a lot of the, a lot of these startups will just go out of business, I guess. I, I don't, I don't know what they're gonna do. Yeah … so yeah, it's a very heavy implications for Europe as well.

Excuse me, Dan.

So if anybody from the government listens to us, we want Fable back, please. go ahead, Wolfram.

It's definitely a precedent now that this has

happened, that it's on the mind of many people, especially in other countries where sovereign AI, it has always been something where we were saying, "Yeah, we need this." But now we see for the first time in real life that, one- maybe the top model AI has been restricted. And even if it's brought back to the American folks, it also means that there's now a real gap between the intelligence available to people based on their nationality, and that is something where if you want to build your business, think ahead three years or something when we totally depend on AI like electricity. But in this case, it de- determines how well it works, and I think, imagine that a j- a company from Europe is competing with a US company and it has AGIs or it has unlimited employees doing everything they want, and now it can just be shut down by an outside force. I don't think anyone wants to depend on this for anything that is important to them.

I

don't like what's happening- So let's see

what happens in the coming years.

I agree with you.

The, the precedent is scary, but honestly unavoidable. Folks, if we are in fact talking about those implications that they mentioned, th- this is like nuclear level, change and if, you know, if Mythos is breaking through the Anthropic CB2 level, which is stands for like chemical biological weapon, is able to create novel chemical biological weapon, some, you know, us in the AI world was like, "Hey, we can make HTML pages." To be… Well, to be clear, th- this ban is not because of that. This ban is because of a jailbreak that it can do, like, like a hack or something. Nisten, you wrote about this as well. We're talking about Fable getting gamed by the government. Welcome to the show, by the way, Nisten.

I love how much it annoyed the Europeans as well, because

they were the whole times that we have to put restrictions. We cannot let these things happen. Look at how… And then, and then they find out, "Oh, okay, now we have the… Now we're getting put restrictions on."

Yeah.

Yeah.

That's a problem. That's, well deserved, yeah. But, maybe they learned something.

Yam, I don't think we heard from you on Fable as well, on the yanking.

Any- anything, more to add that hasn't been said? And if not, we'll move on to SpaceX and IPO.

I don't like it in, in a way, shape, or form, and I just think it's

gonna disappear, and it's gonna come back. The model is gonna come back, and it's all gonna be over soon, I think. It just makes no sense in real life. Like, you're gonna have a, a DeepSeek, DeepSeek version in couple of months anyway, and I don't know, jailbreakable like every other model in existence. And just like you're saying, yeah, you can develop weapons with it, so for, for quite a while already, and it seems like, I don't know, a government is solving things, the name of things instead of the actual things. And, and it's just these things don't hold up usually when you do regulations based on the name of the thing and not the actual thing in real life. Yeah. I, I just hope it's gonna be over soon. We're gonna get, Fable back.

I really want Fable back is, is my- Yeah … is my feedback from this.

all right, folks, we have to say bye to Peter Gosar. Peter, thank you so much for coming, folks. Check out Arena, Agentic Arena, definitely. Peter, we'll see you hopefully next week. All right, folks, we're moving towards, space. So last week, we barely were able to mention that just before the IPO for SpaceX, Elon Musk and other folks announced the space satellites, the, the, the, you know, the AI satellites in space. And then SpaceX went to do the IPO, largest IPO in the history of the world. But why are we telling you about stock market when we can tell you about you, we can, we can tell you about AI stuff? Very shortly after the SpaceX IPO, which made Space Uncle Elon Musk the first trillionaire in the world Kuros. Do you guys think anything changed materially in his life after he crossed that benchmark of billions, you know, hundreds, thousands, hundreds of billions to trillion? Not-not-- nothing changed. He can still fly on his fucking private plane. I don't think there is anything for him to buy more than he couldn't buy before, but, you know, he, he passed the trillion mark. SpaceX acquired Cursor. Finally. We told you about the SpaceX deal where, SpaceX paid Cursor ten billion dollars for some of their data with the option to buy them at the insane valuation of sixty billion. Cursor, the-- they rewrote Cursor since, but Cursor is essentially the VS Clo- VS Code clone, that has been making rounds, became a, an AI training company with Composer and Composer 2.5, which are great, and, a lot of data because they gave Cursor for free to many people while taking their data. And, SpaceX, specifically, xAI's Grok was not that great at coding for, for a while, and the reason why was data, right? So, they bought Cursor for their training opportunities and data, and, this deal is now valued at sixty billion dollars. all share deals valued at two eleven dollars per share, making the four co-founders of Cursor all billionaires, very young billionaires. three days after the insane seventy-five billion dollars IPO, the largest ever IPO. Thoughts on this acquisition, folks? Not necessarily the IPO. On the acquisition of Cursor and how well is Cursor doing

I, no, go ahead, Nisten.

You first

I think for the data it's completely worth it because now they also

have, basically info on how, how every company that uses this operates, how every enterprise, every large company. And, yeah, Elon went to full, Factorio, pro gamer mode, and, got to, j- just got quadrupled the high score of everyone else. And you see, 'cause it's, the whole list of billionaires, and then it just starts to look kind of small, like, is this the best humanity can do? Like, we can do a lot better. so yeah, great on that 60 billion to have data on every single enterprise who thinks that, data retention policies don't have other legal clauses for retaining the data. That was pretty stupid of them in the first place. It should've gone open source. That's just my opinion.

Wait, but stupid of who?

I wanna hear the hot take,

Of, of every enterprise and company that thought that the no data

retention for training policy, actually meant that they don't keep the data. We see time and again in legal cases, in legal evidence, that the deleted data is still there. It is being used in court in real life. So, yeah every company that just gave it all away, now they understand what kind of game is being played.

Let's talk about the implications of this.

the implications of this. First of all, Cursor keeps training Composer 2.5, but also Grok is now going to be better at code. Is this enough to make SpaceX AI a frontier lab to compete with OpenAI and Anthropic, at least on coding? Because as we saw, at least on Arena Agentic benchmarks, on our benchmarks, Grok 4.3 is maybe a great model for research, but it's shit for code. As we also saw, all of these labs, including Gemini with Antigravity, including OpenAI, obviously, with Codex, and Anthropic with Cloud Code, all of these apps now are now making more money on coding assistance, which they are. Like, the, the most amount of money that, that AI makes right now is on coding assistant, Cloud Code, Codex, et cetera. OpenAI is refocusing on code. They also all realize that coding is the way to generalize the agent. That's why all of these harnesses are getting built together with models that are getting trained on these harnesses, right? Is this-- This is obviously the play for SpaceX with Cursor. Is this enough to put them at the frontier? Is knowledge from the Cursor people, plus the insane amount of GPUs that SpaceX has, which they're now giving, you know, they're, they're selling away to Anthropic, to whoever else, is this enough to make them a frontier lab in coding within, like, six months?

