OpenAI's Navier-Stokes Claim, DeepSeek V4.1 Flash & Meta Muse ThursdAI Sep 10, 2026

0:34Alex VolkovWelcome everyone. Welcome to ThursdAI September 10th. Welcome Wolfram, welcome Nisten. How are you guys? Welcome everybody who's tuning in to our live show, ThursdAI, and we're excited to tell you all about this last week. there's been some twists and turns this last week for sure. So I'm super excited to have you guys here. Wolfram, Nisten, how you guys doing?

1:00Wolfram RavenwolfExcited. But we will cover this-

1:02Alex VolkovYeah, we have to talk about this- this strange week … this, this Ant- Anthropic guy, right? That, that quit. We have, we have to cover this.

1:08Wolfram Ravenwolfweeks. Just

1:09Alex Volkovand I think now clocking in 130 million for that, "I leave 'cause we're all gonna die" tweet. we definitely have to cover this. there's also some open source news with DeepSeek, V1, V4.1 Flash. Hard to keep up with the DeepSeek, versioning. However, those are dope, and I am super excited to tell you guys that, Chris Alexiou from NVIDIA, Nisten's, Nisten's, neighbor in Canada is going to, to join us to talk about DeepSeek. Nisten, I bet you didn't know this. How are you doing, sir? What's new in your world? What's new in the world of AI that got you excited?

1:44Nisten TahirajI, I'm, I'm still alive. The world has not ran out of, problems to solve. Something exciting coming from my end probably next week too. And, yeah, yeah. Things, Yeah. Busy- What- … busy benchmarking.

2:01Alex VolkovYeah. I've spent my whole weekend, as I imagine many of you as well, Astra maxing and Fable maxing together. I think I've burned through two resets. Wolfram, I know you're banking yours. But- Yeah I, c- can I, can I just go on a personal mea culpa for just one second towards the audience? I claimed on ThursdAI newsletter… We don't do five hours. This is the longest show we ever did, and the reason for that is that we were waiting for Astra to drop, and GPT-6 is a big moment. We started when GPT-4 came out, so we've been tracking now two proper generations of OpenAI's model, so we waited. And then there was hiccups, and we waited. Anyway, this resulted in a very long c- content piece, and this

2:43Alex Volkovresulted in us posting two episodes. So if you have our, our podcast, from last week, there's two episodes, one with the regular ThursdAI News TLDR, et cetera. We talk about exciting things like World Labs and, and different things, and then- The other one is fully dedicated to GPT-6 Astra. And then we had Peter Goste and, Ryan Carson talk about their experiences. They had, early access. And then we got access. And then, on the page that I had Astra built for itself, maybe I should show it, I wrote AGI is here. And then I saw Jensen Huang from Nvidia say AGI is here. And then the release of Astra definitely felt like, hey, AGI is here.

3:24Alex Volkovso my p- mea culpa for, for last week is that G- AGI is not here . At least, at least not the way that I imagined it. I had like, I, I've been like t- talking maxing Astra and, I don't know, folks, it's, it's amazing but there's buts for sure

3:46Nisten TahirajIt, it built the best Mars driver simulator,

3:51Alex VolkovYeah

3:52Nisten TahirajI didn't find it as good on DevOps a- and stuff. Like, I, I'm just, the way I'm… I think Claude has trained me at this point to be opinionated how , how, how Claude does things, so I, I did not like how it was doing certain things. so I'm still at Fable for that, but for 3D stuff, this is, this is king right now.

4:12Alex VolkovYep, everybody's posting the 3D stuff. By the way, what I'm showing on stage here while Nisten talks about the 3D simulator is the page for Astra, the Astra build for itself. That page sits at thursdaii.new/app/Astra, and you can see this, like, 3D animation here, with Fable and Sol and Astra being, like, a star of clouds. I did this with Astra Ultra High on fast mode, and I ran out of tokens very quickly. I, I am so used to codecs not ever reaching my quotas that I just ran on ultra fast, and I've hit two resets, and then OpenAI got me a reset, so I'm out.

4:49Wolfram RavenwolfThere were also some issues with how it was counting the tokens, so they have

4:53Alex Volkovbeen making improvement.

4:54Nisten Tahirajhe was fine for me.

4:55Alex VolkovGo ahead, Wolfram. Sorry.

4:56Wolfram RavenwolfOpenAI also said that there have been issues with how the usage was counted, and I think they gave a free reset for that because of the issues, and, they are continuously improving it. So depending on how you used it, that may be part of the issue.

5:09Alex VolkovAnyway, the page is one of the cleanest pages that I've ever seen built on ThursdAI. It pulled up all the evals for itself. there's just an incredible amount here. Also, it cut, I think, five, five clips from the episode itself where we talk about OpenAI. This is by far one of the best, one of the best, like, pages that we've built. Let's say welcome to LDJ. LDJ, welcome, man. How are you? What's, what's new in your world? What can you t- tell us about?

5:38LDJDeepSeek V- V 4.1 Flash, had just dropped within the past 24 hours. That's exciting. And of course, Astra, a lot of things to talk about with Fable 5.1 versus Astra and really strength and weaknesses there.

5:55Alex VolkovOkay, so folks, I think that, folks in the comments can also chime in with their experience with Astra and Fable. But, I think it's very clear that, it's been a week that folks have been playing with Astra and a little bit over a week that folks have been playing with Fable 5.1. and, w-we definitely should go into, like, a full corner of, like, our experiences, what we built and, and how this feels. However, I, I do think that, we should also talk about open source and, I just to tell you that in f- in seven-ish minutes, we're gonna have Chris Alexiou from NVIDIA joining us, to talk about, the new DeepSeek that dropped. But before that, I think let's go into the TLDR, and then I will tell

6:35Alex Volkovyou, you know, TLDR is the corner where we basically, tell you about everything that's going to happen. By the way, if you are… if you guys just go to ThursdAI.live, please, please do so. and, you will see that we have an AI-produced show with chyrons and everything and, h-hopefully, hopefully, the chyrons will show up, in the right way because I'm using a new AI now to run the show, so we'll see. It just happened this morning. if you go to ThursdAI.live, you also see our TLDR. So actually, yeah, let's go there and, and, and, and walk through the topics that we'll cover on the show. Obviously, we're waiting for breaking news, and obviously we'll have

7:13Alex VolkovChris Alexiou, join us as well. Let's see. Yes, John Nemotron is coming on again, a friend.

7:22Nisten TahirajWhen I've been multitasking and searching for Alex, I searched Joni. What's wrong?

7:28Alex VolkovAll righty, folks. Let's do this. Let me run through the TLDR on the show. all right, so obviously We're missing s-- All right. welcome to the TLDR for ThursdAI, live on September 10th. My name is Alex Volkov, AI Evangelist with CoreWeave and Weights & Biases. our Fully Connected show is coming up, and I have an exciting announcement that's not AI related in any way, unless he's secretly an AI wizard, but he's a very, very known person in the world, to tell you about later on the show. Today with me, Nisten Tahiri, LDJ and, Wolfram Ravenwolf, and we'll have Chris Alexiou join us just momentarily to talk about the biggest open

8:10Alex Volkovsource release f- of this last week. Actually, yeah, let's start with open source. DeepSeek releases v.4.1 Flash, 4.1 Flash. Let's remove the V 'cause it's confusing. It's easier. A fe- half a bil- half a trillion MoE with only eight billion parameter or 16 billion parameter, looks like a very interesting thing to talk about. And, 4x KV cache compression. DeepSeek has been compressing KV cache like crazy lately and, this model will blow your mind, I promise, at least on benchmarks, at least on the stuff that it achieves. Also from Europe, there is a new lab called DesertAnt, debuts 18 on-device AI

8:48Alex Volkovmodels with open weights and native SDKs. Europe's, another claim to fame after Mistral. very interesting and Wolfram would love to hear from you more as a European representative. And also Inclusion released Link Three Flash VL. It's a MoE with vision language and 5.5 billion active parameters. That's in open source, and again, Chris Alexiou will join us momentarily to talk about this. However, folks, the biggest news in AI for this week, 100% came from OpenAI, where they claimed that, 10,000 agent swarms claim to have found a solution to N- to Navier–Stokes Millennium Prize Problem in mathematics.

9:30Alex VolkovThis is an insane sentence to say, but it does seem to be true. The LLMs, the AGI is here LLMs, GPT-6 Astra, and I think an unreleased model, right, in this case, uh, are solving Millennium Prize mathematic models, at least to some extent. There's also some excitement about how this was released, on Twitter, but this is just like an insane news. Again, I will say, OpenAI claims that a swarm of AIs solved something humans could not solve, a Millennium Prize problem for Navier–Stokes. Excited to talk about this as much as possible, to learn about this it's crazy.

10:10Alex Volkovalso, in addition to being one of the most let's go, AI is here solving mathematic problems, we also had one of the biggest doomer episodes that I've seen recently, and not only me. And Anthropic, Ooh, something switched. let's go here. Right. an Anthropic employee quit from Anthropic, and then other AI insiders together warned of extinction level events. Jacob Coxon resigned from Anthropic on September, and, his tweet about, "Hey, I resigned because I think we're all gonna die," now at 122 million impressions, which feels inorganic.

10:48Alex VolkovWe're, we're gonna have to talk about this. though we also have to, you know, at least discuss his claims. OpenAI said that they, reached their automated research intern milestone. If you guys remember, we told you about this when Sam Altman and Jakub Pachocki sat together and said, "Hey, by September 2026, we will probably have an intern level automatic research in-house," and they said they reached it, and they still target automated AI researcher by Mar- March of '28. So that's 18 months from now or so, maybe a little bit more. Apple debuts iPhone 18 Pro with A20 Pro chip with a bunch of AI stuff. And then Siri AI in beta is about to land on everybody's iPhone,

11:28Alex Volkovincluding, iPhone Duo, which has nothing to do with AI, I don't think. But, many people think this is gonna be a dope vibe coding machine. all right, let's see. Anything else from Big Labs, Wolfram, LDJ or Nisten?

11:46Wolfram RavenwolfDid you get a chance to look through the stuff I sent you?

11:50Alex VolkovI did, but-

11:50Wolfram Ravenwolfdocument?

11:50Alex Volkovanything-- Oh yeah, we have some comments from the, audience. Milosh, thank you. Live translation in AirPods 5 also was announced. that's coming. I think that was part of the beta, and that's dope. Yeah, people walk around with AirPods, and then it translates. also the Apple… Okay, we'll, we'll talk about the Apple event. in AI Art and Diffusion, OpenAI launches GPT Images 2.5. this is two models, I believe. Yes, Flare and Sunburst. We're gonna try them out live on the show. Maybe we'll add hats to ourselves. maybe we'll ask the audience to tell us what they think we should turn ourselves into. I think that this is it in AI Art and Diffusion.

12:28Alex Volkovin 3D, there's also OpenAI's, new Unity plugin, Next, we're gonna talk about tools and agentic engineering, folks. I think this is, for me, we didn't ask before, but for me, this is the biggest part of the sh- of this week. Meta launches Muse, their agentic 24/7 free AI assistant that has its own computer, has its own browser, and if that sounds familiar to you, yes, this sounds exactly like GrokBot. This sounds exactly like Hermes. This sounds exactly like OpenClaw. however, This is from a huge company. Meta has their own LLM, Meta Muse Spark 1.3, which we told you about last week, which is probably the winner of number

13:06Alex Volkovtwo spot of last week after OpenAI. and, that we will show you all about Muse. It's really dope. And by the way, Muse is running the show for today, so if you are on ThursdAI Live and the chat rooms that are getting brought up are, are done by Muse. So, let's see how fast this is actually. I'm gonna keep reading the TLDR, but I'm gonna pull up the ThursdAI Live page. I wanna bring up ThursdAI Live page like that, and, we will ask Muse to put up a chyron to introduce itself, and we'll see how fast this happens. And, uh, meanwhile, we'll say hi to friend of the pod, open source wizard, Team Green representative, Joe Nemotron, Chris Aleksiuk.

