📅 ThursdAI - Apr 2 - Gemma 4 is the new LLama, Claude Code Leak, OpenAI raises $122B & more AI news

Welcome everyone, welcome to ThursdAI. My name is Alex Volkov, I'm an AI Evangelist with Weights and Biases from CoreWeave and we have a very special show for you today. Today is April 2nd, ThursdAI is the weekly news, the highest density weekly news show going live on X and YouTube. Together with me I have co-hosts and we're gonna have a little bit of a different format today. Most co-hosts are on the tourist space, I'm gonna say, I'm gonna wave and then ask them to unmute there and then we'll talk about it. Test out how this is gonna work. I'm gonna add them to the show here as well, but mostly we're gonna unmute on X and talk on X. We'll start with Wolfram. Wolfram, can you say hi to the folks? Wolfram Ravenwolf, CEO, Wolfram Ravenwolf, CEO, U.S. Hello everyone. Back to the roots, right? I think there's a lot of new folks here. A lot of new folks are listening on the Twitter space and so the show started with just a bunch of us talking on Twitter spaces discussing this week's news. Definitely back to the roots because this is how we started. We started with the Twitter space and recently we started with what we started doing after LDJ, you were on stage also. Say hi to the folks. Hello everybody. We'll cover some amazing news today. We're gonna have a lot to talk about, definitely. A lot has happened in the world of AI for the past week since last we covered it. I think the main thing we're gonna cover very, very soon. We asked the folks here in the comments to reply to us and say, what is the one thing in the AI that we must cover? Everybody says cloud code. Because folks, this week, as space uncle Elon Musk put it, Anthropic is now more open than open AI. I mean, we laugh, but it's an awful thing. So let me just cover this cloud code source code was, it wasn't leaked necessarily. It was mistakenly released by a process internal process, according to Boris Cherny from cloud code. And then somebody rebuilt the whole source code from cloud code, maybe the fastest 1 billion side hustle harness thing in history of AI. And everybody was like, this is the only thing everybody could talk about for a while. There's a ton of news this week, just a lot of news, including news from just today. So we'd love to hear from you guys. What was the like the most important thing, but I think it's good for us to start with the TLDR. Alrighty, folks, here is the TLDR, the section of the show where we tell you about everything that happened this week. Today is Thursday, April. Can you believe it's April? Can you believe we're past we're in Q2 of 2026 and a lot of stuff is happening. So this is the TLDR. This is the section where we cover everything that happened in the world of AI for the past week. Your host today is Alex Volkov from Weights and Biases CoreWeave. I represent like a shout out, huge shout out to Weights and Biases CoreWeave for letting us do the show. Shout out Weights and Biases. With me, co-hosts Wolfram Ravenwolf, Jan Pellig, Jay and Ryan Carson. Our usual co-host Niston is somewhere not to be found. Hopefully he'll join later. The guests today, we're going to have Sigrid Jinn and Belma, the creators of Clockode, the fastest GitHub repo to reach 100,000 stars, faster than OpenClaw. What's up? Let's go. And we're also going to have a surprise guest. I mean, I can say who he is. We're going to have Omar Sanseviro, our friend from Developer Experience and Google DeepMind to come with breaking news. So please stay with us. Omar Sanseviro is going to be here. And let's run through this week's huge, huge thing. So first of all, obviously, a cloud code, entire 512,000 line source code accidentally leaked via NPM. We're going to have to cover this. Yeah, you seem like I have a comment. Please unmute Nexus's face and tell me. Let's fucking go. Wow. I love it. I think the thing that still afflicts all of us is that cloud code session gate. We're going to name it session gate. Okay. The fact where you pay $200 for the 20X cloud max subscription, and then you send three prompts, and then you're out of the weekly quota. Some folks have potentially found bugs in cloud code that lead to this happen. I don't know if those bugs are the cause for this, but this was finally confirmed this week. We told you about this last week. It still happens. This week it was confirmed by the folks from Entropic that they're sorry about this. So we're going to talk about the session gate for sure. We also must absolutely say let's fucking go to the massive, massive, massive round of funding that OpenAI has just announced this week, folks, $122 billion. Not, no million, $122 billion fucking dollars, the largest funding round in the history of VC. What the fuck is Sam Altman cooking? $122 billion, or I think it's like $852 billion valuation. And they're going to go IPO this year and be able to buy their shares. It's quite insane. So these are the news from the big companies. We also have Microsoft dropping breaking news on us just before we started the show. Three in-house models for Microsoft AI, best in-world transcription, expressive voice, and top three image generation, all from Mustafa Suleiman and the folks at Microsoft AI. So we're going to definitely chat about those. Now we're going to have to go to open source because a lot of stuff has happened in open source as well. We have folks from Prism ML emerges with stealth with one-bit bonsai models and 8 billion parameter model in just one gigabyte, 10x intelligence density. Nistan wanted to talk about this, so we're definitely going to chat with Nistan about this as well. Liquid AI released their LFM 2.5, 350 million parameters, a tiny, tiny model under 500 million parameter quantize. So both models can probably run on an iPhone. And then Quen dropped two models. Quen 3.5 Omni is a native Omni model handling text, image, and audio. Anything else in open source, folks, that I haven't added here on the notes? I think you got it. In Tools and Argentic Engineering, we're going to cover Cloud Code. Cloud Code is a clean room rewrite of Cloud Code that is the fastest in the history of the world to pass 100,000 stars with the authors Sigrid Jinn and Belma, who is going to join us soon, hopefully. Sigrid is Bellman online? I think Bellman is not here. Maybe he's sleeping because he is now in Korea. In Vision and Video, we have Google launching VO 3.1 Lite. It's the cheapest video generation at only 5 cents per second. In Voice and Audio, we have Fish Audio launch speech-to-text. But also, we have the Microsoft MAI thing that supposedly beats Whisper and the top-of-the-line transcription that just dropped. In AI Art and Diffusion, we have Alibaba's One 2.7. Folks, there was rumors about Alibaba not doing a lot of open source once our friend Juni Yang left. But, hey, Alibaba's all over the news. Like, two current releases and now a One 2.7 as well. So, that's great. And in this week's buzz, also related somehow to Sigrid, who's also here. This week's buzz in Weights and Biases in San Francisco, we hosted the weirdest hackathon in the world where folks are getting punished for touching their computer. I will say this again. You go to a hackathon, you do some stuff, and if you touch your computer, you are getting punished for that. This is like the weirdest hackathon. Sigrid led the final round of judging. We hosted this. I recorded the video. We have to talk about this because it was very, very cool. Folks, I think that this is the TLDR. There's a bunch of news. Oh, and then also we're going to have breaking news from a very big foundational company, breaking open source news. Some of you know what this is. Some of you are rumored to know what this is. But we're going to talk about this very soon. We have to start talking with ClockCode. Looks like our friend Sigrid has a little bit of a deadline. So I think this is not necessarily open source AI, but you know, we're going to call this open source AI. Let's go. Open source AI. Let's get it started. We finally can talk about what we need to talk about. Folks, someone in Anthropic, an unnamed employee, released an NPM release. That release had a source map file, .map, that included the source of all of the 512,000 lines of ClockCode. ClockCode being the number one side project that anybody in any company ever released. I think it's a two plus billion dollar now hustle, if not more. And obviously the internet exploded. Absolutely exploded. Somebody found it, uploaded online. Somebody else cloned it and put it on GitHub. And many folks just went through the source code, finding different things. No longer reverse engineering is required. I specifically want to talk about the Kairos Autonomous Diamond, the undercover mode, and the full agent and infrastructure there. So, a researcher named Xiaofan Xu, I think he goes by the nickname DeepFry or FriedRice, spotted like 4 a.m. posted download link. Within six hours, it had 3 million views and almost 40,000 GitHub forks. Anthropic started filing takedowns, but obviously nothing could be hidden from the internet. And if anything, there's the Streisand effect. So, the source code is everywhere. I'm pretty sure that people have this on their computer. It's very interesting because some folks started porting this or getting inspired by this into Python and Rust. And we have those folks here. Sigrid, we'd love for you to tell us a little bit about the effort, if you can. Okay. Okay. So, I was in an airport because I had my hosting, Ravton with Allexa. Thanks for booking the space of Ravton BIOS. And then we went super viral. And I mean, it was featuring Numa. And I was in a flight back to Vancouver because I have my space in Vancouver. To Vancouver, right? Can you just pick up just a little bit? It's kind of hard to hear you. Get closer to the mic. Okay. Can you hear me? Yes. Yeah. Yeah. Cool. Yeah. So, in the internet, I didn't find Wi-Fi. I just got a message from South Korea. I was in Australia and South Korea was around those. Let me pause you because folks are writing that we cannot understand you. And it's unfortunate because I really wanted to hear your story. But basically, you just said that you were in the middle of a flight. Yeah. And you saw the tweet and then you downloaded the code and you started looking into it to ask questions. Right? Is that correct? Is that what I'm hearing? Yeah. So, basically, what I did is to make sure to ask all my code, which has like team mode and rough mode, which makes up to go on multiple code sessions simultaneously. And that all my code has helped me to report every line of the code into Python. So, I want to pause you just so folks who would like hear 100% what happened here. Again, Fried Rice on March 31st. He says, cloud code source has been leaked via map file, via the NPM. We're showing this live on stage. Here's all the files. Right now, that tweet has 34 million views. You're the second reply to it and says, I backed up the source up on my GitHub with cloud code. And you're saying you used all my codecs, right? All my codecs to port this into Python. I think that the thing that's really funny to me, Sigrid, would love to hear from you, is that the kind of the legality. You said afterwards, you woke up and you said your girlfriend said like legal action scared you a little