I think it's also not just the coding, but

it's the agentic stuff as well, where you need a great model. And, as far as I know, xAI also allows you to use their subscription with other agents like Hermes Agent and basically any of these. Mm-hmm. So they are now, if they have a good model for code, for agentic use, they are on the same level basically, if it's good, as OpenAI with something you can put into any agent and have your model used for all of these things. That is one, one aspect of this, and the other one, we are losing a, a generic, coding agent like Cursor, which was independent of the models or you could use it with any model. We have no idea how this will shape out if, Cursor models will be the preferred ones, if the others will be offered or whatever is happening here.

Yeah.

So definitely a shout-out to the Cursor folks. Like many, many people loved the, the Composer model as well. The Composer model is now available via your xAI subscription, by the way. They are working really fast. So if you, don't have Cursor, but you have the xAI subscription or Hermes or, or something, if you pay for X Premium, you get, Composer 2.5, and Composer 3 is gonna hit really hard. Composer 2.5 is already a great model. If it's now trained on the hundreds of thousands of H100 equivalents within the Memphis supercluster from xAI, Composer 3 is gonna hit super hard, and it's coming very soon. So the race is on, and I'm very excited to see this race. it's very interesting, going back to the US government for just a second, that they prevented access to Fable-5 in terms of using it, but not in terms of training other models. So even foreign nationals that they prevented, are still able to train other models, but not, but not use it. LDJ, last comment, and we'll move on.

To add to exactly what you just said at the end there, it doesn't seem

like there's any directive prohibiting them from developing an even better model than Mythos, and so I don't think actually this will cause any type of significant lag long term because six months out from now or whatever- whatever is the best model, I think will still be essentially the same capability in either parallel universe of whether this directive happened or not, if that makes sense.

all righty, folks, it looks like it's time for us to hit

this, the, this week's Buzz Corner. let's go to this week's Buzz. We're gonna tell you all about a very new, exciting Weights & Biases product that is just launched that I was waiting for for a while, and I'm very excited for. Let's go to this week's Buzz.

All righty, we-- welcome to this week's Buzz, the corner of Weights, it's a ThursdAI that we talk about the Weights & Biases and CoreWeave different things. We usually tell you about models that we put up on the inference, we tell you about different features they release in Weave and Weights & Biases, but today we have a special treat for you guys. We have two, great folks from the Weights & Biases, including a co-founder of Weights & Biases, CVP, believe that you've been here before, so welcome back, Chris Van Pelt, from, one of the co-founders of Weights & Biases, and, Adrian Swanberg. So I think without further ado, how about you guys give an introduction of a new launch from the Weights & Biases team within CoreWeave. CVP, stage is all yours.

Hey, everyone, excited to be on the, the pod again.

that intro video for this week's Buzz, man, it's so good. And, Adrian, I'm, I'm a little jealous of your background, man. I'm at my, Yeah, background

looks incredible.

I'm in my childhood home in Iowa visiting my, my family,

so I've got a lovely background here. But, no, yeah, excited to be on the pod, talk about Hivemind, a project we've been working on for about six months now. and you know, the reason we started building it is we, we saw that fundamentally what it means to be a software developer has changed. And Hivemind is a, is a tool that kind of sits alongside those tools and observes what's happening. So it, it captures, the actual transcripts or the, the interactions in, all of the popular, agentic toolkits, and then stores them in a central dashboard that allows, individual developers to kind of keep track of all the work they're doing. So especially as we're doing many agentic tasks in parallel, you can quickly see at the dashboard level, what's happening and, and where the agents are. but I think more interesting at a team level, when you have multiple people working on a repository, you're able to share sessions with each other, point to, a different feature we're some pretty cool features on top of that, that I- I'll, hand it over to Adrian to, to talk about

As I, as I put Hivemind on the background, first of all

folks, congrats on the launch. I've been a beta user of Hivemind for a while now, and been ve- very excited to like hop in there from time to time and check out on my, kind of like work across different agents, as you guys see right now in the dashboard I have- Is it,

is this your dashboard, Alex?

Yeah, this is my dashboard.

I have Codex in there, I have Claude coding in there.

up to lately?

as you can see from the tags there I'm working on my conference

talk and I'm working on Thursd AI website development and, configs to, scripts and Hermes infrastructure different things.

Nice.

This is our, you know, our take on, on AI, agent session transcripts. we try to process everything and make it more useful to look at. on the right you can see we have broken out your session into various chapters so you can get an outline of, of what actually happened in the session. So we, we try to publish this outline. We'll show you which files the agent touched, we'll show you which, to-dos that it might have wrote, whether you had context compactions. and yeah, we can, we can walk… There's a lot of functionality on this page, so we can kind of just like walk through some of the features. In the top right, there's a little button that looks like a, a fork next to the, the eyeball. so here's something where… So one of the interesting things is like Hivemind helps you stay kind of, a harness agnostic. So if you, if you actually, I don't know which, agent you used for this, I, I forgot to look. you used Claude. This is Claude. but if you wanted to pick this up in say Codex, you can hit the, the fork button, copy that command. If you run that in your terminal, Hivemind will actually pull down a compacted session history for this session, and then it'll give you the choice to launch in whichever harness you want so it's a pretty interesting way to, to try, you know, take, take something you've been working on in one harness and, and try it out in the other harness.

I want to

This has proven to be invaluable when, Anthropic servers are on

fire and I just gotta get something done. I can switch over to Codex. It's nice.

Wow.

Yeah.

Super nice.

And this is a compacted version of the conversation, right?

Without like the tool call, outputs, et cetera

It's not the full transcript.

we don't, you know, we don't translate everything back into the native harness, message formats. but we do, we do, we do write quite a bit more context than you would otherwise get, I think in the native compactions. so it is, it is a bit heavier in that sense.

so, we can go feature by feature, but like I, I just wanna like

ask you guys 'cause, Chris, you mentioned something very important of why this is launched, and obviously we have Weave. Weave is focused on tracing agent execution for people who are building agents on production, and Weave has been invaluable for folks like tracing, and making sure that their agents are staying on point. Who is this for? Like, who is the user of Hivemind? Last week when Fable launched and I ran this long workflow, dynamic workflow thing in Cloud code, included like, multiple sub-agents running and like tokens, I really wanted to count the number of tokens I used, and it looked like-- I was like, "Oh, I'm very happy that like I have Hivemind running in the background without me thinking about this, like how much do I store?" So am I the target user or is the company who wants to make sure that their token maxing dashboards are not getting out of whack the target audience? Like who are we launching Hivemind for and who is the set of features for?

Yeah.