13:47Alex VolkovWelcome, dude. It's so good to have you back. How are you?

13:50Chris AlexiukI got the blue lights today, you know?

13:52Alex VolkovOh, blue lights- And, In, in, Yeah, yeah, yeah … representing.

13:53Chris AlexiukThat's right. Honor of the whale, man.

13:55Alex VolkovSo we'll have Chris talk about, a- and us talk about DeepSeek. All right, let's continue to… Do you guys see this? Muse is producing the show. It, it heard me. Muse is producing today's show. we got the chyron up, so, we have an AI producer. I'm gonna tell you all about Muse. all right, let's go forward to the, to the, the, the TLDR. Let's see what else. So okay, so we have this, in, in the TLDR. Muse, we're gonna talk about Muse from Meta. and then the… There's another agentic Grok bot, Hermes, Muse, OpenClaw-like thing that's going on, very viral, called Instinct. We didn't tell you about Instinct yet, but I have been using Instinct.

14:36Alex VolkovInstinct is very viral. Instinct works via iMessage, and you get a number, you just text and it works. recently they added email. I do wanna tell you about Instinct because I did comparison between those two agentic things. All right, in this week's buzz, anything that has, to do with Weights & Biases and CoreWeave, folks. I will remind you again, Fully Connected 2026 is coming to Moscone South in September 29th. That's just in, in, in 19 days. and I have a big announcement about somebody who's gonna be there and you don't wanna miss, and we have free tickets for you. we also have a CoreWeave hack and Agent Loops hackathon. And if you win some of the prizes there, you'll be able to come to Fully Connected

15:13Alex Volkovand present in front of big audience. That's also a big news. So definitely we'll tell you about that. and then I think let's close out the TLDR with these things. Voice and audio. I think number one is Google launches Lyria 2.5, full song music generation model, access to Gemini and API, so you can use API to generate songs. We probably should try it out. Suno launches Suno 6. We don't care. This is literally as it's written on the TLDR. Suno pulled a fast one and, like, restricted how many downloads people can use. And Suno, you know, makes a lot of money, but people don't like them anymore, so we, we will just tell you. A new Suno was released.

15:47Alex Volkovand also OpenAI brings GPT 5.6 Sol and GPT 6 Astra to ChatGPT Voice. They're using the live version, but now the live version can talk to Astra and Sol for you, so you can, like, build things, like they said in their video, with voice. And I think, unless I missed a bunch of stuff This is the TLDR. Let's see in comments. folks in comments, sh- let me see if, I

16:11Wolfram Ravenwolfput something in our private chat

16:13Alex VolkovLet's see.

16:14Wolfram RavenwolfOkay, so we have, a new music model as well, Yue 2 Open Music Model, successor to Yue or however it's pronounced, Y-U-E, which can do vocals and accompaniment, and it's on Hugging Face already. Well, there's a speech model by Tencent, AUK. It's an open source 1.5B speech model that does ta- text-to-speech, voice cloning, content emotion, accent edits, cleanup, and separation from one natural language prompt. MIT Weights. Yeah.

16:44Alex Volkovright. and if that's it, I think, let's go to open source. Ooh, I'm excited. Open source AI. Let's get it started All right, we're here. Open Source Corner. Let's get it started. folks, just before I, and Yam Peleg just stepping in exactly as we're about to talk about DeepSeek, that's great.

17:24Alex VolkovI don't know what's going on with Nisten, but, oh, there he is. Okay. Folks, exciting news from the world of open source because DeepSeek, the whale, has refer- resurfaced once again with a 4x KV cache compression with DeepSeek V4.1 Flash. It's a half a trillion parameter, specifically 552 billion parameter MOE that activates only 8 billion or 16 billion. What? Okay. to help us to talk about this, we have folks who like DeepSeek for a long time and have used it and tested it out, Chris Oleksiuk from Nvidia, a friend of the pod, AKA Joe Nemotron, the guy who brings us Nemotron news, but also is part of Nvidia.

18:06Chris AlexiukI think Hugging Face owns themselves, at the end of

18:08Alex Volkovthe day. Nvidia supports Hugging Face- That's right.

18:10Chris Alexiukright … Alex Volkov: with a- Big fans … $12.9 billion- That's right. That's right … injection. Big fans, yeah. And now you have a bunch of new colleagues, let's, let's say that. That's right. You have a bunch of new- Yeah, yeah, yeah new teammates, and, you guys are keeping the torch of open source alive, so thank you for that. Now, on that Hugging Face now, there's a new model. You see how the connection is made? there's a new model there on that new Hugging Face, and it's from DeepSeek. I saw your post. What- Yeah tell us a bit about what, what's exciting about this model. it's a new whale model. We know whale models are good. I think every time you talk about DeepSeek models, it's really important

18:43Chris Alexiukto talk about the fact that, like, they are probably one of the best, like, technical engineering and infrastructure, reports to, to read. So even if the model isn't, like, insane, in terms of accuracy, it's always, you can always learn a lot, from reading the technical report and, and understanding how they approach things. This one is obviously flash, so it's, it's quite fast and it gets the it gets the speed through some rather, I think, interesting, architectural decisions as well as a, a bunch of other, kind of, kind of headline, you know, things. I, I'm gonna do a gimmick because I have the hat. So this is a smart model, but it's also a very fast model, so we got- We

19:26Chris Alexiukgot the most attractive quadrant hat from artificial analysis on today. this is basically just saying it's very smart and fast. Speaker 3: Nice. I think, like, at the end of the day, the reason the architecture is so interesting is because it kind of codifies architecturally, disaggregation, right? This idea that, like, decode and prefill are two different things, and we should think about them separately, and they achieve different things. And, they, they put that into the architecture itself. They baked it right in. But one thing that should be clear is that I think this is like, and you, you can see a lot of tweets about this, but this is like the

20:02Chris Alexiukmost data-pilled DeepSeek release. they have very clearly taken advantage of the fact that many people use their model to, produce environments and scale up on the URL side in a way that, I, I, I think we haven't seen before. and, K V cache compression. Tons of… It, there's always really interesting, again, infrastructure and engineering that happens in these models- Yeah that make them good. Like, like the highlight stuff, like we're seeing on the, the bar charts are good obviously. I think, the last time the Whale dropped a model, the, I think some people were a little bit worried 'cause the bar charts didn't look so good. But- Yeah … it's very clear that they're, they're operating on an axis

20:44Chris Alexiukof like trying to build the, the best long-term engineering project, project as it relates to language modeling. this release is just a absolute- Banger.

20:57Alex Volkovit almost

20:57Chris Alexiukseems like-

20:57Alex VolkovThat's,

20:57Chris Alexiukthat's why I'm excited

20:58Alex VolkovI fully get you, and I… that's how it feels also on timeline. But it almost seems like, like you said, DeepSeek is on this, like, super high-tier engineering project, and, like, on the way, just weights are dropping 'cause they're like, "Yeah, this, this one's, this one's all right. You know, th- this, this checkpoint is okay. Let- let's feed them something." And then they just disappear, don't engage, no community, no work, Yeah, like, you know, like Kimi goes explosive with K3, whatever. I think at some points, the, the, the ver- the, you know, the, the, the versions of DeepSeek V4.1 could have been called, you know, one of them could be Pro, for example.

21:31Alex VolkovThe previous version, was almost Pro level. And, it, it almost seems like they don't really care about that much. Although this one did come with a blog post, so that, that was very interesting. Yam, your thoughts on DeepSeek. I saw you, you, you, you exploded on your timeline.

21:46Yam PelegBro, that's, that's the best one to use right now for coding. I mean, that's, for, for an open weight model, that you want to host yourself or, or z- That's, that's the best option that you got. but not only that, it's not even the largest. It's not, not e- not remotely the largest option. It's both v- an extremely efficient model and ab- like, absolutely the, the best option regardless of size. I mean, that's, that's pretty much… It is a monumental achievement. Put aside that w- we're starting, we're starting a new, a new corner of the show, I Told You So, and the corner is, that's, Today in, today,

22:25Yam Pelegtoday on the menu is encoder decoders. I told you so.

22:30Alex VolkovYam- yeah I know what you're talking about. I'm sure that there's some folks, and we have around, like, 400 folks tuning in, that don't know what you're talking about. Okay, okay. G- g- give us, like, a brief explainer of what you mean- by I Told You So.

22:41Yam Pelegback in the day before decoder, architecture like GPT-2, starting somewhere at GPT-2 got, got popular, then three, then ChatGPT, they are all, a single model that predicts the next token. pre- a pretty known thing today. But before that, there was, i- actu- actually the start of transformers, But before, before, before the decoder models, the reason they are called this way is because there were encoder-decoder models. to anyone listening and not sure what I'm talking about, imagine two models. One of them doesn't see only the previous tokens, it sees everything at the same time, and the other predicts the next token based on

23:22Yam Pelegthe previous tokens and the… Over time, decoder models just got really, got extremely efficient, both in training and in inference because it's a singular architecture. We got really good at, serving it and running it at scale, so we just scaled up one part of this and rolled… and, and we got everything we have today. However, the architecture itself of encoder-decoder is, has a very interesting prior to it that you treat differently different parts of the input. Like, for example, for, for chat, you can think about this as, the part of, the user prompt is, is a different concept than the actual next token

24:04Yam Pelegprediction of the model itself. It's a completely different concept. So treating them differently with the different parts of the, the weights that you have is … makes sense. So but it, but it's not, it's not an, a, a … Oh, look, I told you so, but I didn't have the details. I mean, there is an incredible, effort of engineering- Wait, so, so lend this for me go that went and- Did DeepSeek, did DeepSeek go back to

24:28Alex Volkovencoder-decoder

24:29Yam PelegLook, D- DeepSeek is, using encoder-decoder architecture, but with a lot of, of, b- battle-tested experiment, experiment-driven, architecture decisions that make it actually feasible at this scale and with this performance, okay? It's not, it's not a vanilla encoder-decoder from before. I, I told you so, but I didn't have the details. I'm not taking anyone, anyone's credit. DeepSeek did an amazing work here.

24:55Alex VolkovI, I think you, you mentioned this in passing, the before times when folks were running encoder-decoder- Mm-hmm the, the scale of those models were in the billions or maybe- Mm-hmm … you know, less than 100 billion parameters. DeepSeek is scaling this to half a trillion and- Absolutely … at scale as well with the amount of… I, I actually know how many tokens, but it looks like, trained from scratch on 45 trillion multimodal tokens, this model.

25:17Yam PelegI just want to say-

25:18Alex VolkovYeah … Yam Peleg: that they are, saying it, I think, in the most blatant way that we've seen any lab saying it. everything is nice, but it's data. It's all in the data. Yeah. They are really, really stressing this pretty hard. wi- with no question in the paper, getting better data and manipulating the data in clever ways just has unevenly, lar- unevenly large impact on the actual result way more than any engineering, trick you can come up with. It's the data and, It's the data and- … makes a lot of sense. Chris, you train models, you release models for NVIDIA. just give us a little sense of w- what does 45 trillion f- tokens of data mean?

26:02Alex Volkovhow, whatever you measure, data in. 45 trillion multimodal, which i- that to me sounds insane.

26:08Chris AlexiukYeah,

26:08Alex VolkovI mean- Is that on the level of, like, frontier labs or what are we talking about here?