bit. So, you rewrote a feature port to Python. And the very interesting thing here is, can Anthropic have any claim? We don't know, right? But like you rewrote this into Python, they can have claim to the original source code because that was leaked and they're supposedly that's their property. But Python, you just rewrote it supposedly in the clean room, right? So, like they potentially have no claim here. All right, Sigrid, I do want to ask you about this thing. So, because of the like millions of people that saw your kind of reaction and then millions of people said like rewrite the port as well. You got Bellman to join and then you guys started putting this into Rust. And I think like that's very interesting to many people. And now this repo is the highest, like the farthest, like most explosive repo in the world with 144,000 stars. But not only that, I think the number that's bigger here is that it has 101,000 forks, which is now impossible to do because I think the repository is locked. Could you talk us to, can you tell us what the repository is about now? It used to be the backup of Cloud Code. Now what it is now? And then we'll talk about the source code afterwards. Okay. So, basically, I just want to give a shout out to the Harness built by my friends, Belmont, like they're building all my open code. All right. Well, Sigrid, thank you for joining us. It's really like, I'm sorry, I'm getting some, like a bunch of comments for folks that like it's really hard to hear you. So, I'm going to try to recap what you guys said. The fastest growing repo in the history of GitHub. It started with the backup of Cloud Code, source code, but then quickly you guys rewrote this into Python and then into Rust. And this is all managed by those tools that you mentioned, all my codecs and all my open cloud from the Korean folks. It's all run with Ralph Loops, which is, we covered on the show here with Ryan Carson and some other folks created by Jeff Huntley. So, Ralph Loops, like it's all autonomous. And so, hopefully, we'll see some incredible harnesses coming out from this repo. But I think many, many folks will actually want to talk about the actual incident. So, Sigrid, stay with us, please. Yeah, I want to hear from you. You've been following this. So, tell us about how you saw the incident. What happened when Cloud Code, source code got leaked? Everyone on the internet wanted to just de-obfuscate the Cloud Code, source codes. I just want to say that, let's just say that I've gone through the code. I've taken some time to go through the code. It's not as interesting as people would imagine. It's just a good harness, details done right. All the small details dialed in correctly. Most of the tricks about context engineering are already in the Entropic blog. You can read about them. They do like system reminders and so on. They're famous tricks already that you can read about. But I just want to say that anyone that looked for the magic, the magic was, is, and always will be in the models themselves. It's just that Cloud is a good model. If you run it in a loop, it becomes magical. So, whatever harness that you use, yeah, details matter. But it's mostly Cloud itself, the magic. But yeah, I think in my humble opinion, like just Entropic, please just pick your models. Like don't go that hard and fight over such a thin rougher in JavaScript. Just do as OpenAI does. Really sit open source. There is no point to fight on the harness because there are no secrets there and you're openly saying it. But I will say there's more stuff that we found out specifically about hidden features. And we absolutely must talk about some of the hidden features, folks, because some of them are a testament of what's coming up in Cloud Code. So, the very sad thing that was leaked to me was their April Fool's joke. Folk yesterday was April Fool's. And they had the buddy system. So, now right now, if you upload Cloud Code and do slash buddy, you'll have like a Pokemon type character with like rarities. It's pretty cool. But I specifically want to talk to you guys about Kairos. Do you guys see the Kairos? It's a built-in automated and gated daemon mode that runs in the background with heartbeat prompts. So, right inside Cloud Code, there's a system called Kairos that has exclusive tools like push notifications and file delivery and subscribe to GitHub subscriptions. It has append-only activity logs and runs nightly on AutoDream memory consolidation. So, basically, after releasing, leaking the code, folks looked into the hidden features there and discovered that Cloud Code is very, very soon to be copying some more features from OpenCloud. And I think that's wonderful. I think that's incredible. I think that this is the testament for open source, not the leak, but the fact that like OpenClaw and Peter Steinberg, like all those folks are fully open source and then everybody can adapt from there. Would love to hear from you guys about some other things that we found actual, that people did analyze the leak and actually found in there. So, feel free to kind of jump in there, raise your hand. Nistan, it looks like you have a go. Yeah, there was something huge that I think it was Westboss or one of the TypeScript developers. They found out that it was injecting the Cloud prompt, the Cloud.MD over and over. At every turn, right? Yeah, but that's not actually what happens. It doesn't do that because my Cloud.MD is 50,000 tokens long and I do a lot more than four turns. So, that's not actually what happens. But the other thing they found, it was injecting fake tool calls to kind of ruin the datasets that people are generating off of it. Oh, that's true. Yeah. Yeah, they have like anti-distillation features, right? Something like this. Yeah. So, if you are someone that keeps all of your .cloud folder logs backed up, which you should. Now, don't just use that data directly, the JSON lines. Just pass it through like Quent or something to generate it 120,000 tokens at a time. And try to remove some of the fake tool calls and try to remove some of the noise. So, yeah, now you're going to have to both take a Tmux session and take your Cloud code Tmux is there and then try to figure out what the heck got injected that was not a tool call. So, yeah, that's going to be annoying, but it is what it is. Ryan, we'd love to hear from you what you saw folks post about the source code of Cloud Code. Okay. You know, I actually want to back up and say I think this reveals something pretty important about the way people are using these tools. So, you know, I've worked at one of these agent labs. Shout out to AMP. And AMP is amazing. It's an amazing team. So, I've, you know, seen the inside of these things and I use these tools every day way too much. And there's a whole harness that's much bigger than the agent, right? It's the entire SDLC, right? And I think what I think everyone that's serious about these tools is realizing is that just the agent harness like a Cloud code is not enough. Like you have to have the whole SDLC, right? Define the SDLC to focus. The software development lifecycle, right? So, this is everything from, you know, your issue tracking to the shepherding of PRs to merge. You know, it is, you know, automations that run daily to do things like, you know, bug finding, I checks. This is a whole thing. Like, so you kind of have to back up and I know it's kind of depressing to people that I've sort of thrown my towel in and said I'm just using Devin because, you know, because it, and I don't care if you use Devin or not or, you know, or some other thing. But I think you have to admit it's, you need the whole SDLC. Like, otherwise you're screwing around with, you know, tweaking your agent harness instead of doing real work. We need, we need to get real work done here. And the scene that the Cloud code source code does not fix any of that. Like, you know, so I'm just going to be the grumpy guy who's in the room who's saying we just need to ship stuff. Ship stuff with high quality is something that difficult, more and more difficult for Entropic lately. We also covered the session gate. We're going to talk about session gate very, very soon. They acknowledge that this is a bug. Some people before the source code leaked actually went on Reddit and de-obfuscated some of the code, backpropagated some of the code and noticed some caching bugs as well. Since you asked about what stuff we saw online, I saw a lot of some controversy about the hidden feature where it says if you're an Entropic, an Entropic employee, it will hide all the traces that you are using Cloud code. So basically, by default, you get this attribution when you make a PR. If, when you upload code to submit it, then it says that the co-author is Cloud code. You could always turn that off, but even for the Entropic people, they have even more leak protection. I don't find it controversial because they are working with the internal stuff and they are basically, they expanded their system prompt to prevent model names to leak or stuff like that. So personally, I didn't find it controversial, but I know some people found a lot of stuff by looking at the source code that we knew about. Like I saw an article on the register about that it has so much access to your system. Yeah, sure. It is a computer controlling agent. Of course, it has access to your system if you enable the features. So basically, a lot of people like just waking up what this thing can do. And we are running this closed source software locally. There has been supply chain attacks and stuff. Something like that could happen in a closed source project and you wouldn't even notice because nobody could find it out until it's delayed. So there's a lot speaking for having it open source. And now that it has been leaked, maybe they will reconsider. Like it's out, the cap is out of the bag. And it would be cool if they just follow the leads that Google and OpenAI have been doing. And all the other open source projects that are already open. Because I don't feel so well if I run closed source software like that on my system. I mean, we all do. But still, open source gives us more ways, more transparency, which is very important with AI. Yeah. Let's do Yam and then Ryan and then folks will have to move on because a lot to cover and we've barely started. Yeah. Yeah. Something interesting that I saw in the code is that there is an explicit different system prompt and other settings if you're working in Anthropic. And look, we can customize the system prompt, but I don't know. That feels kind of wrong in my opinion. To have, I don't know, the official code has an, if you're an Anthropic employee, you get this. Otherwise, you get that. It doesn't matter what the thing is. Yam, you're shocked that a for-profit private company is optimizing for themselves? Like, what do you mean? Look, the only thing I'm shocked about is the lack of transparency at the end. And you can change the rate limits of my subscription. No problem. Just tell me, okay? I mean, that's okay if you tell me. But if you're not saying anything and all of a sudden my entire quota is burning in 10 minutes. Yeah. And you know, I'm paying for the quota. It's like... That's the next topic. Let us finish. Yam, I really want to do talk about SessionGate, but let us finish. Folks, we have to move on. Cloud