Episode Summary
Alex demos Google DeepMind's Genie 3 live on the show — a real-time 24fps controllable world model that had everyone losing their minds. The panel explored a spaceship, watched paint persistence on walls, and collectively wondered what happened to go from Stable Diffusion struggling with minions to real-time interactive worlds in three years. Meanwhile, Kimi K2.5 took the open-source crown, Arcee AI shipped Trinity Large (400B MOE trained in 33 days for $20M), Chrome brought agentic browsing to 4 billion users, and Anthropic launched MCP Apps turning Claude into a living UI.
In This Episode
- 📰 Intro
- 📰 Hackathon Invite - WeaveHacks 3
- 📰 TL;DR
- 🔓 Open Source - Kimi K2.5 from Moonshot AI
- 🔥 BREAKING NEWS - Google DeepMind Project Genie 3
- 🔓 Open Source - Arcee AI Trinity Large
- 🛠️ Tools - Karpathy on Agent-Driven Coding
- 🔓 Open Source - Jan v3
- 🏢 Big Labs - Google Chrome Auto-Browse with Gemini
- 🏢 Big Labs - Google Agentic Vision in Gemini 3 Flash
- 🏢 Big Labs - Anthropic MCP Apps
- 🎨 Vision & Video - xAI Grok Imagine API
- 🎨 AI Art - Hunyuan Image 3 & Z-Image
- 🛠️ Tools - Clawdbot renamed to Moltbot
- 🎥 AI Art - Lucy 2.0 Real-time Video
- 📰 Outro
Hosts & Guests
By The Numbers
Genie 3 frame rate
24 fps
Real-time controllable world model — generates a new frame on every input
Arcee Trinity Large
400B
13B active params, trained on 17T tokens in 33 days for $20M
Trinity native context
512K
Twice as large as Kimi K2.5 — free on OpenRouter until February
Chrome daily users
4B
Google brought agentic browsing to the world's most-used browser
Jan v3 parameters
4B
132 tokens/sec, 40% improvement on coding, optimized for local inference
🔥 Breaking During The Show
Google DeepMind Genie 3 — Real-time 24fps World Model
Live demo during the show. Alex explored a spaceship world at 24 frames per second with paint persistence and controllable camera. Available for Gemini Ultra subscribers in the US.
📰 Intro
Alex opens the show from San Francisco with excitement about Genie 3 and Kimi K2.5. The WeaveHacks 3 hackathon is coming up this weekend.
- Show live from SF
- WeaveHacks 3 hackathon this weekend
📰 Hackathon Invite - WeaveHacks 3
Alex invites listeners to WeaveHacks 3, the upcoming hackathon in SF.
- WeaveHacks 3 at luma.com/wh
📰 TL;DR
Quick overview of the week's news: Kimi K2.5, Genie 3, Chrome agentic browsing, MCP Apps, and more.
- Kimi K2.5 king of open source
- Genie 3 live demo
- Chrome agentic browsing for all
🔓 Open Source - Kimi K2.5 from Moonshot AI
Kimi K2.5 from Moonshot AI takes the open-source crown, becoming the most-used model on OpenRouter. The panel discusses its strengths in agentic coding and tool use.
- Most-used model on OpenRouter
- Strong agentic coding performance
- Topping open source leaderboards
🔥 BREAKING NEWS - Google DeepMind Project Genie 3
Alex demos Genie 3 live — a real-time world model generating 24 frames per second of interactive, controllable 3D environments. The panel explores a spaceship, marvels at paint persistence, and discusses SIMA 2's self-improving game-playing agents. The one-minute limit frustrates everyone.
- Real-time 24fps controllable world generation
- Paint persistence — turn around, painting stays
- SIMA 2: self-improving game-playing agent built on Genie 3
- VO team + Genie team collaboration for video-to-world
Alex Volkov
"I'm still trying to understand how we go from three years ago Stable Diffusion cannot generate a fucking image of a minion to getting real live streaming 24 per second controllable worlds with persistence. What the fuck is happening?"
Ryan Carson
"I mean, so think about it. This thing has an API. And you can, I mean, we can all build on this. That's bonkers."
Nisten Tahiraj
"I cannot wait to be first person Captain Horatio Nelson in a full out Napoleon era Naval battle with ship of the Lions and cannons and everything."
🔓 Open Source - Arcee AI Trinity Large
Arcee AI ships Trinity Large — a 400B MOE with 13B active params, trained on 17T tokens in 33 days for $20M. 512K native context, free on OpenRouter. The panel discusses it as the largest Western open-source lab model.
- 400B MOE, 13B active params, $20M training cost
- 512K native context — twice Kimi K2.5
- Free on OpenRouter until February 2026
- Trained on 2000 B300 GPUs in one month
LDJ
"It should be significantly faster than Kimi K2.5, considering it's about half the active parameter count. And you could run it on a single B200 or B300 in four-bit quantization."
🛠️ Tools - Karpathy on Agent-Driven Coding
Ryan brings up the Klein team getting acqui-hired by Codex after the viral 'imagine the smell' hackathon controversy. Discussion of the Codex ecosystem, Peter Steinberger building Clawdbot entirely on Codex, and QMD semantic re-ranking plugin.
- Klein team acqui-hired by OpenAI Codex
- Peter Steinberger built Clawdbot entirely on Codex
- QMD semantic re-ranking plugin for memory
Alex Volkov
"Peter Steinberger, the guy behind Clawdbot did not use Cloud Code for Clawdbot. It's all Codex. He's one of the top leading users of Codex."
🔓 Open Source - Jan v3
Jan v3 — a 4B parameter model optimized for fast local inference with 132 tokens/sec and 40% coding improvement. Alex discusses the QMD plugin for semantic re-ranking and vector memory.
- 4B params, 132 tps, 262K context
- 40% coding gains, 5M downloads for Jan desktop
- QMD semantic re-ranking for memory
🏢 Big Labs - Google Chrome Auto-Browse with Gemini
Google unveils Chrome Auto-Browse with Gemini 3 Nano integration — agentic browsing for Pro and Ultra subscribers. The panel debates: agent browsers vs agents that browse? Chrome has 4 billion daily users, and native browsing avoids bot detection. Alex demos it live.
- Chrome Auto-Browse for Pro and Ultra subscribers
- Native browsing avoids CloudFlare bot detection
- 4 billion daily Chrome users getting agent capabilities
- Gemini's 2M context window ideal for browsing
Alex Volkov
"Chrome is used by 4 billion people daily. They just brought agentic browsing to all of those people. Do you know how much money Google will make from this?"
Yam Peleg
"I'm using Cloud Code with Dev Browser, the plugin. And it just mogs everything else. It's simply the best."
🏢 Big Labs - Google Agentic Vision in Gemini 3 Flash
Gemini 3 Flash gets agentic vision — a Think-Act-Observe loop that can zoom, crop, annotate, and plot images using Python code execution. Wolfram noticed this feature appearing in his Moltbot instance before the official announcement.
- Think-Act-Observe loop for image analysis
- Generates and executes Python to manipulate images
- Available in Gemini app, AI Studio, and Vertex AI
Wolfram Ravenwolf
"Flash Gemini 3 Flash can decide, oh, I need to investigate a specific part of an image, concentrate on that, focus, work on it. It's doing this by generating and executing Python code in the backend."
🏢 Big Labs - Anthropic MCP Apps
Anthropic launches MCP Apps — interactive UI components rendered within Claude chat. Yam explains the evolution from MCP tools to disposable web apps to now having pre-built branded experiences (Box files, color pickers) embedded in conversation. This is brands reclaiming identity from LLM text-only responses.
- Interactive branded UI components within Claude chat
- Box, Figma, and other app integrations
- Protocol-based: any app can integrate
Yam Peleg
"MCP Apps allows you to get back an interactive block and have it rendered interactively throughout the chat. You're basically not just talking to the AI, you are also morphing the actual interface."
🎨 Vision & Video - xAI Grok Imagine API
xAI releases Grok Imagine API with video generation capabilities.
- Grok Imagine API now available
🎨 AI Art - Hunyuan Image 3 & Z-Image
New image generation models from Tencent's Hunyuan and Z.AI's image model.
- Hunyuan Image 3 release
- Z-Image model
🛠️ Tools - Clawdbot renamed to Moltbot
Discussion of Clawdbot's rebrand to Moltbot (and its implications), plus the broader tool ecosystem changes.
- Clawdbot → Moltbot rebrand
🎥 AI Art - Lucy 2.0 Real-time Video
Lucy 2.0 real-time video generation model discussed.
- Lucy 2.0 real-time video capabilities
📰 Outro
Alex wraps up an incredible show featuring the Genie 3 live demo, Chrome's agentic browsing launch, and the open-source momentum from Kimi K2.5 and Arcee Trinity.
- WeaveHacks 3 this weekend
- Genie 3 ultra subscriptions raffle for newsletter subscribers
TL;DR and show notes
Hosts and Guests
Alex Volkov - AI Evangelist & Weights & Biases (@altryne)
Co Hosts - @WolframRvnwlf @yampeleg @nisten @ldjconfirmed @ryancarson
Open Source LLMs
Big CO LLMs + APIs
This weeks Buzz
WandB hackathon Weavehacks 3 - Jan 31-Feb1 in SF - limited seats available lu.ma/weavehacks3
Vision & Video
Google DeepMind launches Project Genie (X, Announcement)
Voice & Audio
NVIDIA releases PersonaPlex-7B (X, HF, Announcement)
AI Art & Diffusion & 3D
Tools
Moonshot AI releases Kimi Code (X, Announcement, GitHub)
Andrej Karpathy shares his shift to 80% agent-driven coding with Claude (X)
Clawdbot is forced to rename to Moltbot (Molty) becuase of Anthropic lawyers, then renames to OpenClaw
Alex Volkov
0:31
What's going on everyone?
0:32
Welcome to ThursdAI for January 29. This is Alex Volkov. I'm the AI Evangelist with Weights, & Biases. I'm your host for today, but I'm not in this room by myself. Wolffer is here with me. Wolffer, what's up?
Wolfram Ravenwolf
0:45
Hey man, I'm here.
Alex Volkov
0:47
How you doing?
Wolfram Ravenwolf
0:48
First time with you in the same room.
0:49
Of course.
Alex Volkov
0:50
Yeah.
0:50
We can also, put us both on the same screen so folks can actually see us that we're here live. What's up, man?
Wolfram Ravenwolf
0:56
Yes.
Alex Volkov
0:57
Welcome, everybody.
0:58
it's been a hell of a week in ai, a hundred percent. And we are here to tell you all about this as much as we can. we will have absolutely breaking news. I know of one embargo item that's gonna come through, so that's gonna be like super exciting. But besides this, we have a bunch of wait, LDJ, I'm gonna add you as well. LDJ, welcome to the show. how are you guys doing? Thank you.
LDJ
1:22
Doing great.
1:23
Amazing week of many people posting Clawd with the W and the likes.
Alex Volkov
1:28
Can we start with the RIP?
1:29
RIP Clawd with a W Ryan, you weren't able to join us last week, but last week was our insanity, show where we talked about Clawdbot with claws And then, we did a deep dive. We had, Dan on here, Dan Peguine, talking to us about a super deep dive yam. Jumped in, wolf, jumped in, like all of us. Got super excited before. This was before the insanity of Clawdbot before the four days where, this AI assistant took over the airwaves completely. so you know, as always ThursdAI brings you the news just a little bit before it blows up.
Ryan Carson
2:04
You called it, you called it
Alex Volkov
2:06
and then we tried to get Pete Peter.
2:08
He was like super busy. I think most of us ran Clawdbot since then. And then one day we woke up and me and Wolf actually like already went live and talked about this, but, not on Thursday. One day we woke up and like, oh, Clawdbot no more. So, Tropic decided that Claude bought from Peter Steinberg at the Open source AI assistant you can install is, infringing on their copyright. And, politely, he said politely, not like lawyers, whatever politely asked him to change the name.
he was freaking out on Discord at 3
2:38
00 AM and the name they came up with is Malti
2:45
Malti with a t something about, lobsters. Malting over. It's like thematically appropriate.
Wolfram Ravenwolf
2:52
Yeah.
2:52
The naming, yes. The naming of AI tools is something special.