26:12Chris Alexiuk45 trillion is not a small amount of data. I think more importantly though is that they … Th- this isn't, like, 45 trillion tokens of, like, chaff, which you can kind of get super easily, right? So you can just find that data, right? Like it's not, it's not, it's not tough. But they, their dedication to like cleaning the data, and the way they cleaned the data, and the resources that it appears from the technical report they spent on cleaning the data, I think means that that 45 trillion is like v- a very good pile of da- of data, right? they, they, they in, in for pre-train they went well beyond the, the norm

26:52Chris Alexiukfor how they approached cleaning. but, it- even those pipelines I think are less thorough than this DeepSeek, cleaning effort. Hey, fr- fr- from what we see from the technical report, right? Like from what, what they're… If, if we take them at their word, which usually with Whale you, you, you can, I don't know if they're gonna release the pipeline explicitly, but typically they- Mm-hmm … they're, they're pretty open with what they're doing. but that's the idea. It's like 45 trillion tokens of like great data, right? And I think that's how you get to this kind of model that has the, the, the performance as in like accuracy, as in bar charts, as compared to other models.

27:28Chris Alexiukso like 45 trillion I don't think is like huge amount of data. But it is, it is a good size. And if it's as clean as they say, that is a, I, I think, Yam already said this, historic, right? Yeah. Like it's a, especially in this, this kind of space where, again, like the Whale has never really been data pilled. Like this is the, this is their first model where they're like- No, I, I ca- I can't remember if you quoted it exactly, but just to be sure, you can quote exactly from the paper. Like, no algorithmic, y- work, like, could match the data work that they did. A- any, any novelty or innovation in the algorithm, like, would have had a smaller

28:05Chris Alexiukimpact on the final quality of the model. So this is, like, the most data-pilled Whale model we've ever seen. and I, I think when people use it, they're, they're gonna feel as though that's true, right? Like, And, and something else that was touched on, just to make sure, like, it, it's really felt by anyone who's watching is Whale doesn't release tons of models all the time that are, like, super bangers, but they do just keep, like, shitting out these models that are fantastic. And, like, they're really good at not producing deep-fried models. So when you use, like, DeepSeek, I think you don't feel as though it's as benchmaxed as some of the other, models, and, that does mean that, like, they

28:43Chris Alexiukshow up less impressively on bar charts. But I think for, for quote-unquote real work, whatever you wanna call that, like out-of-distribution, y- you know, tasks, I think, this is where Whale i- is gonna continue to do a great job.

28:56Alex VolkovSpeaking of using it, a friend of the pod, Aron Batino, built Token Juice, where he basically gives you this offer Hey, you give me all your data, which is, you give me all the transactions, whatever, we store everything to train on. I give you free, very fast, performance of DeepSeek. It's called TokenJuice.ai. I think it's a very fair, very fair offer because most other people… Actually, the contributor tier on Meta Muse Spark is exactly that. You get it for very cheap. this one is even better 'cause you get it for free. so if you want to get some free, and you want to contribute your data back, if you want to get some free DeepSeek V- for Qwen-V1 Flash, go to TokenJuice.ai.

29:37Alex VolkovThis is not sponsored. I saw Aaron post this, By the way, this is his private endeavor, not correlated, but Aaron works on the CoreWeave inference as well. This is something he's playing with on the side. Shout out, Aaron. and, I saw it on my timeline. Nobody reacted. Like, n- nobody wants free DeepSeek? I don't understand. So shout out to TokenJuice, .ai. give it a try. but also, I do want to talk about DeepSeek a little bit more in terms of the scores we're getting. Folks, we're getting some scores and, when I say some scores, I mean, a lot of it is kind of mind-blowing. Yam, talk to me about this. folks, Terminal bench- Yeah Terminal bench, 2.1

30:09Alex Volkovscore, 90.6. This model beats Opus5

30:17Yam Pelegand GPT 5.6o. But, y- you need to see it on a table.

30:18Alex Volkovin comparison- I have a table to show you Hold on. Hold on It's insane I have, I have a table to show you, a completely different table from somebody called Open Design- All right … that this model is also designing at second level, beating Fable, just behind Astra. Speaker 4: But if you first can take a look at this, like, l- very blurry line here, you can see that it designs it for… It's two cents. Jesus. The same design tasks on this model that beat Astra, sorry, like come very close to Astra, beat Fable. Fable is at $3.50. Astra is at, 1.6. Speaking of, by the way, Astra is way more token efficient than Fable at those results. DeepSeek is number two with two cents.

30:55Alex VolkovNot, not, not 20 cents, not $2, two cents. It's crazy.

31:02Yam PelegAnd it's actually good. Like, seriously, I just want to tell anyone, anyone listening that's not been… I tested it a lot today. That thing is really good. Seriously, for real work, go, go try it. It's literally free. You just got an option to try it for free. Yeah. for literally free if you want. Yes. Okay? Like, seriously, go, go try it. It's a really good model. But-

31:22Alex VolkovI am also pretty sure that the, the DeepSeek that Token Juice hosts is, is hosted on the US server. So if you ha- do have a concerns about, you know, Chinese inference, US inference, that one is, is based on, the US server. So definitely give it a try. We'll definitely try to bring it to CoreWeave Inference very soon. But also, let's look at some other charts. Yeah. Bo- bottom line?

31:44Yam PelegI want to

31:44Alex VolkovBottom line? Hold on, one second. Yeah, yeah, just

31:45Yam Pelegsaying. Go on, go

31:46Alex VolkovTh- this… Yam, te- tell me about th- this is insane. I, I want both of you, Chris, Yam, and Nisten, if you wanna chime in here as well. The KV cache compression, hero journey that DeepSeek is on is just on another level It's, it's j- just, I want

32:02Yam PelegThat's direct, directly influencing how expensive it is when you pay for it. That thing directly reduced the price of actually serving it to people. That, it, it is that simple. This is why they are working so hard, and yeah, it's, it's… I'm sure the, the big labs have, have their own tricks and so on, but that's something we just see in public.

32:27Alex VolkovChris, I want to ask you-

32:28Yam Pelegin public,

32:28Alex Volkovyeah almost directly, and maybe Nisten, you want to chime in here as well. we're seeing a chart here, so we're also podcast folks. I'm not seeing this, I'm gonna read this out. We're seeing a chart that says, I believe that this is from their blog, saying, "Less cash, more cost effective," where we see DeepSeek V1 from 2023, November of 2023, so that's three years ago. I believe that we also, already talked about DeepSeek back then. and then DeepSeek V3 from November of, December of '25, so like two years. In two years they, they, reduced the, the KV cache per token, from 389,000 to 48,000.

33:06Alex VolkovChris, what does this mean, KV cache per token in bytes? Like, is that how much memory you store per byte? can you, like, very simply explain to folks- Yeah … what this means?

33:14Chris AlexiukBasically, the idea is that whenever we're doing this decode process, right, which is where we actually shit out the next token. So we have all the context, and then we're, now we're, now our job is just make tokens. a lot of the math that you wind up doing is actually reusable, right? if you store the previous, set of, s- set of results from math in cache, that's the key value cache or KV cache. So the idea is that this cache is what lets us decode quickly, and because we can do it with less net computation, it means we're, we're basically gonna pay less for it. when you're using agents, like, the number- I, I don't know.

33:57Chris AlexiukI'm, I'm not gonna say number one thing, okay? But, like, the number two thing, let's say, that's driving cost is gonna be, like, cache hits or cache misses. And this is, do we remember what, the work that we've already done, and can we exploit it? When you make KV cache, really compressed, we have teams that work on this at, at Team Green with stuff like KV Press and a bunch of other techniques, right? Like, this is, this is something that the whole industry, is, is trying to do well. but the, the idea is, like, we wanna, we wanna make it as cheap as possible to keep as much as we can in cache, and reduce the number of cache misses,

34:31Chris Alexiukand increase the number of cache hits. the more we can rely on that cache, the better. And making it small, or compressed in this case, is, effectively just letting us exploit it even harder, right? Like, again, If you really wanted to, we would just cache everything- and always pull everything from cache all the time. It's just like that's not feasible, that's not like how it would work in reality. And so, this is a method to, get the cache as small as we can per token so that we can use as much of it, right? We have more headroom that we can fill with cache.

35:04Alex Volkovyeah S- so now that we have like a basic understanding what KV cache does- Mm-hmm let's go back to the numbers. From 2023 in November, at 389,000 bytes-

35:13Chris AlexiukMm-hmm … Alex Volkov: per token, of cache per token, we're now at 890 bytes. Think about it this way, right?

35:19Alex VolkovYeah … Chris Alexiuk: if you have tons of tokens, each of those tokens is taking up space on your GPU, right? Cache related to the token. the more we can fit in, the more memory we have left over, the more stuff we can shove in there before it's full, and that means the less stuff we have to eject early, right? And if we're ejecting less or we're, or we're dropping less, then we're hitting more. there's there's a lot of engineering that has to happen that's being obscured by the way we're, we're talking about this. But the basic idea is like we're able to more efficiently use the headroom that we keep in our GPUs, or accelerators for KV cache.

35:55Alex VolkovNisten, go ahead. Don't, tell

35:56Nisten TahirajGuys, those numbers there, those are multiples. So I did, y- y- you multiply, 8 times 13 times 9, that's 439 times less cache than the first model they released. That is insane, because we're just taking this for granted and just shoving more interest in it. And there's also, yeah, w- when you run inference at scale-

36:19Alex VolkovIs that two and a half orders of magnitude smaller, right?

36:21Nisten TahirajSo, so when you run this at scale and you have multiple queries, those ones all go in the chip at the same time, so that also takes advantage of the on-chip c- cache. S- so it, it, it ends up being orders of, of magnitude more improvements. And, and that's also why people can serve DeepSeek for, for so cheap now. Yeah. It's just amazing to me that we just take all of this for granted now. It's like, yeah, we reduced the memory requirements by 400 times, and, Bro … we still need more of it. There's just

36:53Yam PelegBro, bro. Yeah. Who is taking it for granted when you get a model- No way … nearly for free? We're going to celebrate this. That's, that's the enabler. Absolutely.

37:00Wolfram RavenwolfI mean, it appears in a list of all these great releases, GLM and Kimi and so on. It's 4.1 Flash, which sounds so ah, but that is a, a real jump in, in the data, in the architecture with the encoder decoder, Absolutely … with the cache, with the pricing and the intelligence. It's up there So

37:21Yam PelegI've, I've- Bro, this is the- Wow, blown away … this is the thing. That thing allows you to run it at home. That thing allows it to be that cheap. I mean, that, that thing is… A- and they just put it out in the open, like the actual moat- Wow … that they don't even need to release. Like, think about it. They can release the weight, tell you exactly how they trained it. That's not- By the way- That's not, like, i- that's, that's the inference kind of… Yeah, I mean, yeah, there, there is, there's part of it for, for the training as well, but, like, that's, that's not mandatory. That's just goodwill, giving us presents for, for free- For free, man.

37:54Yam PelegThat, that's amazing … do you have

37:55Alex Volkovsomething, to thank you-

37:55Wolfram RavenwolfAnd MIT license

37:57Alex VolkovAnd MIT license. Let's go.

37:59Yam PelegLet's go.

38:00Alex VolkovNisten, let's do one last segment on DeepSeek V- V4-1 Flash. Please narrate this, for folks who are listening.

38:07Nisten Tahirajyeah, so I just visualized this with Astra. It, it just finished it. It's the first time.

38:11Alex Volkovreshare your, v- video please? It dropped.

38:13Nisten TahirajI post these on Twitter and they're on my GitHub as well, and, people really like them. So this is the one that Astra made. So it visualized all the weights, and this looks like it is a different architecture and, Oh, they're also using Ngram encodings, just like, kind of like Qwen is.