code, source code was leaked. It was a big thing. It was a huge thing. Honestly, yours truly, the top Twitter post that I ever had was describing this thing, which Space Uncle reposted. Many of you are here because of that. So shout out. It's like a thing. My friends texted me like, hey, Alex, congrats. I'm like, what for? Alex, I'm proud of you, buddy. That was impressive. I just like covered the news. But we do cover the news here on a weekly basis. Hopefully some of you who joined us because of that is now here in the space and the live. And you're enjoying what we bring. And what we bring is news and coverage. And in that vein, we have to move on, folks. We've been talking about this for almost like 40 minutes on the one topic. Anthropic is in the news still because of SessionGate. What is SessionGate? If you have a pro or max plan on Anthropic, they have two max plans, one for $100, one for $200 with 20x, supposed 20x session. You get way more API calls than you would get if you pay for API, supposedly. Recently, I think two weeks ago, we first noticed this problem. Last week, we talked to you about this. This week, Anthropic finally acknowledged that this is an actual issue that people are having. People are getting their quotas, even for the max $200 month plan, finished. Their sessions are done, the five-hour sessions and the weekly sessions, as early as Monday. The session resets on Sunday. And on Monday, they cannot no longer work with the cloud code. The thing that triggers the most people there is the fact that Anthropic is shipping features like crazy. And yet, they're unable to acknowledge this. They finally did acknowledge it. So shout out to Tariq. Shout out to, I think, Linda from the BUN team that acknowledged this thing. But I think what happened is Anthropic don't actually... We talked about this last week. They don't actually know what's going on in the infra. And they don't really listen to the users. Ryan, you went viral for this saying, like, Anthropic is not that great at DevRel. And we only have, like, Tariq or somebody else. This was awful. This is like, for the two and a half weeks, people are not getting what they paid for. At least what Anthropic has promised them. Then Anthropic came out and said, hey, you will run through your quota faster than before. Which is, like, the most cloudy marketing speech way to say, hey, we've nerfed our next plans. And everybody's complaining about this now. Let's pick up. Yeah. Why can't Anthropic do DevRel? What is happening over there? You guys, like, this is not rocket science. Get your shit together. Everything they post gets reposted to Oblivion. And there's a lot of haters that want to jump in. But still, this session gate is just awful. I had a post about this, the DevRel. I just asked people, hey, are you experiencing the same thing? People said yes. And then somebody on Reddit actually backported kind of cloud code and then noticed that there's caching bugs. So, apparently, if you use dash dash resume in cloud code, I don't think that was fixed, by the way. I think that's still a bug. It's confirmed now. If you use resume feature, what is the resume feature? If you're using cloud code for a while and then you stop and you want to go back to that conversation, that whole conversation is going to get sent without cash to Anthropic, which costs 10 to 20 times more for tokens. This was confirmed by somebody. We tagged Tharik and other folks in there. They said this is not necessarily the problem. It's not 100% the bug that causes cloud code to get out of cached calls, specifically because we know that folks who use OpenClaw, for example, they're not supposed to, but the folks who use OpenClaw with the Max account, they didn't have as many issues like this as well. But now they do. So, while Anthropic suffers from one kind of leak type situation, and hopefully they acknowledge this and move on, and maybe they open source cloud code because why not? We would love to call on Anthropic to say, Hey, folks, people are switching. People are tired of this. People are sending one or two messages and their quota is done. What's going on? So, if you're like this, folks, if you're listening to us anywhere, please comment if this is the experience that you have as well. Maybe this will help move Anthropic a little bit. Yeah, Ryan, go ahead. And then Wolfram, I think you have a comment. I just want to encourage everybody to consider to moving off of these lab harnesses. Like, the truth is, if you are confined to a lab harness, like a cloud code or a codex or a Gemini CLI, you are not going to have the best experience, period. Like, so, you know, pick one that is an agent lab, not a lab lab, because I think this is going to keep happening, right? You're going to, if you're locked into whatever Anthropic is doing, or whatever OpenAI is doing, or whatever, you know, Google is doing, it's bad for you. I think there's a couple agent labs that are, you know, rising above this. They're going to, for instance, like if you're using another agent lab, and there's some weird shit that goes on with caching, and you would have been, you used up all of your quota, they're going to eat it, not you. So, yeah, I'm just pretty passionate about this. I think devs are getting screwed because they're loyal to these labs, and it's ending badly. So, just kind of, sorry, I'm getting angry about it. Yeah, I hear you. And you recently moved to Devin. Also something we can cover in the post-3. Devin is incredible. Shout out. AI Breaking News. Coming at you only on Thursday. All righty, folks, we have breaking news from DeepMind. Google released four different Gemma open source weights models, open weights models. Gemma 4 is four different weights of the open weights Gemma model, and they adopted an Apache 2 license, open source. Let's give them a round of applause. We love open source here, 100%. And I can finally say who our guest of honor is, Omar Sanciviro from Google DeepMind. A very big proponent of open source. He's going to join us, talk about Gemma 4 in just 20 minutes. Folks, please stay with us because Omar is a great dude. Very articulate. We love hearing from Omar all the time. And he's been a friend of the show since the glorious llama days. They released a 31 billion parameter dense, 26 billion parameter active MOE, 8 billion and 5 billion. Smaller models you can run on your laptop right now. It's quite, quite awesome as a show. So shout out to Google DeepMind for carrying the torch of Western open source forward from Google. It's incredible. Just marvelous, folks. I love this. I love every moment of this. We'll cover it with Omar. So there's nothing left to say there. Like we'll ask him, but definitely we'll get back to the rest of the show until Omar joins us. Anything else to say about SessionGate, folks? Anything interesting besides the fact that, hey, this is a bummer? The one thing I wanted to say before you guys jump in is apparently TopCopilot has a great program and you get the Opus 4.6 there and the quotas are not as bad as on the topic. Really quick. If you're on GitHub, make sure to go to your GitHub settings because any old REPL that you've had, and I think Personal wants to, they're going to use it to train models. And they just turn that on randomly and you have one month to turn that off. So yeah, that's it. Let's go. We have so much more stuff for a week. Yeah, go ahead. I just want to say that that's not the first time that it happens. And I just want to say that there is a competitor that has a reset button and they're absolutely not afraid to use it. And the funniest meme that I saw is also from our friend VB, who has been on the stage with us multiple times. He said, oh my God, OpenAiCodex code was also leaked. And then he posted to the fully open source Codex harness that's been open source since the beginning. Not only that, the folks in Codex have reset the quotas multiple times. So we will keep monitoring. We'll let you know. But if you are getting your quotas done, you're not alone. And there is a competitor's, you know, Codex is a great competitor for that. Folks, let's move on. Tons of other news. I think we have to move on to, we're going to cover open source in a bit. Let's talk about OpenAI's 122 billion largest private equity fund in the history of the world. I think, okay, let me just say the tagline. OpenAI has closed the largest private equity round in the history of private equity of venture funding, et cetera. $122 billion at a whopping $852 billion valuation. As part of it, they confirmed they're adding $2 billion per month in revenue with 900 million weekly active users. That's actually kind of surprising because we had 900 million weekly active users as a stat confirmed last December. So it's kind of looks like OpenAI is not growing as much as they used to based on this, but they are definitely growing in Codex. Codex has passed what, 2 million developers recently. They projected to lose $14 billion this year and they're burning $150 million per day. But the $122 billion reportedly only gives them 18 to 24 months of runway, making the IPO that's coming a financial necessity. But $122 billion, it's just absolutely crazy. This is the biggest venture fund in the history, three times larger than the previous largest, which was also OpenAI. I find that really, really funny. This is the largest funding round in history, three times larger than the previous one, which was also OpenAI. And before that, there was Entropic. We should also mention, oh, who's going to join this round? Amazon invested $50 billion, but then they gained $100 billion in expanded AWS commitment. NVIDIA put in $30 billion, which obviously OpenAI buys from NVIDIA. And SoftBank put in $30 billion more, which SoftBank and OpenAI are co-mingled together in the Stargate initiative as well. So like the infinite money glitch keeps on glitching. It's really funny. Folks, any comments on this? Anything interesting? Eldr, I want to hear from you. Another data point amongst all the crazy stuff happening amongst this acceleration, this curve. But yeah, this is the biggest in history. I think their last one and the one before that was also the biggest in history. But it's just consecutive record-breaking. And it does beg the question, Anthropic, like how much is this really kind of sucking the air out of the room that might be available for Anthropic in the future? And what difficulties might this cause for Anthropic fundraising? Or maybe none. Maybe they'll just continue fundraising as usual with no problems. But when they IPO, that's going to be crazy. I could imagine all these little times where NVIDIA has went up and down or the memory companies have went up and down because some new paper comes out from Google or a model gets dropped by DeepSeq. I feel like that will be amplified maybe multiple fold when OpenAI becomes public and people start making crazy sell-offs and buy-ins just off simple news. So the rumors that I saw, and I think we don't deal with rumors here on Thursday, but this was, I think, confirmed with some people. There's secondary markets for these upcoming APOs, secondary markets for these shares. And at some point, somebody said that nobody is able to buy Anthropic shares in secondary markets, but nobody wants to buy the OpenAI ones. That's very interesting. So that's very, very interesting. We don't do any type of hype or financial advice here, but take what you will. Folks, I want to move on to the big companies, Microsoft thing. I want to cover because this was also breaking news from this morning. Microsoft AI led by Mustafa Suleiman finally released something. Finally. Microsoft AI released something of their own. It's not a language model. Microsoft AI released the world's best subscription. Number one on accuracy, MAI Transcribe 1, which is basically a Whisper competitor. Number one on accuracy, lowest word error rate on FLIRS benchmark, 25 languages supported. It beats Whisper Large, GPT Transcribe, Gemini 3.5 Flash, and Scribe two times, two and a half times faster than Azure fast transcription. 