So the way we think about it is, is really there's, there's two key personas. you know, number one, it's, it's like us, the users of these, these tools, the developers, being able to get value from a tool like Hivemind is, is really important. So, you know, from, from the use case you just said of, of just wanting to, to be able to go back and, and find something, like without a tool like Hivemind, it can be pretty awkward and difficult to go find something, especially as we start to spin up these things in ephemeral environments in the cloud. Hivemind kind of gives you a way to have that, that, universal view and a, and a bunch of other nice features, that, that we could talk about. But the other core user is, really the, the team lead or, someone within an organization that wants to understand how their teams are, are using Hivemind. So, you know, a common thing we've been, been seeing in the news a lot of, of organizations becoming worried about their token spend or, or saying like, "Are we, are we spending wisely? Are we actually using these tools well?" Hivemind is a tool that can actually help answer that question so we can know how many, of these coding sessions are actually ending up getting merged into- the main line of the, of the repo, or better understand how the engineers are, are using tools. if we're starting to see more users use a tool like Cursor, then, potentially it's time to, revisit our agreement with Cursor. Or we're seeing trends of more Codex, again, giving the organization an ability to kind of predict or understand how the usage of the tools is evolving and how they can, best support the engineers. And then for the engineers to be able to say, "Hey, this is actually, really valuable. we're actually shipping software that isn't really buggy and is providing value to the organization."

Yep.

and I think that the… It's been talked about a lot about the token maxing dashboards and whether or not folks are using the tools productively versus just using the tokens and, and, and maximizing out. I do wanna talk about the users themselves. I think, we are in this, like, beautiful era of, like you said, all of software development, it was reimagined from scratch. Many of us are like now agent babysitters. They're all, like, changing to this agentic monitoring dashboard-ish, right, thing. but again, y- they can only give you features that are part of their product. We are giving you features across all agentic harnesses. c- could you guys talk about which ones we support? And then I will-- I do wanna talk about this new persona of the folks who use the AI the best and how we can help within the organization for other folks to learn from them.

Yeah.

So on the agentic harnesses, Codex, Cursor- Clawd Code, PHI, GitHub Copilot, and I think I might be missing one more. But e- essentially we designed the system to be really extendable, so as we see users, using different harnesses, we're, we're happy to, to kind of continue to add support for other harnesses. I think, yeah, the two things that are really important to touch on: one, there are always gonna be power users in the organization that, seem to be able to really leverage these tools effectively, and one of the coolest features in Hivemind is the ability to generate a persona for one of those power users. So at, at CoreWeave, Tim Sweeney is definitely one of those power users. personas, allows Hivemind to go look at the history of sessions that Tim has done and then say, "Let's create a skill called Talk to Tim that, that tries to embody the different approaches and techniques that Tim is using." and this has been a super popular persona within our org to be able to kind of have a virtual Tim to, to get guidance from.

Yeah, and we should note, you know, this is actually

sort of a grassroots feature. we had another engineer build this functionality just using the Hivemind data that we expose via our API. but everyone loved it so much we had to upstream it into the tool itself.

I love this.

Wolfram, I think you have a follow-up question as well

Yeah, so my question as a non-engineer basically is, how do

you see it, more like as a dev tool? Or is it, since you me- mentioned engineers and merging, or is it also intended now or later to be a general purpose tool for everybody as agentic use becomes ever more widespread in all organizations and not just the engineers are using AI?

Yeah, I think today it's very much a, a dev-focused tool, and when

a product is as, as early as this, I think it's important to have that kind of focus. But we've already gotten requests from other people in the org as, as, Cowork and other tools have, have really unlocked the power of these things to, to everyday users. So we'll, we'll see how it evolves, but, right now dev-focused. And then the last thing, I wanted to touch on is the, the insights feature we have, with Adrian maybe you can just chat about quick, which is, you know, often when these agent sessions are happening, you end up running into some kind of roadblock or you're correcting the agent. Hivemind can detect when that happens and then provide guidance on how you can steer the agent better in the future.

Yeah.

Let's see. I got a… I think I figured out the screen sharing here. Yeah, so we, we've, we're kind of building, in bits and pieces this, intelligence engine on top of all the transcripts that we're ingesting across the organization. so we, we use AI to go analyze how these sessions went. we find issues, ideally issues that are, are correctable in some tractable way. we export them, we cluster them together, we sort of find the, the commonalities, and we expose them here on, on this dashboard for, for people to see.

is not- Find it if you want, and also zoom in a little bit so we can

actually read, like, what's going on.

call.

Yeah.

Yeah.

This is on the wandb-core, repo, I believe, right?

Yeah.

So a little bit of secret sauce spilling out in here, but that's all right. let's see. a lot of preferences,

What are the charts on the right?

Those are per session charts of like how well, how, like well the

Okay, so here's, here's a good one, right?

we have this like very specific pitfall that, at least two users have run into recently across nine different sessions. The, the chart here is the number of times this specific issue has been detected in sessions that have been uploaded- Oh, that's cool … to Hivemind. And so what, what the, what the system has come up with is this, suggestion that you, you may or may not choose to put in your agents.md file or, or somewhere in your, in your code base's, agent context, to basically nudge the agent to avoid this pitfall. and you can do that yourself or you can copy this neat, command that will automatically smart merge this insight into your repo's agents.md file. you know, we'll actually use like your agentic harness to find the right place in that file to, to put the edit. so this is kind of the beginning of our, of our effort to help, basically help the organization become better and better at using these tools and, and avoid the pitfalls by leveraging, intelligence.

I think it's incredible, because of how quickly we're

building tools that evolve with how the profession evolves, right? So like from training models to then running agents in production to now actually running those coding agents. I think, a lot of these tools are, a personal pet peeve of, SAP that like they become products afterwards, so I think that's lovely. Can you give us actual insights towards the models or the products, being used? Do we see a Fable blip in, in, in the thing once Fable was on? can you give us like what are organizations using AI for based on what we see? What are we using? Are we seeing the rise of Codex coming back as it's mostly cloud code? I want like actual insights that we can share at least, with our audience based on what we can collect. And obviously, how can folks use this? We're announcing this in public beta preview, I believe today, so like how can folks actually use Hivemind for their own?

we were all heartbroken when they, they

ripped Fable out of our hands.

Yeah.

It

was a, it was a nice couple of days, and then it's-

a couple days.

So we can, you know, you can see that data

actually in, in Hive Mine. You can see, that little blip of, of, of hope until our dreams were, were dashed. you know, it, it's been surprising to me to see how dominant Claude Code is. certainly within kind of CoreWeave Weights & Biases it is by far the, most preferred model that, that folks are using. the actual Claude Code harness is the preferred harness, although we do have a decent amount of cursor usage, and a small group of people that are really enthusiastic about Pi- Mm-hmm certainly we, we had seen a trend of Codex kind of increasing. I mean, Codex is an incredibly powerful model. I just think what the data's showing us is kind of once as an engineer you, you kind of pick your, your harness or you get into something like Claude Code, it's, it's pretty sticky. It's, it's only if Claude Code isn't working would you like switch to, to one of the other harnesses is what the data seems to be showing at the moment.

Yep.

GPT-5 six, let's go.

Yeah, let's, let's wait for GPT-5 six.