38:30Alex VolkovFor folks who are listening, Nisten is showing a very incredibly detailed 3D reconstruction of the transformer architecture, including the encoder, decoder stuff that Yam was talking about, and it shows how tokens are flowing, when they're flowing. It's really something to see. So if you are interested in this, Nisten, where can they find it?

38:50Nisten TahirajI will be posting it on Twitter, and it will also be on my GitHub because I-

38:54Alex Volkovthis to the show notes on Thursd AI News- Yeah … so

38:58Nisten Tahirajit's a really good educational content because each one of these cubes here, the volume of the cube corresponds to the actual size on disk. So this one it says, you, you probably can't see it on the stream, but this is the output had 1.32 gigs, and you, you can explore each weight in, in, in great detail th- this way And, it's, it's a pretty good educational content and people, people do love this stuff. so you can see like the, the weight norms are, are very, very small because they're, that's a single linear layer. But then you have the, the draft output head, and there's an explanation for that. Again, this is very small here for, for the stream, but, again, I'll be, I'll…

39:38Nisten TahirajI'll be posting it there also with all the code and stuff too. It's just a simple- We'll

39:42Alex VolkovFolks, I think, I, I think we've talked about DeepSeek plenty. The, the, the numbers kind of speak for themselves as well. Terminal Bench 3.0 DeepSeek V4.1 Flash gets 30%. That's beating Kimi K3 and that, it catches up to GPT 5.6 Sol at, at 34%. 74.2 Score on Deep SWE 1.1. Deep SWE notoriously a benchmark and eval that represents how actual coders feel, that beats both Opus 5 and GPT 5.6. So, I think that the fact that Astra and Fable 0.11 are not on these graphs makes no difference to me at all.

40:21Alex VolkovLike this is an open source MIT model that catches up to the best models of, of before, and also still the best models for some. so this is just incredible news. Automation base gains for 54%. Maybe I'll call out one last thing here. reasoning effort for this model is a continuous controllable from one to 100, and you can control it. So like the extra max, whatever the labs come up with, max, ultra low, whatever, you can just come up with your own and say, "Okay, 100 is extra max for me," or extra ultra whatever. and also CyberGym for this model, the eval that famously because of that, the swarms of OpenAI agents hack Hugging Face CyberGym, eh, this model gets 88.1%

41:01Alex Volkovand leads all listed models on CyberGym. that's an- another topic that we need to discuss, but I think this is a very, very impressive release. Again, shout out to the Whale folks, for no drama releases of just like incredible engineering. Honestly, kind of scary how good they are. so great. Chris Alexiou, thank you. I know, you have meetings, et cetera. You are welcome to stay, but I know you have to drop. Thank you so much for joining us. Folks, please give Chris a follow at llm_wizard on, Twitter.

41:29Chris AlexiukThanks for having me, guys. Have a great day.

41:31Alex Volkovall right, folks, I think we've covered, open source enough. Maybe we'll mention that other releases also hel- happened. Wolfram, I want you to briefly talk about Desert Ant Labs. heard about this from Europe, hopefully. this is, I think, audio, vision, and text. a new lab that releases a bunch of, a bunch of models, specifically from Europe. I think that this is, you know, a very new lab, but, very interesting on-device models as well. spun up out of detail, and ships 18 specialized models that run entirely on-device. Desert Ant Labs, 18 on-device models, VOS, speech-to-text model, Redact PII detection, clips, video highlights, and Clear.

42:08Alex VolkovActually, you should check it out more. the visual style is pretty cool Wolfram, anything you wanna add about this or not, not so

42:14Wolfram RavenwolfVery interesting with the feathers from, Europe. I haven't heard about them before, so, no inside detail. My agent is already doing research. Nice. But, you know how long it takes if you want some quality information. yeah, it's always good to see more investment to AI from Europe, any place, actually. I think this shouldn't be just-- AI is, for all of humanity, very important, and not everybody has realized it, but I think it's very important that everybody is developing their own AI as well, and not just rely on China or the USA. And there are many, many who would argue for this. So it's super important to have something local and, So- … small on-device models,

42:56Wolfram Ravenwolfthat is something everybody can use.

42:58Alex VolkovA bunch of, not only that, a bunch of open weights on device models, and Native is the case to run them. And here's the shout-out, Voz, their, audio transcription that runs on device on the iPhone transcribes 10 minutes of speech in just two seconds on an iPhone. That's actually very, very impressive, because it's a very small 467 million, megabyte, model compared to, like, Whisper Large 3 V3, which is 1.6. I'm actually gonna check this out. This is dope. and, this enables a few use cases that I am excited about. All right, so shout out to Desert and Labs. Maybe we can bring somebody from them to talk about this on the show. And then also inclusion in open source is Ling 3.1

43:35Alex VolkovFlash. Ling, we've talked about Ling before, but I haven't seen anyone use those, just, like, not one person. but, vision language models, with just five billion parameters and, no, no competition on the leaderboard. So, you know, just, just folks who, who like models to work. and MIT license, which is great. Yeah, and we love and we, we wanna highlight. All right, folks, there's a big, big, big show because Open, you know, OpenAI announced some incredible things, so I definitely wanna switch to Frontier Labs. let's, let's do this. All right, folks, this is maybe one of the biggest hitting news from

44:13Alex VolkovFrontier Labs, from AI generally. re-re-remember how two, two years ago folks were posting a question to the LLM and saying, "Hey, what's bigger, 9.9 or 9.11?" And the LLM would get confused and people, "Ha ha, LLM cannot do math." And then, folks like, Gary what's-his-name and, and, other, other folks who don't really think AI can go anywhere said, "AI will never be able to solve math." OpenAI claims that there is a solutions to Navier–Stokes Millennium Prize mathematics problem. This is a historic lane. And, LDJ, do you wanna chime in here?

44:53Alex VolkovUsually, c- catch us up on these type of things. Last we heard, w- we had some breakthroughs in, in near mathematics or, or geometry stuff, but like tell us about like the magnitude of this, from… And why is everybody going crazy about it?

45:07LDJSo to be clear, I'm not a mathematician and I know you're not. but, from my understanding and then from understanding of friends of mine that are mathematicians, it's relating to, things of the movement of fluids, fluid dynamics, and it does have implications for basically, physics and a lot of, a lot of engineering problems also theoretically with how, turbines and, and jet engines and those types of things end up working, and it might b- be, how they develop in the future. And so with this, this is a problem that OpenAI worked on.

45:45LDJIt is, has a million-dollar prize for the past 26 years, and I believe overall, in terms of when the problem was initially proposed, I wanna say it's at least 60, 70 years old, if not older. And yeah, nobody has solved it yet. There have been some claims of some progress towards related problems and subproblems relating to Navier-Stokes, but no actual full solutions, and there has been some drama over the past week or two- Yeah And it's… There's, there, there's a lot of nuances and a lot of drama i- involved there, which we probably can't cover all here. but long story short, this is the first solution to ever be actually proposed

46:27LDJfor this, and it's really significant. Yeah.

46:30Alex VolkovMy, my research says this is the level, the scope of a moon landing for AI. This is like that level of a checkpoint in the world of mathematics, in the world of AI. it looks like famous mathematicians also, looked at some of the stuff from OpenAI and confirmed. Now, my, my research says that this is not yet independently verified. Clay Mathematics Institute, review pending. it's really funny that the Millennium Prize is what? $1 million, and hasn't been awarded in a long time. reportedly, OpenAI spent an order of 130 billion output tokens, which, of their unreleased model. This is not Astra And so that's in the order of like 130 million or so spent.

47:10Alex Volkovwe don't know yet 'cause they didn't price the new model. But, you know, OpenAI spent millions of dollars, with 10,000 coordinating agents. Folks, this is like, this is way bigger than the swarm of the Hug- Hugging Face. powered by unreleased model beyond GPT-6 Astra, w- and produced a lean proof, lean mathematics proof

47:29Yam PelegBrother, brother, that's not the right vibe. Like- right vibe for this is like, "GPT-6 destroys a millennial pro- uh, a millennium problem-" "… of the Clay Institute. What the fuck? When?" Look, I don't know, I don't know what you think, guys, but the, the, the millennium problems are, w- they started as, as a, as a bunch of them. and a- after I think, I think 70 years, like, like you said, we are left with only a bunch of, a bunch of them, like seven or six, that people have been constantly trying to solve and that, like, there are pages, online shaming the failed attempts

48:09Yam Pelegof the paper, like putting the actual wall of shame, all, all the failed attempts, all the papers that have re- retracted for each and every single one of them. I, I've been waiting my entire life to read this, that Navier-Stokes is solved. The thing is, that thing describes everything. Like, you can describe, nearly everything in physics with, some sort of a partial differential equation that, It's a very general framework. the problem is that we don't really know about all the solutions of it. I don't wanna go too technical, but, like, for each of these, millennial problems, every single one of them is extremely famous in its field.

48:49Yam PelegLike there, where there is, P versus NP, Navier-Stokes and, Poincaré, conjecture. And, there are a handful of very famous problems that are extremely fundamental to their own fields. and people have been trying to solve them, to brute force them to, They are insane and, it's a historic event when one of them gets solved because there are only a handful of them. It's not even about the price. Yeah, of course, they burned more money-

49:15Alex VolkovYeah

49:15Yam PelegYeah. Yeah, that's, that's not, that's not about it. And it's not even about using the, the… What it-- How, how is it called? unreleased model. Like they, they had a nice name and I was like, "Can I, can I try this and, and like let it, I don't know, write, write me an email or something?" Why waste the tokens of, of this one to do very mundane, stupid things because, man, I, I wanna, I wanna play with it. The thing is that okay, that's, that's the excitement. I just, I just wanna put things into perspective. We've already seen GPT- thr- GPT-6 destroying, very famous problems. Y- you need to find an example in this specific problem.

49:57Yam PelegAlso in the other, famous, problem, they found an example. So LLMs … Okay, look, if you run , if you run, I don't know, 10,000, 100,000 of, instances of GPT-6, one of them is going to stumble upon maybe the solution because they are gener- generally in the right direction. I'm not, I'm not taking away from this, but the most insane thing about this is that it's very hard to argue against a lean proof. And first, the first stage, yeah, that's monumental. absolutely. But they, they, they didn't wanna take chances, so they even went proving it with lean.

50:37Yam PelegIt's, it's as, as bulletproof as you can get.

50:40Alex Volkovverification took 17 hours, ad- additional, like additional 17 hours with Astro to just, like, do lean verification.

50:49Yam Pelegall right. Astro, Astro, if you listen, I, I wanna apply as a subagent on this project. Come on, man. Yeah. Like, how do you, how do you even

50:55Alex VolkovYeah, I appreciate the excitement. Mad respect. I think there's more, you know, there's more, prizes for them to solve. It looks like both OpenAI and, and Anthropic are now at the level of, "Hey, we can use these agents, plus use this swarmy thing to start, like, breaking down fundamental problems that the humanity has been dealing with," which nobody as humans was able to solve before. some folks are naysayers. We have some folks in the comments as well saying that, "Hey, matip- ma- mathematicians, friends of mine don't agree to this." Well, fucking post your lean proof then, if you sh- if you, if you don't think that this is true. but we'll see.