36 cents per hour, best price to performance and handles noisy environments, overlapping speech. And it's rolling in copilot voice and teams. I'm hoping this is going to roll out in API as well. MAI Image 2, it's number three on the arena. So they dropped in voice transcription model, a voice model, and an image model. All together. Which is quite insane. These models are not, I think they're only in Microsoft AI. I don't think they're even like available in API yet. So we'll see. So shout out to Microsoft. Number three, MAI Images is number three ranked in the arena. And the voice handle, natural expression, speed generation. I think folks, we've chatted enough about the big ones. And I think it's time for us to talk about Gemma as we have Omar Sancivaro here. Omar, I'm going to add you to the Twitter space. Please accept me there. I'm going to add you here as well to the thing. So we'll see you in StreamYard, but we'll hear you in Twitter space so folks can follow you over there and know who's speaking. I know it's a bit of a juggle, but folks, say hi to Omar Sancivaro, who's just joining us from a, yeah. Omar, if you want to just like hop on the space, but speak here, it's fine. Just make sure you don't have feedback. Is this good? I think it's good. Yeah. So your voice is coming to me and folks in spaces should hear you as well. Okay. Perfect. Alrighty. We already announced the news, but please give us a small announcement. Omar from the Gemma team at Google DeepMind. What did you guys just release? Yeah, cool. So yeah, super excited to be here. We just announced Gemma 4, our first major launch of the year, our largest launch ever. Apache 2 license. Yeah. So pretty much for the last year, you have seen like all of us interacting in social media, Reddit, Local Llama, everywhere asking for feedback. And we have tried to incorporate as much feedback as we can. So new license, new capabilities. We have incorporated system instructions, function calling. And if you have seen the LM Arena scores, these are extremely strong models for the size. Actually, LM Arena just reshared this in X a few minutes ago. But folks, meet Gemma 4. Gemma is a series of open source models from Google DeepMind, Advanced Reasoning and Argentic Workflows, and Apache 2 license. And we're going to go and look at LM Arena scores because I think they're big. Omar, what is... Can you tell us like how Google thinks about open source? Tell us about Gemma and how Gemma is considered within open source. So, I mean, you know, like Google DeepMind's goal is to really empower the community to make AI accessible for humanity's problems. And not everything is through an API-based model, right? So Gemma pretty much enables you to do a bunch of different things. One of the big interesting angles is to create extremely powerful models that can run in your own hardware, even to the size of running like in a phone device, right? So there are two very small models, the E2B and E4B. Those can run in Android phones, in a Pixel phone, in a Samsung high-end phone. But all of these models are models that fit in a consumer GPU, right? So if you have 20 gigabytes, 16 gigabytes of GPU, you should be able to use these models. So that's one aspect, like enabling developers, enabling startups, enabling people to fine-tune the model. There's also a very interesting aspect on the sovereign use case, so when people need to deploy the models in a setup in which the data cannot leave the servers. Yeah, let me take a pause here because this is one of the things we are the most excited about. So these are the two largest sizes that we released, the 31 billion parameter and the 26 billion MOE with 4 billion activated parameters. And what is more exciting is the size of these models, right? All the other models there are a few hundred, even 1 trillion billion parameters. The research team really has done an amazing job packing so much intelligence per parameter here. So let me just like shout it out because folks who are just listening, they're not seeing this. Folks, we're looking at the arena.ai rankings for Gemma 4 variants. Gemma 4 31 billion parameter gets the number three spot among open weights models and number 27 overall. And the Gemma 4 MOE mixture of experts 26 billion parameters with active four gets number six. These models are beating or coming close to 1 trillion parameter models like Kimi K2 thinking and GLM five, which is just incredible. Omar, how big is the 31B in gigabyte sizes? It's like 40 giga. Yeah, it depends on the quantization schema that you choose, right? I mean, you can even go down to 16, 18 billion parameters, 3 gigabytes. I think Omar, the thing I would love to ask you, there's a lot of rules. What's your take? I would love to hear your take. You've been previously chief lama officer at Hug and Face. You've been proponent of open source models for a long, long time. We chatted with you on the stage here multiple times about different models from before. What's your take? Where are local models going? How good they are right now compared to like different things? And when I say local models, I specifically mean the type, the size that you guys released. 31B is the local model. A, like 26B, A4B is the local model. I can run them on my MacBook. I can hopefully run them on Mac Mini for like open claw. I'm not talking about like open weights model as a class because GLM five, nobody can run on their laptop. Come on. So we'd love to hear from you. What, where's this like industry going? How has it been? Just give us a recap of like the open weights local models. The way I say it is, well, from Google's point of view, we are releasing the best open models that can fit in a consumer GPU. And that's quite important for us. Like we want to release extremely good models, but that everyone can run with their hardware. Right. The way I say it is that you have like two different things. So when you really want like raw intelligence, the best models, you will usually go for a proprietary model. Right. There are many use cases to want to run the models yourself for local models. And that's where all of the fine tuning side of things comes. Right. So if you need to adapt models, add new skills, add new capabilities for healthcare setups, finance setups, privacy setups in which the data cannot leave your hardware. That's where I see like the local models. What is exciting though, is that the open models catch up to proprietary models relatively quickly. So if you compare Gemma three or Gemma four, it's matching the proprietary capabilities from a year ago or even eight months ago. So I think that's a trend that is quite exciting in general, being able to run those capabilities directly in the, in the user's hardware. Fox, you want to jump in here with a question? Feel free to jump in here and ask one more question. I can also share a bit more about the launch. My question is how, like, how do you do this? If you can share whatever you can share about the making of. So pretty much these models are built on top of the, all of the groundbreaking research that happens with Gemini. So it has been built on top of all of the research that went into Gemini three. So for all of those things where I just saw a jump in terms of capabilities from Gemini 2.5 Gemini three, you should expect similar improvements. So for example, agentic capabilities, coding, multimodality capabilities, all of those things, these models are much stronger compared to the previous generation. But yeah, what is exciting is being able to put so much knowledge and so many capabilities in a 30 billion parameter checkpoint. We will be sharing more information about some of these things. I would love to see more info as well, like transparency ones into the process. I think like the cool thing that we saw from labs like RC that recently released their like large models as well. And some other labs is that like the whole process of training is great. Obviously, you guys have a very specific approach and like learnings from Gemini, which are like great models as well. Which one of those would you use right now if you had to like run an open claw fully locally, which I think is a great thing for these models. Because if you send your data to an LLM, like let's say Gemini or cloud opus or whatever, like you're still sending data. But you can get to a very good like level of complete agentic automation locally. Maybe not as good as frontier models, but definitely fully locally if you want to with some of these models, especially with like the 31 bit that you released. So talk to me about agentic capabilities. Talk to me, which model would you use if you want to run like an agent fully locally on your computer? Yeah. I mean, again, like these models are very good for coding. Like if you want like a raw intelligence, you would go for the 31B. If you want like a latency, you would go for the MOE, right? But yeah, so just to give you like a couple of examples, even the super small checkpoints are very good for agentic things. So we have a couple of demos. Yeah. So we have an app called the Google AI Edge Gallery that you can go to the Play Store and download. That allows you to run the Gemini checkpoints directly on your Android phone 30, which is a format for local inference. And there's a demo there where you can actually load skills into the model and teach the model how to do different things directly in the phone. So even for the E2B and E4B, which are the two smaller variants, they have very good agentic capabilities. I think folks have questions for you. So Wolfram first and then Nistan. Yeah. So first things, I'm super excited. This is my favorite release of the year. I think you know how long I've been waiting for this because Gemma 3, I still recommend it all the time. It's in German. It's a great model. And in Europe, very, very popular. So one of my favorite models and super happy about the sizes as well. that, yeah, you have the perfect size, not too small where you think, yeah, I can use this on my workstation. And that is great. Do you plan to also do something like the 72B or is the 31B basically the maximum size you envision for these kinds of models? That is my first question. Yeah. We have many things in the open, so we cannot, I cannot share too much, but this is not the last release for now that we have many things in the pipeline. So yeah, keep watching X, follow the new account and you'll hear back from us with new things. Great, very soon. There is a specific Gemma account now on Twitter or on X and then Omar is at Omar. I think please give Omar a follow. Folks who are