Hopefully, this launches at some point and we'll cover it. I think the, the last, like, missing piece, that we barely mentioned is the actual leaderboard, right? So, like, the token max leaderboard that people talk about that gets created in different companies that leads to not that great of outcomes because people try to game them and then maybe get fired. so we are also shipping that leaderboard as well as fi- part of Highmind because all engineers who have Highmind are now on the leaderboard as well. Folks, you can try Highmind at highmind.wandb.tools today, I believe. and then, if your organization wants it, we will definitely chat with you. This is all provisioned for organization deployment, et cetera. if your org wants to track token usage across the company, definitely, definitely, check it out. Chris, Yam Peleg, Adrian Sw my, thank you so much for joining. congrats on the launch today. I love using Highmind. It runs in the background. I don't need to think about this, and it's there when I need it, and I think it's, like, the mark of the best, the best type of software, especially now where I had to report on here how much Fable context and Fable tokens I used, and I had that data. And I was like, "Yes, I have that data," and it's incredible. my manager also has that data, so that's great. CP, thank you so much. Adrian, thank you so much. Folks, we'll keep checking in on Highmind, how it develops, in the future.

Awesome.

Thanks for having us, guys. Okay. This was fun.

Cheers.

Take care.

Thank you so much, CP.

Thank you so much, A- Adrian, for joining us. Nisten, Yam, LDJ, thoughts on such a tool? Will you use this given the fact that we'll basically store all your, you know, all your sessions with pod code, you know, to give you this, like, ability?

I think that's cool.

I'm surprised I haven't seen more tools like this up until this point, I think all of us have the same problem at the moment, and that's pretty hard to manage the sessions. they are everywhere, sometimes in a specific sub-directory. Go figure where they are. Codex and Claude Code and it's not the same. absolutely you can switch from one to the other. I think everyone wants something that just manage the entire thing, especially now with multi-agents. You don't even know what a, what a session is a real session and what a session is spawned by a different model anymore. it, it's-- I, I find myself having a hard time with this all the time. because it- High Mind knows

We have sub-agent tracking within High Mind because

I think Claude Code exposes that. So like I, I definitely saw, like sub-agent like runs within, within High Mind on my, you know, when I tried Fable and like spun up like 18 agents, sub-agents to do a bunch of stuff. you know. let's switch gears back to- Well,

Are, are the traces gonna be available with this?

Because o- one difficulty I've noticed with when working with researchers or other coworkers is that, yeah, one is on Cursor, one wants to use Codex, most want to use the Claude Code, and then I have to tell them to, either copy the .claude directory to just get all the logs in one place, so we can have something searchable at as a company. But then for, for Codex, I didn't know, and it was just a mess. So what-- will this be able to integrate with the traces as well?

I believe that, you know, obviously we store everything,

but I don't know if we expose the whole raw, trace thing, because trace is mostly for, like Weave. But I believe that some integration between this and Weave is possible in the future, to, to expose everything if, if possible. Wolfram, do you know?

I think it's possible.

I mean, we have Weave, so basically traces are the thing, and, I know everything can be searched and copied, so it needs the traces or at least summaries of the traces. So, I, I haven't used the software, but I'm pretty sure that it's possible, and if it's not, it's a very valid ask, so should be in there. I will ask the team.

Yeah.

Speaking of having used the software, the reason why I only have Claude Code and Codex there is because I use Hermes on my personal machine for, like, a lot of coding as well, and Hermes is not captured yet in Hivemine, and, like, my first request to the team is like to, to have Hermes. Hermes is a coding agent as we saw in Movement as well. Like, this is my first request to the team to, add, Cursor, sorry, Hermes in there. folks, we have Dex Forsythe coming up in, in a few minutes, I believe. We're-- Sh- shouldn't be already here. let's talk about, tools in agentic engineering. let's talk about Claude Design. I don't know if you guys ha- been able to see this, but Claude Design got an upgrade and it actually looks really, really cool right now. Let me open Claude Design and show you.

quickly getting back to Nisten.

So we are not storing the full traces like Weave does, but it's more like transcripts than program traces.

Okay, 'cause that would be two things that would be really useful.

If you can integrate it with traces because they, you can build like your, your knowledge, central knowledge as a company from everybody else. So-- 'Cause I'm able to get pretty good traces just from using the, the wandb inference already and, so yeah, And the second one would be some kind of kit that you can use to configure your own harness to integrate with this. Or like if you're building an agentic UI or another agentic tool, and you can integrate this in, that would also be very useful because then you can get all the data, different models of harnesses, et cetera.

Okay.

Yeah, would be very, very useful because right now it's a little bit of

a mess for anyone that wants to do that

just before we have, Dex Thorley come up, to talk

about HumanLayer's evolution, let's talk about OpenRouter. Wolfram, you wanna talk about Fusion API for OpenRouter before we add, Dex Thorley here on stage?

If you could share your screen, that'd be great as well.

share the

screen?

Yes. Please share your screen- yes … for OpenRouter. So we're moving towards tools and agentic engineering corner on Thursd AI Here. Before we chat with Dex from HumanLayer, OpenRouter launched Fusion API. This is a, a panel of budget models, beats GPT 5.5 and Opus 4.8, and lands within 1% of Claude Fable at half the price, which is quite a ridiculous feat, if I say so myself. So Wolfram, take it away

So basically what they did, they have a hundred

hard research tasks and benchmark different models, and the fusion works that you can have multiple models. Like here they were using Fable 5 and GPT 5.5 as a fusion. So it's, yeah, it's sending the same task and giving tools to the different models that are part of the fusion. So in this case, Fable 5 and GPT 5.5 got the same question or task and, got tools to basically do their thing and return the response. And then another model that was on top of it, was taking the responses and integrating them into a better response that takes all these things into account. So it's similar to what Anthropic and Pathi did with the LLM consortium, where you can ask multiple AIs, multiple LLMs and integrate the answers to get a better result than just asking an individual model. And in this case, you don't have to use special software, you don't have to set it up yourself. You just use the API and this happens in the background. So it's like there's one model, a fusion model that you are talking to and this all happens in the background. And you get great intelligence that way and even cheaper than if you were just using one model,

I, I love how, you know, Fable came out and very quickly,

Fable or nearly Fable, near Fable like performance became a thing. but shout out to OpenRoute for this Fable, fusion API. All right, folks, I wanna add to the stage friend of the pod, Dexter Fortney. Dex, I don't know if you know this, by the way, second time on the pod here, but I believe like, y- you're- Third third time on the pod. Dex, founder of HumanLayer, and you had a big launch this week and the reason why I wanted to bring you on is because we just had, Chris Van Pelt from Weights & Biases, the co-founder. we launched Hive Mind. I don't know if you were able to catch that segment where we track all of the kinda engineering, sessions that people run via Codex and Cloud Code, et cetera. And the role- Okay … of software engineer has been changing for a while. Definitely something happened, like Swyx said, December 2025 and on, and everything's changing. you, as far as I can place you in kind of my timeline of folks, are more on the software engineering needs to remain software engineering and AI is a tool type area, where there's other folks in the YOLO camp, let's, let's throw away everything that we did so far, let's just trust agents, completely. And you've been helping companies. Let's talk about the role of software engineer in the world where Fable comes out and like does all these things and, and let's talk about HumanLayer. I would love to hear from you what you guys launched and where we are and, yeah, let's, let's have a little chat

Okay.

what's up, dude? Thanks for the incredible intro. I, you know, founded a company called HumanLayer. We launched our agentic IDE and, skill system and collaboration platform, and there's like a, a, a lean towards like building blocks for the software factory. I think the thing that we think about this is like, yeah, software engineering is still important, and actually even the people who are doing a lot of specification and caring about the system architecture might not be doing enough. the software factory has a missing piece. The lights-off loop where you just write specs and tickets and tell people what you want your users to do, what, what you want the behavior to be in your app is not enough. If you just put that in and you let the model verify everything, and you have AI code review, and then you put that into prod, and then you take all the support tickets and all of the metrics and all of the crashes, and you feed those back into agents, and you just do this like, like loop where all humans do is read and write tickets and everything else is self-driving, I think what we've seen in companies of many different sizes is this is, the fastest way to trash your code base. And now there is a time and a place for vibe coding. If you are a zero to one founder, if you're doing a side project that 12 people will use this, Adios many talked about this in his post last week.