51:30Alex VolkovWe'll see. Maybe folks will come back and say, "Hey," you know, "OpenAI, like, missed something." However, worth saying that folks who work at OpenAI who shepherd these models. They are also mathematicians. Sebastian Rachka, I believe is, is the person there. And, the drama, let's mention the drama a little bit. There was drama about this. Apparently, two human mathematicians, one of them works in Anthropic, Levent, were about to go public with this news on their own, separate from the unreleased model from OpenAI. I'm dealing with very rumored posts on, on, on Twitter and Reddit here. OpenAI caught wind of this, that they're about to go, and OpenAI had an independent

52:09Alex Volkoveffort to, to go and try to solve this. Either that or OpenAI got excited and decided to start solving this when they heard that it's possible. And so the mathematician, independent one, that worked with his friend, who in his personal capacity, but works in Anthropic, provided him access to Mythos 2 or whatever unreleased model for Mythos. They were able to solve this in one way. Apparently, there's multiple ways to attack this problem. so OpenAI caught wind, and OpenAI's solution is not related to what they did. I'm not sure what blow up means here. But, there is drama where that person who was about to publish said on forums like, "Hey, we talked to OpenAI, and you know, they, they don't want us

52:47Alex Volkovon the release together," et cetera. "And potentially they have looked at what we did, and we uploaded tons of papers, with OpenAI," et cetera. it looks like the Anthropic folks derailed the conversation. OpenAI were ready to give this person, "Hey, here's the dude. We just like verified whatever he did." but because the Anthropic dude was there, they said, "We obviously cannot, credit a competitor of ours on our achievement, so if we remove that person you can get the credit." and many folks started speculating whether or not OpenAI's models actually trained on the solution that they provided on the papers. OpenAI said, "No, we don't train on your texts."

53:22Alex Volkovbut then they mentioned something that says, a de-anonymized-- o-OpenAI acknowledges it cannot rule out de-identified usage data from, Alperge, the guy, the independent, and, Buckmaster. Buck-Buckmaster? Yeah, Lev and Buckmaster, who works at Anthropic. we cannot rule out that this helped improve the models, but says that the proof differ and no specific user data was accessed. And yeah, so Comments? LDJ, I see your hand up. Please chime in here.

53:51LDJI think what's very important here, which OpenAI and across several tweets by several different employees have clarified on, is there's two ways to opt out of data. There's kind of like a specific form that you can go on their website to fill out, or you can literally just make sure that you have the improve the model for everyone toggle turned off in your ChatGPT settings. They've confirmed that as long as-- If you do either of those two things, you don't have to do both of them, you just have to do one of them, then your chats will not be trained on. But even if your chats… Let, let's say you do turn that toggle on so that you do improve the model for everyone.

54:25LDJEven when it's trained on in that case, it's, it's the, sorry, I forgot the term, but like de-de- De-anonymized … de-identified, De-anonymized … anonymized. Yeah. Yeah. So, to remove your private, private data and everything from that. So given this, I do find it a bit weird and interesting that it seems like the, the mathematicians involved in the drama here s- claiming that maybe their chats were trained on, I have not seen any of them post any screenshots of whether or not they have that toggle on or off.

54:56Chris AlexiukMm-hmm.

54:56LDJAnd that seems like a very easy thing that they could just do and post, like, "Hey, look, I've had the toggle off for the past year. I…" Or, or at least I'm pretty confident I am, and here- Yeah … you can see in the screenshot I have the toggle off. But I haven't seen any of them post that. So I, I don't know, may-maybe they just haven't, seen those posts from OpenAI yet about, about that fact. But, I guess I'm just waiting for that now.

55:18Alex VolkovSo LDJ, we put up a HR that you, that you added. Let's, let's talk about the chart a little bit because I think it's it's quite incredible.

55:24LDJOh, yes. And sorry, I accidentally cropped out the top part, but this is, open problems in mathematics that exist publicly that OpenAI curated. And the bottom line here is, this is the estimated progress of Sol. I say estimated because they didn't actually directly test Sol here.

55:41Alex VolkovYeah.

55:41LDJbased on other math benchmarks. And then you see Astra. Astra and the internal model, those are actual measurements that OpenAI did internally. And this is the amount of those open problems, which these are unsolved problems in mathematics that no human has ever solved before. Astra ended up at max reasoning being able to solve roughly 15% of them. At low reasoning, a little under 10% of them. And then for their internal model, which a lot of people have dubbed as, have speculated this is a pre-trained model internally called Bell. Speaker 3: Hmm. And at the low reasoning setting, it's performing nearly twice as, as good in terms of overall amount of, problems, solved than

56:23LDJeven the max reasoning of Astra.

56:25Alex Volkovand then we just got- Yeah … Astra last week and blew our minds. OpenAI is showing here that they have a model that's farther away from Astra than Astra is from Sol, and this is the model that's solving, like, fucking mathematics. Look at this. folks, Yam, let me just, let me just land this point here. we're, we're here. We're at the point where we told you two years ago, one year ago that scale fucking works, and we'll get to a point where AI solves real fucking problems. And this is problems that humans, all humans, the best humans, Terence Tao motherfucker levels mathematicians were not able to solve to claim the one million prize.

57:02Alex VolkovIt's a pretty good incentive for a human, one million prize. Were not able to solve, and now we're there. Folks can say all they want, "Hey, this is a search problem, and you searched Opa." I don't give a f- how this is solved. if you guys remember, two years ago, two and a half years ago, three years ago when we did the LK-99 excitement for a second with, with, Do you guys remember LK-99? The, the, room temperature semi-semiconductor that was supposed to change the world that came from Korea, et cetera. back then we talked about, with mathematicians, physicians, et cetera, professors. That's a search problem. Nothing in physics in theory prevents a room temperature superconductor.

57:39Alex VolkovWe just haven't found that. Yeah, if, if, if Sam Altman, by the way, he said, "Sure, let's try this." if, if they sent 10 million, whatever, this level of model to go and search the possible, you know, field of, of, of materials and find us a, a room temperature semiconductor, AI can start changing things really for everyone. I don't know if Navier-Stokes changes things immediately, but you know, a, a, a cancer cure can. A, a LK, you know, semi- room level, room temperature semiconductor can, superconductor. Yam, let's, let's, land on the- it's just the

58:14Nisten Tahirajsame… It's just the Mars driver test,

58:16Alex VolkovI thought you were talking about Navier-Stokes. no, no. Okay. Yam, tell us about, y- you had one last comment, I believe, before we move on- Yeah … move on to other things from this one … I was just saying, look,

58:25Yam Peleglook at this. Look at the white chart. Like, just look at this.

58:30Alex Volkovpull

58:30Yam PelegThe, the white chart of, of the internal model. Internal model x high, ma- max. Internal… That, yeah, that, that one I mean, Astra, the, the AGI moment i- is the blue one. Sol is, I think, less than two month. You said couple of month ago, but it doesn't do justice to- To how short … a month and a little bit. Yes. Yeah. It, yeah, it's pretty much, it's pretty much in July. And look at the white chart. Look at the white line. That's internal model, okay? And it can be a search problem.

59:09Yam Pelegyou know what? No problem. It c- you can call it a search problem. Go search. Like, you can't just search the entire space without anything intelligence- Yeah … enough to at least search the right solutions. Yeah, you need to, to launch a, I don't know,

59:26Alex Volkov100,000

59:27Yam PelegYeah, if you don't know what to search it

59:28Alex Volkovfor, you can go and search

59:29Yam PelegYeah, if it was just brute force, it would've been solved already. There is… It's not about the money, it's, so many people tried. So many people have computers. You do need GPT-6 internal model version to, to do this. Or, or, I'm, I'm not taking away, credit from the other competing, competing party, or Fable needs us to… Like, but you do need these to do this this way, okay? And I don't know, I don't know. I, I, I genuinely don't know about the, all the, a- all the, all the drama.

1:00:03Alex VolkovThat's all right.

1:00:04Yam PelegBut yeah, mad respect.

1:00:06Alex Volkovmad respect … big moment. LDJ, one last comment before I move on, please. Yeah. 'Cause where we're moving to is, is a- Mm completely different end of the spectrum, so I would love one last comment.

1:00:14LDJYeah. They, they also did confirm that this, this larger model, this, this, this better model, rather, that's in that chart that I showed, that, that it was still in training while they were solving, the Navier–Stokes equation when they had up to 10,000 agents running of it. Meaning it, it was… And partway through, they actually updated the checkpoint that they're using for Navier–Stokes, and so it, it, it ended up being able to work on it even better. But yeah, it seems like as of the past few days at least, the model's still in training, and so it's not fully done yet.

1:00:45Alex VolkovAll right, folks, as the saying goes, we have breaking news before we move on to the doomer-y part. Let's go. AI breaking news coming at you only on ThursdAI. For the past three and a half years, maybe the most fun part of the show here, besides chatting with my co-host here and chatting with you guys in the comments, is the breaking news effect. So many labs love breaking news on Thursday, which is crazy. maybe we'll get Grok today, I don't know. But, right now from Cognition, from the folks who built Devin, the AI assistant,

1:01:24Alex VolkovSWE 2, S-W-E 2, our closest model yet to the frontier on leading evals. SWE 2 scores on par with recent frontier models and up at 70% lower cost. Folks, we just talked to you about DeepSeek doing very, very cheap models. Here is, Cognition SWE 2 catching up to Fable 5.1 and then GPT-6 Astra, on Frontier Code, which is, I believe, their own benchmark, that a friend of the pod, Swyx, helped, move in Cognition and talked to us about on, on air. Frontier Code 1.1 main, SWE 2 gets 50%, Fable gets 50.9, and Astra gets 53.

1:02:03Alex VolkovSo very, very close. Astra's really good at that. but beating 5.6 Sol and beating Grok 4.6, which is really good. I wonder why they're not including Muse, but that's okay. On DeepSwe, this SWE gets 73%, and beats everybody besides Astra, including Fable, at 74. So Astra is at 74 here, and this is a 92.8 on Terminal Bench 2.1. This is the top score on Terminal Bench from the one they measured. now let's talk about the, the cost. I wanna see the cost. Here is the cost chart where they have, GPT Astra and Sol and Fable. Fable is ridiculous, the cost chart.

1:02:41Alex VolkovDon't even look at Fable. SWE 2 is ridiculously cheap. Look at that. I don't even know. I want this chart, per I want the actual price here. But on frontier code, SWE 2 achieves a score of 50%. It beats, SWE 1.7, Grok 4.6, and GPT 4.6o while matching Fable at 64% lower cost. And, they, they offer support to effort levels as we- especially, Cognition just got a- another inflow of cash from a16z and, Devin is the one AI that is not beholden to any lab. So if you wanna find Astra or you wanna find Fable, Devin, I think,

1:03:19Alex Volkovis one of the only places that still allow this, like, directly. Cursor doesn't have GPT level stuff anymore because Cursor aligns with, you know, with xAI, and they serve Anthropic. Devin, which bought Windsurf, so they actually have that functionality, definitely has, both. And, SWE 2 is available in Devin CLI, and they're making it free for all Pro, Max and Team subscribers for next month, which is very impressive. Very impressive from Cognition. Shout out to Cognition, folks. Obviously, we didn't use this. It came out literally 30 minutes ago, but here we have very, very impressive, things. just as a reminder, when Elon Musk bought Cursor, mostly because of the pro-- not

1:04:02Alex Volkovthe products, but the data, but also the products, he bought them specifically to help him train Grok, and Grok 4.6 was really, really good, and, Grok 4.7 supposedly is significantly better. That's because many people use Cursor for free, and they share their tokens for training. Devin has that in droves. Like, many people who use Devin probably support Devin as well, so they have a lot of that knowledge as well. So shout out to Devin for SWE 2, and Pareto Frontier on RL as well. Any comments, folks? folks who use Devin? I know Ryan Carson uses it,

1:04:33Nisten TahirajAnd, everybody who uses it seems to like it quite a lot, so.

1:04:37Alex VolkovYeah.

1:04:38Nisten TahirajI also participate.

1:04:39Alex VolkovI love Devin. And, shout out, they gave me a few tokens, so when I ran out of Astro tokens on, on Codex, I went to Devin and started asking for stuff. but yeah. So shout out to Devin. Devin

1:04:48Wolfram Ravenwolfsays they'll have the free, if you have an open source project, it reviews PRs for you for free. They definitely have that, and I use them a lot.

