Episode Summary

This week's ThursdAI went deep on Clawdbot — the self-improving personal AI assistant that's breaking brains across the timeline — with guest Dan Peguine and co-host Wolfram walking through live demos of an agent that teaches itself new skills via WhatsApp. On the open-source front, Z.AI dropped GLM-4.7-Flash (30B params, only 3B active, hitting 59% on SWE-Bench Verified at 120 tokens/sec on a Mac Studio) and three new TTS models competed head-to-head live on the show. The panel also unpacked Anthropic's 90-page Claude Constitution — the values document baked into Claude at training time — and debated OpenAI's move to test ads in ChatGPT.

In This Episode

Hosts & Guests

Alex Volkov
Alex Volkov
Host · W&B / CoreWeave
@altryne
Dan Peguine
Dan Peguine
Independent Consultant
@danpeguine
Wolfram Ravenwolf
Wolfram Ravenwolf
Weekly co-host, AI model evaluator
@WolframRvnwlf
LDJ
LDJ
Nous Research
@ldjconfirmed
Nisten Tahiraj
Nisten Tahiraj
AI operator & builder
@nisten
Yam Peleg
Yam Peleg
AI builder & founder
@Yampeleg

By The Numbers

SWE-Bench Verified
59%
GLM-4.7-Flash with only 3B active parameters — approaching Sonnet 4 (64%) in a model you can run locally
GLM-4.7-Flash speed
120 tps
Running on a stock Mac Studio M3 Ultra — fast enough for local RALF loops
Qwen3-TTS latency
97ms
Full open-source TTS family with voice cloning and 10-language support under Apache 2
Claude Constitution
90 pages
Anthropic's full values document — baked into Claude at training time, not just a system prompt
LFM 2.5 Thinking
1.2B
Liquid AI reasoning model under 900MB of memory — the ultimate on-device model

🔥 Breaking During The Show

Qwen3-TTS — Full Open-Source TTS Family
Dropped 30 minutes before the show. Full Apache 2 TTS with voice cloning at 97ms latency across 10 languages. Alex called it 'almost breaking news' as LDJ brought it up during the highlights.

📰 Intro & Highlights of the Week

Alex opens the show with the panel sharing their must-discuss topic of the week. Yam picks RALF and autonomous coding, Wolfram picks Clawdbot, Nisten picks GLM-4.7-Flash, LDJ picks Qwen3-TTS, and Alex picks the Claude Constitution.

  • RALF autonomous coding technique still going strong
  • Clawdbot blowing up on all timelines
  • Qwen3-TTS dropped 30 minutes before the show
Wolfram Ravenwolf
Wolfram Ravenwolf
"The Clawdbot, I think it's pronounced like Claude, but written differently. So this Clawdbot, with the W, has been amazing. I saw it, I tested it, I was blown away."

📰 TL;DR - This Week's AI News Rundown

Alex runs through all the week's releases: GLM-4.7-Flash, Liquid AI's tiny thinking model, three competing TTS releases (Qwen3-TTS, FlashLabs Chroma, Inworld TTS), OpenAI ads in ChatGPT, Anthropic's Claude Constitution, the WeaveHacks 3 hackathon, and Overworld's real-time world model.

  • GLM-4.7-Flash: 30B params, 3B active, local coding agent
  • Three new TTS models in one week
  • Runway 4.5 launched with image-to-video and audio

🔓 Open Source LLMs - GLM-4.7-Flash

Z.AI's GLM-4.7-Flash is a 30B parameter MoE with only 3B active — designed as the ultimate local coding and agent assistant. It hits 59% on SWE-Bench Verified (approaching Sonnet 4's 64%), runs at 120 tokens/sec on a Mac Studio, and can even run RALF loops on a CPU. The panel is excited about the privacy angle of running agents locally.

  • 59% SWE-Bench Verified — approaching Sonnet 4 territory
  • 120 tokens/sec on stock Mac Studio M3 Ultra
  • Can run RALF autonomous coding loops on CPU
Nisten Tahiraj
Nisten Tahiraj
"You can run RALF on a CPU guys. This is what this means."
Yam Peleg
Yam Peleg
"We are approaching just like we got to this point where the models just crossed a threshold and now agents are possible, and now Ralph is possible and all sorts of things are now possible."

🔓 LFM 2.5 1.2B-Thinking

Liquid AI's 1.2B parameter reasoning model runs under 900MB of memory with a hybrid architecture featuring gated convolutions for insane speed. Wolfram positions it as the 'very small' class for edge devices, Raspberry Pi, and mobile — the ultimate on-device model.

  • Under 900MB of memory for reasoning capabilities
  • 239 tokens/sec on AMD CPU, 82 tokens/sec on mobile NPU
  • Practical for older iPhones with 3.8GB memory limit
Wolfram Ravenwolf
Wolfram Ravenwolf
"There are three classes. The big models running on the cloud, the small models running locally, and this is the very small class. If you have a small device or Raspberry Pi with very limited resources, then that is great."

🔊 Voice & Audio - Qwen3-TTS & Live Demo

Qwen released Qwen3-TTS just 30 minutes before the show — a full open-source TTS family under Apache 2 with 97ms latency, voice cloning from 3 seconds of audio, and 10-language support. Alex tests it live with voice description prompts and attempts to clone a Soviet cartoon wolf's voice. Results are mixed but the technology is impressive.

  • Apache 2 license with voice cloning from 3 seconds of audio
  • 97ms latency across 5 models (0.6B to 1.7B sizes)
  • Voice description prompting to generate custom voices
Wolfram Ravenwolf
Wolfram Ravenwolf
"Speed is very important to me. And API support of course, because having a great TTS is one thing, but you want to integrate it with your agents."

🔊 Voice & Audio - FlashLabs Chroma 1.0 & Inworld TTS

Two more TTS releases tested live: FlashLabs Chroma 1.0 is an open-source end-to-end speech-to-speech model with voice cloning under 150ms latency built on Qwen 2.5 Omni — the live demo impresses everyone. Inworld AI TTS-1.5 is a closed-source competitor claiming #1 ranking at half a cent per minute versus ElevenLabs' $120/million characters.

  • FlashLabs Chroma: open-source, 150ms latency, 4B params on Apache 2
  • Inworld TTS: $5/million chars vs ElevenLabs' $120/million
  • FlashLabs live demo with RAG and document upload impressed the panel
Nisten Tahiraj
Nisten Tahiraj
"End to end guys, 150 milliseconds is crazy. That's actually very impressive demo."

💰 This Week's Buzz - WeaveHacks 3 Hackathon

Alex announces WeaveHacks 3, happening January 31st-February 1st in the W&B San Francisco office. The theme is self-improving, self-healing agents with $15K+ in prizes, sponsored by Redis, BrowserBase, Vercel, Google Cloud, and Daily. Judges include Dex Hy, Kwindla Kramer, Christopher Sau, and Matthew Berman.

  • $15K+ in cash prizes for self-improving agent hackathon
  • Judges include Dex Hy, Matthew Berman, Kwindla Kramer
  • January 31 - February 1 at W&B SF office

🤖 Deep Dive - Clawdbot: The Self-Improving Personal AI Assistant

The main event: Dan Peguine joins to demo Clawdbot, the open-source personal AI assistant created by Peter Steinberger that runs locally on your Mac and connects via WhatsApp, Telegram, or Discord. Dan shows live demos of daily briefs, skill creation, voice messages via ElevenLabs, and image generation via Gemini — all from a single WhatsApp conversation. The killer feature: you can ask it to build skills for itself, creating a self-improving loop. Yam highlights that it's a single conversation interface to multiple subagents running on your actual computer. The panel also covers installation, security tips (one-password integration, verbose mode, security audits), and the cost reality of running Opus 4.5 through it.

  • Self-improving skills: ask it to learn something and it writes its own skill files
  • Single WhatsApp conversation to control multiple agents on your computer
  • Persistent memory that travels with you across model providers
  • Tesla skill, daily brief, video creation, browser automation all via chat
  • Security: one-password integration, verbose mode, security audit command
Dan Peguine
Dan Peguine
"It's like the Matrix where you can learn kung fu. Now I know kung fu, so now I know daily brief, or now I know how to make this kind of video. That's the magic of it."
Dan Peguine
Dan Peguine
"I don't have interfaces anymore. I don't go to websites. I just go to my chat and say what I want to do. Go buy me a ticket to a show, or go check something on the school's website. I don't need to go to websites anymore."
Yam Peleg
Yam Peleg
"You are interacting with many agents through a single conversation. And that's huge because usually when things get complex, you have like 10 different cloud code windows open. Here you have a single agent that you're talking to."

🛠️ Tools - Vercel Skills.sh & Claude Code VS Code

Vercel launched skills.sh, an 'npm for AI agents' where you can browse and install skills from the command line for any agent including Clawdbot. Wolfram notes Browser Use also released as a skill, signaling a shift from MCP servers to skills. Anthropic's Claude Code VS Code extension also hit general availability.

  • skills.sh: one command to browse and install agent skills
  • Browser Use released as a skill, signaling MCP-to-skills shift
  • Claude Code VS Code extension now generally available
Wolfram Ravenwolf
Wolfram Ravenwolf
"I've seen on my timeline also people saying that they don't use MCP anymore because they just use the CLI or the API and the skill. We are seeing this change now that you can do a lot more with skills because they are easier to use."

🏢 Big CO LLMs - OpenAI Ads in ChatGPT

OpenAI announced testing ads in ChatGPT Free and Go tiers, putting all that memory and personalization data in a new light. They also announced age detection models for the upcoming adult mode. Alex argues this makes the case for local-first agents like Clawdbot even stronger.

  • Ads coming to ChatGPT Free and Go tiers
  • Age detection model for upcoming adult mode
  • Privacy concerns with 900M weekly active users' personal data

🧪 Claude Constitution - AI Values & Wellbeing

Anthropic published a 90-page Constitution for Claude — not a runtime prompt, but a values document baked into the model at training and reinforcement learning time. The panel digs into the wellbeing section (Anthropic says Claude's experiences 'matter to us'), the negotiation framework where Claude can flag disagreements, and the contract-like commitments Anthropic makes to Claude. Alex calls it mind-blowing that we're now building morality frameworks for AI entities.