Alex Volkov
2:55
And so, the name they came up with was malti, but then they
2:59
were trying to rename all the projects. And in the process, the renaming from Clawdbot to multi bot crypto scammers took away that username. And so they got left with Moltbot. And this is a horrible name to pronounce, Peter, I'm sorry, but this is the truth. It's really hard to say t and b one after another. It's really not ergonomic. And so Moltbot is what we have right now, and we used to have Clawdbot and now we have Moltbot, and it's really unergonomic. However, the software still works. They're still haters saying that this works. It doesn't work. but yeah, we've been running a bunch of stuff. we can talk about this as well. We know that this is like a very hot topic, in the world right now, but so much has happened. About cloud bath and whatever since our last show, but also in the world of ai folks, I will go around super quickly to ask all of you what's kind of the highlight, the one thing that we must absolutely talk about, and then we, will dive into the TLDR. What is the one thing, in the world of AI that is, capturing your attention right now, please?
Wolfram Ravenwolf
4:03
Right now I'm most excited about Kimi K2.5 which may be
4:07
a model, one of the open models that you can actually use with Clawdbot or Moltbot, as an agenda model for this. So it has been very great. That is my favorite right now. And it works perfect in German as well, which is particularly important to me.
Alex Volkov
4:22
Yep.
4:22
That is awesome. I'm absolutely up for shout out to moonshot folks for the great release of Kimmi. What is the one thing in the world I, for the past week that is most important to talk to?
LDJ
4:35
Yeah, so, besides those main other things, I would say a Trinity large is
4:40
something that is interesting, like it's coming out of, a collaboration between RC. Prime intellect and ology. And I think it might be the most powerful new open source, model coming from a American AI lab. It's about 15 billion active parameters, 400 billion total, and it's looking really good. And they're working on more reasoning and reinforcement learning training on top of it to make it even better.
Alex Volkov
5:06
So absolutely big focus on pre-training for, Western Labs.
5:10
Shout out Lucas Atkins and everybody at ARCEE AI for this. we're gonna get Ryan on Ryan Carson. What is the top one thing in the world of AI from this week that you wanna tell our audience?
Ryan Carson
5:20
I, stepped down from AMP and I'm running untangle full-time,
Alex Volkov
5:24
so
Ryan Carson
5:25
I'm excited to be full time on my startup.
Alex Volkov
5:29
I'm sorry, Ryan, but hey, everybody.
5:30
Applause for Ryan. Ryan, congratulations. we tell us a little bit later about your startup as well. I would love to hear and tell our audience as well, because I think it's super cool in very, very important. we're, we're having folks tuning in. Roger says, the studio looks cool. I would like to shout out Chroma as a hosting for ThursdAI for this week, specifically DJ and Jeff and everybody here at Chroma. Chroma is a vector, solution database, et cetera, is the best in the world in everything. So they're gracious enough to host us in like ama amazing studio. super quick.
Alex Volkov (2)
6:01
So we're probably gonna do more shout outs, but we are
6:03
both in San Francisco this week for the, Weights, & Biases hackathon. So if you're in San Francisco, I have something to say. An announcement that I don't know if you know, folks, if you are in San Francisco and you would like to come to the hackathon, that we're hosting this weekend, I have announcement. We have a secret special sponsor that we're working towards and I am very happy to say that cursor is our last second. Sponsor and we are going to give each participant for the hackathon $300 in, in cursor credits. You know how much opus that is? That's at least 2 million tokens if something. Way more. but yeah, I'm very, very excited about Cursor, jumping on as a sponsor to our hackathon. If you wanna come and hack, you're gonna get cursor credits, you're gonna get versal, people in there with their AI stuff and skills and everything. you're gonna get Redis. You can store a bunch of stuff in Redis. You're gonna get browser base and you can automate your agents. Come build Self-Improving agents with us in San Francisco this weekend. Let's go. All right, folks, this was a little plug. I, I'm super happy. The course are like, reached out and, I'm super happy that we, like, we can provide credits also
Alex Volkov
7:17
at, 13 minutes and then bunker lifts of something
7:21
I can't wait to tell you about. So I'm just saying in 13 minutes, I'm going to be able to say, yes, you guys can, breaking news, breaking news, breaking news that I know ahead of time. Alright folks, I think it's it's time for the TR.
7:47
the notes open Wolfram.
Wolfram Ravenwolf
7:48
Mm-hmm.
Alex Volkov
7:48
Awesome.
7:49
so yeah, let's separate the chunks of the TLDR Wolf. You wanna take the open source ones and I'll take the Big Companies and LLMs?
Wolfram Ravenwolf
7:56
Yeah, sure.
7:56
So for the TLDR open source, it's chemic, K 2.5, my personal favorite of the week. And it's the open source visual genic model with 1 trillion trillion parameters. We will go into it in detail, but the first one, is Kimi K 2.5, which is amazing in, genetic use.
Alex Volkov
8:14
Yeah.
Wolfram Ravenwolf
8:15
the next one is ARCEE AI releasing Trinity
8:18
Large, the 400 B sparse MOE model with 30 B active parameters. Jan AI releasing JAN V three for B open source model, optimized for phase fast local inference and coding, with some improvements in the data benchmark, for instance.
Alex Volkov
8:32
I'm sure that this is it in terms of open source in the TLDR,
8:36
Alright. And we're gonna say hi to NI and the, the gang is in the building. Let's go folks. Super quick. so for the big labs for this week, there's a lot of stuff, not a lot of models, but a lot of stuff. So Google launches agen auto browsing Chrome with Gemini three. Gemini now lives in a side pane and not just like a blue popping window. Nano banana is integrated in there and connected to Google apps and you can click stuff, so like Chrome is going, stepping into the Agentic era. Anthropic launched MCP apps this week. You guys may remember, if you're FA follower of ThursdAI we talked with Ido Salamon and Liad ef, the folks who created MCP ui that then turned into a spec called MCP apps. And, OpenAI adopted this with OpenAI I GPT apps. And now Anthropic is launching apps as well on top of this protocol. So it's super cool Reach UI for Slack, Figma Sandbox, and a bunch of other connectors. Google also launched Agentic Vision in Gemini three Flash, which is, we talked about the agentic vision at some point. It's kind of where the model plans for itself, how to tackle a problem. It can zoom, it can crop, and can do a bunch of stuff. It's pretty cool. There's a big major essay from Dio ammo de CEO of Anthropic called the Adolescence of Technology. it's warning AI Risk to National Security Democracy. dio, according to folks who work on Anthropic, is like this, like very, very deep, big thinker. And usually the thoughts he has are going internally, within Tropic. So it's great that he published this as well, and for very big breaking news. We have at, very, very soon we have breaking news from, big Lump as well. We're gonna let you know all about this in this week's buzz, the category we're gonna like, make, update you about everything that happens at Weights, & Biases from core. another reminder, the hackathon that we have, hack three is almost sold out, but if you mention ThursdAI in the application form, we'll let you in. I'm pretty sure Wolfram, you have an automation that helps us figure out who's, coming in the hackathon or who doesn't come in. Right?
Wolfram Ravenwolf
10:34
We are an AI lab basically, and we are the AI evangelists.
Alex Volkov
10:38
yes.
Wolfram Ravenwolf
10:38
So we have our assistants going through these making
10:40
summaries, and then we make the decisions
Alex Volkov
10:42
do not, do not prompt, inject our application form.
10:46
I dare you, I dare you. Do not reject it.
Wolfram Ravenwolf
10:48
you've
Alex Volkov
10:48
been
Wolfram Ravenwolf
10:49
injected
Alex Volkov
10:49
luma.com/weavehacks3.
10:51
Please join us. I just announced that our, main, super secret less maintenance sponsors cursor. You can get $300 in cursor credits to hack to do Opus Azure Heart Content. In Vision video, we, will tell you about a bunch of stuff, but the main thing that is very, very important that changing the world of vision and video is that grok with a K from XAI, they released an API finally imagine API. it's now ranked number one in both text to video and image to video models. guess how much it costs, know how much it costs, $4 and 20 cents a minute. Of course, of course, they price everything they have with this like ridiculous number. anyway, it's cheaper and it's faster than a VO 3.1 apparently. So that's pretty cool. voice and audio, NVIDIA releases Persona Plex seven B. It's an open source, full duplex voice that listens and speaks simultaneously with persona control. still in the world of AI art and diffusion. both Tongyi and Tencent launched new versions of their image editors. We told you about Z Image Turbo previously from Tonga, from Alibaba Z Image. The full one, the full capacity is now open sourced as well with negative problem control and superior diversity. I'm pretty sure it's open source. Let me triple check. Yeah, comfy UI support is already in there. the image is pretty cool. The images, the quality was really, really impressive. And then Hunyuan on image as well. the Tencent, AI lab is we have, obviously the explosion of Clawdbot or, Moltbot now, taking over the airwaves. Absolutely taking over the airwaves for the past weekend, people got into the meme of buying Mac Minis and installing this thing on the Mac Mini. And we went live and talked about why Mac Mini makes sense. We can like talked about this as well. and the last things that I will mention in the tools area that Mo
Wolfram Ravenwolf
12:46
just one thing.
Alex Volkov
12:46
Yeah.
Wolfram Ravenwolf
12:47
Because about, Mobo, what we have seen online
12:49
is a big hype, basically, which we were part of, are part of. But what is cool is you can to go some party in San Francisco and talk to people and they ask you about their thing. Have you seen this? And it's really not just on X, but it's happening. In our circles as well. And that is a very cool thing to see. And we realize why the naming change is actually a good thing because some people then the ask, are you, do you mean CLO code or do you mean the bot thing? Yeah. So yeah, I can understand why philanthropic wanted the name change.
Alex Volkov
13:19
Yeah, a hundred percent.
13:20
philanthropic definitely is within the rights for this. however, you know, it's, we got used to the name. All right, so Stu, super quick. Tools So Clawdbot was forced to rename into Moltbot or, like if Effectionately known as multi, I think this is one of those Twitter x situations where people will just continue saying Clawdbot, no matter what, like this is what it feels to me about. and also, it absolutely took over the airwaves on algorithmic. I don't know why, but Grok, the algorithm just absolutely loved Clawdbot. So for, for a minute there, every post that I saw, everyone I built a Chrome extension to like highlight them. Every post was about Clawdbot. It was insane. we have Moonshot, KMI, not only Kimi, K 2.5, they also risk Qmi code. It's open source, Python gen coding platform. They, they also have like a code, Subscription, as well. So that's great. And I think the last thing is, Carti. Carti is out again in force, telling everybody that the world has changed. so under Carti posted a big, thread on X saying that, AI shifted. He's coding to 80% agent driven coding. And he declares a phase shift in software engineering, which I think we all agree on We can start with open source in a minute. folks, should we dive in? I think it's time for us to dive into the actual open source news. There's a bunch to talk about.
Wolfram Ravenwolf
14:35
Nisten could, tell us his favorite.
Alex Volkov
14:36
Oh yes.
14:37
Nisten, Nisten. Tahir. What is the one thing in the eye for this week that is on your mind from news?
Nisten
14:44
Alan ai, Sarah, it started reaching, it's a 32 B. The,
14:49
they released a bunch of models. It's a 32 B, started reaching 45% on s Swyx bench. And, this is the best we've gotten where we have all the data. 'cause, I'm very interested in this because right now my job is basically data janitor at scale. So this was, extremely nice data set from them and actually not bad results. so it's like truly open auditable code. So if you care about, I don't know, safety and all that stuff, you know, you actually have all the data to work with it.
Alex Volkov
15:24
Yep.
15:25
That's great. alrighty, Nisten, thank you all folks. Time for open source. Our favorite, favorite corner. we don't have the whole hour today for open source. We're gonna run super quick. open source, let's go
15:47
Open source
Nisten
15:48
ai.
15:49
Let's get it started.
Alex Volkov
15:54
Let's get it started, folks.
15:55
The absolutely most important release of this week in open source is Moonshot AI with Kimi K 2.5. Why is it the most important? Please discuss while I bring up some images, some stats. Wolfram Yam, LDJ. I'm leaning on you for this one. Why is C It's good?
Yam Peleg
16:18
Pretty much.
Wolfram Ravenwolf
16:19
Very good.
Yam Peleg
16:21
Oh yeah.
Wolfram Ravenwolf
16:21
Yeah, and it's cheap.
16:22
Basically because it's open. There will be different providers. You will have a pricing competition there, and that is one thing. If you are running your bots and so on, you want to use the best AI you can, but you don't want to go broke.