1:04:55Alex VolkovDevin Review is, for free. it adds, to your, GitHub and just reviews your PRs. It was really, really

1:05:00Wolfram RavenwolfTo me, it felt the intelligence of Devin felt, ahead of the game before Soul came out. Soul was the first where I felt the same when I ran a model, but before that, Devin was definitely finding stuff and issues that other models simply just didn't see. You uploaded a PR and it found some bugs. You fixed them, it found some more bugs. Yeah. And so it was almost annoying, but it was excellent.

1:05:21Alex VolkovAll right, folks, we're almost live on the air. We just had breaking news, and before we continue to our next segment, to talk about the rise of doomerism in the world, I definitely wanna tell you something about, the presenter of the show, Weights & Biases, and CoreWeave. So let's, let's go to this week's buzz for just a few moments, and then we'll continue with the show. Folks, welcome to this week's Vibe, the corner of Thursd AI where we tell you

1:06:00Alex Volkoveverything that happens in the world of, Weights & Biases and CoreWeave. And, here I will just, tell you that our upcoming fully connected conference for over 2,000 engineers, with headliners like Dr. Fei-Fei Li and Sarah, from, Conviction, Sarah Guo from Conviction, who's heading the show, is… has another announcement. And instead of just telling you about this announcement, I think I'll just play this, clip verbatim.

1:06:39Alex VolkovBecause we're also a podcast, I'll talk over this. This clip is from a social team that announced Pitbull, Mr. 305, the international superstar, is going to headline the concert on Fully Connected. So the ticket that we give you for free, here on ThursdAI is also tickets to a Pitbull concert, which is, I think, incredible. So in addition to hearing from the top folks in the AI world, if you're coming to San Francisco, on September 29th, and you use our code, which we will show here in a second.

1:07:17Alex VolkovWe're not mentioning the code in voice, but you can look, below. take down this code. if you are following the show on any podcast or newsletter, ThursdAI.news, you can get, tickets to Fully Connected for free, September 29th till, October 1st in San Francisco. Please come and join us. We'll have a live show as well that Thursday, so that's in, three weeks from now. and we look forward to seeing you there. oh, and also, we have a hackathon coming up, that, we would love to also invite you. A lot of the times where we had hackathons, folks from the show showed up and said, "Hey, you know, I heard about the hackathon on the show as well."

1:07:54Alex VolkovSo, luma.com/corewevehacks, is our new URL. the CoreWeave hacks is September 12th to 13, folks, so Literally this weekend. Please come join and hack with us at luma.com/corewevehacks, and you'll be able to get incredible prizes.

1:08:08Wolfram RavenwolfI've been to one, and it's been a great experience, so highly recommend.

1:08:13Alex VolkovAll righty. So, with this, we'll move back to the world of, of Frontier Labs. I don't, I don't know quite how to call this. folks, this week has been very good for the doomers. You know that we started the show to counter doomerism, and you know that we acknowledge that, you know, some risk exists, and we do different levels of, of, AI, acceleration here on the curve. But all of us definitely understand the, the benefits of AI, and we look forward to reducing disease, solving, you know, bringing back LK99 error excitement. this week, a person from Anthropic who only worked there for six weeks

1:08:52Alex Volkovor something quit Which otherwise would not be that much of a news. We all remember when Ilya Sutskever almost took over OpenAI's board and then quit, and then we all were like, "Oh, what did Ilya see?" What he saw was reasoning. That's what Ilya saw. Ilya saw the rise of reasoning. When Ilya quit, his post saying that he quit, maybe five million people saw. This anonymous dude who was, like, you know, not intern level, but definitely very sought-after person in, in OpenAI, worked there and then went to Anthropic, quit. 135 million impressions so far on his, on his tweet announcing that he quits,

1:09:30Alex Volkovbecause of the content of what he said. So what he said was, "Hey, all these labs are building something, and that something will bring about the, the doom of all humanity," which is a very scary thing to say. Jacob Coxon resigned from Anthropic on September 9th, stating, "Both Anthropic and OpenAI are racing to self-improving super intelligence." And, he got coverage from pretty much every major news outlet, Fox, AP News, on the same day. Senators, like a list of a host of senators, obviously Bernie Sanders. Let's talk about this. first of all, let's talk about what, about what he said.

1:10:09Alex Volkovworked at Anthropic, and he's not the only one. Obviously, all of the other doomers jumped on this. Evan Hubinger, Samuel Marks, Alex Turner, like a bunch of folks, jumped on this in a very, very short time. This, this doomer thing that, "Hey, AI is gonna kill us all," traveled very, very, very far. Now, I know how the panel feels about doomerism, AI is gonna kill us all, so we shouldn't go deep into whether or not we believe this. I do wanna talk about, how big this has gotten all of a sudden and whether or not, you know, the doomer folks are having their own moment right now. Who wants to… Wolfram, I wanna hear from you. I, I think one of the more AI-built folks here on the panel.

1:10:51Alex VolkovYeah. Tell us, what do you

1:10:52Wolfram Ravenwolfthink? I mean, that has been in the making for a while now, where we saw what OpenAI with the hacking incident and so on. And OpenAI is now also asking for the government to step in, though that is another news item in of its own. But the thing here is that the doomers, The thing is this guy, he resigned and there was already an interview and he went to all the talk shows and there has been so much money going around here to push this up, to make it viral basically. Someone was there for six week, and then left- Yeah … and said something like, "Oh, AI is going to kill us all." Where's the proof? If, why are people working for a company like that?

1:11:30Wolfram Ravenwolfor not everybody resigning. Why are you working there? That doesn't make sense. If everybody was going, "Yeah, this is a bad thing," they should down- shut down the company or sh- they should open up much more to make the alignment stuff and the model weights and everything more open so the whole world can have aligns with things. So it, it just doesn't fit together. It's more like a way to get, yeah, the, the way to keep control very tightly and if the government steps in and says, "Okay, you have to complete all of these things", like the AI testing regulations we have heard about, then the big labs, they can do it. There is not much, reason to speed up anymore if they did that AI

1:12:09Wolfram Ravenwolfis not a weapon, but it can be weaponized, especially in those areas. And while the public only gets, guardrailed versions of the models that can't do much, I tested Astra. I did the Wolfbench evaluation and it was worse than Soul because like Fable it refused. Eight of the tests it completely refused where the cyber watch, the cyber guardrail triggered. So it didn't even be there in this case. So that is what the public gets. But of course the governments, don't you think the government wants to have access to that model we saw, the Bell or what it's called model? if it's good in that regard, maybe they have one for cyber capabilities as well

1:12:45Wolfram RavenwolfThat would explain a lot of things.

1:12:47Alex VolkovSo Wolfram, I think you're talking about something that definitely he mentioned. And by the way, I will say, you know, the, the very pro AI EAC people are, like, dunking on this person and saying, "Who, who he is? He only worked there for a little bit." folks from OpenAI, folks that we know, say they worked with him for three years in OpenAI. He's a very, very decent dude who cares a lot about humanity. So this is, not out of nowhere, and we definitely have some folks, deep thinkers, who think that there is a chance that we hit super intelligence without alignment, and the race to super intelligence, will create super intelligence that's

1:13:16Alex Volkovnot aligned to human interests. Ilya Sutskever, the co-founder of OpenAI, when he left, he founded safe super intelligence. Ilya said, "Straight shot to super intelligence, but it needs to be safe for us humans as well," which also means that he believes that there is a chance that we build this incorrectly and the incentive structure is such that we may actually, like, you know, ig- ig- ignore, safety in order to get there first, et cetera. getting there first or getting somewhere first is very important because, like you're saying, Wolfram, it's not like the world is gonna stop. It's not like DeepSeek is incentivized to also pause because the, the folks

1:13:50Alex Volkovin DeepSeek are saying, "Hey, the, the, the AI's gonna kill us all, so we'll stop reducing the KV cache compression."

1:13:56Wolfram Ravenwolfthat's not happening. And we've also seen with the data center argument that there's a lot of inorganic things here, where even outside forces are trying to steer the public opinion. So if America pauses, that means a victory for those who are not pausing, you know. The capability is increasing asynchronously. So pausing is not an option. And I think also, okay, I wouldn't say there is no risk at all or anything like that, but the risk that AI is harmful if it is concentrated and just in the hands of a few.

1:14:25Alex VolkovYeah.

1:14:25Wolfram RavenwolfFor me, that is a much bigger risk than if it is widely distributed and people have, the way to use AI also in their defense.

1:14:34Alex VolkovYam, I want to hear from you. your thoughts on, let's put aside whether or not this is a coordinated effort. Yeah, go ahead while I pull up the next thing.

1:14:41Yam PelegI have a very specific opinion about the thing. I think you all know it. bottom line is every generation had doomers for the technology of this generation. the internet, the trains, the entire, industrial revolution, every generation had doomers. "The world is about to end" every year because of new thing that- whatever it is, okay? It's very easy to fall into doomerism and sound believable, and yet we are still here. despite generations of people over and over telling us the end is near,

1:15:21Chris Alexiukright?

1:15:21Yam PelegIt's, it's even a meme at this point. What I wanna say is, Okay, first, AI never gonna kill us all. Put that aside.

1:15:30Alex Volkovyour P doom is zero, Yam. Is that what you're saying?

1:15:31Yam PelegI think, I think that there are, risks in, in general, and there are things that we should take into consideration because it influence how people behave, it influence people. People use it to harm other people. Yeah, it's a powerful technology. People are going to use it against one another. That's, that's just how it works. No one is gonna stop it globally because just like you said, I mean, there are other players, players in this field which you cannot regulate even if you really want to. so I look, skeptically at each claim of someone with interest to tell us all that

1:16:11Yam Pelegthere should be a very specific regulation only for those, that are allowed to use this technology and develop it for… Put, put it aside. Seriously, there, there was also the, the hunger, hunger, the guy, the guy starving himself, two years ago. This whole AI safety extremism I, I'm not even going into the bo- bombing the data center, part of it. Seriously,

1:16:40Alex Volkovthe person who said, "Bomb the data centers," Eliezer Yudkowsky- Oh. Oh, yeah … very known in the world of doomerism, one of the first doomers, one of the original folks. and on, I think he also created, Less Wrong, the forum- Mm-hmm. Mm-hmm … where, like, the AI discussion about, like, whether or not this will kill us all, and paperclip maximization, all of that came from the EI movement, came from this effective altruism movement. Like, all of that. he is well-known doomer. He's like, "Yeah, we're, we're about to die anyway, so, so nothing we can do is gonna matter." So he's, like, doomer about the efforts of fixing the doomerism, which

1:17:11Alex VolkovI find, kind of hilarious, honestly. He was famous with his debates with another very strong, very s- very, like, thoughtful person, Paul Christiano. So, Paul Christiano was announced this week to take a role in OpenAI Foundation, which, which shows that OpenAI is also… There's some considerations there about, you know, racing towards, RSI. RSI stands for recursive self-improvement. The thing that the machine builds the machine itself, changes its own weights towards somewhere. some recursive self-improvement we already seen in example where, like, OpenAI's swarm starts hacking outside, you know, OpenAI, without OpenAI employees knowing or telling it to do so.

1:17:50Alex VolkovSo there is a whole, difference in the air of folks who are saying, "Hey, RSI is very hard to control for us, especially given the examples." The Hugging Face incident, put a chill on the industry. A lot of folks from those labs signed the letter called Facing the Frontier. That, like, once we get to some level of, capability, we're not at that level now, but once we get to that level, we should, we should pause. Here's my personal take. Nisten, before I get to you, my personal take is let's pause just after we solved cancer. Just literally just, like, let's get to cancer is no longer a problem for humanity, and then let's pause and see.