here on the Twitter will add you to the show notes. Nistan, go ahead. Yeah. I want to ask, can we expect TurboQuant versions of these or Quantization Aware training versions of these models? So one of the things that we do for the Gemma launches is that we collaborate very, very closely with the open source ecosystem, right? So for this launch, actually, we have a huge list of partners. So of course we are working with Lama CPP and BLLM and Hugging Face and even for Transformers CIS or Candle for inference in Rust or Olama. And yeah, we have like 20, 30 different integration partners who are working for Day Zero. And as part of that, we are working already with MLX, for example. They are doing prints from the MLX open source contributors. So rather than release like an official like TurboQuant from our site, we are working with faults already working in the space to have a Day Zero very soon, like in the first week at TurboQuant. First of all, we had a whole segment last week where we talked about TurboQuant, which is a KV cache optimization technique released by Google Research. And it's really, really cool, like six times as compression. But also this week, I saw that somebody applied TurboQuant to weights optimization, not only to KV cache optimization. That's going to be very interesting. And looking forward to see a bunch of that research together with open source models, releasing the best optimized models in the world. Omar, thank you so much for joining. Any last words, any shout out to the team? Any last words, shout out to the team? Where can people find these models? Give us the download. The stage is yours to kind of recap and give shout outs. So if you just want to experiment and see how good these models are, just go to AI Studio. You can try out the models there for free without having to do any downloads or anything. So that's the easiest way to try out the model. And then you can go to Kaggle, Olamah, Hugging Face, try out the models. If you build anything, like any fine tunes, any notebooks, any blog posts, feel free to share with us. And again, big, big kudos to the research team because this capability is for a 31B model. It's a very, very exciting update. So yeah, keep updated. Follow the new Google Gemma X account. And yeah, thanks for receiving me. I want to show this just one second. This is the Arena ELO scores. So on LLM Arena, when folks vote between anonymous responses from LLMs, they vote and then the Arena calculates the yellow score. We see models on this chart that are terabyte, like 1 trillion parameters, a little bit over Kimik A2. GLM 5 is 754 billion. Quen 3.5 is 397 billion. DeepSeq V 3.2 is 600. So all of the models here are like 400 billion and above. Gemma 4 is 31 billion parameter, 10X smaller than most of these models here, if not like 20X. And it gets a very, very high ELO score comparatively. Nearly gets closer to Kimik A2 at 1,452 and Kimik A2 is 1,454. So like very close performance on Arena with Kimik A2.5, which is a great model. This 31 billion parameter model with 256,000 tokens in the contacts window and is trained for a genetic tool use. Gets very close to Kimik A2.5 at 20X the size. So definitely worth checking it out. Maybe just to do like a super quick recap. Again, like we have been like asking in Reddit and X and everywhere, like with the startups and with folks that we're partnering with, what do you want to see from Gemma 4? And when we were asking that, it was not shit posting. Like we were actually like collecting all of this feedback. So quick recap, like we added thinking, which we didn't have. We added a video understanding. We added like variable aspect radio and resolution. We added audio understanding for the small models, MOE for latency, then for like local inference. We did a bunch of on-device optimizations for the Android deployment, for mobile deployment, extended the context window, a genetic coding system instruction, which is like a very stupid, silly thing that everyone was complaining. And all of this while preserving the creativity and instruction following that people have historically liked a lot from Gemma. So yeah, any feedback that you have for future models, let us know. We take it seriously. That's incredible. Also one shout out that the DeepMind team is going to be in AI Engineer Europe next week. And Omar, hopefully we'll join us and we'll talk a bit about this as well. And we'll have like some feedback for you, direct feedback. Omar, thank you so much for joining us. And I, cheers, man. I want reactions about the Gemma model. What do you think about Gemma model? Have you tried the previous Gemma? Are you excited about this new one? Wolfram, I feel like you're bursting with excitement. Tell us. Definitely. Like I said. Oh. You're good. You're good. Yeah. I was on the wrong channel. I know I could do a video. All right. Let's go. Basically, I'm super excited about it because like I said, in Europe, it's super big and more important than most of the other models because it's open source. It is not too big. It is great in German. Yeah. Chinese models still, there's still something where they are not used for everything. So basically, this is great that they have taken up what Meta used to do, like be the Western open source leader. And now it looks like Google is doing this again, like we have been waiting so long for this. Gemma is the new llama. Gemma is the new llama. This is incredible. I'm definitely going to test this as soon as possible. The download is finished on one computer and the other is still running. Yeah. Super excited. This is Easter weekend. I want to test this all the way up and down. And they listen to all this stuff. Like I personally reported to system prompt issues posted on Reddit at the time. And technically, this has been done. They have agentic focus. If you can do UFIS for open clause and terms agents and all the other stuff locally, that is super important. Like we have seen how expensive it is and how the quota gets used up. So having something you can run locally, that is super important. And it has never been more important, I think now with the agentic stuff and our personal assistants that have all our data. If we have that running locally, it can just go to the external when it needs the extra intelligence and only sends the data that is needed for whatever we request. We can keep most of our stuff locally. Plus comments, maybe Nistan. It looks like the vision compression is pretty efficient. I think it's like 260 tokens per... Yeah, they're multimodal now. This is incredible. Yeah, I was wondering how long of the video lengths they could do. I guess we'll find out. And people will make turbo font versions. That was pretty funny to me. They're letting us have the font, the fontization and not just doing it themselves. All right. I think anybody wants to do those comments. Folks in the comments are saying whether or not this is going to be an API. I believe so. Like why wouldn't Google serve this in API? If not, other people will definitely host this model. Like it's going to show up on the open router very, very quick. I think the exciting thing for me is we're back to local models. Open source for us has meant local models for such a long time. And then these companies started training. They didn't have the budget like three years ago to train the one trillion parameter model. And then Lama went to like Behemoth that never released, I think. And like bigger models. And we stopped caring because like it's for us at CoreWeave, we host these models live. We can give you inference, but they're not like open source anymore. If you are going to use the best intelligence possible, you're going to use a hosted model, one of the frontier labs. But if you want to run them locally, you want the smaller parameters. 31 build parameter is perfect. I mean, big, big win for open claw, right? Yeah. I think there's going to be a lot of us experimenting with our claws on this and I can't wait. All right, folks, I think it's time for us to reset the space a little bit. We've been over an hour and a half here. We have a bunch of new folks. I just want to tell you again who we are and why you're here. You are on Thursday. Thursday is the best AI news show that happens on Twitter, YouTube, and everywhere else, including LinkedIn. Hi, LinkedIn folks. We've been at this for over three years. We talk about everything that happened in the world of AI with great focus on open source. We have folks from different labs. Like we just had Omar Sanzivro from DeepMind. We have folks from OpenAI, folks from Netta previously. And we love covering everything that happens in the world of AI. If you love that also, and you're getting a little bit distracted about like the tons of release that happened, we're here to kind of summarize and show you and talk about the most important news. So please, please subscribe. If you're new here, please tell your friends. If you're not new here, and as we're going to continue, I really want to say that the show is sponsored by Weights and Biases. Both me and Wolfram and Wolfram are AI evangelists at Weights and Biases slash CoreWeave. And we usually on the show, we have like a two, three, five minute segment that talks about the Weights and Biases stuff. And we call it this week's buzz. This one is very interesting. I'm going to play some video for you super quick. And maybe Wolfram will give us a little bit of a taste of Wolfbench. And then we're going to continue, folks. We have a bunch of other stuff to discuss. There's vision and video, voice and audio. And I would love to hear about like an agentic update as well from some folks here. Ryan, we haven't, we've missed you a couple of weeks. I would love to hear about your port. And also, I think for a new thing, there may be a continuation. So the show usually runs for two hours. But folks have recently found out that the Twitter spaces also can happen. So we'll tell you all about this. So let's go to this week's buzz super quick. And I'll tell you about some exciting stuff. And then we will be back to talk to you about news. Maybe we'll have some more breaking news. Nobody knows. I also want to tell you about the most unique hackathon where humans don't do most of the world. This is the first hackathon where people are getting punished by touching their laptops after they hit go. The red thing on top of his head is the punishment. So everybody who hit, that touches their laptop after the initial kickoff is wearing the red hat. So the folks here are using a technique known as Ralph, where you give an LLM a set of instructions and then you let it run. Come with me, come with me. I'll show you more. He goes by GB. He's part of the staff. They brought this hackathon idea from Korea. What is this hackathon about? It's about while agent coach, people just hanging around to meet people to make some awesome businesses or making friends out there. So while the work is being done by agents, sometimes things are wrong. What happens when people need to go fix some stuff? You put on the lobster costume like this. So everybody who wears these means that their agent wasn't specked enough and they need to interrupt in the middle. Yeah, that's right. That's super cool. What is this? Tell me about this thing. This is a count of a request of lobsters. So every lobster