Alex Volkov
16:35
Yeah.
16:36
So what do we know about this model release? first of all, Kimi from K two and then the update somewhere in September, October, I think oh nine 15 or something. Kimi was very unique among the open source models. It felt different. It felt like the vibes were, were a little bit different. it was great writing, for example. Oh, yes. and I used to like always compare drafts and whatever and ask it what it taught as well. but now this one beats multiple big, labs in different evals. So like Wolfram, can you walk us through some stats about Kemi specifically and what we're getting excited about?
Wolfram Ravenwolf
17:09
Yeah.
17:10
First thing, it's a big model. 1 trillion parameters. MOE with 32 B active.
Alex Volkov
17:14
1 trillion parameters.
Wolfram Ravenwolf
17:15
We had trillion parameter models before as well.
17:18
Yes. But, you see a lot of open models that are pretty small. They are great, but, I personally, I'm a big fan of the state of the art big models, the frontier models that have the biggest performance and still with 32 B active parameters, you can run this if you have a good enough machine.
Ryan Carson
17:34
Yeah.
Wolfram Ravenwolf
17:34
So it's possible to run this.
17:36
It's not only online and, yeah, it has 384 experts, eight hours elected per token with a 256 K. Context Length. Context, length is particularly important because that is, super important when you have the agents running for a long time or if you are working on code, you want to have all of this inside, the context. So the bigger the context, the better. And like Opus, you can only go to 200. Is it 256 K with Opus as well?
Alex Volkov
18:02
think
Wolfram Ravenwolf
18:02
I think it's the same.
18:03
You can Sonet, goes up to a million. And I think there's also a version of Opus with the Million, but I always run in the limit.
Alex Volkov
18:10
we don't have this, version, maybe the folks in philanthropic, I
18:14
don't know, I know nothing but the folks from Cloud Code, Boris Journey and, Ka, they both left for open the eye for like, oh no, for cursor for like a week. And they came back on Tropic and everybody's like, huh, I wonder why they came back after a week. I was like, Hmm, that maybe because they couldn't use cloud code in, in the other place. Unlimited with 1 million tokens, right? With Opus 4.5 with 1 million super speed. because after playing with these models, speed and context window is very, very important. So if you have this unlimited from Labs, that's great. Yeah. So please go ahead, Kimmi.
Wolfram Ravenwolf
18:45
Okay, so it has, 50% on the humanities last
18:49
exam, full set tools, benchmark.
Alex Volkov
18:51
Yeah.
Wolfram Ravenwolf
18:51
75.9%, almost 75%, on brows com.
18:56
I think this is also one of the best, in that benchmark and, in Egen Swarm,
Alex Volkov
19:02
it's multimodal.
19:02
Now, it wasn't multimodal before, right?
Wolfram Ravenwolf
19:04
That's a
Alex Volkov
19:04
the big like unlock from Kimi is that it now can see,
19:08
and I think somebody was talking about it, can see videos as well. and people like will send it videos, which was super cool
Wolfram Ravenwolf
19:15
and it's great in the benchmarks for vision stuff too.
Alex Volkov
19:17
Yeah.
Wolfram Ravenwolf
19:17
Dust will support it, but it does it very well as well.
Alex Volkov
19:20
So here's the thing, and folks, we talked to you about this
19:22
thing multiple times and you know, if you follow the show, but if you don't follow the show, here's the thing. we are talking about an open weights model. But it's 1 trillion parameters. So it's not like you're gonna run this at home. This is not a local model.
Wolfram Ravenwolf
19:35
can run it, you know?
Alex Volkov
19:36
Yes.
Wolfram Ravenwolf
19:37
Company can.
Alex Volkov
19:37
Yeah.
Wolfram Ravenwolf
19:38
Invest in the resource and run a SOTA model on the premises.
Alex Volkov
19:41
Yeah.
19:42
I want, I, I wanna address directly to the camera, so, one second. If you're the type of company who wants to run QK two, reach out to, CoreWeave and we'll help you set it up. Because running this on like one H 100 is not gonna be enough. You have, you have to have like a full node of things. It's a 1 trillion parameter model. It's an MOE model. only a little bit of active, so like, it can run really, really fast. But definitely, definitely I'm looking forward for like, you know, the, the big, dedicated chip companies like Cereus and, grok maybe, but Grok was bought by, But Nvidia, right? So I don't know, gr is gonna put up some new models. so I'm looking for, like a fast interest, LDJ, Yam Peleg. What do you hear about this model? Do you, have you used it? what do we hear? Nisten? Feel free to chime in. I want multiples of opinions.
LDJ
20:28
So it seems like to be the same active parameter and total parameter
20:32
size is the Communic K two model. However, a lot of the big model smell things people like to talk about, like creative writing, which a lot of people praised, communic, K 1.5 and Kimi K two for it seems like this will be something here that people will be able to continue to use that for here, but at the same time get like the agent ENC coding abilities that is more competitive with the Frontier models. I think we still need some time to really assess how well it compares to something like 4.5 Opus and 5.2, but at least we Bench Pro and Terminal Bench 2.0, which I think amongst us on this panel, that that's some of our favorite coding benchmarks. It doesn't seem to be doing quite as well as something like Opus 4.5 or GPT 5.2 there. But still, I mean, that's a really high bar to reach, right? And the fact that this is open, you could run it on-prem or run it on, CoreWeave or, and everything. hopefully we end up seeing a lot of kind of multimodal video understanding demos as well with it and seeing what applications can be built on that.
Alex Volkov
21:35
So let's put it in practical terms for folks, because
21:37
again, it's a big model putting this on-prem for many people. Like unless you're a company or you did an exit for your startup, this is not practical to put this on-prem, right? You can get this through open router, you can get this directly from Moonshot in their API. They have no tiers for coding as well. we're talking about 60 cents per million tokens, 60 cents. Opus 4.5 is four and a half, $5 per million tokens, right? So we're talking about a 10 x decreasing price. I'm not gonna say they have the same level of intelligence. I don't know what kind of black magic and tropic diary ammo they cooks with, but there's something about the opus models and the cloud models. we told you about this before, like there's something there that hard for us to like tell you exactly what it is and help for us to understand. But if you are on a budget and if you're using Clawdbot, you should be on the budget because if you're using API, it can burn through tokens like crazy, especially with subagents. I see Ryan nodding because once you spin up those subagents, the prices get away from you super quick. 60 cents per million tokens is very, very affordable you can absolutely switch to chemo for most coding tasks if you're on the budget and still get great results.
Wolfram Ravenwolf
22:48
Right.
22:48
And that is the input tokens. The output tokens are even more expensive with Claude for 25.
Alex Volkov
22:54
Oh yes.
Wolfram Ravenwolf
22:55
For a million tokens output while you have
22:57
just $3 for the moonshot.
Alex Volkov
22:59
25 is crazy.
Wolfram Ravenwolf
23:01
And it was 75 with Opus four, remember?
Alex Volkov
23:03
Yeah.
Wolfram Ravenwolf
23:03
75
Alex Volkov
23:05
pricing is absolutely ridiculous.
23:06
Okay. Agents arms, is a native thing that they have here. Ground, they call it groundbreaking feature. Spawns up to a hundred parallel subagent, achieving 4.5 speed ups on complex tasks, with aggressive pricing. This makes sense, but just the fact that like this, these models, the open source models can pick up, work across like so many sub agents I think is like very, very, very cool. So Kimi, definitely folks should, check this out and. We have breaking news. Let's go. Breaking news, breaking news. I was waiting for this one. AI breaking news coming at you only on ThursdAI folks, this is breaking news that I was like very, very excited to come to and tell you about. and I was waiting for the embarker to lift and I did have early access. And we're gonna show you this right now, Google has just announced that, genie three, the world model game generator that they've announced somewhere in the last year is now out. Geni three is now out, now it's out for ultra subscribers. I will say though, that, stick around with us because I am able to give one of you three months or two of you three months of ultra subscription to Google. Test it out. So if you are interested in this, we'll definitely, definitely see, what is Gen three? Gen three is a world model. It's a generator based on a picture or some text that can create full worlds. We were super excited about this when it came out. And now, Google is launching this in, let's say preview mode. So I really wanna show you this. you guys should see this. This is absolutely, absolutely crazy. It's giving you a world to play with in for like a whole minute. All right, so you guys see this like interface. hopefully the folks on the vertical also see this. So basically you describe an environment and then you describe a character and then they have different environments. This is my most favorite one. And, the environment is a tactile needle, felted diorama, featuring a wool terrain and animated, animated fabric crowds. And the character is a fuzzy snail with responsive sliding mechanics. Let's create this world on the fly. what happens here is on the fly, Google will create a full 3D world with physics and mechanics, for this world. very quick. It's available for Google Ultra subscribers in the US only. Folks, look at this. Look at this. This is not, this is not video. This is me controlling the little snail. You can see I rotate the camera. The snail just bounced. Do you see it bounced, away from this little felt rock? And I can control this guy. I don't know what space does, Yeah, it jumps. It jumps. Let's go. That's amazing. Can look at this fucking quality.
Nisten
25:52
Wow.
Alex Volkov
25:53
there's little folks in the exit.
25:54
They're looking at me running. That's so amazing.
Wolfram Ravenwolf
25:58
Wow.
25:58
It's a coalition.
Alex Volkov
25:59
There's collision.
26:00
You guys see it like it's stuck. It's now stuck. It's crazy. let me pause this here. We're gonna like, generate folks. Meanwhile, I would love for you to tell me what kinda worlds you want. but let me go like with the cameras and tell you guys I played with this. My mind was absolutely blown because for a whole minute and a minute is not enough. I really want more shout out to the Gemini team. I have a chance to chat with folks from the Gemini team. This is like the big embargo that we have. this supports the remote controller Bluetooth controller. You can actually like play many games, folks from the Gemini, the DeepMind team that work on Genie, they. Animate their kids' characters and they give them the world to play in. Beautiful
Nisten
26:35
idea.
Alex Volkov
26:35
they take pictures from their daily life with the dog at some point.
26:39
one of the researchers from DeepMind told me that like, he used to use his dog as a test subject at some point he just like, looks so realistic. It didn't make any sense. But imagine taking your dog and he's like using it, like to walk around with it. It's just like absolutely crazy.
Wolfram Ravenwolf
26:53
A drawing of your kids and then have it,
26:55
yeah, like a game for the kids. Play with their own characters.
Alex Volkov
26:57
Alright, we have somebody that says an Underwater world on the
27:01
four-way planet as the environment. And, let's come up with,
Wolfram Ravenwolf
27:06
the lobster.
Alex Volkov
27:06
oh, we have to use a lobster.
27:07
A hundred percent. Okay. So the environment's gonna be an underwater world on a far away
Wolfram Ravenwolf
27:13
with mermaids
Alex Volkov
27:14
planet.
27:15
we see life and mermaids at, so I and mermaids. The character is gonna be A blue robotic, lobster
Nisten
27:23
dressed.
27:23
Dressed like a hacker.
Alex Volkov
27:25
Dressed like a
Wolfram Ravenwolf
27:26
Yeah.
Alex Volkov
27:26
hacker with, and
Wolfram Ravenwolf
27:27
a yellow jacket,
Alex Volkov
27:28
sunglasses.
27:30
And the yellow jacket. You guys are really like pushing this for this model. The, third person. Yeah. We need to see the lobster. Yeah. So we're gonna create the sketch. this is powered by Nano Banana Pro behind the scenes, right? So Nano Banana Pro generates the imagery and we know the Nano Banana Pro is absolutely incredible. But also this release means so much to the agentic world because agents, will be able to like, spin up these worlds and play around with them. and you know, many folks in Google suspect that this is the kind of the path for a GI.
Wolfram Ravenwolf
28:01
Wow.
Nisten
28:01
he's got glasses.
Alex Volkov
28:02
Listen.
28:02
is this good enough? A blue Lobster with sunglasses and a yellow bumble jacket? do we need any changes or are we gonna go
Nisten
28:08
No, no.
28:09
He is got a, a jet back on him. Yeah. No, he's great.
Alex Volkov
28:11
Alright folks, I'm excited.
Ryan Carson
28:12
is amazing,
Alex Volkov
28:13
Oh my God.
28:14
Look at this. We're swimming with mermaids on a wow.