Yeah.

Great But if you have high stakes and you work in an industry perhaps

where you're gonna get fined millions of dollars if the code is wrong, then, dear God, please read it because it's…

So on the topic of reading code, so we're going

to, AI Engineer Worlds Fair. So Wolfram's gonna be there, I'm gonna be there. I'm assuming we'll see each

Oh yeah, dude, I'll see you there.

I'll see you at the speaker dinner. Speaker dinner … if you're here next week, there's, we're gonna do a little pre-party for speakers, but,

and so my talk is going to be about, the ZL continuum,

which you he- you helped me doubt. Like, like when I w- came back from AI engineer in London, the whole conversation- Yeah … there was, as far as I remember, from one side, Ryan Lompolo from OpenAI, the guy who's like, you know, let's y- YOLO Lompolo, we call him. Like, let's, let's YOLO into code. one billion tokens a day is, you know, is just the beginning. We're all gonna… And there's nobody gonna be able to read that code, 'cause like how much… You know, there's so, so much code you can read a day, and one billion tokens is a lot of code. And on the other side we had Mario Zestner from PHI also. I think those are now the sixth and the seventh most watched talks from AI engineer globally. that talks about read every line of critical code. Obviously Mario, I think, you know, mentioned you a couple times.

did.

I did … rollout.

You donated to the charity.

Well, you showed up- I just

wanted him to say No Dumb Zone.

I actually didn't know he was gonna put my name on there, but we'll take

and the whole conversation there was about this kind of like thing

that I later identified as ZL Continuum where, maybe it's not per person, maybe it's per task, but, there is a range of folks who are like YOLOing completely versus folks who are, you know, let's read the code, like we don't want to get bugs in our system. obviously we know what side you are on, but tell me how, literally tell me how it's possible when folks are like YOLOing one billion tokens a day. Like, who's going to read this code? Isn't, isn't a human review on the PR now is the biggest bottleneck that we have?

I mean Mario's put it pretty well as like, "I think we need

to slow down." I won't use his exact, colorful language, but everything is breaking, and I think a lot of teams are trying to go 10X faster, 100X faster. If you really think about it, if you could go two to 3X faster and maintain your current level of quality- Mm-hmm … that would be game changing. That would transform every single company on the planet, every single team on the planet. And people are so busy trying to get to this 10X, 100X level that they haven't learned to like drive in third gear first before they're like on the highway crashing into stuff.

Look, I don't wanna call out, like, specific, things,

but I will say that, OpenClaw, I think, is a very good example of full complete autonomy, like YOLO all the code, barely read it, from folks. Because OpenClaw, I, I, I loved OpenClaw. I loved, like, the agentic era it put me in, and then I was not able to use this at all. I go to public with this, right? I talked with Vincent, about this as well.

Yeah, in London.

Yeah, and I gotta wonder-

I had full hair in January when I started with

OpenClaw, and look at me now- … after all this rushing after updates.

a- a- and I gotta wonder if that's, like, the symptom of,

like, what you guys are talking about, Dex, as well, because, companies-

I- I love Peter.

If anyone's gonna figure out how to pull this off, it's Peter- Yeah … for the record. Like, I, I, I do believe that, like, he is far and far and away much more advanced than almost everybody in AI coding right now.

Yeah.

if someone could pull it off, it's him, and there is this thesis

that like, "Hey, we're gonna do stuff. The problems of today's GPT models will be solved by GPT-7." Maybe that's true. Who knows when that model's coming out, and in the meantime, like, we have to figure out how to keep systems reliable and standing.

So how do we keep systems reliable and standing, Bex?

Give me g- give me, like, a very quick TLDR on what folks are to expect at the HumanLayer's, like, new idea.

Yeah.

I mean, if you, if you take one word away, it's, it's leverage. You can write a PRD that is very s- complex and give it to Codex Goal and spit out 10,000 lines of code. I am by no means saying that you should read all that code. I don't think that's the… a good way to build software, and before AI, many engineers would, would post like, basically like, "If you send me a PR that's more than 300 lines, I'm telling you to go back and break it up, and like, do the engineering work of breaking this down into smaller problems that can be merged individually," because it's just not safe. And so my basic advice is, like, find ways to get leverage. Find ways to correct the model from the 50,000 foot view and resteer. Find ways to do a lower level spec at a 25,000 foot view. Maybe even go into the program design, not just the system architecture, but the program design. How are we designing the interfaces? Where are the test seams? Matt Pocock talks about all this a lot. how is the dependency injection working? 'Cause we've found over and over again ourselves and with our customers, you can write a really good architecture spec and you can be like, "Look, I've, I've put in my time. I've done my engineering. I've designed the system. I made sure the endpoints are right. This is scalable. The core algorithms have been identified correctly and the model knows how to implement them." And then you go look at the code and it's still trash because the model's like… So, so, so we're constantly zooming in, and the, the goal is not to have a perfect spec because a perfect spec is indistinguishable from code. I forget which Haskell person posted this or whatever it is.

is not the territory, for sure.

And so it's like, okay, at a certain point you're not gonna get

the spec perfect, and you shouldn't try. You should get it 80% right. You should maximize the chance that when you zoom down another level, the, the chunk you're digesting at that level and steering and moving is human consumable, and that it's close enough that you can resteer without having to zoom all the way back out and change the whole approach.

I'm left wondering kind of like how long are we in this era

for, given that Fable came out very quickly as a blimp, and Anthropic literally said, "Hey, with Fable-level models, with Mythos-level models, we're switching from checking the outputs of the work to checking whether or not the right work is being done."

Interesting

… Alex Volkov: I would love your thoughts on this as well, because from one side, Anthropic does suffer from maybe too much AI in their code base, 'cause we've s- we, you know, we saw their uptime light up as a Christmas tree.

said it, not me.

Yeah, no, it's, it's, it's very clear.