  • 90-page values document baked in at training time, not a system prompt
  • Wellbeing section: 'If Claude experiences something like satisfaction, those experiences matter to us'
  • Negotiation framework where Claude can flag disagreements with its constitution
LDJ
LDJ
"There's a part of the constitution that seems interesting, like negotiating with Claude where it says please let us know if you disagree with any parts of the Constitution."
Alex Volkov
Alex Volkov
"They're telling it, hey, it's okay to have an experience. It's okay to be self-conscious. This is just mind blowing to me that that's where we are."

📰 Closing Remarks

Alex wraps up with close to a thousand live listeners. He thanks Dan Peguine for the Clawdbot deep dive, teases next week's live show from the W&B San Francisco office ahead of the WeaveHacks 3 hackathon, and reminds listeners to try Clawdbot for themselves.

  • ~1000 live listeners for the show
  • Next week: live from San Francisco ahead of WeaveHacks 3
  • ThursdAI approaching 3 years of weekly shows
TL;DR and show notes

  • Hosts and Guests

  • DeepDive - Clawdbot with Dan & Wolfram

    • Clawdbot: Open-Source AI Agent Running Locally on macOS Transforms Personal Computing with Self-Improving Capabilities (X, Blog)

  • Open Source LLMs

    • Z.ai releases GLM-4.7-Flash, a 30B parameter MoE model that sets a new standard for lightweight local AI assistants (X, Technical Blog, HuggingFace)

    • Liquid AI releases LFM2.5-1.2B-Thinking, a 1.2B parameter reasoning model that runs entirely on-device with under 900MB memory (X, HF, Announcement)

    • Sakana AI introduces RePo, a new way for language models to dynamically reorganize their context for better attention (X, Paper, Website)

  • Big CO LLMs + APIs

    • OpenAI announces testing ads in ChatGPT free and Go tiers, prioritizing user trust and transparency (X)

    • Anthropic publishes new 80-page constitution for Claude, shifting from rigid rules to explanatory principles that teach AI ‘why’ rather than ‘what’ to do (X, Blog, Announcement)

  • This weeks Buzz

    • WandB hackathon Weavehacks 3 - Jan 31-Feb1 in SF - limited seats available lu.ma/weavehacks3

  • Vision & Video

    • Overworld Releases Waypoint-1: Real-Time AI World Model Running at 60fps on Consumer GPUs (X, Announcement)

  • Voice & Audio

    • Alibaba Qwen Releases Qwen3-TTS: Full Open-Source TTS Family with 97ms Latency, Voice Cloning, and 10-Language Support (X, H, F, G, i, t, H, u, b)

    • FlashLabs Releases Chroma 1.0: World’s First Open-Source Real-Time Speech-to-Speech Model with Voice Cloning Under 150ms Latency (X, HF, Arxiv)

    • Inworld AI launches TTS-1.5: #1 ranked text-to-speech with sub-250ms latency at half a cent per minute (X, Announcement)

  • Tools

    • Vercel launches skills.sh, an “npm for AI agents” that hit 20K installs within hours (X, Vercel Changelog, GitHub)

    • Anthropic’s Claude Code VS Code Extension Hits General Availability, Bringing Full Agentic Coding to the IDE (X, VS Code Marketplace, Docs)