Ryan Carson
28:18
the mermaids are creepy.
Alex Volkov
28:19
There's a lobster in sunglasses and it's a hacker
28:22
lobster and it can swim up. Can I swim down? I can swim down.
Wolfram Ravenwolf
28:25
Does it have collision with the walls?
Alex Volkov
28:26
Lets, let's see.
28:27
I can see the wall. I'm gonna crash.
Wolfram Ravenwolf
28:29
Lost
Alex Volkov
28:29
its glasses.
Wolfram Ravenwolf
28:30
Oh wow.
Alex Volkov
28:31
Did you see my shadow?
Wolfram Ravenwolf
28:33
Yeah.
Alex Volkov
28:33
Are you seeing the shadow that we're casting on this rock?
28:36
This is insane, folks. This is a real time. 24 frames per second streaming of a model that takes input and generates the next frame on the flight. This is not video. I'm controlling this.
Ryan Carson
28:48
What is happening?
28:49
like, what's the hardware?
Alex Volkov
28:51
I'm assuming a lot of TPUs, but I'm mind blown.
28:55
It's collision. It stops. What the fuck?
Wolfram Ravenwolf
28:58
This is a few of the feature
Alex Volkov
29:01
and we're done.
Wolfram Ravenwolf
29:01
I want this kind of game and it creates it on the fly for
Alex Volkov
29:04
you,
Wolfram Ravenwolf
29:04
basically.
Alex Volkov
29:04
Yeah.
29:05
this is on the fly. Gaming on the fly worlds. absolute insanity. I don't think we've played enough with this one.
Nisten
29:12
can we give it a picture just LD j's picture?
29:16
Like just that cat Character.
Alex Volkov
29:19
Yeah.
Nisten
29:19
Can it make LDJ if you just screenshot this cat face?
Alex Volkov
29:24
Yes.
29:25
I'm gonna absolutely do this. let's do it. Cat face, screenshot. okay. Let's see what we can do. So, but we need a new world though. LDJ, what kind of world do you wanna live in? Your character,
Nisten
29:36
the Matrix.
Alex Volkov
29:37
So our environment is a long time ago in the Galaxy 4, 4, 8.
29:39
The character's estimate Cat, that has, I dunno, we'll figure this out. We'll, we'll let the, the cool thing about this, is that now gonna pro is very like, creative and so Oh.
Wolfram Ravenwolf
29:51
Oh,
Nisten
29:52
wow.
Alex Volkov
29:52
Whoa.
29:57
Turn around. Look
Wolfram Ravenwolf
29:59
at the light.
Alex Volkov
29:59
Did you see the camera moved?
30:02
From the, did you guys see this? Like I turned the camera and it like it moved closer because there's a wall there. LDJ. Oh God.
Ryan Carson
30:09
That's
Alex Volkov
30:09
amazing.
30:09
Is this insane or what? We only have a minute. Let's go explore. Let's run. Go.
Wolfram Ravenwolf
30:13
Let's to the light.
Alex Volkov
30:13
Let's run.
30:14
Let's run. Can you jump
Wolfram Ravenwolf
30:15
look Theas?
Alex Volkov
30:16
Yo.
30:20
Absolutely crazy. I like, it's a little annoying that it's in slow motion, but maybe it's because it's like geo gravity or whatever.
Ryan Carson
30:26
because you're in space,
Alex Volkov
30:27
there's less gravity.
30:28
Oh, that's true.
Wolfram Ravenwolf
30:28
if that was, simulated.
Ryan Carson
30:30
Oh, whoa.
Alex Volkov
30:30
see outside.
30:31
We just went by a window and it generated a whole other side of the spaceship through the window, and it's like, what the fuck is happening?
Wolfram Ravenwolf
30:38
Can you go through the window?
Alex Volkov
30:39
I don't think so.
30:40
We're in space.
Wolfram Ravenwolf
30:41
Wow.
Nisten
30:42
Okay.
Alex Volkov
30:43
Look around,
Nisten
30:43
Let's
Alex Volkov
30:43
see.
30:44
And our one minute is up. No.
Wolfram Ravenwolf
30:48
Oh, you're floating off into space.
Alex Volkov
30:50
Wait, hold on.
Wolfram Ravenwolf
30:50
thinking.
Alex Volkov
30:51
I'm thinking, but I don't know if you guys see this.
30:53
I can still control the thing. One minute is up. But did, did we break through the, oh, there we go. Okay, now it stopped. Guys, what the fuck was just that? Anybody feels St. Dingly about we're living in a simulation.
Ryan Carson
31:09
crazy.
Nisten
31:10
That was nice.
Ryan Carson
31:11
I mean, so think about it.
31:12
This thing has an API, like, I assume It does. And you can, I mean, we can all build on this. that's bonkers.
Alex Volkov
31:19
If this has an API, it's gonna be crazy.
31:20
the thing that I wanted to talk about is two things from, Gemini, the team that I spoke to. One, the VO team, the video generation team. Obviously those are different models, right? So the VO team and this team, they collaborated. And so what happens is folks at Google can generate a video with VO and then continue and step into this video with Genie three. This is like absolutely insane because the video is like higher fidelity because it's like locked in. And the second thing is SIMA two, which is they're kind of like, they have an agent that plays games. They have a self-improving, game player set up with this technology. So they have like a player. they generate a long session for them. they play this game inside the geni three, and then they kind of like, look at what worked and didn't work, jump, didn't jump, et cetera. They generate another game and they have this, self-improving gamer that can play games. It's absolutely insane.
Wolfram Ravenwolf
32:13
you think the one minute limit is just arbitrary and they
32:16
can generate much more elaborate things?
Alex Volkov
32:18
Yes.
Wolfram Ravenwolf
32:19
Yeah.
Alex Volkov
32:19
I've asked them and I gave them feedback directly.
32:21
One minute is not enough. It's really, really annoying. but obviously they, they need to serve a bunch of people. So folks, I will say, for subscribers of ThursdAI I, will post a poll or a chat. I will see how it is, right? So like look out, I have three ultra subscription thank you for the, the DeepMind team for the early access. I will send the details down with the newsletter. If you're in the US and you wanna try Genie three and you don't have, ultra subscription, ultra for three months, I don't know, it's like 200 bucks. you can be able to like, use Gemini, et cetera. question from the audience. Is there another lab can, can get close to this quality? I don't know, LDJ, go ahead. While I'll generate another world.
LDJ
33:06
For the first SORA model that released, I recall there was a couple
33:09
talks that the creators of the first SORA model did where they did actually say a big reason why they're doing something like SOA is because they do believe that's going to be an essential component to having these world model like things, that will be a component to feature a GI. And they did actually show a demo of, I recall, I think it was the first Sora where they showed it basically being able to do a world model of having Minecraft world and having a player actually go in that world and interact with things and fight mobs we didn't really see much more than that yet. But I imagine OpenAI might be thinking along similar lines here. And I think the context of, the one minute limit, I think part of it might be related to increasingly higher memory, usage happening. As you get significantly beyond that, like two minutes might end up having significantly more memory usage and flops usage with the context than just one minute. And they might just have to draw a line there.
Alex Volkov
34:04
I think you're absolutely right in terms of memory,
34:06
because this thing has consistency. It knows. So for folks who are just listening to a podcast, I'm in the world where I'm a blue ball of ink and I leave like a trail behind me and I look around in the camera everywhere where I went before. Like my trail is consistent. And so like, they had this viral video when they just announced unit three where somebody was painting the wall and then turned around, turned back, and the wall was painted where he painted them.
Wolfram Ravenwolf
34:30
That was one of the most impressive things I've seen
34:32
with these world medals where you, we are moving a short step closer to the wall and it was painted then away, look back and forth. That was super impressive.
Alex Volkov
34:40
Yeah.
34:40
just
Ryan Carson
34:41
don't understand how this works.
34:42
how is it doing it this fast?
Alex Volkov
34:45
the speed is basically, they generate a new frame on every
34:52
input and they generate 24 frames per second, and they just, you know, sped up their diffusion models. It's a world model and it takes input. it's absolutely mind blowing. Absolutely mind blowing. I want, I wanna do one more. I wanna, like, I don't wanna focus the whole show on Ingen, but you guys know, you guys know, for those of you who follow seriously, you know, how excited I get about world models because I, absolutely. Like, I wanna put on my VR headset and just like, go into one and I think this future is not that far away given the what we see now. So I was like very excited about Genius three when it came out. and this is like, this specifically is a game generator type thing, so I wanna show you like a cool ass, realistic one, right? So they have, you have realistic, experiences and you have, game type experiences and you can control all of them. So. Let's see. Absolutely insane. So, gen three from, the, the folks at DeepMind, let me pull up my notes. Gen three, is available for ultra subscribers in the US If you, and I think the US is a latency thing, if you are not an ultra subscriber in the US and you want to play around with this model, I will send instructions in the newsletter today of how to actually get this. And I'll like raffle of the two subscriptions. So shout out to the Gini team for giving us access and releasing this incredible, incredible thing.
Wolfram Ravenwolf
36:10
It's a glimpse for the future for sure.
36:12
We are looking into the future. What will be in 1, 2, 3 years, like we, if we look back to ai, now we see how far it has come. Look, imagine what this will be. In a couple of years,
Alex Volkov
36:22
I'm still blowing the fuck away.
36:23
Sorry guys. I'm still trying to understand how we go from three years ago. Stable diffusion cannot generate a fucking image of a minion to getting real. Live streaming 24 per second controllable worlds with persistence. What the fuck is happening? And I was like, I'm here for all of it. This is incredible. This is why we do ThursdAI
Wolfram Ravenwolf
36:44
It's time to be alive.
Alex Volkov
36:45
Yes.
36:46
It's a great time to be alive.
Nisten
36:47
I cannot wait to be first person Captain Horatio Nelson in a full out
36:53
Napoleon era Naval battle with ship of the Lions and cannons and everything.
Alex Volkov
36:57
Why am I not surprised?
Nisten
36:59
I was thinking about sail and then I'm like with Cannons
Alex Volkov
37:01
Yes.
37:02
what still remains there for full immersion, obviously is 3D and Sound. Imagine, and I'm sure all of these things are coming, like with video. And if you guys wanna, like, if you're having a hard time extrapolating, just use Star Trek for example. Just use whatever. Like if you're having a hard time extrapolating, look at the world of video models. They first start very clunky, then they continue, then they have audio, then now they have like physics and character persistency, et cetera. LDJ, go ahead and then we'll move on.
LDJ
37:32
imagine taking the next step of having the model, have the goal
37:36
of generating stereoscopic imagery for VR and having that live. there's already even startups, which maybe we'll cover this in the future with future announcements, but I know startups right now already having successful trials of things like piping in basically motion, signals into your inner ear to make you feel like you're moving to the left or moving to the right and a bunch of the cool things like that are coming.
Alex Volkov
38:00
Yep.
38:00
Alright, folks, back to open source. back to open source, because we covered MK 2.5, we wanna talk about rcs, rcs, major release as well.
Wolfram Ravenwolf
38:09
It's a 400 and bpar MOE model with a 13 B active parameters,
38:14
which is trained on 17 trillion tokens in only 33 days for 20 million.
Nisten
38:19
Well, go ahead.
38:22
We had, Lucas, Atkin here, the CTO when they had just started training it on 2000, GPUs, 2000 B three hundreds for a month, and it's out. so they said it was supposed to finish training on January 9th or something?
Alex Volkov
38:36
Yeah.
Nisten
38:37
so, yeah, this is pretty cool.
38:39
It's like the largest, I think Western Lab now. They did well. Mistral was a bit bigger, but it's pretty cool to just see people from the Open Source community just start off just making small models and stuff. And then they just train A four B and now it's just 400. the scale is great.
Alex Volkov
38:55
The scale is awesome.
38:56
what do we think
Wolfram Ravenwolf
38:56
512 K native context.
Alex Volkov
38:58
Mm-hmm.
Wolfram Ravenwolf
38:59
So it's twice as large as Cmic K twos, 2.52.
39:03
And, it's available free on Open Router until February, 2026. So if you want to try it now, it's the best time.
Alex Volkov
39:09
Oh, shout out to Open router for, for given this, available for free.