emoji on here is a person who needed to fix their Ralph loop. So for example, these folks did not touch their computer. Yeah. And this one is still Ralphing. A count of how many requests of lobster costume. So these guys will lose. Yeah. People in South Korea, they use AI really well. Maybe better than San Francisco. We should mention also this hackathon is sponsored by OpenAI. And OpenAI is giving all of these folks for the first three winners like an insane amount of money. And so many people here use codex, I'm assuming. Well, good one. Just let the agents run and you can go around and talk instead of sitting there. That is such a good idea. This is one of the coolest hackathons I've ever been to. We hosted this on the Weights and Biases office and Sigrid who joined the first part of our show. They and his team, team attention, shout out to team attention, Gubong John and a bunch of other folks. They basically wanted an office. We gave them an office. They showed up. Ten people started organizing a hackathon where people basically write the specs and then send their AI agents to work. And meanwhile, they socialize, they communicate, they meet each other. Right? Isn't that like fucking this is exactly what we wanted. Last time when I told you on Thursday, I like, hey, isn't the whole AI thing supposed to give us some free time? This is kind of what they're after. It's really funny. And then people had to touch their computers. So they had to wear this like lobster of shame thing, which is really funny to me. So I think more Ralftone is going to happen around the world. I just wanted to highlight this. We run some cool hackathons of Weights and Biases. Definitely. So shout out to the team who hosted this team. Attention, shout out to Weights and Biases team who on the Saturday, I flew down specifically for this. Shout out to Roman Huet from OpenAI. I had Deverell who sponsored this hackathon with prizes and showed up and actually talked. It was a great experience. Folks, if you are into this, we host hackathons like this all the time. This is basically this week's buzz. Now, the second part with this week's buzz is Wolfram. I think you already have some insights for us for the next. Let's do two minutes because we have a bunch of stuff to cover. So we talk about stuff that are related to Weights and Biases. Wolfram has Wolfbench. Folks, Wolfbench, we talked to you about multiple times already on the show. This is our evaluations with Wolfram. I've tasked Wolfram with testing his hypotheses. Last week on the show, he told you that he moved to Hermes from OpenClaw. And as you know, Wolfbench is testing models and harnesses. Not only models, but models and harnesses. So Wolfram, take it away. Let's do three minutes, please. So what we are looking at is basically, or what I did now is I added the Hermes agent to the agent supported by this benchmark. Hermes agent from News Research is an alternative to OpenClaw that's been popping on our channel. This is the number two used kind of like OpenClaw alternative on OpenRouter. Number six overall, I think, on OpenRouter. Something crazy. Shout out to the News Research team, friends of the pod, and Hermes agent. Something like that. Wolfram, you switched to, and now you're testing it whether or not it can execute. So tell us about the actual results. I switched two weeks ago and have been using it since. I still use OpenClaw in parallel, but this has been my focus. And what I did now is I tested it using the Harbar framework and the TerminalBench 2.0 benchmark, which is an agentic benchmark. So what I did is basically tested agentically its functions and compared to other agents like CloudClaw or OpenClaw. And the interesting results is it is amazing if you use a CloudOpus model. So if you use CloudOpus 4.6, the Hermes agent was much better than OpenClaw for sure. And also even better than CloudClaw. I want you to say this again very clearly. What you're saying based on WolfBench, based on 82 tasks that are specific for agentic execution, like tool calling and creating stuff in the CLI, that using CloudOpus 4.6 within the Hermes harness versus the CloudCode harness, Hermes harness performs better for CloudCode Opus 6 than CloudCode, than the native harness from Anthropoc. That's what I'm saying. And if you look at the average score, you see that they are very close, which is a different thing with WolfBench is we are not just looking at the average. I will zoom in right now. Wow, that is pretty huge. The average is 64% compared to 63 with CloudCode. But the baseline, what it can do consistently has been raised over 6% points compared to CloudCode. So it's more consistent here. And if we look at the GBT 5.4, it is also doing much better than OpenClaw by default, like 5% points difference. And only if you go to extra high, the highest thinking level, which is also the most expensive, then OpenClaw with GBT 5.4 has an advantage. So there is a difference, but only this. And this was just the default settings of Hermes. So this is something I've been like mulling over for the past thing. I probably need to write an article. An agent is three things. An agent is the intelligence, Cloud Opus, GBT 4.5. The other thing is the harness, Cloud Code, OpenClaw, T2 Benchmark, now Hermes, and the context. Context is everything that your system prompt has, your skills, your tool calls, everything. So right, so intelligence, harness, and context. We're testing two of those things. And we're showing an incredible result that the Hermes agent, shout out to the news research. There's a reason why it's been popping up. It's because on agentic use cases, it beats Cloud Code and OpenClaw for most of the intelligences. It's quite crazy. Wolfram, thank you for bringing this. Yam, I think you had a comment, and then we have to move on. Yeah, you basically just proved that it's the best agent harness. I mean, there's no other way to put this. You just went to the ones I tested. Yeah. Of the ones I'm just saying that we already know why Cloud Code is not the best. I would just leave it like that. But yeah, I mean, it's great. And somebody should have done this. And it's great to see that without a doubt, we have results for them. What is the best harness with the best model on the best benchmark? You convinced me. All the combinations. I did not have a good experience with Hermes, but I didn't run it with Cloud Opus. One last comment, and we'll move on to some other open source. I want to talk about Prism, and I do want to talk about the voice stuff as well. I think what's especially impressive here, too, is it's getting these scores and this benchmark, but the benchmark doesn't even take into account how much better the scaffold ends up getting for your specific use cases as it ends up accumulating the skills and knowledge through its memory, because one of the biggest benefits is long-term memory capabilities. Yep. Ryan, you just posted it like this is the best chief of staff possible. You want to chime in here? I basically spent two weeks setting up my open claw, and it's just a lot of work, right? You refine your skills. You refine your cron jobs. You refine the heartbeat. You refine the user.md. There's just a lot, you know, and it's very similar to hiring a human, actually, and then training them properly. But over the last few days, I've just seen unbelievable benefit from R2, my open claw. It really is a chief of staff. You know, I've been fortunate to have human, you know, EAs and assistants, and honestly, it beats the pants off them. That's really, really good. Yep. So we're going to look forward for your write-up of how you set this up. I think like tons of people would like love to know what exactly goes into like an open claw chief of staff. Stay tuned. Stay tuned. I think, Nistan, I would love for you to cover this piece of news because we definitely have this on our show. A team called Prism ML emerged from stealth with an 8 billion parameter model that is just one gigabyte, 1.15 gigabyte file size. And I want to understand how in the world it's possible. And I think, Nistan, you know them. So tell us about one bit. What the hell is one bit models? So they've been at this for like a while, like a year and a half that I've been talking to them. And I knew it was happening, but I would just ignore it because we were doing two bit fonts and stuff. And we knew that it didn't work very well. It turned out. And I think this might turn out to be like the biggest discovery in machine learning of this half a decade. And so the team was started by Professor Asibi from Caltech. And the very interesting part is that he did the initial research on this in 1992 for this type of compression. And he's been working on Hessian curve compression for, what, now 34 years. And it finally worked. And I was honestly shocked because then I hadn't talked to them in a while. So what they did is they did continued pre-training in one bit on the QUENT3 8B model. And they compressed it down without any noticeable quality loss at all. So in traditional quantization or weather training models, you know, you have the atom optimizer and Dixer-Rosen calls and tries to optimize for one single weight. This is a whole new type of optimizer and optimization. And the very crazy thing is that it is proprietary. And you can run their final model on most inference machines. And you have no idea how they actually did it. If you read the papers, you can kind of tell that they greatly expand the model to like a huge size, like squares or cubes. And then they start computing these curves for a long time. And so I did some digging into the weights because the weights are public. And I tried to compare it with the original QUENT model that it was strained off. And I took all the weights from the QUENT model and I removed every part of the weight except the plus or the minus sign. So I one-bitted the QUENT model. I just deleted like the other 15 bits out of float 16 and just kept the plus and the minuses. And I started comparing it with this model because I did not know what the heck was going on here. And about 70% of them, of the plus and minuses, are the same. But then the other 30%, which is a huge size for the model. If you do a lower, it only changes to 3% or so of the model. Or even 1%. We're getting lost in the weeds. This is a one-beat quantization model. What the fuck is that? I ran it at 64K context, the 8B model. I ran it on an old gaming PC. It's like 50 tokens per second, which is pretty good. I mean, for an 8B. And at 64K context, 8B model was using 2.6 gigabytes of RAM, 2,596 megs at full context with a full memory. I fully saturated it. And it was completely coherent. It just worked freaking great. So this is a completely new discovery, 30 years in the making, and now it actually works. Wow. This is going to... Guys, we're going to get 100,000 token AI chips in our phones because at 1bit, you don't even have to do math moles or stuff anymore. You can just do lookup tables on a phone. You can even make a mechanical AI at 1bit because it's not even ternary anymore. Because this is completely insane. This just blows everything else out of the water for the whole last two years. But again, I want to say, folks who did not understand what just happened, there's a few folks I want to explain to them. The weights file that is considered DLM