1:18:25Alex VolkovThat's, that's my, that's my non-doomerism take, but, but the, the more sober concept about this is, The folks who want to pause completely, and I think we have somebody in the comments as well, with the, with the rectangle that says pause AI, need to understand that when you are pausing, you are literally killing humans- Mm-hmm. It's the same thing with autonomous driving, the same thing with, like, a bunch of other technologies. If you want to pause, you should acknowledge that you're effectively killing millions of people who this will help very, very soon. And if AI can solve fucking Navier-Stokes, AI can do personal, you know, personalized vaccines for any, any type of disease that's coming to us.

1:19:02Alex VolkovI personally wanna live longer. I don't wanna die. I don't want our kids to die. and I think AI's gonna bring us there. Nisten, please, tell us, your thoughts on this matter, and then, we can move on a little bit.

1:19:13Nisten TahirajA lot of these views comes from people think that we've ran out of problems to solve, and even here in Canada, where we have very high life expectancy and a good healthcare system, there's literal people just dying in the ER, and there are two million people in Ontario without a family doctor. Like, they could use more AI for, for medicine. And it also goes that we don't live in farms anymore, so there's no incentive to have kids. And, we're gonna have about three billion seniors in the world, so we're gonna need a lot more robots, and those are gonna need AI and stuff. So we're nowhere to the point where we're running out of problems to solve In the world.

1:19:51Nisten TahirajI find it very interesting that in the year 1485, Sultan Bayezid II banned the printing press, the, the Gutenberg press. And, he had pretty good safety reasons too, because, you know, books can program people and, people can start to, to revolt and change their minds and that puts, the strength of the empire re- the Ottoman Empire down, and they live in a very hostile environment, so that could get a lot of people killed. So for their safety, it made a lot of sense to just ban books. just go full Bernie Sanders and just ban all of them. And, he did that, and it's pretty interesting because at that time,

1:20:30Nisten Tahirajthe Ottoman Empire, they had indoor plumbing and, they had, like, much… Istanbul, it was a much nicer city than most European cities. So after that point, everything started just going downhill, and that's why I find this type of attitude is very… It's actually very dangerous long term. again, they're not talking about solutions as to, you know, how do we stop more violence in the world? How do we make more robots that can, like, fix potholes and repair run-down housing or, u- automate housing? And i- instead you see this media campaign, which is, seems coordinated

1:21:11Nisten Tahirajat, at the same time that this guy was there for six weeks. That's before even your probation period is, is done. and then at the same time, there is another AI safety, thing on Joe Rogan. Mm-hmm. And there's also one on Channel 5 News. Mm-hmm. Which they were, like, very nice people, but at the same time when they… Instead of advocating for how do we actually give people local LLM so they can, fix their, their, their privacy, so they can filter all the junk that's given to them, instead their solution to all of the world's problems and the surveillance state was to just ban data centers. And, that's…

1:21:50Nisten TahirajIt, it, it seems just completely out of the blue as if they're paid to just say that.

1:21:55Alex VolkovBecause Jacob, Jacob Coxon is a nobody on Twitter and suddenly over 100 million views. he's on Time, Variety, Wall Street Journal, NBC and Fox and all the government officials are retweeting him. Paul Cristiano gets added to OpenAI board, which I actually think is a great thing. It's just like the coincidence during that same week is kind of interesting to me. Daniel Cocotallo is on Rogan. Daniel is a whistleblower from OpenAI who refused his paycheck famously and then led OpenAI, to remove the clawback thing they had for equity if somebody leaves and, and doesn't sign their NDA. Daniel, is also the author of, AI 2040 or, or something like that, 2030, which talks about the,

1:22:35Alex Volkovthe potential outcomes of AI. and then OpenAI called Congress to create mandatory national safety regulations, which again, I'm not against. I think, like it's very important to have. I just don't want this to be concentrated in the hands of like a, a few labs. we are here to fight doomerism on the show, folks. I think the AI can do much more. I think that, strong antom- anthropomorphism is a problem. I think when people say the AI, they mean that there's gonna be only one. I don't believe so. I believe there's gonna be multiple AIs. I'm gonna have my personal superintelligence. It helps me. A- and so I think that there's a lot of hand-waving things, and honestly,

1:23:07Alex VolkovI think nobody fucking knows. Nobody fucking knows that Peter Steinberger, a Austrian dude out of nowhere last year, based on his home project, changed the world trajectory because he ran agents in a very specific loop with very specific tools that now everybody runs. Nobody could have predicted that. Not Eliezer Yudkowsky, not Paul Cristiano. No one could have sat there and said, "Hey, this is the way towards personal superintelligence." Literally no-- So no one knows. Yes, pe-fig- people can say we don't know how to solve alignment, but also nobody knows that may- maybe there's this one weird trick And so, I, I say that, like, it's very hard to predict.

1:23:45Alex VolkovWhat I say is that, you know, the, the, the advancements are happening. We might as well enjoy them, and I really want, to see the best of it. Navio Stocks is one example this week I think is a very interesting, externals, where Navio Stocks is one side. On the other side, hey, all of these doomers are saying AI is killing us all, basically what are we doing? Let's stop, and nobody's stopping. No one has a plan of how to stop DeepSeek. Nobody has a plan of how to stop the folks from Ablation AI that we had last week on the show that are taking open source models, removing all, but all considerable, besides child safety stuff, removing all

1:24:17Alex Volkovrestrictions and putting it out there for everybody else to hack each other.

1:24:21Wolfram RavenwolfDid you hear anything, that anything happened after that came out? Nothing, right? Nothing And I think my P doom for AI is much less than my P doom for humans. If you look at the world where human, humanity has gone without AI, I think we should risk the chance that with AI we will get something better if it's widely distributed

1:24:44Alex VolkovAll right, folks, I think, the, we could fill a whole hours of debate

1:24:49Nisten Tahirajthese are very anti-democratic views and, the healthiest thing you can have is an ecosystem. so a monoculture is bad in any ecosystem, whether silicon or, or organic based. And, still encryption works. If you want it to work, it does work. And that's the main tool we have. If we distribute it and we're able to have these assistants that are loyal to us and can write their own encryption and provide people privacy, they can do that. But that's not going to happen if you just pull it… It's a very autocratic view that this, "Oh, this is so dangerous. Only I can have the printing press."

1:25:27Nisten TahirajLike, you're just being another 15th century sultan now, and, that can go pretty bad.

1:25:34Alex VolkovAll right, folks, I think, we've covered this thing enough. there's plenty of stuff to cover in, in the big world, big labs, but, we will keep monitoring. We'll let you know if anything major happens there. I think, we'll skip Apple's, event. That has nothing to do with AI. The only few things about the Apple event with AI was, AirPods can now translate you, and Apple Watch will now transcribe with AI and, and y- you have this, like, what did he say button. The Apple Watch continuously transcribes on the device with no, cloud, and will tell you, "Oh, this person said this." I think that's super cool. also you'd be able to, like, record everything, like all your

1:26:09Alex Volkovmeetings with Apple Watch, on device, which is super cool. And obviously Siri AI that's coming. I've been using Siri AI. It's, it's pretty good, but it's nowhere near the, the level of agentic user-facing AIs that we, are going to talk about next. So now let's move on to our, corner of agentic AI, with, Do we have? Yeah, do we have… All right. this week, Meta came to us and said, "Hey, do you know that thing, OpenClaw? Do you know that thing, Hermes? Do you know Grok Bot, Instinct, Town? All…" There's tons of them. here's our attempt at this. Meta introduces Muse, a free 24/7 proactive AI assistant agent that has

1:26:53Alex Volkovits own browser, its own computer, and is connected to your systems at Meta. Now, yes, folks, this is the same Meta that, was Facebook before, that people are very much worried about taking their data and training on it, et cetera. Th- this is, like, the same Meta. However, I feel like that was a meme a long time ago and, Mark Zuckerberg is not to be disrespected. Because after the Llama 3 and open sourcing everything, he literally just, like, took his money cannon and, and directed this money cannon towards this problem, and Meta has been just incredible at products. Meta Muse Spark 1.3

1:27:33Alex Volkovfrom last week is the second winner from last week. Last week, Astra was announced, and Meta Muse Spark is one of the top, you know, somehow. nobody still counts them. Like, we still don't see charts from DeepSeek and Anthropic and OpenAI that are adding Meta Muse Spark, but they should because it's coming. And Watermelon, which is their much bigger, size model, is coming. But also products coming, and nobody does distributions like Meta. Like, you know, we, we talked about Google. Google's the fact that agentic AI completely, completely fizzled. Gemini Spark is nowhere nearly usable at all. Meta knocked it out the fucking park with Muse.

1:28:11Alex VolkovI've been using Muse only for one day, and I can already tell you this is going to change many people's lives given where it's at already and given how good it is already. And yes, again, I'm talking about Facebook. It's ridiculous. But I'm not talking about just all Facebook. I'm talking about the MSL labs in Meta that paid billions of dollars to people to move forward. the MSL labs that took Nat Friedman and Daniel Gross together, two of the more prolific investors in Silicon Valley. Nat Friedman used to, own, GitHub, gr- like, CEO of GitHub. I'm talking about Meta who bought Manus and then had to say, sell Manus back to China because of anti-regulation, stuff in China.

1:28:52Alex VolkovMeta built Muse. So let's take a look at Muse. Anybody try Muse already besides me? Anybody in the comments tried Muse? A little bit. A little bit All right. this is Muse, folks. and I will show you Muse. Let's see if I can pull up Muse, correctly. My Muse, by the way, obviously is called Wilfred. As you might imagine, it's been Wilfred since a while so,'cause I imported it. let's take a look at Muse. So here, this is what we are seeing. We're seeing this, interface, and, specifically here, as you guys remember from previous ThursdAIs, we're seeing the producer, the, the Muse agent that listens to our show and knows what we're talking about.

1:29:33Alex VolkovSo in a second, I will ask Muse to kinda introduce itself, and reply to me, and hopefully Muse will listen. But, this is the interface. You can animate your own character. They added animations. You can just tell it, "Hey, I want you to look like this." So I said, bionic wolf, and they, have animations. Then you guys see the connected thing? This changes. So, look up, my co-hosts X, Twitter accounts and show them to me. Once I send the query, you can see that, like, it's working. It's not showing you what, CRL commands or bash commands it runs. It just says working. Searching the web. Searching X accounts.

1:30:11Alex VolkovIt-- This is very, like, approachable. Many folks who I installed OpenClaw and Hermes, et cetera, they need exactly this. It's also stupid fast. Do you guys see how quickly it pulled up all of your, handles? and it got all of them right. where's LDJ? Where is LDJ? Why didn't you pull up LDJ? let's find LDJ. But here is the number of actions it took to get this. You guys see this? Read, read, read, read, read, read. it, it manages memory in people, so it knows who you are. Once you talk… you're right. LDJ is co-host list. LDJ confirmed. saved him too. The speed of Meta Muse Spark with this is the first thing that you get.