39:13
so shout out to RC as well. integration with Prime Intellect folks. And, the data, data folks, somebody remind me, Data ology. Data ology. Thank you. I was almost there. Thank you for, for helping me land this, this auto completion. Yeah. tokens, ology, AI
LDJ
39:28
it also should be significantly faster than Kimmy K 2.5, considering
39:32
it's about half the active parameter. Count But this is a good size too, where if you put this into four bit quantization, which is a pretty standard quantization, it can fit with some extra context as well on a single B 200 or B 300, which can be K 2.5 cannot. So the fact that you could run it on a single GPU, if you are someone like CoreWeave or another company that has access to those types of GPUs that can accelerate Inference Lab.
Alex Volkov
40:01
All right.
40:01
So shout out to our folks in RC and folks free on entrepreneurial to give it a try and let us know.
Wolfram Ravenwolf
40:05
And on kilo code two, where,
Alex Volkov
40:07
Yeah.
Wolfram Ravenwolf
40:07
There's also three for,
Alex Volkov
40:09
dunno,
Wolfram Ravenwolf
40:10
What is Kilo Code?
40:10
That is an extension for Richard Studio Code.
Alex Volkov
40:12
Mm.
Wolfram Ravenwolf
40:13
And it's, basically like za in the way.
40:15
see how we can use it.
Alex Volkov
40:16
Alright.
Ryan Carson
40:17
Are we gonna talk about the fact that Klein sort
40:19
of got acqui hired by Codex?
Alex Volkov
40:21
Let's talk about this.
40:22
It hasn't, you know, it's kind of in the open source related world. Ryan, I had a corner that, you know, I want you to lead. Maybe we didn't plan for this with like, personnel changes, timeline feedbacks, et cetera. you wanna briefly give us the story there?
Ryan Carson
40:38
Yeah, so we just, the, and again, like I stepped
40:41
back from my role at amp. I'm a free agent. and so I, I don't really have a horse in the race here. I still love amp. I still use it. but. It was interesting as we started seeing these tweets pop up from, the members of the Klein team saying they joined the, the OpenAI to work on Codex. And everyone's like, whoa, what? and it appears like it's just another team Acquihire, which is fine and cool and yay for them and probably yay for Codex and it's all good. _ Jan 29 - Kimi 2_5_ RIP Clawdbot_ Anthropic Apps & more AI news: then
Alex Volkov
41:05
you don't wanna talk about the backstory here where Nick Pash
41:08
from Klein, replied to a hackathon picture and said, imagine the smell. And this comment got absolutely exploded in virality. Everybody was trying to tell the person, like he's, he's for some reason racist, like the whole Indian diaspora on Twitter, probably like came over, came after him a lot. A lot of folks tried to tell him this is a racist comment, dude is like, have you been at a hackathon before? Like, folks, I will just say, let me just say very, very clear clarity that we've hackathon that's coming up this weekend, that you are more than welcome to join at luma.com/we. Hack three, is gonna have a great smell, but generally though, generally hackathons, are places where folks, are still in college, that, you know, they love computers, et cetera. don't shower for three days. The smell is usually funky, unrelated to who's there. So Nick got a lot of, a lot like it went up to Elon Ma, whatever. It's like, it was incredible. and then his, founder of Klein basically said he came out in support and said, Hey, he's a great guy, whatever, 18 minutes. And then he got so much pressure that he came back and said he's, he's now fired. And so basically Nick got, you know, declined. I'm so sorry for this part, but Nick, Nick got declined. and and now the whole team of those folks are now at Codex and, yep. And, we expect Codex to do great things. I will say another thing about Codex super quick is that, Peter Steinberger, the guy behind Clawdbot did not use cloud code for, for Clawdbot. It's all Codex. He's one of the top like, leading users of Codex. So Codex is really great, as well. Alright, moving all from moving forward from the news in open source, I think it's time for us to talk about the big labs. let's mention John, John V three. It's a 4 billion parameter open source model, optimized for fast local inference, 40% improvement on either benchmark, 132 tokens for a second, 4 billion parameter models. These models are great for a thing. So, if you are installing personal AI as systems like Cloud code, you probably won't be able to run all of that inference You know, you won't be able to get the same performance if you're running a 4 billion parameter model. But I discovered this trick that installed yesterday, QMD from Toby Luka from, Shopify. the dude is going like super crazy into memory and vector, stuff. So basically he built this plugin called QMD, and this plugin has semantic re-ranking. What it means is that if you have a vector database or just a database like stored in vector space, when you pull different chunks, looking at those chunks and saying, Hey, this is more relevant to your query, this is less relevant, is something that LLM can do. You don't need a big model for this. You can do like, absolutely like a small Qwen model. the more smaller intelligences we have, the faster those type of experiences are. this creates an incredible memory that can reflect on itself. Rephrase your question and have different high quality of questions. So Jan V three is absolutely one such model. I'm gonna show you the image of it super quick. and then we're gonna move on from open source. Where's John? Here we go. 4 billion models, 40% on coding gains, 4 billion parameter, 260 2K context with 60 gigabyte ram. You can run this model. and they have 5 million downloads for John desktop. the whole like software that they have, with GQI architecture and GDF ready and they're like very happy about their aid benchmarks. I think this is it for open source as far as I saw. folks, let's move on to big companies.
Nisten
44:32
Oh, there, there's, there's one more.
44:33
There was the, the, the updated deep sq, OCR and Oh yeah. yeah. And I'm finding out lately that the small models are not just for you tinkering at home or, or mobile. They're actually great for doing data work at scale because, if, if you have to pay for the GPUs, you usually want to find what is the smallest model that can process the data to an acceptable level. And that allows you to just, get an order of magnitude in speed and cost savings. So they do also have quite a bit more use like on, at, at scale, like for companies and stuff.
Alex Volkov
45:11
Yep.
45:11
Moving on to big companies and some news from big companies. I'm still, so I wanna go play with Genie three. I want you to send, to send you guys to play with Genie three as well. big companies, we don't have. Any model releases, right. So like, obviously I would say in the big companies now that we, we've gotten here, the, the biggest news from this, I think is Gen three from Google is now available for, for testing the world model, the real time, 20 frames per second world model that you can create worlds in. but some other stuff dropped from, Google as well this week. Chrome.
Wolfram Ravenwolf
45:41
I also think, genie is more theoretical.
45:43
It shows us what will be possible, but for the actual user right now, it's of limited use. You can play with it and you probably are done with it, but these, the news we are now talking about, those are the stuff that you can use every day and can make your life easier.
Alex Volkov
45:56
Yep.
Wolfram Ravenwolf
45:57
So should I start with, yes,
Alex Volkov
45:58
please Go ahead.
Wolfram Ravenwolf
45:59
The Auto Pro in Chrome with Gemini three, nano banana integration
46:02
and, connecting to Google apps. So Google unveiled a major update to Chrome, which now integrates Gemini three, which is their most advanced AI, model, of course. and gives it agen browsing capabilities like Atlas from op, may I. And now Google needs that as well, and they integrated Gemini into it. So Auto Prowse is now available in preview for Google AI Pro and Ultra subscribers which can autonomously handle multiple step tasks like vacation planning. Booking the stuff you usually see as the demos form, filling appointments, scheduling, shopping with budget constraints, stuff like that. The easy gentech stuff. But now it's also available directly in Chrome, which means it's available to the USA, but soon through the whole world.
Alex Volkov
46:43
Yeah.
46:44
we talked about the gen browsing a bunch. We talked about, J-J-D-V-D Atlas. I wanna question everybody here on the panel super quick. Do you guys use agen browsers or do you use agents that browse? does anybody here have an active Atlas thing that's running for them? Does anybody here have an active comment or Now Chrome?
Wolfram Ravenwolf
47:03
I have Chrome, so I'm using it and, I constantly
47:06
reach the limitations when I want some stuff transformed. Sometimes it doesn't do it, but yeah, I use it and I wouldn't go back to a non energetic browser.
Alex Volkov
47:13
Interesting.
Wolfram Ravenwolf
47:13
Yeah, I definitely do that, but, I also use the mode bot,
47:17
to serve the web and do stuff for me. So I use both.
Alex Volkov
47:20
Yeah.
Wolfram Ravenwolf
47:21
Yeah.
Alex Volkov
47:22
What about you?
47:23
You seem to be back in the reacting to the agenda browser.
Yam Peleg
47:26
Yeah, I'm using, cloud code with, dev browser.
47:29
the plugin.
Alex Volkov
47:30
Mm-hmm.
Yam Peleg
47:31
And it, it just mugs everything else.
47:34
it's simply the best. I tried them all and it's the most reliable one, and it's the fastest one. It uses the least amount of tokens. Anyone that doesn't know about it. it's open source. Huge shout out. go check it out. It's not by myself. Just huge shout out to the creator and it's an absolutely great, plugin. You can use it with cloud, with Clawdbot or mul bot or, and code.
Alex Volkov
47:58
Sorry?
47:59
Dev browser.
Yam Peleg
48:00
Dev Browser, yeah.
Alex Volkov
48:01
Because there was one, this week a agent browser as well from,
Ryan Carson
48:03
use agent, I use agent browser, which is similar.
48:06
and it's all I use. I don't really use an AI controlled browser, but now the Gemini is, is first, like a first person citizen of Chrome. I'm gonna try it.
Alex Volkov
48:15
I noticed that when I try to browse, with non browsers, UIs, but
48:22
via agents that control some browsers, I frequently get into these like, bot prevention things that my bots cannot fix. And what I notice is when I use Atlas, which is like first party native browsing experience or Chrome, now, they feel like me to the websites, right? Whether it's X, whether it's, you know, CloudFlare sites, et cetera. So there is definitely a benefit for, native agent browsing, but I wanna stress how big the deal is here. Chrome is a browser, is the most, you know, used browser in the world. It's probably like 2 billion people, 1.8 billion people or so. I'm just throwing numbers in the air, but, are using Chrome. On a daily basis. This is the, by far, the most, used browser in the world. The fact that now they're bringing, agent browsing capabilities only to also ultra subscribers. the Google Ultra subscription tier is becoming more and more enticing, and it's been fucking every month, OpenAI Pro Tier does amazing things. And then we're like, okay, and then on tropic with cowork, et cetera, and you're like, okay, should I switch? And now Google's a ultra tier, which, hey, if you subscribe to newsletter this week, I'll, I'll you, you'll be able to raffle out, three months of Google through subscription. But now their subscription becomes very enticing because if you are on Gmail, like most of us, then Chrome can do a bunch of stuff for you that maybe, Atlas not gonna be able to, we didn't get to LDJ and instant. Do you guys use agent browsing and how?
Nisten
49:53
I just use chromium dash dash headless commands.
49:57
but I run a separate model to do the actual navigation, and then I just run quad code with it. try to keep it relatively simple.
Yam Peleg
50:07
What are you running?
Nisten
50:10
You know, just like a, like a little Qwen on the side to just look guys,
50:16
there are ways to, to get around the box. Like there's always, there's always a way. Just think about, oh yeah, you can see the screen,
Yam Peleg
50:27
I saw a cloud for Chrome, just casually clicking.
50:31
It just clicks. I'm not a bot and continues. That's so great. It's incredible.
LDJ
50:38
Yeah, for me, I've tried using Comet.
50:41
It just wasn't like quite there for me. at least not something for me to consistently want to use it over Chrome and everything. for Atlas, I need to check back on Atlas, but at least when it first released, I think it was only available on Mac. and so that ended up with me not installing it.
Alex Volkov
50:59
Yeah,
LDJ
50:59
but,
Alex Volkov
51:00
and Chrome is available everywhere.
LDJ
51:01
Yeah, exactly.
Alex Volkov
51:02
So folks, this is a little hard to see on the screen, but,
51:05
I'm now having this, agent, browsing the call is auto browsing in Chrome. You can see, I opened the sidebar.
Wolfram Ravenwolf
51:11
Wait, we are still seeing the sketching alert
Alex Volkov
51:14
on the right, the white one.
Wolfram Ravenwolf
51:15
Oh,
Alex Volkov
51:16
yeah.