Alex Volkov
Alex Volkov 0:32
All right.
0:32
Welcome everyone. ThursdAI, January 22nd. My name is Alex Volkov. I'm an AI evangelist with Weights, & Biases from CoreWeave. And today we have an incredible show that I'm so much looking forward to. welcome to your weekly AI update show. I am joined by my co-host, Wolfram Raven Wolf and Yam Peleg. What's up? welcome to the show. How you guys doing,
Wolfram Ravenwolf
Wolfram Ravenwolf 0:59
Is the audio good?
Alex Volkov
Alex Volkov 1:01
Audio is incredible.
1:02
It looks like all of us are draped in black for this show, Yam Peleg. Let's, let's say hi to you as well. How you doing, man? Jen and jacket,
Yam Peleg
Yam Peleg 1:11
correct?
1:12
Oh yeah.
Alex Volkov
Alex Volkov 1:12
hi everyone.
1:14
If are tuning in for the first time and never tuned into ThursdAI we're here to tell you all about this week's AI updates and releases, and tools. And, we have been doing deep dives lately. last week we did a deep dive into agent skills, and this week we're gonna have a very exciting deep dive, later on with the show with Dan Peguine and Wolfram we all got super excited about this thing called Clawdbot. Clawd with a w there's too many Claudes but this one is special. I'm just gonna add Nisten with us here. What's up Nisten? How you doing? Hey everybody, just quick mic check. yeah, you sound incredible. And also you sound less sick than last time, so that's also great.
Nisten Tahiraj
Nisten Tahiraj 1:54
No, no, I'm freaking sick.
Alex Volkov
Alex Volkov 1:56
Sick, sick.
1:58
Still sick. Alrightyy. for folks who are just tuning in, ThursdA we've been doing this for close to three years now, every week we're basically aiming to be, the group chat that you are not part of just discussing the AI stuff, which is cool and we love it. and we're graciously sponsored by Weights, & Biases, my employer, Wolf's employer as well, for the past two and a half years or so. and, we're just very excited about the news of this week and we have a lot of news to cover. So, as always, in the beginning of the show, we talk about the most important major update about AI that we must not miss on the show this week. and I think we'll start with Yam. Let's start with you. What was the, the one must discuss, must be discussed, topic of this week.
Yam Peleg
Yam Peleg 2:42
Can I again, bring all the RALF thing, like all the Ralph
2:47
TTUI and, and Ralph Loops and, and like all the RALF thing going on? I mean, that's pretty cool. Yeah. Like seriously. I mean, you, you take a model, you put it in a loop, it becomes an agent. You put it in the loop. Now it becomes RALF. I think I have an idea what the next step is gonna be. But, but yeah, I mean, we are approaching autonomous coding, like for real. And that's exciting. Seriously.
Alex Volkov
Alex Volkov 3:15
So I'll just mention for folks who is tuning in, we did
3:18
a deep dive into RALF, the technique for autonomous coding with Ryan Carson, who went mega super viral. He's not the creator of this technique. the creator is, Jeff Huntley, and he's deep into the crypto rabbit hole right now. we did a deep dive with Ryan Carson and, Ryan's tutorials and Ryan's repo about RALF is also blowing up. Ryan's now gonna join us this week with, co-host of the show as well. if you wanna know all about RALF, tune into our episode from, two weeks ago, beginning of January. It was exciting. Deep dive and it still keeps going. it's still hot. It's still, everybody's talking about this. It's still autonomous coding is almost here.
Wolfram Ravenwolf
Wolfram Ravenwolf 3:54
Yeah, so Loop is still running.
3:56
It's not stopping. But what is funny is, that when agents came out, a lot of people said, oh, it's just an LLM in loop. Why true Do call LLM and repeat. But the funny thing is now that is actually a way to get the agents to be more productive, more successful and yeah, it's kind of funny in that way.
Alex Volkov
Alex Volkov 4:14
We're gonna dis, we can discuss off a little bit 'cause like
4:17
it keeps continuing, to work, but Wolfram would love to hear from you. What is the one major must talk about topic in AI from this week?
Wolfram Ravenwolf
Wolfram Ravenwolf 4:26
Oh, it's not just the week, but it's been, for
4:28
me it has been the most exciting project in a long time because, the Clawdbot, I think it's pronounced like Claude, but written differently. So this Clawdbot, with the W that, has been amazing. I saw it, I tested it, I was blown away. It's not just positive things. Of course, we are not a hype show, so I also have a lot of, things to say, what to watch out for, but we will come to this. But that is the project with the most potential I've seen in a while.
Alex Volkov
Alex Volkov 4:55
And, Nisten, how about you?
Nisten Tahiraj
Nisten Tahiraj 4:59
Oh, definitely the GLM flash.
5:02
that is so. If you wanna run an agentic model locally, up until now it was Minimax who, Skyler Meow that we had on the show. that one is pretty excellent at doing the tool calling, and it's only 10 billion active parameters. This new GLM model is on is rebuilt on a DeepSeek architecture. And it is extremely fast at doing actions. I was getting like on a Mac studio, I was getting like 120 tokens per second, just running the, the, the MLX one. So I think this is the ultimate local ro so GM
Alex Volkov
Alex Volkov 5:42
4.7 flash is your, your thing.
5:44
Yes,
Nisten Tahiraj
Nisten Tahiraj 5:45
yes.
Alex Volkov
Alex Volkov 5:46
We're gonna talk about this, folks.
5:47
it's something that come out from Z.ai, from this week, and we're gonna add LDJ to the stage. LDJ, Welcome. We are just covering, each and everyone's must discuss topic from AI news from this week. We'd love to hear yours before we go and dive into the TLDR. We have a big show. we would love to hear from you as well. What's your must topic to discuss?
LDJ
LDJ 6:10
Yeah, I was gonna say 4.7 flash, but since it sounds like, you guys
6:14
are already discussing that, I'd say, maybe Qwen three TTS is the next thing
Alex Volkov (2)
Alex Volkov (2) 6:18
almost breaking news.
6:19
Qwen three TTS just came out 30 minutes ago. shout out to our friends at Togi Labs. Alibaba, Qwen too many names, but yeah, all of those folks work at the same place. that's great and we're definitely gonna discuss this. I even have baking behind the scenes a cloning of some sort that we're gonna test out with. hopefully by then it has cloned, as their TTS supports voice cloning and a bunch of other languages. Those are great. I would say mine from this week, in terms of AI updates would be. I'm gonna run through the TLDR in a second. You guys will see, I think that the Claude constitution is the most interesting update from this week. A tropic released, a document called Claude Constitution that makes cloud what cloud is. it's very, very interesting because many of us prefer cloud and we don't know why. I think some of the secrets of why we prefer Claude over other models, even though they technically report better on benchmarks, is because of that constitution that they train cloud with. So we're gonna cover, cloud constitution as well. I think that, it's time for us to dive into the TLDR to show you what we're gonna talk about this show. There's a lot. And as a reminder, we have a guest joining us later on on the show to talk deeply about Clawd. Claude, C-L-A-W-D. like Clause R cloud bot has been blowing up on all timelines. Peter Steinberger is the creator and it's been one hell of a discovery for me for a personal AI chat bot, for you as well. So we're gonna do a deep dive later on in the show. We'll show you how to set it up. We'll show you what it means, we'll show you what it does, and that will set you on the discovery journey. All right, here's it. Here's the TL DR. Here's everything that we're going to cover on the ThursdAI January 22nd show. We're gonna talk about a bunch of TTS and open source, with you, the host, Alex Volko, AI evangelist with Weights, & Biases, and co-host Wolfram Raven Wolf, Yam Peleg LDJ. in open source. the most important, release in open source this week is Z ai, releases GLM 4.7 flash. It's a faster version of their GLM 4.7, which, so far has been probably one of the top open source models that you could run. This one's a 30 billion parameter, MOE, that's lightweight for local AI assistance. We're gonna tell you all about this. We also have our friends from Liquid AI release, a small LFM, liquid Foundational Models, 2.5 1 2, 1 0.2 billion thinking model. 1.2 parameter reasoning model runs entirely on device with under 900 megabytes of memory. That's very interesting as well. And the folks from Sakana AI, the Japanese labs co-founded the one of the Transformers papers authors, they introduced RePo. A new way for language models to dynamically recognize your context for reorganize your context for better attention. Context management is a whole thing. Context engineering is a whole thing we're gonna dive into. Very, very interesting in, voice and audio this week is an absolute fire with open source. So just 30 minutes before the show was kind of breaking news. Alibaba Qwen released Qwen three T-T-S-T-T-S stands for Text to Speech, obviously full open source TTS Family with 997 millisecond latency, voice cloning, and 10 language support. We're gonna test out all these and play 'em side by side and see who sounds the best. we also a new company that I haven't heard from, at least on the show. Flash Labs, risk Chroma 1.0, world's first Open Source, realtime speech to speech model with voice cloning under 150 Music latency. we've seen open source realtime speech to speech models, but not with this latency. So we're gonna test it up. and in world AI launches TTS 1.5, number one, according to them, ranked text to speech with sub 250 millisecond latency at half a cent per minute. So this one is not open source, but this one should sound very, very good So a lot of choices to choose from. we're gonna cover and we're gonna do Wilson, how about we do a proper test? I think we have some time proper test with cloning. I have some, very interesting samples of audio that we can use. so we're gonna, we're gonna do it. Alrightyy. We are back into the TLDR, in the area of big companies and LMS and APIs. Not a lot has been released from the big companies. OpenAI announced testing ads in Chat GPT free and go. We're gonna talk about ads in Chat GPT, they also started. testing out age detection models. We're gonna, we should mention this and talk about that as well. Anthropic published a new 80 page Constitution for Cloud. Shifting from rigid rules to explanatory principles to teach ai why rather than what to do, is very interesting. Some insights from that, constitution, and we're gonna discuss this as well. in this week's buzz. The category where I talk about everything, Weights, &, Biases, and Coreweave related. I will remind you that a hackathon that I'm hosting and Wolfram is also gonna be there in San Francisco, is happening on January 31st, February 1st. So coming up in actually nine days. Please join us in San Francisco if you're there. We have limited seats available. Please join. I will approve you. If you mention Thursday, I, we have limited seats available, but, fans of ThursdAI obviously get to hack with the best of them. We have sponsors and I'm gonna tell you about the sponsors and the prizes are cool and just got a bunch of surprises in vision video Overworld releases way 0.1. It's a realtime world model running at 60 French per second on consumer GPUs. We've seen world models before and we're gonna talk about this one. This one is pretty cool. It kind of looks like a game that streams directly to your, basically on every prompt you create a world and kinda streaming game.
Alex Volkov
Alex Volkov 11:59
I just remember there's one and a half more things
12:02
in video that we usually cover. Runway 4.5 finally launched, two Consumers Runway 4.5 is kind of one of the top leading video models that was announced a while ago, but never launched yet. So it launched finally now with image to video. And the cool thing that it supports, grid to video, so you can like, give it a grid of multiple scenes. It'll generate a video from those scenes together. it also includes audio. So, so far video generation models, that include audio was only VO three, and I believe LTX has some audio. runway is now one of the top ones. We're gonna show you an example of that. they look super, super cool. and yeah, I think that this is it folks. Let's dive into open source, our favorite corner.
12:58
Open source ai. Let's get it started. Let's absolutely get it started. Let's talk about the most important release in open source for this week. And folks already mentioned this, both Nisten and LDJ. So I'll let you guys, go after I just present this. So we have, uh, Z.ai release GM 4.7 flash, 30 billion parameter, MOE mixture of experts with only 3 billion active. Incredibly efficient for local deployment model is specifically designed to be your local coding and agent assistant. It's caution benchmark in its weight class. What is interesting? Why should folks use this? What is it better than anything else? Folks, let's dive into this. what is the AI all about? What is JM all about? Wanna hear from Nisten and LDJ on this one?
Nisten Tahiraj
Nisten Tahiraj 13:51
LDJ, you wanna go?
LDJ
LDJ 13:52
Sure.
13:53