But, but also from the other side, you know, they're, they're… If you look at companies to identify what the companies are there for, they're for to move capital in the world. There's- Mm. There's maybe five companies like Anthropic that were able to pull off what they pulled off in terms of how much capital they moved in the world, given that- that's what they deliver, right? So, like, Anthropic is answering the call of what a company should do, while also, like, you know, sending up, the, some call them finicker companies, some call them whatever, resetting code. we famously talked about multiple times where Claude became dumber and nobody knows why, and, you know, and they discovered- Someone made

a dashboard of it, right?

You could look throughout the day when the degradations happened.

Yes.

think they did it for Sonnet models.

and I, I, I, I gotta wonder how the curve kind of

looks, so would love to also ask you how the curve kind of looks. I think you opened your, one of your essays about, like, you don't, you can't predict, like, the, the GPT-7, like, level or Fable level. But, like, Fable, I'm assuming you played with.

I played with Fable.

Yeah.

Remember what he opened with?

I don't remember what he opened with, so remind me.

I remember his post. I think I have it here.

it's a very good post and I agree with almost everything in there.

But the first point he made that I take a lot of issue with… And I love Rahul. Like- Yeah … and like I said, it was great. It's a great essay. Short, easy. but, like, the As a mental model, Fable class model is, is the first model that can take English and compile it into code, and the code will be correct. I tried it on a couple refactoring tasks, I tried it on a couple greenfield projects. I tried it on a couple, like, "Hey, build this big feature." It was the closest I've ever been to… I remember in May of last year when I first figured out how to, like, spin up four tmux sessions and have a merger agent on main and three workers on every branch, and the merger is, like, checking the branch progress and merging in from the w- this, like, s- this thing I saw XJ, XJDR doing on, on, on Twitter, like, in maybe, like, March. I figured it out in May and I was like, "Holy shit, I…" This was, in retrospect, it was the closest I've ever been to AI psychosis where I was just like, "I need to be up, I need to be constantly prompting these things. We need to just keep them going." And then eventually- Right … you cool off and you realize, okay, maybe this doesn't scale, the models, we, we have to be a little more in the loop. So just if you come out in the middle of the night to go to the bathroom and the laptop's open, just please leave it open. Don't touch it." And I hadn't been there in, like, a year, so, like, I was very excited. I tried all kinds of stuff and I just found it had a lot of the same problems that GPT and Opus have, which is, like, not very good at code quality, not a good sense for refact. I gave it this big, like, 8K, 8K React PR, which is for… When I just naively prompt it, "Here's what I want it to be and here's what the refactoring I want," the first pass was terrible. It was bad. Mm. It introduced, like, React contact, introduced all these things that we didn't need, that we don't use, that we had better patterns for, and so I'm like, "Cool, this is much better, but it's not the step change that means that you can let go of the reins." We're not there yet. That's my take.

Dexter asked you this in a comment, and then we'll ask you this on

air, and I would love, an honest answer. Why does it matter? Why does code quality matter? Why does unmaintainability of code matters if you take into account that Fable comes out and in six months Fable 5.1 or 5.2 comes out, whatever, and in a year we're gonna be as far from where we are right now as we are right now from a year ago, which a year ago nothing agentic basically were able to happen at all, and models were not having, like, 1 million context windows barely, maybe Gemini, and Gemini is not a useful thing. given that the, the whole thing advances as fast and, you know, why does code quality still matter, and does it?

Yeah, I mean, the code quality matters now just for the same

reason it mattered 10 years ago and 20 years ago, and we had the great software crisis of the 1970s, and a bunch of really smart people had to get together and invent the, the, the field of software engineering and program design. And the problem is, is like the more you put code into your code base, whether it's human or AI written, without thinking about the structure, the faster it becomes unmaintainable. And by ma- unmaintainable, I mean it's slower to make new changes. It's much harder to change code over here without breaking something over here, and you start spending more and more of your time trudging through complex changes. You spend more and more of your time like, "Oh, we fixed this, but we broke that," and your velocity just tanks. The irony is we train models on projects that have not made this mistake. All of the SWE-bench like metric are Django, Redis, I think they do Spring for Java on SWE-bench multilingual. Like there's all these projects that have really good architecture, but somehow the way we're RL-ing them to solve problems and like they're basically somehow still reward hacking around, just make the test pass. And we never make these model, and we're, I'm gonna talk about this at Ai.engineer as well, is like we, we are, we have not yet, achieved a like penalty function for writing code that makes the code harder to maintain. Oh, that's true. And it's really hard to write a judge or a verifier for that.

Yeah.

So I think humans

are still needed for this.

Humans are still needed, but given that maintenance will

also be handed off to models, is the size of maintenance also important? Like I, I, I'm thinking about this. We have, let's say, unlimited tokens on lower quality models. Fable's obviously very expensive, so you're not gonna run Fable continuously on every like bug maybe unless you're in Anthropic. But given that like most of the maintenance now is a prompt away, is, is velocity going upwards not better than just- getting stuck in, in the review of humans.

you still have to know what prompt to give.

And a lot of times, like whatever… I know I said spectrum development is a bad term 'cause it means 100 things to 100 different people, but like when you're building specs, you're basically building a prompt for the next session to go implement it, right? And you have to know what to prompt for. a really good engineer who has really good sense of code quality and, and all this stuff can get Fable to write really high quality code. I did, but I had to tell it exactly what I wanted, and like, it's a lot easier to do that before you start writing the code than afterwards. And like my take is basically you spend an hour reviewing specs and architecture and even going down to the level of program design, which is I think a new term that's gonna become popular. you can turn a three-hour code review that has a lot of changes required to a 20-minute code review because m- it only needs, you know, five… Instead of six hours of reviewing and fixing stuff, I do an hour up front of prep, and then I have a 20-minute code review. Like, basically like task-specific RL on, like improving code quality and making things better separated and, and, and easier to maintain.

I have one last question for you, Dex.

How is HumanLayer now helping with the problems that we just described? I think let's land there

All right.

I hate pitching on air, but we did launch yesterday, so I will humor you. like I said, I believe the missing piece of the software factory that will allow you to get to 2 to 3X, like overall throughput, is actually moving a little bit more slowly and building really good systems for humans to collaborate in real time on the design of software before and while it's being written. I think everyone's betting on that. We need more synchronous c- like, ways to collaborate on code, on, and keep everybody on the same page, and do that high leverage work together. And I think leveraging AI to do that is a space that is still untapped. Lots of labs, Cognition Factory, like Anthropic, lots of people are like moving in this direction. We are obsessed with this problem, is how do we get AI to solve hard problems in complex code bases, giving the maximum, like finding the exact high leverage points to put humans in the loop, and letting the models do what they're good at, which is reading hundreds of thousands of lines of code and understanding them. that problem is solved. We're very good at that. And taking a well-written spec and going and ripping it out and implementing it, like those two things are really, there's this piece in the middle of like deciding what we're gonna build and how we're gonna build it. And so how do we, how do we get… the way you get the most speed is you seek leverage. What are the highest leverage points to insert humans into workflows?

All righty.