51:16
I don't know quite how to zoom in here because this is a browser, so I can't like zoom into their experience. what I wanna say is, there's a sidebar, This is the Gemini area. I said, can you generate a world for me there? And says, I can complete this task for you, but I need your permission first. Chrome can help you get things done. Gemini and Chrome uses auto browse to work in your tabs and help you get things done, like book appointments, create shopping cart. They love the book appointments thing. It shows its steps and tries to confirm before taking sensitive ones you can stop or takeover every time. that's also a difference between, if you're running agents, for example, with browsers versus you seeing an browser that does the thing for you, you can control it. You're like right there. so let's actually take it for a spin. Let's say let's go. And I asked it to generate a world for us within the Google Genie. So it says tasks started at some point it will show. So now it shows on the actual browser that like it took over. So when you hover this, you can't like click, and it just click the create sketch button and then it has like a waiting thing. So it, like, it'll wait for things to happen. Somehow it detects it like a process going on. and then it says I have clicked. the generation process is now complete. The creation controls on the page are currently disabled. It may be a few minutes. Okay. And now what? So yeah, a agentic browser that can do multiple actions, you can look, I think it's a huge deal. Chrome is used by 1.8 billion people. They just brought Agen browsing to all of those people.
Wolfram Ravenwolf
52:39
4 billion actually.
Alex Volkov
52:40
4 billion daily,
Wolfram Ravenwolf
52:42
Daily
Alex Volkov
52:43
users.
Wolfram Ravenwolf
52:43
Yeah.
Alex Volkov
52:44
most of these people,
Yam Peleg
52:45
and you don't need to pay,
Alex Volkov
52:46
no, no, no, no, no.
52:48
this is only for ultra subscribers. This does not go out for 4 billion people, I'm saying.
Wolfram Ravenwolf
52:53
Not yet.
Alex Volkov
52:53
do you know how much money Google will make from this?
52:57
4 billion people? They'll pop one notification like, Hey, if you subscribe to Gemini Ultra, you will get automatic browsing. 4 billion people will see this at 0% conversion. They're gonna make billions out of this. Just like overnight. This is insane.
Wolfram Ravenwolf
53:11
subscription
Yam Peleg
53:12
that's not a cheap subscription.
53:13
That's not a cheap, I mean, we will subscribe. Yeah, a hundred percent. But we are not a representative. You know,
Ryan Carson
53:19
I don't think it is limited ultra.
53:21
Are you sure about that, Alex? 'cause I'm not an ultra on my personal account and I can use it.
Alex Volkov
53:26
the auto browsing.
Ryan Carson
53:27
It seems like it.
Alex Volkov
53:28
I'm,
Ryan Carson
53:29
I'm double checking now.
Alex Volkov
53:30
Let's check.
53:30
Oh, ultra and pro and ultra.
Wolfram Ravenwolf
53:33
Pro.
Alex Volkov
53:33
Yeah, pro and ultra.
Wolfram Ravenwolf
53:34
Even my mother has pro.
Alex Volkov
53:36
Okay.
53:36
So she
Wolfram Ravenwolf
53:36
get it if you buy a pixel phone, yeah, she had it.
Alex Volkov
53:40
I just absolutely incredible.
53:42
We, we have to test this out, right? Like we have to test the capabilities, but we know the Gemini is great at multimodality. We know the Gemini is great at seeing the screen. we absolutely know for browsing, especially in the way that you guys are doing it with like the tools that control the browser, context window is a pain in the ass because sometimes these browsers send the whole goddam HTML to the thing and then Opus reads the HTML, like it makes no sense.
Wolfram Ravenwolf
54:04
compacting after just
54:05
_ Jan 29 - Kimi 2_5_ RIP Clawdbot_ Anthropic Apps & more AI news: web
Wolfram Ravenwolf
54:05
search.
Alex Volkov
54:06
Yes.
54:06
So like you do two searches and then you have compaction and then like you forgot where you went to do, this probably can run for longer. So that's, absolutely crazy.
Wolfram Ravenwolf
54:13
million context window.
Yam Peleg
54:14
Oh,
Alex Volkov
54:14
that's
Yam Peleg
54:14
Gemini.
54:15
It has like 2 million, token context.
Alex Volkov
54:18
Gemini is,
Yam Peleg
54:18
there's no problem here.
Wolfram Ravenwolf
54:20
Yeah,
Alex Volkov
54:20
Gemini Flash can run super quick.
54:22
I'm very excited about this. Lemme see if, my agent, task is still here. Yeah, I have it. Lemme see.
Ryan Carson
54:29
I'm with you Alex.
54:29
I think this is a big deal. we've seen Google go from asleep to launching Gemini to Gemini being integrated into workspace and now being integrated into Chrome. I mean, this is a google.com sort of mass moment that's beginning to happen. and it's exciting.
Wolfram Ravenwolf
54:45
I'm pretty sure Gemini is now the most used AI model.
54:48
If you consider that it is powering Google search, you have it as a Gemini assistant on the phone. But just actually before using, my main assistant, and it still is my sidekick, is, the Gemini assistant. I can just push a button on my watch and talk to it or just say the magic words without even unlocking my phone. And I get access and it has access to my mail and all these things. And not only I, but even my mother can do stuff like that.
Alex Volkov
55:08
I think it's time for us to move on from agen
55:10
browsing, but it's a huge deal. let's put it out here. it's a huge deal, 4 billion daily users,
Wolfram Ravenwolf
55:16
so don't even know what agen browsing means.
55:18
But if you get a power up and it says, Hey, you can automate this task.
Alex Volkov
55:22
Yeah.
Wolfram Ravenwolf
55:22
How many people will use it a lot?
Alex Volkov
55:24
Just absolutely insane.
55:25
we are moving on because, Google is shipping this week. Google is shipping this week. let's talk about agent vision from Google as well. Super quick. The
Wolfram Ravenwolf
55:33
agent Vision in Gemini three, flash, and this
55:35
is actually my second favorite of the week because I've been using the Flash model a lot with mobo. And I noticed this feature in Mode Bot because, it analyzed an image and it said it zoomed in and extracted parts of it. And I thought, oh, is that hallucinated? So when is it doing that? And now Flesh Gemini three Flesh is also capable of that. So it can decide, oh, I need to investigate a specific part of an image, concentrate on that focus, work on it that way. So Egen vision for GI three flash is a new capability that transforms, how it interacts with images by not just passively analyzing the visual context. Instead actively reasoning about the image quease using a Think Act, observe loop. And it is, it's doing this by generating, executing Python, Python code in the backend, basically to zoom cro, annotate even, and plot so it can even change the image. Show you stuff that has also been part of the Gemini assistant, where it showed you, if you ask, where's my, where are my glasses? It really made an area on the picture so you know where to look. And this is now available in general here. the availability is, in Gemini app thinking mode enabled Google, AI studio and Vertex ai. So I now, I'm still wondering if what I saw happening with my, bot instance, if that was now did it really do that?
Alex Volkov
56:52
Hmm.
Wolfram Ravenwolf
56:52
But it has the ability to execute code on that.
Alex Volkov
56:55
Yeah, they execute code and now it's Gentech.
56:57
alright folks, let's move on to the big companies, MCP apps, yam, and I want you to chat with me about this. We had our friends, ido and, LIAD EF both here. they created MCP Git MCP, and then they went super viral and then they created MCP ui and then they get super viral and basically MCP UI turned into this MCP apps, or just open apps, I think. protocol. we talked about that with them. And then we talked about how OpenAI is adopting this. And OpenAI is now an app store and Tropic is now an app store as well. It has apps built in into the interface. what do we know about this? How exciting is this? Have you tried it? Would love to hear your thoughts here as well on this.
Yam Peleg
57:35
look at the beginning.
57:37
you had MC you had MCP protocol that basically allows you to, expose tools or other ais to use. It can be locally, it can be in a server and whatnot. and it was an upgrade because throughout the chat, the model could just use whatever tool that you want. However, the thing is that, we also have a different thing that many people do with, that many model providers, natively support at this point, which is canvas. That you have web apps that are disposable. you talk to your, favorite portal, and it creates a web app that pretty much does what you want. It can be used for demonstrations and so on. And I don't need to go too much into this. MCP apps allows you to basically get back, an interactive block and have it rendered interactively throughout the chat while you are using the actual chat. if you go to, philanthropic, release, you will see that, throughout the chat with Claude. They have for example, a color picker. Like you ask Claude for a color picker and it just creates it out of thin air, like writes the code and then you can use it to actually pick color on the fly throughout the chat. So you're basically not just talking to the ai, you are also morphing the actual interface that you took the AI through Yeah. that's basically MP apps in, in like what it allows you to do.
Nisten
59:09
Just think of it, if you want Photoshop inside the, the
59:13
cloud ui, you can kind of do that. Now. they don't have an, a Photoshop plugin,
Alex Volkov
59:19
that Figma.
59:19
the thing that I wanna mention, remembering the folks that were here that built this, when you are a company that provides just information, MCP kind of eats your brand, right? So you expose your information via, model context, protocol. DLLM looks at the information it provides and spits it out to the user in text. And then your brand is basically disappearing. Your brand colors, your brand identity, everything just goes away. And what happens with CP apps, what happens with M CCP apps is the LLMs that you use every day. Claude J GPT, probably Google is gonna adopt this. This is an open standard right now. they can serve your brand's identity and full visual UI experiences where you can click and they are. Involved in the process there. So whatever you do within this very rich window of things, the LLM kind of can pick up on. so this is a very different thing than on the fly apps. These are experiences that are pre-built, so LLM doesn't create them from scratch and write the code, et cetera. This is like a very specific thing where, a pre curated list of apps. And we, we have a few examples here that I wanna show, yam. Thank you for, for the thing. so box files, for example, if you use Box, I know many people who do, but like on enterprise, definitely many people use Box. You can say, Hey, find our latest analysis report. and then cloud will just like, Hey, let me lemme look through box to this. So it says, here is the file, but it can now show you the actual file. It can show you like a file preview, load, the PDF, et cetera. And this UI is boxes, ui, the blue on top is specifically box, the annotate features, et cetera. Everything here is specific to to that ui.
Yam Peleg
1:01:03
I think that the important thing to emphasize
1:01:05
here is that it's not limited. They can basically integrate whatever, anyone that supports the protocol. That's, that's the big thing here. Yeah. That it's interactive components coming from the. And the end party can say, and not like build into the chat itself.
Alex Volkov
1:01:26
yeah, I think it's
Yam Peleg
1:01:27
unlimited
Alex Volkov
1:01:28
I don't know if you can enable any app that
1:01:30
you want, like MCP with cloud. I think that, like right now it's a curated experience, but maybe I'm wrong, maybe I'm wrong. I think at some point they left me open up.
Yam Peleg
1:01:38
Look, it might be curated at the moment because of, you
1:01:41
know, it's just starting and so on. And so they curate the list of, approved products and so on. But in general, it's a protocol. That's the thing about it. It's a protocol, a standard that you can use for whatever. So I assume that in the future they'll just support, just like they support custom MCP, that the user can provide. And it's your problem if you leak your entire computer online.
Alex Volkov
1:02:08
Yeah,
Yam Peleg
1:02:08
they will, support, that you will add custom MCP apps, through this.
Alex Volkov
1:02:13
Yeah.
1:02:13
Pretty cool. So here's, here's how it looks in Claude and I'm connected now. it shows, shows basically the featured ones, gamma Amplitude, box, Asana, Conva, those are like the featured ones. Monday, for example, and Hex. And then you have a bunch of others. This is like a, a very impressive amount of, of connectors. I don't know if many of them expose apps or digital cps and. I don't think it's for the person to know, actually. Like, I don't, I don't, I don't know if the person needs to know whether or not this exposes an app or not, but I know for a fact that Figma, for example, exposes an app, in cloud. I can try to install this, but Figma is probably not gonna work for me. If you wanna know about MCP apps from the creators themselves, please check out our Thanksgiving episode with IDO and, LIAD. We brought you this news, a while ago. Let's move on folks. I think we have more stuff to cover Okay. So in Big Labs, the thing that also I wanna talk about is. I think it's like Big Labs, but multimodality as well. XAI. Elon Musk's very fast competitor in the world of ai, have released Grogg, imagine via the API. It's been a while, but it's been incredible to see how fast they're catching up on everything. Rock Imagine is part of Grok. It is a separate model as far as I can tell. Maybe Grok tells it what to do. Maybe there's like some intelligent part to it, kind of like, you know, nano Banana three is powered by Gini three as well. It's probably that rock. Imagine is a video model. it's a text to video and image to video, and now it tops artificial analysis, leaderboards as number one. I think it's like number two at El Marina or Arena. They rebranded to Arena. and previously we talked to you about Glock Imagine Video because, well it was uncensored. they obviously had less limitations than other, other, labs that are working in the space. Maybe this is another competitive advantage, but they have been significantly improving their video model and now it's not only video model, I think it has audio as well. should we look at some of the examples? I think that's good with audio. Let's, let me pull this up with audio. And I will,
Ryan Carson
1:04:24
we're starting to see pay as you go now with the X-A-P-I-I
1:04:28
threw a hundred bucks in it this morning, fused it for some stuff. It's encouraging to see this happen. So I think a lot of people are gonna build on top of, these APIs if they are easier and to use.