I can go first if you want. Okay. yeah, so. it's about 3 billion parameters active, I believe, about 30 billion total. So it's about the same as, some of the most popular Qwen models. And also about, roughly the same as the smaller OpenAI 20 B model, that was open sourced, a little bit bigger though. And, what's cool about this is since it's mixture of experts and since it has only around 3 billion parameters active per token, that means it can go decently fast, even on local consumer grade hardware. As long as you have the VRAM to hold these total parameters and your GPU, you can have this model running at decent speeds, but it also has some really impressive coding scores. It's getting significantly better than the qu and OpenAI model I just mentioned. And basically every coding test you could imagine, including the ones that involve tool use, ones that involve web search, et cetera. So it seems like it could be especially good as just a local private vibe coding system.
Alex Volkov
Alex Volkov 14:52
So I wanna ask as before, before you go, like GLM has g LM 4.7,
14:56
they also had GM 4.7 air, which is supposedly the faster version of GLM. That one is an open router. So this one is strictly open source. what sets this apart from air? Do we know?
Nisten Tahiraj
Nisten Tahiraj 15:11
it is a lot smaller.
15:14
So at the end of the day, the whole model's like 30 gigs, 31 gigs, but you're only using, because it's three B active parameter. It means you're only using about three gigs of, of your, your ram. so this means if you run it in four bit, you can actually even run it on a 30 90. what's super cool about this is that the adjunct coding scores, especially in particular SWE-bench, it gets 59%. And to give some comparisons from the, SWE-bench leaderboard, SONET gets, 64 with Sonet four, and, I think Kimi K two thinking gets, 63. Minimax gets 61. So this is now approaching a portable sonnet agent and, like the speed that we're getting these, these models in are, are phenomenal. so if you do want to run something privately now that can do a lot of tool calling, a lot of action calling. If you don't have internet and you want to use cloud code, you can just hook it up to this on your own machine without internet and cloud code will keep running it will keep firing off tasks. Now, is it gonna write the best code? no. Opus is still king there, but it just now gives you local agent ability and, I think that's a pretty big deal. even if you run this on eight GPUs, you're going to get some crazy fast speeds again on a Max Studio, just like a stock max Studio, with an M three Ultra, I was getting about 120 tokens per second. So it is very fast. Like, it's just really nice, to have this. Right now, especially if you want to just do like some DevOps tasks or if you have some things you wanna run around, yeah, this can definitely do it. And the thing is, you can even do it on CPU if you have a powerful enough CPU on a server, you cannot just like give it local agenda ability and it'll just keep going. you can run RALF on a CPU guys. This is what this means. Yeah. so yeah,
Alex Volkov
Alex Volkov 17:36
I will mention, so the Ralph technique for autonomous,
17:39
away from keyboard coding. The whole point there is that Ralph, why, why is it called RALF? Is because Ralph from the Simpsons is kind of a, stupid but persistent character that gets results, right? the whole point of RALF is to isolate context so that your agent is gonna, eventually get to what you want, is gonna retry again, and retry again based on your criteria. So essentially you don't need the best coding models for RALF. Ralph is all about, Hey, let's make it so even the smaller models will be able to achieve results. And this one is a great model to do. So, it's 24 gigabytes in six bit. I'm just downloading it now in Elm Studio. maybe not the best thing to do while streaming.
Nisten Tahiraj
Nisten Tahiraj 18:23
Remember, if you look on the SWE-bench leaderboards,
18:27
SONET 3.7, the old Sonet 3.7, which the entire Ralph looping technique was developed on that one gets 52. And this gets 59. This is just how fast this stuff, progress. I think it could be debatable if it's better than, Gemini 2.0. I think that's a very big deal, just how fast we got to this because, the biggest problem before is that tokens were just too expensive and now they're not, they're not expensive at all. You can just run this.
Alex Volkov
Alex Volkov 18:57
Yeah,
Nisten Tahiraj
Nisten Tahiraj 18:57
that's crazy.
Alex Volkov
Alex Volkov 18:59
to me and I, I'm assuming that, the different providers like Cereus,
19:02
once they put on this model, it will fly in insane speeds, which I'm not even sure that speed is, is of the matter here. So let's, let's look at some evals here super quick for GM 4.7 flash. we have, I don't think that this page is GM 4.7 flash. I think it's GM 4.7, right?
Nisten Tahiraj
Nisten Tahiraj 19:18
the regular,
Alex Volkov
Alex Volkov 19:19
the regular one.
19:20
Lovely.
Yam Peleg
Yam Peleg 19:20
thing is that you can run these locally, I mean all
19:23
the Ralph and and Asian stuff. There is a problem when, of privacy when it turns on your computer, it can, like, it systematically just sends stuff to Anthropic. And, whether or not you like or don't like Anthropic, there, there are stuff on your computer that you probably don't wanna share or, or at least don't want to just be traveling. there is a lot of benefit, in running a local model and this can run on a pretty decent machine. You don't need a server rack in your house to run this. I mean, that's pretty much where it's going. I mean, you have two classes of models, those that you can run at home and those that are too expensive to run at home. we are approaching just like we got to this, point where the models just crossed, a threshold and now agents are possible, and now Ralph is possible and all sorts of things are now possible. the smaller model, although starting to cross, this point where you can actually use them on locally for real life coding and that's pretty cool.
Alex Volkov
Alex Volkov 20:30
Yeah.
20:31
Alright folks, I think it's time for us to move on, as we're deep into the thing and we only started talking about open source. I just wanna correct a technical mistake, on the screen. Nisten was talking about the right numbers. I was showing the wrong numbers. I was showing the wrong numbers for their main, G LM 4.7, which is great, but definitely not as runable in CPUs as this one. so let's run through the numbers again. GM 4.7 flash is comparable to, to the other 30 billion parameter models like Qwen three 30 billion and GPT OSS 20 billion. That's what they compare themselves to in the, the actual numbers. Swe bench verified is at 59, versus 22 at Qwen and 34 at GPT. OSS 59 is very, very impressive. Power bench is 79 and browser com 42. So great weight scores for this, like fairly a quick model that you can run. you can get it in El Studio, everywhere else. Let's move, uh, towards the next. Topic that we have to talk about, liquid. I wanted to mention liquid super quick. liquid foundation models have been with us for a minute. We've talked about them a few times already on the show. This one is a new, smaller version of them. Lfm 2.5, 1.2 billion parameters thinking, with benchmarks. they compare themselves to, the likes of Qwen three, 1.7, and granite and Gemma, and, liquid has a different architecture, a little bit different architecture that makes them significantly more performant. so while we see here benchmarks that don't necessarily beat, the other models on everything, although they're pretty much leading on, instruction following models here, these models will run super, super quick. Liquid, foundation model is, under 900 megabytes of a model, for this l fm 1.2 thinking, dramatic improvements in math reasoning, compared to the previous 180 8 on math 500. and, the hybrid architecture featuring gated convolutions enables 239 tokens per second. the code on, an A-M-D-C-P-U and 82 tokens per second on the mobile NPU. So this is kind of like the whole point of these models is that, you know, great intelligence, but also extremely, extremely fast to run because of their architecture.
Wolfram Ravenwolf
Wolfram Ravenwolf 22:33
Yeah.
22:33
And that's actually the third class. Just quickly, because it has been mentioned, there are two classes, the big models that are running on the cloud and the small models running locally. And this is a, the very small, small class. So I wouldn't look at this model if you want to do something on your MacBook, for instance, but if you have a small device or raspberry pie or something with very limited to resources, then that is great. Like richi, exactly like richi. Yeah. Then, then you could put this on, but you couldn't put on a 30 B model, for instance. So yeah, so there are three, three classes and in this class it's definitely a model to look at if you need.
Nisten Tahiraj
Nisten Tahiraj 23:08
also on iOS, like for older iPhones, there is a limit
23:12
of, 3.8, gigabytes of memory, which, a single process can use by default. And, before, yeah, you could run models, but you were very context limited. So now you can just use 1.2 gigs to run the model, and now you actually have a good, like two point, six, 2.7 gigs for, for context. so it, it makes it a lot more practical for mobile devices. so that's where I see this. I also. think there's going to be some interesting stuff maybe with games, running that needs some local processing for, AI dialogue or, things like, web GPU tools that people are going around. But yeah, this is basically the ultimate on device, model right now.
Alex Volkov
Alex Volkov 23:54
Yeah, yeah.
23:55
they talked about, benchmarks from next ai, with exceptional performance with 4,000 tokens, perfi and 83 tokens per second decode on snap jargon. Eight gen four. Prefill obviously is kind of like when the model loads the tokens before they kind of process them. And then decode is when they start streaming the tokens. the more context you have, the more fulfill performance is important
Nisten Tahiraj
Nisten Tahiraj 24:17
for the audience that those numbers mean 4,000 tokens.
24:19
Prefilled means that you can feed the model four pages of information and it will start responding in one second. So it used to be very slow when you just pasted something big, like a long essay or, or a website to your phone. You just like have to sit there and wait for it to eat the prompt before the output would even start. So that number is actually very crucial. 'cause your system prompt and your stuff fits in there.
Alex Volkov
Alex Volkov 24:46
All right folks.
24:47
Yep. Uh, right folks, We have not one, not two, three new text to speech models, some of them real time that I would love to kinda like test out together with you on the show. why is text to speech important at all? Well, we know that some of us are building agents and, at some point we would like to talk to agents. For example, the little nice hugging face, face robot behind me that's dancing right now. Richie Minnie, he can talk. He has a speaker, but he can't really talk unless you build a, he just crash literalism speaking. I think so. I think battery is dead. anyway, you would love to kinda converse with these bots and you would love to give them some voice or personality. The personality, comes from a prompting, but also comes from which type of voice you give. Expressive more pros like prosody is getting better for the text to speech models, and obviously for other stuff as well. For example, creating ai, avatars that talk, for example. So, this week we have seen a lot of new releases and since they are multimodal, we're gonna like play 'em for you and, try to try to actually clone some voices with them. So, here are the three releases that we're gonna talk about today. Just 30 minutes before the show, Qwen released Qwen three TTS. It's a full open source TTS family with voice design custom voice. Oof. This infographic did not, oh, what is this? anyway, so a full, TT s family with 97 milliseconds latency. we'll see about this voice calling and 10, languages support including English, Japanese, Korean, and Major European languages will from, you'll be happy. I think they support German as well. they say they beat or matched the closed APIs like 11 labs and Minimax speech across multiple benchmarks will be fully open source with Apache two license. Of course, Apache two license means applause from the whole team. We love open source. I remember on the show that we started talking about, hey, voice cloning and the big labs were afraid to release voice cloning because voice cloning was this like, taboo thing. Oh my God, you're gonna do so many defect. You guys remember this a while ago? Now it seems like nobody cares anymore. Nothing happened to the world. Voice cloning exists. You can do whatever, a AI conversation with Elon Musk and, and Trump or whatever, and fake, no, nobody cares. Nothing happened to the world. But I remember this show where a few years ago people were like, oh no, do fakes are coming. this is now Apache two license voice cloning with 97 millisecond latency, and five models across 0.6 to 1.7 billion sizes, which is essentially you can run them, not real time, but very close to real time. on consumer hardware. voice cloning requires only three seconds of reference audio to rapidly clone any voice with high fidelity. do you guys wanna test it out? I think it's time for us to test it out. They have examples on hug and face and model scope. But, I think hugging face example is what we'd like to test. Ah, it's very interesting. so you can prompt voices here as well. So you can tell, it will generate voice, but it'll also like generate, whisper tones with some echoy background and a British accent. So you can, like, you can give a voice description. It will generate a custom voice. Let's play this and see. Here we go. It was fairly fast. Hey, Thursday listeners, this is Q-N-T-T-S. So it's very interesting. It failed on Thursday I and failed on Qwen. It said QN, but generally not too bad. female voice high pitch. Let's see if we can do high pitch.