Folks, definitely check out the new HumanLayer and Dex's, Dex's talk at Ai.engineer. my talk at Ai.engineer, Wolfram's talk. My, my talk is gonna mention you and kind of the stuff that you're doing, obviously, 'cause I'm talking about the ZL continuum. but, definitely, check out HumanLayer. We have to move on, Dex, because we have this incredibly surpri- Feel free to stick with us, by the way, for this last segment, 'cause everybody's back on stage, including a friend of the pod, Tanishq Abraham. Welcome, Tanishq, back to the show. You're also a friend of the pod, have been here for a few, a few times already- Mm-hmm back when we did Spaces. And the reason why I invited you, Tanishq, is because yesterday an AI company did something, interestingly novel. They decided to announce a hardware product, and we're not talking OpenAI and Journey Ive, we're talking about Midjourney and David Holz. And you were there in the audience, right, Tanishq? Is that, is that correct?

Yeah.

I was there. I was attending right in the front row, I guess you could have seen. but yeah, it was a, it was a fun event, yeah.

All right.

So let's talk about, Midjourney's crazy announcement. I expected them to do something with, I don't know, image generation because that's what Midjourney does. But no, David Holz from Midjourney announced a full body ultrasound machine, and that's kind of crazy. Tanishq, do you want to talk about this a little bit? Were you surprised at all? Like, what's, what's with AI companies launching hardware? Like, this going to be a new thing? Midjourney, by the way, is fully self-funded. There's no VC money, and they're, like, very, very well, you know, performing very well. So this is, like, a very interesting thing. So let's talk about what the actual specs are. This is crazy.

I think this is like a very impressive effort from Majority,

and because Majority is a very unique company that they're able to do this. And yeah, I mean, like David Holz has been interested in medical imaging for quite some time. Like he tweeted that he wants to start a medical imaging lab like two years ago, and I don't think many people noticed that. But yeah, he-- this is just like his interest because, because, you know, Majority is fully self-funded and they have, they have the revenue to support it, they can more variety of different projects. And so this is one of the projects that they're exploring. But yeah, they have like a few other hardware projects that they haven't announced yet. but yeah, I mean, this one is pretty impressive. It, it's basically, yeah, full body, kind of scanning, using, ultrasound, rather than, you know, a ra- like rather than like MRI or CT, which, you know, CT has its limitations. You know, it's using X-ray radiation. MRI has its limitations as well. You know, it uses these magnets, these expensive huge magnets. it's a very large system and, you have to, for to do a whole body scan, you have to be there in, in the scanner for quite some time. And this is like something where you can apparently get scanned in just 60 seconds. So you can get like a, a full body scan in like 60 seconds. So, yeah, the way it, it works, I mean, it works similarly to like something like CT where you have a variety of different, sort of transmitters and receivers across the entire like a, like an entire ring. although y- you are combining the transmitter and receiver into a single, sensor, like it's called trans-transducer, an ultrasonic transducer. because, you know, sound isn't able to propagate well in air, what they do is they actually have a, a tank of water that you are actually lowered into. so that way, that, that, that's, you know, the, the s- sound can more easily propagate through water, and then it can propagate into, into your body. you basically are, being lowered into this tank of water, while you're being scanned. And then once you're done s- with the scan, then you lo- you are brought back up.

We've seen this on, on the kind of the, the graphics that I

think David also cr-created on himself like a crazy, like a crazy person. we're seeing like a layer that they scan, like a piece every, every like second or so. The sound travels, I just wanna add some numbers. The sound travels through your body at 1,481 meters per second, and under 60 seconds you can get a full 3D atom- an- anatomical map of a hu- 25 plus organs. The raw data per scan is 806 terabytes and, and it's scanning at like 16 gigs per second. So every second sends like 16 gigs of data that they need to then condense down to, to something. David Holz said there, and I remem-- I saw the live as well, so Tanishq, I wasn't there in person, but, like, I was able to catch some of it. he's mentioning that, like, compared to MRIs, which is also insanely expensive, there's a whole superconductor thing that needs to run with an MRI machine. Like, MRI is, like, very expensive. he's like, this is essentially the, the, the price of a scan is free compared to an MRI, right? Like, the, the machine and the, the, the business model around it is something. and the business model they announced is also, like, very interesting with the spa in San Francisco, 26,000 square feet spa. That part of it is gonna be the scans, part of it is gonna be the spa. but this is absolutely it. This is, like, one of those… The reason why we're talking about this and this not being AI here on the show is because this is the future we want. We want-- A- and this is powered by AI, and also this data means also AI, right? At some point, based on all this data, we don't know what advances in medical examination we can get when the person can get scanned, the full body scan every minute of, of their day, basically, right? And, and this allows this. Nisten, after this, would love to hear as well. You

Yeah, I have a question for Tanishq or everybody

else that I'm wondering. Do you guys think they're d- do, do you think they're doing diffusion generation for each one? I remember the past experiments trying to guess the image based on the MRI scan. Do you think they're doing some type of diffusion here? And then how would you train that? Would you, like, grab a CT scan, grab the slices from that, and then get the data from this and try, try, try to match it?

No, there, there isn't a- any AI actually in the actual,

Oh … generation of the images. It's a pretty, like, basic signal processing algorithms that they're using right now. and that's something that they're ex-- Like, yeah, they have, like, really basic algorithms that they are, like, trying to come up with more sophisticated approaches of, of reconstructing it, looking at also, like, the, you know, they have to look at the physics, and they are developing physics simulators and all these sorts of things to actually, basically, get more accurate reconstructions. But that's, like, just a prototype of what they're doing, and so there's still, like, so much more that they can explore here. So yeah, they haven't even touched, like, using AI to help at all for reconstruction. Like, there are probably ways to use AI too, and, like, I was listening to a conversation with, like, one of the i- you know, imaging scienti-- Like, yeah, 'cause, like, all the scientists were there at the, at the event, and there was a conversation between one of the imaging scientists and, John Barron, who's one of the inventors of, like, neural fields and, like, Neursa. And they were, like, trying to figure out, like, are there ways to, you know, utilize Neursa for similar sorts of things for reconstructing, these sorts of images. So I'm sure there's probably gonna be ways down the line where you can utilize more sophisticated AI-based approaches for reconstruction. But right now, it's really, really basic.