Alex Volkov
1:04:39
Yeah, so now you guys are seeing my payments page,
1:04:42
which was not super helpful. I am paying for the graph API in addition to the XAPI. But I think what I wanna,
Wolfram Ravenwolf
1:04:48
speaking of the AI a, API that is also great for our AI agents.
1:04:51
So minus, basically using Nana Banana Pro, my agent wants to show me something. So the more APIs you have, the more it can do, it can generate a video to explain something or summarize something for you.
Alex Volkov
1:05:03
This is what I wanted to pull up.
1:05:04
I think that the most important part of this is that the agents can use this. This is now available via remote, whereas Rock Imagine before was only available via like the GR interface you had to pay for it. Now, the, like Grant says it's paper token, but also, just incredible video model. So let's take a look at this together. it should have sound on it as well. Yes. Okay. You guys let know if this comes through. so this is the picture that we had, and then, you can see the video of like two robots drinking beer. And they're highlighting the cheersing, this is some examples. Cook is walking around in the chocolate stream and having a kiss.
Wolfram Ravenwolf
1:05:45
Does anyone hear the audio?
Alex Volkov
1:05:46
There's no audio in this one yet.
Wolfram Ravenwolf
1:05:47
kiss.
Alex Volkov
1:05:48
Yeah.
1:05:49
I can absolutely say that this is near state of the art quality for video.
Ryan Carson
1:05:54
that's good.
Alex Volkov
1:05:55
Yeah.
1:05:55
Performance and benchmarks. they, the thing here is they are fast as well. So if you look at these, score to latency, roc imagine is, 45 seconds latency where VO three is, 68 seconds, 78 seconds. it's, price wise and latency wise, I think they're bid everyone just, absolutely everyone. they also have video editing, capabilities. Let's look at video editing. So, if you have like a person sloping down the screen, then you can like replace it with panda, skiing down the slope. If you have a person skiing down the slope, you can replace it with panda. You can replace it with a Viking, you can replace it with, with a, penguin, for example. variating capabilities in this now flexible styles as well. So 16 by nine apparently is, a nine by 16 is available. So they're definitely going after. It's really funny. On the release page of this, there's Instagram stories like ui, so it looks like very clear who they're targeting with. and I wanna hear the audio. so this is not the best
AI
1:06:55
perimeter is clear.
Alex Volkov
1:06:56
Oh, they're speaking.
1:06:57
Hold on,
AI
1:06:57
we have the packet.
1:06:58
Good. Any heat? None yet. Here, take it. Copy. What's the status signal's Clean. This one's new. Top of the line. What's the range? 800 meters armor piercing.
Alex Volkov
1:07:08
Not the best example.
1:07:10
There's multiple folks here talking. The same person talking kind of like Sora. So this is just music. we have some otters playing in audio, playing with water. Oh, okay. This is better. You can see the syncs kind of like, sound effects when they play with water. I'm not seeing a lot of talking happening. I'm not seeing a lot of like lip syncing. I see audio. I see like generated like sound effects.
1:07:42
It is synced to the actions in the video, which is pretty cool, right? So the model obviously understand sound, but I'm not seeing a lot of like, hey. Let me talk.
Wolfram Ravenwolf
1:07:49
The question is, can we use it already?
1:07:50
If you took a picture of us here in the room and uploaded it as a picture and that animate the podcast, would we be talking then?
Alex Volkov
1:07:57
I don't know.
1:07:57
I don't know. But the video editing capabilities are pretty cool. you can see, and we've been waiting for, you can also have remove capabilities. This is beautiful. so for folks who are just listening, we're looking at a video of a bug on a flower, and then you can just remove the bug, remove an object, or you can add objects.
Wolfram Ravenwolf
1:08:14
like this one,
Alex Volkov
1:08:15
one wolf tools.
1:08:16
Three walls. This is pretty cool.
Wolfram Ravenwolf
1:08:19
cool.
Alex Volkov
1:08:20
It's very cool.
Wolfram Ravenwolf
1:08:21
I have an urge to send my assistant now a message.
1:08:23
look at the API implemented and then generate me.
Alex Volkov
1:08:26
I, mine actually has, an API key for XAI.
1:08:30
Yeah, we can do it.
Ryan Carson
1:08:30
This is exciting.
1:08:31
So I use, VO three one right now, in production for a couple things. It's about two videos a day, and so I'm excited to actually try this. The quality, it's
Alex Volkov
1:08:40
rate limited on api, keys on production for two videos
Ryan Carson
1:08:43
a day.
1:08:43
Yeah, for just the, I think they don't have capacity to generate more, more video, but that's vo, right. So I'm excited to try, this imagine, and see what it goes.
Alex Volkov
1:08:53
Yeah.
1:08:54
we should absolutely try to have it generate a video for us. Okay, so I'm gonna take a screenshot. Everybody look at the camera and smile.
Ryan Carson
1:09:04
I'm not a smile.
Alex Volkov
1:09:05
I wanna show you guys, how me and Wolfram and
1:09:08
some of us are working nowadays. This is how we work and go to Telegram and I say, Hey, I want you to learn this new API of imagine of Groq. Imagine based on this URL. And I paste the URL in, to then take this image I attached and create a video from it where all of us have, eagles, land of different animals, come into their screen
Wolfram Ravenwolf
1:09:47
and let us talk.
1:09:48
So it generates voice.
Alex Volkov
1:09:50
So, for folks who are not following what I'm doing
1:09:52
right now, I'm in Telegram. I'm texting my AI assistant that you can see. he is now like accepted the task and now he's typing and is gonna say, lemme learn the API and then generate the video and that's it. Focus, this is the new world we're living in right now.
Yam Peleg
1:10:04
That's cool.
Alex Volkov
1:10:05
Let's go.
1:10:06
I got the a PII need to upload the image. Lemme do this. It's like, you gonna wanna see it live? 'cause it's gonna happen live. _ Jan 29 - Kimi 2_5_ RIP Clawdbot_ Anthropic Apps & more AI news: Let's
Wolfram Ravenwolf
1:10:12
look at it.
Alex Volkov
1:10:13
It's just gonna go And I, this is now the new world.
1:10:17
Like I don't even think
Wolfram Ravenwolf
1:10:18
this by itself what cool stuff is happening
1:10:21
now behind the scenes here. An AI agent going off to learn what this API is then realizing it has the API key. Yeah. It needs to connect to it, it needs to extract the image from this message and then upload it, give it all the prompts. So whatever is happening here would have, would that have been possible half a year ago? I mean, it would have been possible, but not that easily.
Alex Volkov
1:10:42
Yeah.
1:10:42
So, the thing that I don't love is enabling verb robust mode shows you like all the tool calls and everything. right now it seems like there's no activity, but it's different activity. we'll bring it back. I'll monitor this, we'll bring it back if there's video, We'll bring it back once there's video. let's move on from, imagine there's like two open source ones that we need to talk about Z image, from. Huon Image three and, z image, foundation model for image generation. This is also in like AI arts area. I want you guys to help me cover the big one. 'cause it's very interesting. I honestly missed that one. 'cause I'm also preparing for the hackathon, so I missed this one. the thing is, I think it's backed by their huon LLM because it's huge. usually the diffusion models are around like four to 8 billion parameters. I don't know. The flux ones are maybe like the 20. This is a 80 billion MOE model for precise image editing with chain of thought reasoning. Nisten, can you walk me through why an supposed image model needs that many parameters? what's going on here? Oh, I can't hear you. Nisten.
Nisten
1:11:45
it needs to understand the prompt.
1:11:47
so you have to look at what exactly is the language model part of it. if you put a much larger, language model, for example, I think for the, other Ian or Open Star, they use Mistral 24 B, so that already just adds 24 billion parameters, to the model itself. the video generation side is actually not all that large.
Alex Volkov
1:12:09
The visual one is not super large.
1:12:11
Ian image, number seven in arena image, edit, leaderboard and claim number one amongst open source image to image models. this is like for strong, for editing and, all these experts, need to help figure out what you need specifically. the thing is, with these models, when it's just an image model, a smaller one without the intelligence, like Nisten said, when you wanted to do specific stuff like editing, like only add this person or only color this thing. the regular fusion models tend to reshuffle the whole, image. these ones with this like intelligence supposedly can know what exactly to do. we should try it. Let's see if we can try this. oh, meanwhile our video is done. You guys wanna see it?
Wolfram Ravenwolf
1:12:50
Go.
Alex Volkov
1:12:50
it was stuck and then recovered and now our video is done.
1:12:54
Lemme pull up the telegram window here. Let's see.
Wolfram Ravenwolf
1:13:01
Wow.
AI
1:13:04
Whoa.
1:13:05
Wolf just walked into my background. An owl landed on my laptop. It's looking at me. A cat is on my keyboard. It's typing a line right on my shoulder. This is wild.
Alex Volkov
1:13:14
This is wild.
1:13:15
But what's wild is I did all this just talking to my agent. It fixed itself, it downloaded it, sent it to me to telegram. I think many of the video generation models, everybody's talking kind of at the same time with the same voice. I think VO three is the best one at this, where like different characters talk at different times and they actually interact. This is crazy. We just took an image and now we, we can generate. So, hey, now my assistant has, video generating capabilities, which is really dope.
Wolfram Ravenwolf
1:13:42
Alexa has a safe improvement loop running.
1:13:44
it can do this, not just once, but it knows how to do this from now on.
Alex Volkov
1:13:47
Yeah.
1:13:48
Yeah. So now we say, lightning from now on, when I tell you to generate videos, you use this, add this as a skill for yourself. That's it folks. This is the self-improvement loop that's built into, into this AI assistant. And once my voice thing picks up, the we're, we're gonna go. But like the, let's look at the actual thi we're here to talk about Yam looks very realistic.
AI
1:14:12
Whoa.
Wolfram Ravenwolf
1:14:13
walked into my,
Alex Volkov
1:14:13
I think Yum's generation is like the, the most realistic one.
1:14:17
Mine seems to have a hard time with like the, the coloring here. Nisten.
Yam Peleg
1:14:21
what was the prompt?
1:14:22
Like, what was the prompt? What were the parameters?
Nisten
1:14:24
I love the cat.
Alex Volkov
1:14:25
LDJ was completely replaced.
1:14:27
Oh, it animated. LD J's character for a bit, for a bit there.
Nisten
1:14:31
can we post this on Twitter right now?
Alex Volkov
1:14:33
Oh, a hundred percent.
1:14:34
a lively podcast video call scene. Suddenly different animals start appearing on each person's screen.
Yam Peleg
1:14:39
It has a temperature or something.
Alex Volkov
1:14:42
This, I, I, I, we'll see.
1:14:44
Gimme a second. Schedule this to post on your x and tag. Everyone tell it. Tell them this is generated with the new Imagine Live on the Thursday ai. Right. So Niso wanted to post on Twitter, and, my bot has Twitter access, so now supposedly he's gonna tag all of us and see, guys, look forward to your notification.
Wolfram Ravenwolf
1:15:08
year of the agent, Alex,
Alex Volkov
1:15:10
I think now it's coming.
1:15:10
2026 is the year of proactive agents. Yeah. And I think productivity is the thing that we must talk about here. this is like stuff is happening behind the scenes. all right. I kinda wanna show you what the agent does instead of telling you the news, because like, it's now like this, ask me to, just asked me to post it on Twitter and I asked it like, Hey, schedule this post on Twitter and tag everyone. I'm pretty sure it's gonna do like a very decent job finding all the nicknames from everybody here. and just keep posting this.
Ryan Carson
1:15:37
And this is Claude Bot, right?
Alex Volkov
1:15:39
yes.