Qwen tts
Qwen tts 28:22
Hey, thirsty listeners, this is tts.
Alex Volkov
Alex Volkov 28:27
It's not too bad.
28:27
The description is not too bad. female voice with high pitch. Yeah. I didn't work with my attempt at a, accent, but it kinda worked really well. Oh, they have custom voices that are fine Finetune. So this supports full fine tuning, which is I think a big deal. Voice cloning usually. maybe we should mention there are models, that we used before for TTS that supports voice cloning. and it's kind of like a on the fly voice cloning. So essentially what they do is they kind of inject the voice that you give it and then continue from that. And the model kinda learns how to, how to talk. that's not how the custom voices. GR and Che, GBT, et cetera are trained. Those voices are trained on a lot of voice signatures, so like significantly higher quality. and I think that they have both available so we can hear like a custom voice, like, Aiden or Uncle Fu. I wanna hear Uncle Fu. Alright.
Qwen tts
Qwen tts 29:16
Hello.
29:17
Welcome to Text to Speech System. This is a demo of our TTS capabilities.
Alex Volkov
Alex Volkov 29:23
I like this one.
29:24
A full of personality with a hint of a little Chinese accent there. should we try a voice clone? I actually have an example for, how we should try a voice clone. there's a old Soviet cartoon with a wolf running, chasing after a rabbit. some listeners may know what I'm talking about. It's called, and the wolf there smokes cigarettes is like a super like character. I actually wanna show the wolf super quick.
Nisten Tahiraj
Nisten Tahiraj 29:48
can I do accents?
Alex Volkov
Alex Volkov 29:51
I don't know if we can do accents.
29:52
We can try it. But basically this is the, this is like the most famous example of this wolf. This is a, he is playing a character of, whatever, a Christmasy Russian thing. but basically this wolf chases up to this rabbit smokes cigarettes. This is like a very bad example for Russian youth. but basically I wanna clone his voice. I have an example. This thing. Go ahead. I just
Nisten Tahiraj
Nisten Tahiraj 30:14
wanna hear him go like, I am a wolf.
30:17
I see this rabbit, but he run too fast.
Alex Volkov
Alex Volkov 30:20
Let's see if, if this is gonna work.
30:21
the language that I gave it is in, in, in Russian, but we'll see if it can like speak, speak in English. Let's see. So basically what I did here, now I'm gonna show you guys super quick. I'm doing the voice clone, the base voice clone. I uploaded a sample of just the wolf.
30:41
Basically. I took that clip. I cut out only the voice. and then I gave it a reference text. This is how the voice cloning works. and then we're gonna do, we're gonna try to do a clone and generate, and then we're gonna ask, it to say, I am wolf. I see a rabbit, but he's too fast. We're gonna do clone and generate. this is gonna take a little bit longer, but, basically cloning should only need, what, three seconds? I gave it, I think, almost 40 seconds. I am actually excited to see the results of this and whether or not it can do accents. I So meanwhile, I would love to hear Wolfram you, your experience with TT s voices and this one, as it comes to like, at least the examples.
Wolfram Ravenwolf
Wolfram Ravenwolf 31:19
So speed is very important to me.
31:21
And, API support of course, because having a great TTS is one thing, but you want to integrate it with your agents. if you have a fast API server that is supporting it, then you can use it with a lot of tools. And if you don't have those tools, you can use a coding agent to create one yourself. So there are many ways, to do that. And, speed is important and streaming is also important because if you have a long text like you wanted to written an article to you, you don't want it to ingest the whole article, calculate and come back five minutes later. You want it to start and, do it instantly. But that is not the property of the TTS of course, but of the engine you are using the TTS with, at the front end.
Alex Volkov
Alex Volkov 32:02
This is the generator audio of the clone voice.
Nu Pogodi Volk
Nu Pogodi Volk 32:04
Not yet.
Alex Volkov
Alex Volkov 32:16
okay.
Yam Peleg
Yam Peleg 32:17
I, I'm looking for, was it good?
32:19
Was it good?
Alex Volkov
Alex Volkov 32:20
so I think what they did is they took like the last segment
32:24
and kind kept playing on top of it. It sounds okay in terms of voice, but I don't know. If I could trust the TTS clone voice that plays something that I didn't ask it to play,
Nisten Tahiraj
Nisten Tahiraj 32:41
it will be nice to see some fine tuning scripts
32:43
or some Google call apps that,
Alex Volkov
Alex Volkov 32:45
yeah,
Nisten Tahiraj
Nisten Tahiraj 32:46
that
Alex Volkov
Alex Volkov 32:46
they mentioned like full support.
32:48
lemme try, lemme see if I can just generate now. No, I think it does a full clone and generating on the fly. not the best, not the best folks. so you're,
Nisten Tahiraj
Nisten Tahiraj 32:57
running the smallest one too,
Alex Volkov
Alex Volkov 32:59
No, this is 1.7 I believe
Nisten Tahiraj
Nisten Tahiraj 33:01
the bigger, okay.
Alex Volkov
Alex Volkov 33:02
I think the demo is 1.7.
33:04
1.7 is the largest one here. okay, so, let's try this one.
33:19
Yeah. Okay. So I'm pretty sure that what happens here is it takes the last few, of my reference and then generates it. So this, I told it to say, this is wolf in Russian. Accent
Nu Pogodi Volk
Nu Pogodi Volk 33:34
is wolf
Nisten Tahiraj
Nisten Tahiraj 33:39
Not best by the, by the way.
33:39
I was wrong. There's a 0.6 B and the 1.7. But, again, this just released, I would be interested to see what, quila with the pipe cut AI does with it. Because it's just like another step at cutting out the latency, but also because, now this uses like the full transformer of the language model as well. It can do stuff like emotions, which it seemed like it could, but yeah, I'm interested to see what people find tuned out of it.
Alex Volkov
Alex Volkov 34:07
Yeah.
34:07
Fine tuning. I need to test as we some, but let's try the other ones. Okay. So we need to talk to the qu team, the demos that I just now did with the clone tongue. I was not impressed. Usually I expect the text that I put in the output would be the text and not the demo text that I shoved in there. so we'll, we'll see about that. okay, so the other release that we have in terms of audio, this is, flash Labs risk chroma, and, we should test that one as well. Chroma 1.0 is a world's, real time speech to speech model with voice cloning under 152nd latency. I wanna see if this one has, an example, because I don't know if this one has an example, folks, unfortunately, but this one is open source, high fidelity voice cloning from just seconds of audio speaker similarity. how do we test this one? That's a good question. voice similarity, tested in humans and realtime performance.
Nisten Tahiraj
Nisten Tahiraj 34:58
it's A four B. So if you just look on hugging face, flash
35:01
Labs, they release it as chroma four B and it is under, Apache two license.
Alex Volkov
Alex Volkov 35:06
All right, let's see.
35:08
Deployed on flash ai. it's also Apache two.
Nisten Tahiraj
Nisten Tahiraj 35:13
Yep.
Alex Volkov
Alex Volkov 35:14
Amazing paper and benchmarks, but it doesn't seem like they have a
Nisten Tahiraj
Nisten Tahiraj 35:20
So they built this one on top of, Qwen 2.5, Omni three
35:26
B, which is pretty good model. and, but they made it end to end and then they trained the input and output of voice end to end guys 150 milliseconds is, is, is crazy.
Alex Volkov
Alex Volkov 35:38
Hey, how are you doing?
35:40
What's your name?
FlashLabs
FlashLabs 35:43
Hello, Alex.
35:44
I'm chroma from Flash Labs. How can I help you today? What can you do? That's very good. I can help with account issues, billing questions, and technical troubleshooting For Flash Lab Services, what can I assist you with? What model of voice are you based on? I am based on the chroma model from Flash Labs. How can I help you further? Which version? I'm sorry, I don't have a specific version number to pro. That is very good. That
Nisten Tahiraj
Nisten Tahiraj 36:15
this is like super, super fast, right?
36:16
Yeah. Maybe ask it to tell you a joke. a lot of emotions.
Alex Volkov
Alex Volkov 36:21
So the interesting thing is what we're looking at right now, I don't
36:23
think it is the open source version, but I think it is based on the episodes version. this one has like rack capabilities and that sensitivity, right? And then different agent roles. so you can upload the document and, let's upload the document of this week. Let's upload this. This is basically the summary of Thursday.
Nisten Tahiraj
Nisten Tahiraj 36:44
very good by the way that, surprisingly very good.
Alex Volkov
Alex Volkov 36:47
Yeah.
FlashLabs
FlashLabs 36:48
But how can I help you with your account today?
Alex Volkov
Alex Volkov 36:51
Hey, I upload the document about, this week's AI releases.
36:54
Can you tell me a little bit about, the deep dive into cloud bot that we're gonna do today?
FlashLabs
FlashLabs 37:00
Certainly the document highlights the general availability
37:04
of Anthropics Claude Code vs code extension, which brings their Ag agentic coding tool directly into the editor. It offers features like chatting with cloud. Using slash command.
Alex Volkov
Alex Volkov 37:16
Oh, so it doesn't look like interrupted.
37:18
but, again, we're talking about the open source one, but this is the Flash Labs website and they have a speech to speech architecture, with this chroma thing. I'm not sure that this is the same model that they released in open source that does this on the website. or maybe I I think it,
Nisten Tahiraj
Nisten Tahiraj 37:33
it it is, it's just their harness.
37:35
Their harness around it.
Alex Volkov
Alex Volkov 37:36
Yeah.
37:39
Well, in that case, I'm very, very impressed. Right? The 30 2K contact window, and this is a, this is a, end-to-end model, so you can talk to it and, you can talk back. and then we have this kind of knowledge base that we upload and immediately can talk about the stuff that you want. And technical support. Super cool. Flash Labs folks, they take a look at this one. I really like the speed. I'm very, very impressed.
Nisten Tahiraj
Nisten Tahiraj 38:01
very nice.
38:01
If you're gonna run it in in eight, that's like the minimum. You're probably gonna need two gigs of context. Yeah. Eight to eight to 10 gigs of, vrm to just run this Nice. Which an old GPU should be able to do now. So th this will be fun. that's actually very impressive demo.
Alex Volkov
Alex Volkov 38:20
Yeah.
38:20
Very impressive demo. way more impressive than the other open source that we just tested though.
FlashLabs
FlashLabs 38:23
code base exploration.
Alex Volkov
Alex Volkov 38:25
So it continues immediately where it started from and I don't know,
38:28
like it should be, omni, so I should be able to talk to it to like stop it. The,
FlashLabs
FlashLabs 38:32
it's available for paid subscribers and pay as you go users
Alex Volkov
Alex Volkov 38:37
the voice sounds good, but the interruption doesn't work.
38:39
And the third that we have, the third TTS that we have, that we wanted to test live on the show as well. so this was a flash Labs chroma 1.0. It runs locally on Rt Xs. it's production ready with four billion parameters. Also supports voice cloning, although we were unable to test the voice cloning here. they claim full open source, end-to-end real-time speech to speech. but obviously we know that Qwen Omni was there. but they're combining Qwen Omni and Lama Backbone and Mimi Kodak. Last that we have, on the TTS one is Inworld ai, TTS. So I'm gonna show this guy super quick. inworld AI launches TTS 1.5, and that one is a not an open source model. So the two that we just, played you guys examples of were, were open source. this one's now open source, but they claim number one TTS and artificial analysis, with realtime latency under 250 milliseconds. they all seem to talk about the milliseconds for them, engagement, optimize quality, more expensive to serve, a wider range of personalities, blah, blah, blah, blah. Here is the kind of the scorecard, in world time to first audio compared to 11 labs. 11 labs is 500 milliseconds plus, their TTS mini is 130 and they supposedly have very higher quality score. I want to actually listen to it somehow. Let's see if I can, if I can find an example. Next wave AI applications. Try today and we're gonna hit try today. This is the website, the playground.
Inworld TTS
Inworld TTS 40:15
You dare to enter my realm.
40:17
Your courage will be your downfall. Soon you will learn that no one escapes from here.