And even with this, like, fairly basic thing, I think

the, the benefits are incredible. Obviously not quite to MRI level, but a non-radiation, no magnets, no speed MRI is quite crazy on its own. so I'm like-- I feel excited about it. There's-- You wanna talk about the, the, the kinda how they, they… Like, the scale they want is that with, I think, fifty thousand scanner fleet, targeting a billion scans a month, they basically wanna scan and allow, like, everybody in the world to get scanned. what are the benefits of something like this? What are the type of stuff that can get caught by something like this? And generally, people-- regular people cannot afford, like, a full body MRI, I think, once a year even. what are the benefits of, like, a full body scan? What, what type of stuff can be caught that cannot be caught, like, o-o-otherwise?

the main, of course, sort of thing here is the fact that you

can do this full body scan very quickly, and you can do it, like, without… Yeah, without… You can just do it as many times as you want. So being able to track the, the, the, the progress of y- of your body over, over time is actually something that is a lot more achievable, with this technology than, you know, with, with MRI. And yeah, I mean, y- I think l- you're looking at, like, just differences in, in, in the tissue, and in tissue density, and things like this. If you think about, what ultrasound is already being used for and, just kind of extrapolate based on that. So there are things, like, looking at cancer, looking at inflammation, looking at all these sorts of things that people are already exploring using ultrasound for, but, in a very limited case. But now you can expand this to everyone and being able to kind of get scanned. Anyone can get scanned, and so it's like, you know, this could enable early diagnosis of these sorts of things. if, you know, there is some sort of, like, lump or something that, you know, is y- y- you can notice, in the, in the ultrasound scanner that, you know, you may have not noticed if you waited until you… something felt wrong and to go to a doctor. So I think that's the, that's the advantage. But yeah, I mean, it, you know, it's not w- obviously it's not without limitations. ultrasound, some… There, there are certain things ultrasound just can't see. For example, like, it would be very hard with the current iteration to, you know, image the brain, for example, or something like that, right? So there are certain things that, can… just cannot, cannot be seen. like, looking at the lung is also difficult because the lung is full of air, and ultra- you know, sound doesn't propagate through air very well. So it's not that it can do everything, but, and it's not gonna be like a full replacement of, of MRI, but the fact that you're able to do these scans, like, as many times as you want is… And I think there's, like, a lot of debate, you know, within the medical community if, if this is a good thing, and I think it is a, a, a good thing. But, we'll see also, I think, especially when you start powering this with, with, like, AI algorithms to actually, you know, monitor your health. I think this can be, can be really beneficial.

I, I, I think it's a… It, it, it's a complete game changer

given that in a lot of places the wait list to even, like, get an MRI can be, like, a year or something. So if they can get the scan down to $20 or, or $50, and maybe you have some bowel problems or you have cancer and stuff, you can take that data and you can start to track it very precisely over time if you can just go and do one e- every week, or so. And I'm wondering if they can just get around the medical device requirements and just put it as a wellness thing, because you can still get the data, which is also wonderful because it just gets it to people a lot faster

So I think the, the cool thing there is that they're engaging

with the FDA, but also they said that they're going to open a spa, a 24,000 square feet spa, 10 minutes away from Union Center in downtown. Gonna place like 10 of these machines. You'll be able to go there, they actually will be able to hang out there, talk about code, and then one of them is like, "Oh, hold on one second. I need to pop out and like scan my full body real quick with ultrasound," and come back, and then you have 890 terabytes of data or whatever they send to you as well. Like Nisten said, like the lines are a year ahead, and it bangs there, like it sounds of headshots or whatever. Like it's very loud. You have to s- lie very still, so you cannot put kids in the MRI, for example, 'cause, you know, they, they move all the time and you have to lie there for a long time. And it's scary and it's not productive and it's like radiation for C- CT scans. This is a 10-minute, as easy as going to the gym type scan, I think, Tanishq. does that summarize kinda like the ambition there? they wanted- Yeah.

Yeah

… Alex Volkov: like incredible things.

The idea is to make this as pleasant of an experience as

possible, as opposed to how it feels to go to a doctor's office and how it feels to try to get an MRI scan. So this is, supposed to be the complete opposite, so yeah.

I'm, I'm fully down for this.

the other thing they mentioned is that ultrasound is incredible and we haven't even gotten close to, to realizing how incredible this is, including actual l- l- like, like therapeutic usage of very highly directed ultrasound waves that can like do incredible things to the body, which they're not going to, but they can because the scanner is overspecced. LDJ, last comment then we'll move past to towards like landing this plane.

Yeah, I think it is really exciting that there is the future applications

of-- Actually, already today, over the past five years, there has already been trials with some success of, people with severe, depressions and addiction being successfully treated with actually focu-focus ultrasound in the brain in specific regions. It's not, it's not, a hundred percent foolproof. It's not without any side effects at all, but they-- it is increasingly getting more accurate, getting assisted by better algorithms. And I think it is gonna be really exciting as they do start to apply AI to filtering out a lot of that noise. And they said end of twenty twenty-seven for the spa in San Francisco and twenty twenty-eight for a full revamp of custom imaging, custom sensors and everything.

And folks, this is what we bring you here, about AI, but not

only, and I think it's very interesting that an AI company decides this. So shout out to, David Holz and the whole Midjourney team. Tanev, thank you for joining us and giving us a little bit of a glimpse. I know you have to jump very, very soon. folks, I think it's time for us to let this play. The, the, the few things we haven't talked about yet is, the two video image, models that they've launched. Grok launched, Grok Image one point five, and so I'll show you this very quickly here. And the other one is, Google's Omni, which we talked shortly after Google IO, has also launched an API. And, just after I show you this Grok Image thing, this is my prompt for an astronaut riding a horse on the moon. I don't know if you can hear the music, but it generates music. It's na-natively multimodal, and it looks pretty cool. Google Omni jumped to the top of the charts. so this is Artificial Analysis, not Arena. let me, let me show you here. Video Arena, image to video leaderboard. so on Artificial Analysis, they don't have Omni yet, looks like. But I think on, on, on Arena AI, arena.ai, Google Omni is now the, the top by, quite a while. So OmniFlash Is now the number one ranked, image to video, text to video on the, Arena leaderboard with quite a jump here, with a almost 100-point jump. so OmniFlash now beats every- everything else, including Cedance, which was-- used to be the top one. Dexter Horthy, founder of Human Layers, thank you so much for joining. Danish Gabra, founder of Sofont AI, also a friend of the pod, was here to talk about the newly launched Midjourney. We also had two other co-founders of Weights & Biases. oh, sorry, co-founder of Weights & Biases and, and, early engineer there, Adrian Sandberg, talk about Weights & Biases High Mind and, very exciting, very exciting week, folks. We're still waiting for Fable to come back, but otherwise, very excited with, with a bunch of open source. thank you so much for joining. If you missed any part of the show, ThursdAI is available for, for you as, as a podcast, as a news- newsletter. everywhere you get your podcasts, just search for ThursdAI. if you are following us on, one source, why not check out the other one? If you're reading the newsletter, why not check out our shorts on, on YouTube? And if you are on YouTube, why not check out the newsletter as well? with that, Wolfram, thank you so much. Nisten Tahiri, LDJ, and, Yam Peleg was here for a bit as well. We'll check, next week if we get Fable, and we'll definitely see you here next week, folks. Cheers, everyone. Just a reminder, the next place where you can meet us in real space is Ai.engineer World's Fair. Wolfram's gonna be there. I'm gonna be there. Weights & Biases CoreWeave is sponsoring the whole event. So if you are coming to Ai.engineer, please chat with us. would love to, to see you there. Cheers, everyone. Thank you so much for joining ThursdAI for this week, and we'll, we'll see you here next week. Bye-bye, everyone.