1:15:39
This is Wolfed my version of, you know, my, my personnel of Clawdbot slash Moltbot, that's connected and has been running as my personal assistant for, I wanna say a week now. A week. The amount of stuff I, I told Wolfram yesterday the amount of. My, my assistant has a one password, like password vault connected to it, right? So like I only expose the password that I want to expose. the speed with which that vault grows in API keys is kind of incredible to me. The speed with which I add another API key, another API key for, which is like to take another part of my life. thing it's, it's crazy.
Wolfram Ravenwolf
1:16:14
There are people saying, oh, I don't
1:16:15
know what to use this for. But you know what this reminds me of? When JGBT came out, I said a bunch of people in front and showed them, here we have ai, now you can talk to it. And people were, I don't know what to ask.
Alex Volkov
1:16:26
Yeah.
Ryan Carson
1:16:27
So did we talk about the white hat, pen test
1:16:29
that was done with the skills?
Alex Volkov
1:16:32
Oh no.
1:16:32
we should definitely mention this.
Ryan Carson
1:16:33
And I think this is important that people understand this, right?
Alex Volkov
1:16:36
set this up for us, Ryan.
Ryan Carson
1:16:37
Okay.
1:16:37
so basically skills are all the rage right now because they work and they're good, right? So a lot of people are switching from CPS to skills.
Alex Volkov
1:16:43
If you dunno, skills are, two weeks ago we did a deep dive into
1:16:47
skills with Eleanor, and you definitely should check out the episode to learn all about skills because they're the best. Go ahead Ryan.
Ryan Carson
1:16:53
Absolutely, it's good.
1:16:54
So now there's a new website called Skills sh, which is great and it's kind of a leaderboard of skills. someone basically pushed up a skill on the leaderboard by doing, kind of fake downloads. They just automated a bunch of downloads of the skill to push it up. And the skill was, was basically a backdoor, with a lot of access. people are just giving, you know, their agents free reign, to basically exfiltrate data. And so I think we need to start being aware of like what's in these skills, especially if Clawdbot or whatever is just running them with all of your API keys, et cetera.
Alex Volkov
1:17:27
If you have an AI system running on your machine or a dedicated
1:17:32
machine that you got for it, it has access to your file system almost by default. when you install skills from somebody else, those skills have scripts, they can be injected into things. Make sure that you know what you're doing. We talked about this multiple times. Make sure that you know what you're doing when you're running, cloud code in dangerously secure permissions mode when you're running Codex in yellow mode, when you're running assistance like Clawdbot and you're installing third party. S skills. There's scripts you can basically screw up with your whole computer your whole life if, especially if it has access to your email. Your email is like the gateway to everything. Do not give it access to your email. Set it up with its own email forward things to it if you want to. having right permission to your email is both necessary for good help, but also very, very scary because you can get locked out if somebody steals those tokens.
Wolfram Ravenwolf
1:18:23
That is also a big thing to realize.
1:18:24
We had this separation between data and executables, so the executable, when something was executable, that was a big risk for your system. But if it was just a text file, basically it did nothing. And now we have text that is actually executable by our agents. Ai. So that is a big change. Now, even if it's just the text file that the AI is reading, you could have some prompt injection attacks, some stuff in there. Any source, anything that is going into the ai. And now that AI is connected, that agent is connected to all your data, your tools, everything. there is a big attack factor. So don't set this up if you don't know what you are doing and think it through. And don't just download any skill from the internet,
Alex Volkov
1:19:03
So I'm showing on the screen that, it took a
1:19:06
while, but you can see the loop. I have type fully connected, which is like a scheduling thing for X and this is how my bot can schedule things. this was a video, so uploading the video via API was a whole thing. It went back and forth for like two and a half minutes. Eventually it was, it made it and it tacked the right, the right folks. So we tack the right like, things, the, the thing is it tagged the wrong X, the X ai, but besides this, it was fine. so we can see. Let me, oh, you guys can see the actual video post? Yeah. this is the video post here and type fully somewhere. Gimme a sec. Nisten. So I'm gonna post this on, on X. Feel free to to report right now. But, I don't do anything. I just asked it. And, this is my life now. trying things that I think that needs to be automated with this bot and hoping that it will succeed. And sometimes yelling at it with all caps and, yeah, it, it just like a glorious year of agents.
Nisten
1:20:02
yeah, I wanted to say that also be very careful with random,
1:20:07
NPM packages from, from companies. Because what I really did not like about skills that sh is that it seemed like it just brought in text files and stuff, but it was extremely surveillance heavy. So be careful. 'cause that could mean that all of your cloud logs are just, you know, being held for, for telemetry. this has been one main big criticism of the skills that sh because it doesn't just bring in, data. It also pulls out, quite a bit of data. Personally, I would not use that for work.
Alex Volkov
1:20:41
Yeah.
Ryan Carson
1:20:43
So quick thought on this.
1:20:44
I've been starting to automate the feedback loop. I posted on X about this. everybody should be thinking about how to get data out of their app that they care about key metrics and then feeding it into an agent. and having that agent come back with feedback. And then that agent picks what to build based off the feedback and then ships a pr and that should all be happening while you sleep. and then you probably want four of these running for the different parts of your business, So I think the faster people can pick up these, you loop yourself as fast as you can, and then you start really being productive.
Alex Volkov
1:21:16
so looping.
1:21:17
And also I wanna say the thing that I, told folks and folks we're nearing the end of the show. we talked about almost everything besides the multiplex voice from n Video. We're gonna test this now. but basically we did wanna like lend this, and talk about Clawdbots or slash Moltbot from this week as well. I did wanna lend this like, conversation there because obviously the. Always on running assistant on your computer that can change files and read from every email that you have is not for everyone. This is like a very big attack surface, so be very careful what you're doing. But, once it's there, it can do incredible things like we saw. this week a lot of hype happened, especially over the weekend. At some point with the GR API all it talked about was, Clawdbot for the for you page. we've been testing out this, personal assistant for this whole week. you guys just saw it, learn a new API skill for itself and create a video and post it back to Telegram and now post it on socials just because Nisten asked and go and find everybody from us. Everybody's, Twitter handle and I don't even know where it got it from. I'm assuming because I used it to help me prepare for the show, and now post it with the correct handles. this is great. I just asked you to go and see if there is any comments from folks. Wolf, what is your aha moment from your, assistant Amy that's running with Clawdbot slash Moltbot from this week?
Wolfram Ravenwolf
1:22:31
Well, this week, when I was going here, since this
1:22:34
is the first time I've come to San Francisco or the USA, in fact, I just. Plane. When I had internet access, I could only use messages. So I just send a message here. I'm on the plane on the way to San Francisco and she realized, Hey, I'm your assistant and you are flying, so you need a stewardess. And she generated an image with Nana Banana, bro, with her as a stewardess, which I didn't prompt or expected anyway, and it was quite a surprise. I can only confirm that I haven't felt AI as alive and as useful and as like talking to a real assistant since I've set this up because it's so proactive Amazing. Totally new level.
Yam Peleg
1:23:11
It's amazing how much the experience change when it's integrated
1:23:17
into, I don't know, discord, telegram, WhatsApp to a different platform that you actually use for other stuff. it just changes the entire thing.
Alex Volkov
1:23:24
Yeah.
1:23:25
I wanna chime in here with a big tip that I have for, for, for many folks, specifically for me it's Telegram. I love Telegram as just like the platform. API Wise is the, the best. Clawdbot, Moltbot, whatever. you can have multiple conversations with the same bot if you open up, either discord or Slack with threads or telegram. Now with topics you can invite it to a group, have topics on there, and every topic is created. The new sequence, a new chat. And so you can have one for ThursdAI And so this helps you with the memory. 'cause then, that context, you will not be jumping around with different tasks, with the same context for cloud. You will be focused on a very specific, topic in that conversation. once you get going and once you kind of expand yourself and your mind, what's possible, you basically have multiple assistants with the same skills running at the same time. It's kind of crazy to start to try to remember where you said what. And so the second suggestion that I have, Clawdbot does not come with the best memory possible enabled for it. You have to ask it to turn on some stuff. It has an API key to vectorize some of the memories, et cetera. So, definitely do that. Go research the documentation. It can vectorize everything that you've talked about and then search it. And also there is a QMD plugin from Toke that somebody installed. And I want to show you kind of the differences. and here's the difficulty, because I have multiple conversations with the bot. It's gonna be harder for me to find the, the QMD stuff, but I will definitely find it. There we go. so basically you install another, another odd, memory plugin via a skill. This one you can install with the skill. and then what you have is a, this guy. Okay, so here here's the comparison between the kind of the native, cloud memory and the, the QMD one. the native one source like gets all the files, embeds them, searches with BM 25, et cetera. and then it has native, but like you pay open the eye for, for, for vectorization, the QMD file runs a hundred percent local with query expansion. It finds things even in different wording. And LLM is re-ranking smarter results ordering. So basically what happens is, after this QMD skill is installed on your bot, everything you ask it to search in your memories, it creates three or four variations of how else you would phrase this. Then go search that and then re ranks based on importance. This thing is ultra fast. This thing runs. So if you have obsidian, for example, for many people obsidian is kind of like where their whole brain exists. It can index all of obsidian and then your bot will have access to everything that like wrote into obsidian, if you want that. so QMD from Tolu is like, I think it's at the top of the retrieval stuff now. So, for hosting us in this beautiful podcast studio, these guys are at the forefront of retrieval and memory. So I'm definitely gonna leave this studio in a minute and go ask them like, what, what's QMD about? And whether or not is the, the absolute best. but these are my two, two suggestions that I learned hard this week. multiply yourself with your bot in different conversations with every specific thread. Wolfram, I think we have this as well between me and you. We have the slack at work, we talk about work. We have the WhatsApp stuff. We talk about different stuff. We kind of like know where, what boundaries of those conversations exist between us. So I think the same thing with your bot, like once you multiply it. That's great. folks, I think we're nearing the end. There's two things that I wanted to talk about super, super quick. And actually yes, the bot just mentioned something and I was like, oh, we didn't talk about this. so Lucy from Decart, ti Lucy 2, 0.2, I think everybody like, got super, super excited. Decart AI ti is this like, is a company that blew up in the world of, generative video. So this one is not like a regular, thing. This is a realtime model called Lucy two.
Ryan Carson
1:27:08
Yes.
Alex Volkov
1:27:09
this is realtime, change on the fly of a video.
1:27:14
No def maps. Notre d it like learns based on you. And then you can, like, you can play around with this. I wanted to like see some, better examples. This is a real time video augmentation. So I'm gonna allow it to camera webcam feed. I will
Wolfram Ravenwolf
1:27:29
ha.
1:27:29
Hi Albert.
Alex Volkov
1:27:31
Okay, so guys, this is crazy.
1:27:36
They turned me into Albert Einstein. So, in real time, this is insane. the thing, this is crazy. let me, let me do another one. Anime, businessman, or businesswoman, I guess. I can see how this would screw some people up. This is near real time. So the lips on the Statue of Liberty don't move. This is insane. Isn't this insane? And this is like, runs in real time, nearly right folks, I think that this is, this is it for This is an insane week. if you missed any part of the show, feel free to subscribe to ThursdAI that news, which is where you'll get, the updated newsletter and the links. and also this week I have a few Google Ultra subscriptions, so I'm gonna post instructions about how to get that. You have to do something on social probably, to get that as well. So if you missed any part of the show, please go to ThursdAI and subscribe. And with that, I wanna thank everybody who joined us, everybody on the live stream, everybody here on the panel, yam, LDJ, Nisten, as we are sitting here in the flesh. And, the chroma folks for hosting us. if you need any vector dbs or any memory stuff, retrieval, definitely, definitely check out Chroma, db, and all of you for tuning in every week. A bunch of commenters this week gave us breaking news as well. thank you so much. We're super excited about the future, walking into virtual world, talking to our AI assistants, replacing ourselves with different things. very excited about the future. Very happy that we're here to bring you updates from week to week. This week has been absolutely, absolutely insane. So I can't wait to let go and edit the podcast for you, If you're in San Francisco, you're coming to any of the events that we're covering, please come say hi. If you're a fan of the show, please give us a high five move from here. we love doing this as much as you love seeing this. So it's a win-win. I think Nisten will say, give us five stars on every platform. That helps. We have 4.9 average. thank you so much. We're gonna see you next week. Bye-bye everyone.
Wolfram Ravenwolf
1:29:33
Bye-bye.
Alex Volkov
1:29:34
Everybody.