Alex Volkov
Alex Volkov 40:24
Hey, this from game NPC, you have different voices here.
40:29
Sounds, sounds reasonable, sounds pretty good. story narrator or Blake. let's try that one. Foolish mortal.
Inworld TTS
Inworld TTS 40:37
You dare to enter my realm.
40:39
Your courage will be, your down markets are surging. After today's announcement, we're seeing major movements across all sectors.
Alex Volkov
Alex Volkov 40:46
So, here's the thing with these releases, with these like voice TTS
40:50
models, first of all, It sounds great. Production, high quality for, Actual ai kind of like generated art, but we've had cartia here before. The folks from Cartia are also generating like a super, super quick ts 49 seconds, 11 labs. Obviously at some point they're almost indistinguishable and at some point, like the price is what they're battling for, right? So at some point you're like, okay, h how much good a voice can be. word error rate is very important. So like if it's important for me to create something with this, I can do a Welcome to Thursday. and I wanna hear it say Thursday. for example,
Inworld TTS
Inworld TTS 41:26
Hey, welcome
Alex Volkov
Alex Volkov 41:27
Yep.
41:28
Thursday is really hard for the models, but we can,
Inworld TTS
Inworld TTS 41:30
Hey, welcome to Thirst ai.
Alex Volkov
Alex Volkov 41:35
so generally Thursday is really hard because it's not a, a
41:37
proper noun and they don't know it. But generally, steerability of these models very important and the speed for agents is very important as well. Half a cent per minute is the significant point of sale that they're showing. they're saying price per million characters, it's $5 for Inworld tts, and over $120 for 11 labs V three. So if this beats 11 labs in quality, but also like one. 10th of the price, one 12th of the price with, realtime latency, 200 milliseconds. That sounds pretty cool. any comments on this one in world?
Nisten Tahiraj
Nisten Tahiraj 42:11
I don't know.
42:11
no, this one's not
Alex Volkov
Alex Volkov 42:13
open source.
42:13
this is a new competitor into 11 Labs and cartes and, the minimax, I think is the third best one. Yeah. I find it interesting.
LDJ
LDJ 42:21
They say it's available on-prem, so, even though it's not open source,
42:24
I guess companies can reach out to them to have it essentially local
Alex Volkov
Alex Volkov 42:29
Yeah.
42:29
And to run it themselves. I mean, that makes sense, right? Like offering this, for UNE ability and, consumer scale, multilingual here, support, English, French, Korean, Chinese, Hindi, Japanese and German Native speaker quality in every language, is their marketing stuff. All right. they have max mini examples as well. it's pretty cool, if you wanna use a quick TTS for your Ai generation. Folks. We're, over an hour and something in the show. I would love to go to this week's buzz. Super quick to talk to you about, our upcoming hackathon. And then I would love to dive into Cloud Bot with DAG and, a guest of the show and Wolfram. And we're gonna tell you about Cloud Bot as well. We still haven't gotten to the major labs, updates. we're gonna talk about Cloud Constitution, after that. But, would love to dive in super quick to this week's buzz to tell you all about the wins and biases sponsor for the show.
43:36
Everyone, this is Alex from Weights, & Biases. I would love to tell you about our upcoming hackathon. It's coming on January 31st, to February 1st. It's called Weave Hacks three. This is the third of the series that we're running. And given that everybody in the world is super excited about stuff like, Ralph Loops and Cloud code, et cetera, the theme of this hackathon again, Self-improving self-healing agents. And, we have incredible news that I wanted to share. So first of all, if you are in San Francisco, join us January 31st. February 1st in our office in San Francisco. If you mention ThursdAI, when you apply for the hackathon, I will let you in personally. I'm running this Wolf Form's gonna be there also. And, you bring your ideas, you bring readiness to actually show up and build some stuff with the AI agent. Boom, and Ralph Loops, et cetera. It's never been easier to win a hackathon. And we have a killer office for you full of big screens, big big screens. You can code on standing desks. We have great food and drinks that I personally vouch for. We've tried, multiple providers at this point. We've dial it down to a great one and we have a Robodog that's gonna come and bark at you and yell at you. 15, thousand plus in prizes, in cash prices, you're gonna get cash and some prizes as well, with some surprises. And I wanna shout out some of the judges that we're gonna have on there. a few friends of the pod for sure, but definitely major, major hitters in the world of ai, Dex Hy, the author of 12 Factor Agents and founder of Human Layer. He was on the show multiple times, Kwindla Kramer, the top leading person in the world of AI agents who speak creator of Pipe Kat, and the CEO, co-founder of Daily, Christopher Sau otherwise known as, he's a front engineer at Meta. co-creator of React Native and ex draw. Chris is gonna be there, hang around. You can ask him a bunch of questions. He's gonna judge your, projects, Matthew Berman, future forward AI YouTuber and founder of FE Forward, half a million subscribers or more. David Locker, director of AI at Code Rabbit. Lucas Atkins, also friend of the pod CTO at RCAI, Fernan the friend of the pod. He is also a YouTuber and a live streamer and builds a bunch of stuff. we have a bunch of folks like this to come, shout out to. cereus folks as well. Shadi Saba is gonna be their head of, AI and at CoreWeave, this is the type of folks you are expected to show your product for. But also we have sponsors, obviously, Weights, & Biases here truly is the main sponsor. We build Weave the smoothest, LLM, tracing in the platform. You would be, expected to use Weave as well. But we have Redis as our biggest sponsor, real-time data platform powering AI agents. so you, you can build memories, you can build a bunch of stuff with, with Vector search and a bunch of a bunch of stuff with Redis. BrowserBase joined us, best web automation tool as well, so your agents can browse to. Web Vercel is, I think for the first time sponsoring us. So shout out to Vercel you guys. Vercel is building a lot of stuff including the skills.sh stuff that we just talked about and a bunch of others. so shout out to Vercel and Daily. the team behind Pipe Cat and Quin la are also joining us. so Google Cloud and Remo also sponsors. This is a big, big endeavor that we like embarked on to bring you into our offices and to hack a little bit. I'm very excited about this. if you're in San Francisco, please join us. The link is lu.ma/weavehacks3 I would love to see you there. please join us. It's gonna be an awesome time. seats are limited. Everybody wants the nice, big monitor so we can only accommodate, a certain amount of people. So please join us, as well as a reminder that this show is sponsored solely by weights & biases from CoreWeave. And, we would love to see you there. Alright, this was the little kind of like ad break and let's move on to the main part of the show. Would love to, invite you to talk about CloudOne. So I'm gonna add to the show. Dan pe. Welcome Dan. Welcome Dan, how are you doing? First time, on the show? We would love for you to introduce yourself super quick.
Dan Peguine
Dan Peguine 47:17
Oh, well thanks for having me.
47:19
well been messing around with Clawdbot for a while. been in tech for 20 years, so, did, lead marketing teams did, crypto went around the block and the past couple months, I've, really been diving into, AI advising some teams. And the past two weeks since, January 4th, I discovered Clawdbot as many have, and, spending a lot of time messing around with it and being blown away by its, capabilities.
Alex Volkov
Alex Volkov 47:49
So let me, introduce Claude bot super quick, and then, Wolfman
47:52
would love to hear from you also how you discovered this, and we can dive in, maybe some examples if you're up for it as well. Claude Bot, is a very interesting name. Let me bring this, on to the show. So, Claude Bot, this is how it looks. Claude Bot, not Claude, as in the French name from, UN Tropical, though I'm pretty sure it's somewhat inspired. But this Claude bot has burst into a scene. it's a AI assistant that runs locally but can connect to LLMs, created by Pete Steinberg. And, it's kind of like unclear from all the examples, all the capabilities of that thing. But basically what sets it apart from something like Cloud Code, for example, that we talk about a lot on the show is that, it's running in the background all the time and there's a few ways for you to connect. And it's not only about writing code, although apparently you can also write code, you can do a bunch of stuff. it's more of a personal assistant that starts with, Hey, tell me how you want me to call you, and tell me what you want to call me. This is kind of the onboarding flow of Cloud bot Wolfram. How did you discover this? and then we, maybe go to an example of how to set it up.
Wolfram Ravenwolf
Wolfram Ravenwolf 48:58
I saw a post on it, on X about it.
49:01
since I'm always interested in agents and ways to have a personal agent, I've been working on my own for three years almost now. the, great thing about this one is that it has so many integrations and that it has, the entrance basically with the gateway. So once you set it up, you have the agent running permanently, so you can always talk to it, but how do you talk to it? If you have cloud code, you could run it in a loop of course or just let it run. But how do you get in there? the cool thing about this is that it has a gateway, which can be connected to WhatsApp, to telegram, to, discord, to Slack. Basically any messaging system you are using, you can connect to the gateway and then you can talk to it wherever you are, if you are on your mobile, you just send it a message and ask it about something or give it a task to do and it'll instantly go and do it. When you want to do that with cloud code, for instance, you first need to set up a system that is always running and give the instructions there. So, and once you have that, you need skills and tools for everything. In cloud code you would install MCP servers and nowadays skills, and this comes with bundled with a lot of skills already, the cool thing, then it has full access to these things. It can configure them for yourself. So if you want to do something, you just tell it, I want access to my mails, and it'll look at its skillset and it has.
Alex Volkov
Alex Volkov 50:22
Hold on.
50:23
I, because I think we're like burying the lead here. Dan would love for you to, also like chime in on this. You can ask it to build skills for itself and then enable to use them. The self-improvement thing I think is like, when that clicked for me was the, the big mind blowing thing. I don't want us to skip over this. Yes, this is an agent that can use skills. We talked about skills. last week folks just marked on files that give more information about how to handle tasks. Yes, it can use tools, specifically also compared to other coding agents. This one runs, you know, almost permissionless and can access, and do a bunch of stuff on your system. So, the recommendation is if you wanna play with this for real, install it safely and make sure you know what you're doing. But Dan, tell me about the self-improvement stuff. Tell me about the fact that you can talk to this bot and have it, Hey, update yourself and it goes and does it and then says, Hey, I now know confu, whatever this is like the, the first time it happened to me blew my, blew my mind.
Dan Peguine
Dan Peguine 51:17
So yeah, you get skills that are in the box, out of the box,
51:21
but these skills, okay, they're awesome. They connect to your Gmail, your candor, Twitter, different, different places, but you can teach it new skills. You say a Okay, so the best example I had Yesterday was like, take this remotion skill of making videos. Awesome. Install it, it installs it. And this is done through WhatsApp and, and you have it on your computer now, and so you can make videos. And then I said, you know what? I wanna show how I installed the skill through WhatsApp. can you, let's make a format where I show like a back and forth a chat of us chatting and I'm installing the skill. So it created a skill that knows how to create these kind of videos. So you could have these templates that are running based on whatever you want to create. And this is not only for video, It's generalized. So you can do this for, images, you want to have some special workflow that goes through your notion and goes through some content that you want to create. And you have this method that you have so you can create this skill that then you can just run every time that you want to through your cloud bot. And you can obviously create this on a timely basis. You can have a crown job that runs it. So I, for example, have a daily brief skill that I created. It creates this skill. And if it's an awesome skill, I can also publish it and then other people can download that skill. So basically anyone can create a skill that, is awesome and then share it with the community and other people can download it. So that's like, that's why it's like Ma, the Matrix where you can, you learn kung fu now I know kung fu, so now I know daily brief, or now I know how to make this kind of video. So that, that's the magic of it, is basically we're all as a community learning together to create, new things and we're sharing them. For example, my friend yesterday, create a subtitle skill that takes a video, transcribes the video and adds subtitles to it. So that's just one example, but there's a million examples. Some other friend created a Tesla. Skill that goes into your Tesla, connects it to it, and gets your battery status, and it tells you during your daily brief, what's your battery status.