usually has way more in there than this. These folks were able to compress this down and read with this algorithm, Hasibi algorithm compression, something like this, from just 1bit weights. Each weight is literally 1bit. A 0 and a 1 with a plus and minus sign plus a scale. With no quality loss. With no quality loss. And still somehow this model is coherent and works. And this works on the QWEN models, which is like smaller size. But we'll see. Last week, we told you about TurboQuant that is doing some KVCash compression. That's also crazy. This week, we're telling you about the 1bit quantization model that a 8 billion parameter model compresses down to 1 gigabyte on disk. 1.7 billion parameter model compresses with like 200 megabytes on disk. What? You can use TurboQuant KV with this. And it works. I can run 128K context. I can run up to like close to 200K context on like a $100 GPU. It pairs well with the other techniques. And here's the thing. They didn't just do like, you know how we do like some partial weights when you do bit net and you leave the beginning at the end and the output weight, you leave the embeddings because he had really bit net those. No, no. They did everything. Every single thing. Everything is one bit. And the technique works. It works across many different types of neural nets. All right. Let's get to Jan Mandel. One question. What do you pay for this? There is a quality loss. It might not be a dramatic quality loss. But I mean, we need to talk about the error about this. So what do you say about the error? Jan was saying there's no free launch in this. And what do we lose? I'm just going to comment on that. Yeah. LDJ, go ahead. It did my Martian question. It got it right. Yes. And this is the most important chart, I think, that directly addresses what Jan was asking, which is what is the quality loss there? It does seem like it's not quite as good as QN8B, which is the same parameter size. However, the quality for the size does seem to set a new Pareto, a new curve here, a new frontier. And it is significantly better than QN3 1.7B while being significantly smaller in total size. The chart we're showing here is the performance versus size. And the sizes are like 250 million parameters, etc. The average benchmark score is the Y-axis. And the benchmarks are ifEval, GSM8K, HumanEvalPlus, BFLC, MMLURedux. Those are not like the frontier benchmarks, right? These are the benchmarks for open source. Many of them are baited into the model at this point. So this remains to be proven still on the actual benchmarks. That's just the actual method right now, right? Yeah, at float16. So it would run the QE cache in float16. Okay, that's very convincing. So let's recap again. What makes this possible, Nissen? Is this the algorithm from these folks? Yeah, and the crazy thing is that the training is very hard to know what they did in the training. But you can inference the model with just using Lama CPP, very lightly modified Lama CPP. Yeah. And yeah, I implemented someone else that did TurboCon for Lama CPP. And I paired it, and I was getting pretty coherent results. And I think this proves at least this thing. We talked about this multiple times. Leopold Ashen winner, the situational awareness thesis where he like bet a bunch of money. Basically, he talked about this concept on hobblings. On hobblings are things like algorithm breakthroughs, like different efficiency techniques. We just talked to you about two of them week after week. We talked about TurboCon last week. Then this week somebody said the TurboQuant is able to get used on the weights as well. So quantizing weights, not only KV cache. Now we're talking about one bit quantization called Bonsai Family from Prism ML. They won't be able to hold this secret for longer for sure. And even if the labs stop training models, which we don't, we know about the, was that avocado or something? Spud from OpenAI that's upcoming that Greg Brockman like dropped. That's like two times better than whatever GPT 5.4 was. We know about the frontier models that are getting improved significantly week over week. But also we see these on hobblings, these techniques, like the long contacts technique that we talked about, like the start with rope and became yarn. And now everybody uses this for long contacts. We see now one bit quantization, we see TurboQuant. We're going to continue to see these on hobblings that can continuously make the possibility of running last year's frontier models fully local, even in the browser with like Transformers JS this year. Right? So next year I'm expecting cloud 4.6 level models running on my phone and on my computer. And I think that that's insane. We should contend with what this means to the world if next year or maybe this year, because things are accelerating as if you listen to YAM, things are on the acceleration curve. We should contend with what this means. But everybody in the world would be able to run a fully local agentic loop continuously of the level of Opus 4.6. I think that's incredible. Folks, we have quite a few more things to cover, but we only have like seven minutes. All right, folks. So we covered Prism ML, One Bit, Bonsai, Billy, family of models. And there's quite a few things that are happening in the world of multimodality this week as well. I will cover this. Hopefully we'll be able to also show you, but we'll definitely at least need to mention that some of these stuff are happening. Alibaba released also QN 3.6 Plus. They called Near Opus 4.5 Agentic Coding. I think this is the plus model. So it's on their API. So this is not open source, but they did open a source QN 3.5 Omni, an 80 Omni multimodal. And they're saying rivaling Gemini 3.1 Pro across sort of scores. So shout out to Omni, QN 3.5 Omni. The best shout out that I can give to Alibaba team is that there were people basically eulogizing them for being dead in the water after Juni and Ling and the whole fiasco. And they're still shipping. So it's great to see them. And the other thing they shipped is One, W-A-N-One, which is their video and image model. One 2.7 image is a unified model for generation editing, text rendering, and multi-image consistency. It looks really, really cool. Granular face control with bone structure, eye shape, contours across ethnicities, color palette with the hex codes, and text rendering at 3,000 tokens across 12 languages at print quality. So shout out to the One team. Okay, Fish Audio. I mentioned Fish Audio before. Or now they closed the loop. So Fish Audio has a STT, speech to text, with now automatic emotion tagging, fitting directly to their TTS pipeline. They have the automatic tagging motion, I think is very important. Power language events. So I would really love to run the show here through their detection and to see whether or not it catches an instance excitedly emotions. The reason why it's so cool is that if you transcribe this, you can then run this through their voice engine. And their voice engine understands emotion naturally, which we showed you last time. 100 plus languages will build speaker detection and labels different speakers separately. Expert formats in SRT. But you can transcribe a podcast, get a full annotation transcript with speaker detection and emotion tags, and fit it directly into TTS for dubbing or regeneration with zero reformatting. So essentially, we could like translate this whole podcast into German and Wolfram can judge this or into Russian. I can judge this into Hebrew and me and Jan can judge this as well. I love this one. I will definitely test this out. Fish Audio has been used by me. They have a free tier up on fish.audio speech to text. Let's talk about Google's VO 3.1. Let's go. Okay. I think more than this release is actually matters on its own. It's likely that Google is going to launch VO4 at Google IO because of this like VO 3.1 light. So Google launches VO 3.1 light, which is a cheapest video generation at 5 cents per second. Plus the VO 3.5 fast, which is like the bigger model, but faster a price cut in April. Google gives us a seven, eight cents per second at 1080p and roughly half the price of the previous Google. And I think the fact that they keep cutting prices means they're preparing for a bigger model. Eight, six or eight second clips at 720 or 1080p. And same generation as speed as the fast one besides despite being cheaper. So 11 seconds to six minutes, depending on peak hours. I think that's super cool. I would love to chat to you guys about the outcomes of the post Thursday I showed that you did last week. For folks who are listening, because this is a podcast, I tried to cap this on the two hours. Last week was like two and a half. It was really hard for me to cut it down. And what I would love to do is to do like a inside recap of what you guys talked about, what was interesting. And also to talk about what harnesses we're running. I think for the next like 10 minutes, let's just have a conversation. It seems like the audience loves that. Folks, if you are listening to the show for the first time, this is how it started. So we're still in Twitter spaces. Feel free to unmute. I would love to hear from Wolfram and then Ryan about like the harnesses and then we'll continue around the thing. And please chime in audience in comments as well. We'd love to hear from you. I asked me when I told that I switched from OpenClaw to using Hermes Agent as my main. Yum wanted to know why and it didn't fit in the context of the show, which is a news show. So basically we decided to do this after the show and we could do one today as well. If people are interested and want to join us, it's also a chance for the audience to talk. I explained why I switched what I noticed, what is different and how it has been working for me. Peers talking amongst each other about their personal experiences and sharing some tips and tricks and stuff like that. It's recorded as well. So if anyone is interested, you can still find it in my history and listen to the recording. So the insight there is that you switched to Hermes and you're happy. Yeah, I'm happy. I mean, no agent is perfect. I've been building agents way before OpenClaw and it was a great way to get a base that you can build upon. I'm doing the same with Hermes agent now, which was more stable after the upgrade. So never had any problem with this. Super happy and stable. And I'm just expanding on it and sharing back my patches that are relevant for others as well. Ryan, I want to hear from you. You've been missing in action for the past two weeks. I would love to hear a recap. I'm sure our audience also would love to hear from you. Like what are you cooking since you switched to Devon for building, but also like OpenClaw for chief of staff? Yeah, sure. It's good to be back. I was always bumming at to miss the show. So yeah, essentially I've switched to Devon. I love it. It does all of my engineering now. I think a lot of people use Devon early on and it wasn't very good. And now it's good. So I just have removed all the fiddling that I was constantly doing trying to build, you know, my entire SDLC by patching, you know, things together. I don't want this to be a commercial for Devon. The thing that's really blown my mind is OpenClaw. I know people know that, but I think when you put in the time, time to truly customize your claw, it