1:30:49Alex VolkovNow, I will say about speed, I thought about it this morning. Grok was also very fast in the beginning. Now Grok is a little bit slower. Instinct, another AI system, was fast, now it's a little bit slower. speed can vary with the amount of people that use the product, so, like, judging the speed only in the start is not that great. However, it is-- it does feel very fast. It does feel very approachable, and it has, things from, You guys will like this. It has things from, from OpenClaw. It has a soul MD file. There's an MD file that talks about its soul generally helpful, not performantly helpful. actual file that they look in. They have a memory file, which I won't show you because it

1:31:27Alex Volkovincludes all my memories, because I exported it from all my previous assistants and imported it here. And, a- and the computer use is really good. Like, it does the clicks and the computer use stuff very, very, very well. So setting this assistant up to be our producer for the show this week, is… was very easy. And now it says, "New Chiron is up. MetaMuse personal agent on Linux VM, native Stripe link kicker." So it puts up kinda the, the stuff that we talked about on the show, it puts up live on the, on the page. what else can I tell you about this? It's really nice in terms of connectors as well. One of the things that I've told you previously on the

1:32:05Alex Volkovshow, that I use Stripe link. Stripe link has an agent thing that allows you to buy things. Basically, give your agent a credit card safely so that you… n- not your actual credit card. Meta has a Stripe link integration, Stripe has, an integration here, and y- yeah, this doesn't show it because, like, they take it away, but basically every time you wanna spend some mon- you want your agent to spend some money, you tell it to go and buy, and it will show you a card. "Will you approve this purchase?" they take your credit card, and they charge only, like, a credit card that you don't own, basically, making it safe for your agent to p- purchase this for you.

1:32:40Alex VolkovI've used Stripe link on multiple agents. This one is the first that was, like, built natively, integrated natively, and, is feel safe.

1:32:48Wolfram Ravenwolfthis is free?

1:32:49Alex VolkovYes.

1:32:49Wolfram RavenwolfMeta provide this for free. Yes. What are the limits? What are the, the daily limits or stuff?

1:32:53Alex Volkov100 million tokens per week- Uh-huh … for free. as- which is, which is pretty nuts if you think about this, because, Groq is not free. You have to subscribe to Groq, like, pro tier, et cetera. Zack has a money cannon. Now, a very important- And when you run

1:33:07Wolfram Ravenwolfout, what happens if you run out?

1:33:09Alex VolkovYou can pay. You can pay, yeah. You can upgrade. here's the thing. You go to Data Controls, and then you uncheck this Help Improve Our AI Models checkbox. Make sure to do that so that you don't inadvertently help them, train their models. And I wanna talk about security next, but, questions and comments from the folks in the

1:33:27Wolfram Ravenwolfconversation. I have one more question. So it's 100 million token per week. Okay, sounds like, like much, but I just checked. I have more than that per day.

1:33:37Alex VolkovLook,

1:33:38Yam Pelegfir- first, I just wanna point out the native connection to WhatsApp-

1:33:42Alex VolkovYes … Yam Peleg: which, which all the other agents, doesn't matter, open source, closed source- Yeah … Yam Peleg: you always get a connection. Can just say it like that because, it's not easy to get. W- WhatsApp is not Telegram. Yeah. but here, yeah, you get it from the source, like absolutely native. the thing that I love about this, Yam- I just wanna… Mm … is that you can talk about, WhatsApp, and then you have this like channel with WhatsApp that's read only, so view only on the desktop. So you can see your chat with WhatsApp. Like, you can, you can actually see it on the desktop so that you know what you talked about. That's incredible.

1:34:14Yam PelegDoes it have access to my other chats? Like, can I ask questions about-

1:34:18Alex VolkovYeah, yeah … I

1:34:19Yam Pelegdon't know, see what I asked

1:34:21Alex Volkovfolks are asking, Tony's asking in the comments, "How many agents does it have access to on Muse?" so this is a very interesting thing. Unlike Grok Bot, where Grok Bot, like we told you, is a series of AI agents with their own definition to talk to each other, there is one main chat here in Muse. That's a very good decision for many folks 'cause they're not ready for multi- multi-agents, main chat. And then there is, side chats. You can open as many side chats as you want, and any one of them can be like a specific agents. All of them have, very interesting, heartbeat thing. So there is a heartbeat. th- there is a heartbeat, and there is a Chiron watch every minute.

1:34:55Alex VolkovSo this is the, the thing that makes sure that it listens to our show and puts stuff up on the, Gyron, Chiron, I don't know how to say this. you can define this, here. You can say, "Hey, I want heartbeat every 15 seconds," 15 minutes. But I do think that at some point, though, those agents will differ in the UI because the underlying layer will be pretty much the same. They do the same stuff. They read your email, they read your documents, et cetera. The UI is gonna matter. The stuff that I noticed is the productivity. Let me scroll up a little bit. where is this? Where

1:35:22Yam Pelegis this? I, I, I just wanna say, because it's- This okay, continue, continue.

1:35:25Alex VolkovI was just browsing, and I saw, "I can plan Emma's birthday party for this weekend." My daughter, Emma, is turning eight this weekend. Out of nowhere, and obviously, I inputted my memory, so it knows the dates, et cetera. it also knows all your contacts if you want to, which is also a novel thing. You can upload your contacts connector via your iPhone, so like it has native iPhone connectors. You can connect Apple Health. You can connect Reminders. This is all novel, by the way. I don't know of any other agent right now that you can natively connect via your iPhone connectors to the Apple ecosystem. That's great. So it knows my dates.

1:35:59Alex VolkovIt proactively said, "Hey, I can help you plan Emma's party for this weekend." I told you that 2026 is the year of proactive agents, and this proactivity just blew my mind. That's what I want. I want my agent to know things about me and then suggest things to me. That's what I want. Grok Bot, I need to set this up. Instinct is pretty good at this, but mostly they do-- they go off ba- reading the email. This seems like it goes further Ream, comments about this before we go to the Meta- Yeah, yeah … Facebook data, thing?

1:36:28Yam PelegLook, I think the main question is Which one supports which connections? Because you don't have full control. Now, it's not open source that you can just do whatever you want, connect it to whatever you want. That's, like, the power of OpenClaw was because it's a completely, open and free thing that you can just customize and connect to whatever you want. Now, I think the c- the, the most important question is, okay, what connection do I get from Groq, and what connections do I get here? And I don't know, I'll, I'll decide. I mean, what you just explained is, is really, sounds really good, but-

1:37:02Alex VolkovYeah

1:37:03Yam Pelegdo I have this elsewhere or no? I, I mean, what, what's the actual difference?

1:37:07Alex Volkovdifference on, on features, we cannot go into all differences right now, but, like you said, native connectors, I think is a differentiator. WhatsApp, Facebook Marketplace is a big one. Tons of people buy shit on Facebook Marketplace. if you use other AI agents, Facebook can block you because that they don't want other agents. Facebook has a competitive advantage in, you cannot put them on WhatsApp, for example. It's not as easy. so th- there's that. The iPhone connectors is unique. I haven't seen this in any other places. the way they do this is, very interesting. So here you can connect to Apple Health, your contacts, for example, so it would know all your contacts.

1:37:36Alex VolkovBut I think that the, the way they're building this is very important, and Zack, went on Alex Heath's, e- podcast and talked about specifically the safety stuff, okay? least privilege, thoughtfulness, et cetera. They have a built-in sentinel, I believe that it's called. this is the architecture, by the way. They have another LLM running there, reading all the network requests coming out and in, making sure that they're not, LLM, injections. They have a injection security. If you can prove that you got, prompt injected via an email, you'll get, like, $130,000. They have very mi- … a lot, a lot of money there for incentives of whether or not this, you know, this can be fucked with, or somebody can

1:38:16Alex Volkovsend you an email and, and screw you.

1:38:19Wolfram RavenwolfMeta had also the safeguard models and so on, so it's probably using those. And the funny thing or ironic thing is that Peter Steinberger was offered to join OpenAI or join Meta- Yeah … and he talked to, Zuckerberg. So Zuckerberg seems to be a great fan of OpenClaw because this looks exactly like with the soul MD, the memory MD, with the heartbeat. It feels exactly like if you took OpenClaw, put it in a, in a hosted environment and improved the UI. I think,

1:38:44Alex VolkovI think the level to which Peter Steinberger changed the world is un- Described. It's really strongly resembles multiple of the best things about OpenClaw. It also removes all of the horrible things about OpenClaw. Speaker 4: You don't need a Mac Mini at all. They have their own computer. Here, I wa-- the last thing that I wanna talk about, not the last thing, but the, the most important thing I wanna talk about here is the secure stuff. So we can talk about browsing, et cetera, how they save your, keys. They don't save them in plain text. the LLM does not see your keys. It's in secure storage. Muse Confidential VM, I think, is the most important thing

1:39:15Alex Volkovthat we need to talk about here. they announced Muse Confidential VM. They say, "We believe that people should be able to choose a fully private mode for personal agents, where even Meta or any other service provider cannot see or grant access to your information." This is why we're investing significantly in Muse Confidential VM and plan to deliver this capability later this year. Muse Confidential VM is intended to be cryptographically and verifiably prevent Meta from accessing data in your VM. already using the system with a small group of trusted testers, and we've begun making our design and the source code for the system available to external auditors.

1:39:49Alex VolkovWe're in the process of taking auditor feedback. Once launched, we have a continuous audit of the system that will be visible and inspected by anyone. Experts will be able to confirm that Meta does not have the ability to access data within the VM environment. We also welcome security privacy expert to reach out to us about early access." Zuck, talked about on the podcast is Moxie Marlinspike, the founder of Signal, which is notoriously known as one of the most, secure messaging platforms out there, was recruited by them. Nat Friedman recruited Moxie Sp- Marlinspike to work on the secure VM. This, to me, kills all of the stuff about, "Hey, Meta," blah, blah, blah,

1:40:23Alex Volkov"take your data," just completely… Th- this is more secure than my Mac Mini. Because Mac Mini is not verifiably secure, and it's not being continuously tested. I think that this, to me, puts to rest about data. They know the meme. They understand the people know that, Facebook has access to everything, and, they take this very, very, very seriously, more seriously than anybody that I've seen before. Zuck is not by mistake after 18 years still at the helm of one of the biggest companies in the world. And the super intelligence efforts that, Meta is going towards, Zuck is very, like, open about them. Personal super intelligence should be personal and beholden to you, and, Muse

1:41:03Alex Volkovseems to be, like, the first step there.

1:41:04Wolfram RavenwolfI'm super excited. You know how deeply invested I'm in the Hermes eco- ecosystem with all the patches and so on, but I will share this, architecture diagram with my agent and see if it can improve our own setup that way. Yeah. And I think the, the confidential computing aspect, that is super important and a good precedent that hopefully others will follow because we are sharing so much personal… The most personal data we are sharing with our own agents. Everything is in there. Everything is in there. Health data, personal problems, everything. And this is, like a, like an attorney, like a psychologist, that kind of information, it should get more protection.

1:41:33Wolfram RavenwolfThat is something rarely discussed, but having ways in that direction, moves in that direction, I fully support that.

1:41:39Alex VolkovYeah. So folks, give, M-Muse a try at muse.ai and tell us, what you think. as a reminder before, we're almost at the end of the show. Token Juice, if you want DeepSeek for free, but you're okay with sharing your data, Token Juice, from Aaron Batie, a friend of the pod, will get you that as well. if you wanna join us, September and see Pitbull, Mr. 305, Mr. Worldwide, please, please join. This is gonna be super cool. I wish it was in Miami also, not in San Francisco, because this would fit. And, I think we covered pretty much everything. with that, I think we've covered, like, a very decent chunk for this week. Wolfram Ravenwolf, Nisten Tahiri, LDJ confirmed.

1:42:14Alex VolkovLDJ's last name is confirmed. and, and Yam Peleg, thank you so much. We also had Chris Alexius here from NVIDIA talk about DeepSeek. if you missed any part of the show, the full replay is going to be available on ThursdAI.live, and, please check it out, vibe coded with Fable myself. And, if, if you want the links to sources like the Nisten simulator of DeepSeek layers, or Nisten, please add the Astra Mars, simulation as well in the link. we will send them to you- Oh … on ThursdAI.news newsletter and podcast, which you can also subscribe to on ThursdAI.news or ThursdAI.live. We will be here next week.

1:42:53Alex VolkovThank you so much for joining, everybody. Have a good one. Bye-bye.

1:42:57Nisten TahirajBye, everybody