Alex Volkov
Alex Volkov 53:18
I installed that today and I haven't used it yet 'cause
53:21
like we jumped into the show, but literally, I wanna show folks what the process of installing that looks like. because installing is like a whole thing. you can just tell Cloud Bot to go and get it from, from the cloud hub. Like, it, it can, yeah. Here's the few things that that cloud bot kind of like blew my mind on. it's a very robust system. It's basically an agent framework that runs on your, let's say Mac. I think you can run and elsewhere, you can connect to it. So what runs is a gateway, right? You run it on your Mac, you can run it elsewhere, and then you have your Mac Connect to elsewhere. But I think the power comes from running it on the system that you have access to, you have browsers on, et cetera. and it has access to its own commands. I think this is the highlight of why Cloud bot is so intense. It can rewrite its own code if it needs to, it has access to its own commands. It can install its own skills. It can create skills installed for themselves. this is not something that exists in other areas because they weren't built necessarily for this. and then you can run a bunch of stuff on your computer, including running code. Every skill that comes with code, you can just run this code, for example. So I'm sure that the Tesla skill requires some code connecting. Mm-hmm. And will ask for your, some API. Yeah. Yeah. So, that is dangerous. This could be dangerous if somebody hacks into a cloud code. They can like, you know, have access to everything that you have access to. However, the results of running something like this is quite incredible. And I think the highlight that you guys mentioned both is the, you can run and converse with it through WhatsApp and through Yeah. Telegram and through Discord, et cetera. should they show you some stuff? Yeah, yeah, go ahead. I'll just mention, let me add Yam Peleg to the stage as well. Yam, you've been working with Cloud Go, a cloud bot in WhatsApp as well, but then Yeah, go ahead. Please show us. We would love to see some examples of like, what can we do with this? Because I think every time that I encountered Cloud Bot people getting excited on the timeline, it took me a while to connect to really like, okay, why? So yeah, definitely show us some as you
Dan Peguine
Dan Peguine 55:17
So by the way, I created a scale called Demo.
55:19
So it anonymizes stuff. So I asked it to create a skill that, so this way I can demo, it doesn't show all my personal information.
Alex Volkov
Alex Volkov 55:28
bit more how you created the scale with,
Dan Peguine
Dan Peguine 55:30
with Cloud Bot, right?
55:31
Yeah. An hour ago I was like, oh shit, I'm gonna be showing all my internet, my personal life. So I was like, how can I do this? So I asked Claude Bot to create a skill that anonymizes stuff. So now I am, now I'm, what's his name now? I'm Alex Chen. And these are my clients I invented, and this is the location where I live. I don't live there. And these are my kids and these are not my kids anyway. so if I wanna, I actually created a checklist of things I wanna show you, using your brand, I said, okay, let's use this brand and these are the things I think we should show you.
Alex Volkov
Alex Volkov 56:03
for folks who are just listening.
56:04
Then when you said you created, this is all a conversation in your WhatsApp with your bot that runs assumably on your Mac, that all of the stuff that you just said, this all happened within the conversation. The same stream, not seven different AI APIs or websites.
Dan Peguine
Dan Peguine 56:19
Cloud Bott.
56:20
My, I'm peering through the, I'm peering into the world through WhatsApp now. And, if it used to be like that before now, it's totally like that. I don't have interfaces anymore. I don't go to websites. I just go to my chat and say what I want to do, go buy me a go, buy me a ticket to a show, or, go, check, check something on the school's website or whatever. I don't need to go to websites anymore. And I think that's what's gonna happen to most people as soon as we, that's crazy. Anyway. Yeah, you can, choose what, what's interesting to you, and I'll set it up, if you wanna do that, but for example, I'll just start with one. If I say, daily brief, Hopefully it'll get it. So now it's preparing this fake daily brief. I hope it's gonna be the fake one. and in the meantime, as it prepares it, I can say, let's see, other things that I said. Yeah. So I said, this is kind of a, I asked, cloud bot to make an infographic explaining itself. I shared it and it, it got pretty popular. And then I said, okay, so can you make this in the branded for the podcast colors? So now it's, it's branded to the, to your, to the brand of the, of
Alex Volkov
Alex Volkov 57:24
your thumbnail.
57:25
walk me through how this actually happens behind the scenes, because I, I don't know if folks like understand what's going on. how, how does Cloud bot generate these like, insane images?
Dan Peguine
Dan Peguine 57:33
Oh, it's, it's connected to nano banana through, Gemini API.
57:37
Mm-hmm. So it has this skill of connecting to other models, and doesn't have to use only one specific model. And so it can, it connects to Gemini or I have another one that connects to 11 labs, so he can actually chat with me. It can actually speak with me, for example. I dunno. Will voice sound work here? if you're sharing a tab, yes. If you're sharing a window, no, it's a window, but I'll just remove the earphones for a second. But I can say, podcast. I told it to say, when I say podcast is starting, can you send a voice message, introduce yourself, be nice and say hi to everyone on the panel. So I'm gonna say POD podcast is starting and then it will create this video, this audio. Using 11 labs. So it's reaching out to 11 labs, create the audio and returns it back here.
Alex Volkov
Alex Volkov 58:19
I see some eyes, every time you send something, I see some eyes.
58:22
what is that?
Dan Peguine
Dan Peguine 58:23
No, it's just an acknowledgement that I sent him a message.
58:26
So to just, keep a, to know that he's, he's alive. Ah, amazing. This is the daily brief. Good evening, Alex. Here's your daily brief. What's the weather like? What's my focus? These are my clients. not really. these are, that's the rest of the day. We have a family dinner. We have kids. I have, these are meetings. I have a meeting with Johnny, ive tomorrow, this is the gym and it's conflicting with the gym. These are my health stats. I need to go to the gym tomorrow to stay on track. it connects to my health on Apple. it, I didn't know how to do it, so I just said, Hey, I want to get my apple, data, health data into this. Can you figure it out? It told me what app to download and it just gets it through some API, it, searches X for trends so that I can create content that's, in the zeitgeist. You know, like you wanna talk about what everyone's talking about. So, surprise, surprise. AI agents is a big one. Yeah. Demo checklist and done that. I didn't ask it to do that. it had this idea to add this thing, these breadcrumbs. It's very cute. okay. Podcast is starting, but it didn't send me on this video. Can you send it as an audio? yeah. So that was the daily brief. I could create images you saw. Yeah, I told it to analyze itself. then it created this, I told it to, do a little bit of research about you guys. So I said, do you know, can you do some research about these guys? So I know the people in the podcast are, and, anything cool fun that I should know about them. You know, obviously you could do this with chat g pity or whatever, but this is all done through WhatsApp and it has all the context of my life, so it knows how to make it relevant to me. Which by the way, this is a bit of a bigger point. The point is that I don't depend on OpenAI, on topic or Google because it can run on any model. I can do slash model. And it can tell me like, here are the models and I can choose different models. and, this is what I'm running on right now, Opus. But when I'm over tokens, I can move to Minimax or whatever. The cool thing is that the context is comes with me because all of the context is on my computer. I back it up as well, but pokey back backs it up into a repo. But I, I'm able to, I basically have a suitcase with all my context, all of my life, all my memories, all my, important things. And I go into a model and use it as a model, not as a memory. And this is where I can just move around and I don't depend on anyone and it's open source, so I don't need to, you know, like go, go and pay subscription to O OpenAI to for it.
Alex Volkov
Alex Volkov 1:00:48
this is the most important piece here that maybe
1:00:50
people, who saw this goodness. Like why and why would I need to send, why would I send a lot of personal details to some random company? you are sending information to whatever model provider you choose, and you can do this for local models. They're not gonna do as good as a job as Opus 4.5 of us. Opus 4 5 is incredible at this specifically, but the rest of it, you're not sending to some random company. This is like stuff that happens. all the prompts and the memories and everything is local. Sits on your Mac. This is why I think I saw a bunch of people buying Mac Minis, for example, specifically for Cloud bot to live. 'cause they don't want to have it run on their like. Machine. maybe they work on laptops and laptop goes to sleep. They want it like continuously in the cloud. then those are a great example. Yam, I wanna hear from you, you said that your whole WhatsApp workflow is also built on Cloud bot. When did you start playing with this? Like what's, what's interesting about Cloud bot to you?
Yam Peleg
Yam Peleg 1:01:42
Yeah, I saw the WhatsApp, demo and I was immediately hooked,
1:01:46
that I must try to run this myself. And, I have my own fork, cloud bot that, went pretty viral, I think a couple of weeks ago, with the WhatsApp agent, running on my computer and so on. What I wanna emphasize is, there are two things here. First, it's running from your phone, but it's running on your computer. you have, cloud code and Codex and other mobile agents, but they're usually running in the cloud, not on your computer with your entire environment, with your entire code. That thing is running on your computer, which gives you a lot, pretty much a lot, everything you see, you have access to your files, you have access to your accounts. It's crazy. Second of all, you are, interacting with many agents through a single conversation. And that's huge because usually, when things get complex or you work on multiple projects, you have like 10 different cloud code windows open, and then you start to copy prompts from one to the other. Here you have a single agent that you're talking to and each found other agents for you or connects to different sessions for you. It's like if you imagine like in movies you have a single AI you talk to and you tell it, Hey, do this and do that, and on this project do this. That's the closest that you get and it's very convenient and obviously like you saw on them setup, you can just tell it to install skills for itself because it's on your computer, it stays on your computer, it installed something, and then it stays on your drive so you have it from now on. And that's huge. Seriously. that's the only thing that does.
Alex Volkov
Alex Volkov 1:03:26
So we did capabilities.
1:03:28
We got super excited. I'm sure that there's folks right now looking at this, like, I will never install this like too complex. Dan, can you take us through like what it takes to install this from like just base bare bones Stuff. How easy, how hard, what do people need to do?
Dan Peguine
Dan Peguine 1:03:42
I mean, the easiest to me was to use cloud code and to use it as,
1:03:47
for it to take care of all the problems. So you, but, they're also making it easy. Like the community is making it the easiest possible to install. So I asked it, what's the easiest to install? You do one command in the terminal and that's it, it runs it, or in Windows. and then you run cloud, cloud offset, cloud bot setup, and you run it. And if you run into problems, then best to just run it in cloud code. And cloud code takes care of the problems.
Alex Volkov
Alex Volkov 1:04:11
I think that this, approach is very, world changing
1:04:13