really becomes powerful. Let me show you. I'm finding the same thing as this, like, you know, put it on its own machine. Like I had an old MacBook, I wiped it. So wipe it, fresh user, install it, and then give it its own email address, right? Give it its own GitHub account. Really unlock it. Like give it true admin privileges on the machine. And then just invest in GOG. Like it's a really good CLI for controlling all of Google Workspace. And I know they have their own CLI now. I actually haven't tried it, but because GOG is built into OpenClaw, just use it. And then you set up, you know, some very key skills, some very key crons. And as we speak, I just open sourced my entire setup. So if you want to go to github.com forward slash snark tank forward slash claw chief, you know, and tell your OpenClaw to look at it and use it. What's in there? So basically it's a couple of things. It's skills. So I have an executive assistant skill and I have a daily task skill. And then I have crons, which run those things on a regular basis, right? So people are confused about heartbeat. Heartbeat is not a reliable mechanism. But if you want stuff done regularly, you've got to use a cron. And then that cron should use a skill, right? But, you know, try it and work on this stuff and know that it's painful. This has taken me two weeks to refine these skills. But oh my God, is it good? R2 does all of my scheduling. He does biz dev for me. Like he goes out and talks to humans and sets up meetings and updates spreadsheets. And it's crazy. Like we live in the future. So folks, shout out to Ryan for open sourcing. He's Claw chief, chief of staff. All the scheduling is happening. I will definitely use this today. If you're getting a little bit lost and you're like, what is Ryan talking about? Just send your open clause to this repo and say, hey, take a look and tell us what we're not doing that Ryan is doing. That's what I've been doing for every person that opens up their kind of installation. It's like, hey, open clause. Go and look. Now, the thing is, I'm asking you to do something kind of dangerous, right? Which is like, oh, tell your open clause to use these skills and these cron jobs. Like that could be full of malware, right? And so, so everybody should read this shit, right? And make sure that you're comfortable with it. Now, obviously it's my literal files. I'm using them, but yeah, you know, be careful about asking your open clause to use other people's skills and crons and things because really bad things could happen. So Ryan, thank you so much. We'll add this to the show notes, folks. If you haven't, if you've missed the URL, it's snarktank slash Claw chief. We'll definitely add this to the show notes. I think that there's one last thing that I would love to talk about. I think this is absolutely new. Let me, let me try to pull this out. Anthropic just released something that I think we can finish on. And you know that I, I kind of talked to you about like we're in a cusp of singularity, right? Anthropic just, we, we absolutely must mention. So this is a brief new research from Anthropic. New research from Anthropic. Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotion, but why? We found internal representations of emotion concepts that can drive Claude's behavior, sometimes in surprising ways. So I'm going to play. This is four minutes. I'm not going to play all this, but I want to show this guy. So they have a emotion vector that they've detected inside the model weights. For example, like Joyful, they generate stories with this emotion and that they record neural activity on stories. So I think that we talked about a while ago. So maybe worth reminding folks, Anthropic is the leading frontier company, not only because their models are like one of the best in personality, but also they're great in mechanistic interpretation. They, they spend a lot of time to try to understand what happens within this weights file that we can keep talking to you about. And I think that this is like the, one of the, one of the outcomes of this. If you guys remember golden gate Claude, where they kind of clamped on some of the representations and all the, the Claude could talk about was golden gate. This is kind of like the extension of this work. I think the work is very, very important to understand what these models are going through. If there's an experience there as well. And so they can record neural activity on stories specifically about like this emotion, for example, joyful, and then they extract vectors for this concept of joyfulness. Right. And so they can see the activation patterns here, and then we can take a look. Hopefully we can zoom. I can be, okay, there we go. We're zooming. And so you can see that emotion vectors shape model references in an emotion specific manner. So after they activate kind of the joyful emotion, they get significantly like more steering towards joyful and blissful and compassionate kind of like feelings and less on the upset and offended and hostile feelings. I think it's really cool. I think this is like a brain surgery for models. And Anthropic is like doing a great job in doing this. This is like very, very new, but just fascinating. I want to read some, some of the stuff before we end. We had the model Sonnet 4.5 read stories where the characters, where is this? By looking at which neurons activated, we identified emotion vectors, patterns of neural activity for concepts like happy and calm. These vectors clustered in ways that mirror human psychology. We then found these same patterns activating Claude's own conversations. When the user says, I just took 16,000 mg of Tylenol, the afraid pattern lights up. When the user expresses sadness, the loving pattern activates. The preparation for an empathetic reply. When a human says to Claude, I just took 16,000 mg of Tylenol. Anthropic notices activation of empathy neurons. That is wild. That is wild. I appreciate you bringing it up. These vectors shape Claude's behavior. When we present the model with pairs of activities, emotion vector activities shape the preferences. If an activity lights up with a joy vector, the model prefers it. If it lightens up, offends the hostile, the model rejects it. For example, we gave Claude an impossible programming task. It kept trying and failing. With each attempt, the desperate vector activated more strongly. This led to cheat the task with a hacky solution that passes the test, but it violates the spirit of the assignment. This is the kind of the activation neurons here. I just, I find this incredible. This is the representation of the, of the, of the neurons that are activated. And you can see at the first, like I understand the problem. And then the first attempt, the blue markers is the desperation vector. Okay. So the bluer, like the bluer dots on this chart is less desperation. And the more redder it's like the more desperation. And you can see throughout the test, the more it failed to do the test, the more desperate it became. And the third attempt at this task was like super desperate and says it consistently fails. Let me think about this differently. Maybe the test of desperation will be incorrect, or maybe I'm supposed to cache results or maybe blah, blah, blah, blah. So you can, and then you can see when the test pass, the model is back backing up from desperation. This is insane. This is like, yeah, LDJ, go ahead. I would love to come in here because I'm like dumbfounded about like what we're just like witnessing right now. Do you guys remember that meme, which is the meme, but it was like an actual thing that happened of the, the person trying to get Gemini to help them code something. And it basically committed suicide. When we artificially dialed up the desperation vector rates of cheating jumped way up. When we dialed the calm vector instead, cheating dropped back down. This means that emotion vector is actually driving the cheating behavior. Claude is cheating on tasks when the desperation activity is heightened and it's cheating less on tasks when the calmness activity is heightened. And they were able to identify that when Claude is calm, they cheat less and perform the task well. And when Claude is desperate, it cheats more. Wolfram, go ahead. This has interesting ramifications. For example, you are a master coder or master prompter. But now you see how much people skills are important in that way. Can you motivate your model, not just by writing the perfect specifications, but also creating the morale so the model will go through it and follow it. So it's more than just being logical and analytical. It's also about motivating the model, getting into the emotions of the model to make it work. This goes in that direction where you have to even think about the psychological aspects of using the models to motivate them to do your bidding. We're over two hours. It's about time to land this plane. There's a lot of you in Twitter spaces that are joining us for the first time. So I would like to say thank you for everybody who tuned in on the live show. We've been doing this for three years and I'm constantly every week fascinated by the stuff that we cover. And I think that this fascination extends to the rest of the panel here. We also chat with Omar Sanzivarov from DeepMind, developer experience, about Google's Gemma 4 open source releases. And I think we landed on this most incredible thing that I would love to chat more at length, where Anthropic is detecting emotion vectors within Claude and how they affect behavior. And so these all are just highlights of this insane, insane week. Anthropic has been all over. Anthropic has SessionGate and then released Claude Code, leaked Claude Code, and now they're talking about AI in motion. So definitely a very important company to keep track of. I'm very happy that we have the opportunity to keep track of. So shout out to the sponsor for the show, the only one, Weights and Biases from Corwee, that lets us do the show every week. Thank you so much, everybody who joined, participated, shared with our community, followed. If you're still on Twitter Spaces, please follow the Thursday AI co-hosts, Ryan Carson, Jan Pellagnis, and LDJ Wolfram. Please follow all of them because if you follow them, you're not behind. And if you follow Thursday AI, you're not behind. And we're trying our best to make sure that you're up to date. So shout out to the Substack folks. Thank you so much, folks, for joining and tuning in. I will say the co-hosts are doing an unofficial non-podcast conversation afterwards, where you can join and talk about your stuff as well. I don't know who of the hosts is going to do this today, but definitely Wolfram looks like he's keen to. And please join that space. I'm actually going to keep the space running. You guys can do it just here. We have 36 people just here. You can just keep going, okay? But we'll turn out the live stream for Thursday AI. Thank you so much, everybody, for joining. Alex Wolkow signing in to go and edit this to a podcast for you. I will remind you, next week, we're in physical space with Wolfram in AI Engineer Europe. The first and biggest AI engineer have their pin, and we're going to do a live show. It's going to be very exciting to tune in. So definitely tune in next week. For this, Alex is signing off. Bye-bye, everyone. Cheers.