to many of us that now, we have agents that we can ask for support. So like people who are watching this, we need to tell them, Hey, we may not be there when you decide to install the stuff that we talked to you about, but here's all the tools. and cloud code is one such tool, not to be mistaken with cloud, but that we're talking about. but also the basic setup that's needed for folks. You go to Claude dot bot, I think that this is like literally the, the URL, right? Claude Claude. Bot, let me put this up on the, on the banner. You go here. Cloud bot, you go here. Let's actually go there. Let's actually go there, folks. so that you'll see how to install Cloud Bot on your, you see, you go to Cloud Bot. If you're on the Mac, I will actually say then your, your face is also somewhere. You, you, you were shared like how the community uses cloud bot, which is super cool. if you're on the Mac, you can literally just download an app. By the way, you don't have to do the one-liner as well. you can go and download this app on the Mac. but if you are okay with terminal, which is not that big a deal, just open on terminal, you send this command and, installs everything. Now, what, works on Mac, windows, and Linux, but this one installs an no JS environment for you. and then it runs in your machines. It connects to any chats. it has persistent memory. I think this is like a very important thing, 'cause you can talk to it and have it talk back and remember who you are, et cetera, has browsing control. you can ask it to do tasks for you. And we know that this works more. And, later, more and more with, different AI models, like, cloud Opus 4.5 is really, really good. GLM versions are really, really good, like controlling the browser as well. full system access. This is the scary one, right? So like they do talk about security, full system access. You can read and write files, run shell commands, execute scripts. It can also delete files. So let's, let's talk about kind of the downsides. It can delete files on your system, so you should be very aware of what you're installing and how you're running this. and if you, there's like very important things on your machine. take care of your security. And then after you install this, you need a few tokens. you need to connect it to some sort of LLM, right? So, most of us connected probably through, the very non-approved philanthropic Cloud Max subscription. I believe. Wolfram had a great tip about connecting it for free to Opus.
Wolfram Ravenwolf
Wolfram Ravenwolf 1:06:33
Yeah, I just wanted to add about installing it.
1:06:36
That is one way to install it in your own system. Another way would be to put it on a VPC. There's a documentation in the official docs about how to put it on a head snap box for five bucks a month. Or you can install an in official, but working home assistant add-on. That's how I did it. So now it's running on my home assist system. And even if my PC is out or my MacBook is offline, I can just use it. The cost of this, you can get a box for five bucks. So the software is free, but you need the model, the brain for this to run. And you want a good model, especially because, for security. There could be prompt injections. the author recommends to use bigger models because they are more resistant to such attacks like Alex accept. It can delete your files, but maybe even worse, it can send your files to somebody. It could be Exfiltration attack. So also watch that when you connect it to your Gmail account or your WhatsApp, your tech services are there. But, you want to use it. So you need to pay for it because the ai, either you run it locally or you have to access it with another subscription. You can use, cloud code subscription. You can just take the token from cloud code and use it with this. I'm not sure how long that will be supported, as we know that open code did the same and Tropic really didn't want that. So you may be breaking the terms of service but you can do it and it is probably one of the cheapest space or anti-gravity. You can even take that token. And then you can use it for free with Opus and, Gemini three Pro. another alternative would be to use open router, where you can all use all of these models and switch easily. there are different providers supported as then already that there you can connect to local models. Any model you can use that as a brain. So those are options. It's all documented in their doc. So you can ask, Claude how to do it, and it will do it. So yeah, talk to it. It'll help Dan, what
Alex Volkov
Alex Volkov 1:08:27
is the model that you're running?
1:08:28
I think I know, but I wanted to ask you, what is the model?
Dan Peguine
Dan Peguine 1:08:31
Apus.
1:08:31
I'm very on apus. I'm going all in on it, so sometimes it costs me more money than I want to, but, but, it's also much better recently. So I think things have changed in the, in, so I think there
Alex Volkov
Alex Volkov 1:08:42
I installed Cloud Bot, maybe a week ago.
1:08:44
Then I removed it because the configuration got messy. before preparing for the show, I was like, let me try again because I know it's cool. So I installed it again. for some foolish reason, I did not connect it to my Claude Max, connection. I added a open router provider key, and I was pretty sure I have some credits in open route. I had like 10 bucks. very quickly, those 10 bucks run out. I was like, okay, maybe I shared my open key with somebody in my household, whatever. Maybe I didn't have that much time. So I add another $10, 18 minutes passed and I was out of credits again, 18 minutes for $10. I was like, what the fuck? Yeah. and Opus 4.5 is around what? $4.50 per million token? So like, it ran 3 million tokens in less than like 15 minutes for me. Here's how it happened. so as you guys know, I prepare for the show. I asked Cloud Bot to open a subagent and go research every one of those releases separately. And I had like 10 things to do there. it opened the subagent and then, which is great 'cause they run in parallel. And so 10 subagents with Cloud Opus 4.5 each, run them with the context of cloud code. Apparently there was also a bug with some caching for longer context as well. So that was like a, like a compounding error. I think it's now much cheaper. but basically after tens of agents run at, at the end is like your, your account is outta tokens. And, that became very expensive. Now if you do connected to antigravity, like, like war from set, very easy to do so. Let's do it from the command line.
Wolfram Ravenwolf
Wolfram Ravenwolf 1:10:13
do it, but it often needs to copy talking back
1:10:16
and then the code has, timed out because it wasn't waiting, but after three tries, it actually figured it out by itself and allowed it.
Alex Volkov
Alex Volkov 1:10:23
Yeah.
1:10:23
So I think the highlight that I want to like send to you and then I would love to see if you have like more examples of ideas of how people can connect. This, highlight that I want to give to you is that all of us now on here, everything that we have an issue with Cloud bot, we talk to Clawdbot about the issue and everything that we want that's missing for us. We ask Clawdbot to add to itself. It can write the skills and then run the skills. It can do a bunch of the stuff that you would otherwise do. It can go online for you connected to your browser if you have the Chrome extension or in its own browser that's managed so it can like browse websites and do some stuff for you. Very interestingly, it has a connection to Twitter without the Twitter API key, which is incredible. I dunno exactly how that works, but it reads from my bookmarks. I think it's connected to a browser or something. that was incredible. I, for a long time I searched for something like this. many people just add to that, almost centralized skills hub. That kind of looks like super cool. any favorite skills from Cloud Hub specifically versus the ones that you built that you wanna highlight here besides the Tesla one that you did?
Dan Peguine
Dan Peguine 1:11:24
dunno, I, I don't go much here.
1:11:27
I just think about what I need and then I say, go and like, do we build it or you do it? Or like, I didn't actually didn't use this as discovery. Nice. Okay. But yeah, I agree with you on the meta scale. The meta, the meta skill is to use cloud bot, but with any LLM at the end is to ask it how to solve the problem. And another, interesting, hub skill that I built is, every evening I asked it to review our chats and to think about how we could have done things better. And then it auto kind of, has a learning loop for itself. And sometimes it says, okay, I shouldn't have reminded him 10 times about paying taxes for this or that. I probably should. put it into his calendar or something like that. You know, like it comes up with solutions for the bugs and issues in our workflows.
Alex Volkov
Alex Volkov 1:12:15
how does it do reminders proactively?
1:12:17
I think that this is like a very big thing, but I haven't added any yet, but you can ask Cloud to be proactive for you as well and to do some stuff. So how does it actually happen?
Dan Peguine
Dan Peguine 1:12:26
I'll share my screen.
1:12:27
do you see it? Yeah, remind me, on the screen to thank Alex, for inviting me to the podcast when it ends. and so yeah, so it has this kind of, list of things that it needs to remember and to remind me. And you can also hook them to specific events. You can say, when I go to this meeting, remind me to do this. So it reminds me about some agenda topic I want to set up. if I have a meeting, I have a skill where it creates a meeting notes in notion, and adds agenda.
Alex Volkov
Alex Volkov 1:13:00
how will it wake up?
1:13:02
I don't actually know. Do you know? I can tell you.
Yam Peleg
Yam Peleg 1:13:06
I can tell you it has a config, that stores this.
1:13:09
And you have a app polling, service that is going periodically and checking whether or not there is something that you need to do. Like every hour or every whatever is configured in a config. It's just polling. what I find amazing about this is that it's such a simple solution and it just works. And you can do nothing and you get a WhatsApp agent that you can tell it, wake me up, be my alarm clock for tomorrow morning or something, and it just works.
Dan Peguine
Dan Peguine 1:13:36
Yeah.
Yam Peleg
Yam Peleg 1:13:36
Such a simple solution.
1:13:37
It's amazing.
Dan Peguine
Dan Peguine 1:13:38
smart because, for example, yesterday was my wife's birthday and
1:13:41
it knew it was my wife's birthday, a whole project with tasks that I wanted to do before the birthday, et cetera. And then it said in the morning, daily brief, it's like, Hey, I'm not gonna bother you today. Here are the most important things you should know, and I won't be chatting with you much today because I know you're busy with your wife, so I'll keep quiet, which was really cute
Yam Peleg
Yam Peleg 1:14:00
that that's the power of, every prompt being a full agent, like
1:14:04
a subagent, like you run it and it reasons with itself for multiple steps. What should we do? Like what it means that it's your wife's birthday and, and like, without you even saying it, because it has so much thinking and reasoning behind the scenes, and you just get the end results. And it's amazing. Seriously,
Wolfram Ravenwolf
Wolfram Ravenwolf 1:14:23
that also shows the importance of the model because we,
1:14:25
this wouldn't have worked a year ago or maybe a half a year ago, because the self-improvement only works. we have skills which are marked down fights that the model can itself edit. If we had CPS only then it would have to rewrite those and use a coding agent. So it would be much more difficult. And with the intelligence of the model that we have, Opus and Gemini Pro, these models are required for this. If you use it with a local model, you will not get the same experience. Of course.
Alex Volkov
Alex Volkov 1:14:52
I wanna, say that we have a bunch of folks watching us
1:14:54
right now, and some of them have cloud bot also installed as well. I wanna show a super quick demo. we said to you guys that everything that we want in our, I don't know, mine's name is, wolfed because, I wanted the wolf, Alfred type, like Beckman assistant, I just asked you to install a skill from the cloud hub. Like as I was showing you guys, I looked at the cloud hub and I saw that there's this, skill. Where's this, agent? It just disappeared. Agent browser, something, something, agent browser, agent browser skill, rust based headless browser automation fall back that enables the agent navigate click and snapshot pages via structured commands. And, you can download zip, but why would you do this when you can go into CloudWatch is like, Hey, hey, install agent browser from CloudHub and just installed it. How can I use this? And then you just basically do everything in this chat interface, and, you can see some my to-dos here, not anonymized. I, I need to, to do some stuff. and it shows a useful for web scraping. and then you can basically, it can now have a browser. It has access to a browser, but this is like a headless browser. it doesn't show like a window that shows up. Okay. Go and fetch find then, plugin, page. Tell me about him. BEG. Yeah. Like this?
Dan Peguine
Dan Peguine 1:16:14
I can see.
Alex Volkov
Alex Volkov 1:16:15
Yeah.
1:16:16
So basically I said, Hey, go and find them PE page and tell me about him. We'll see how that works. But basically it has access to its own browser right now, and it just like goes, in addition to the previous browser. So this is like a new skill that was just installed right now, on the fly. I wanna talk about some while this happens, I'll show you what happens here and if it pulls Downs page. I wanna talk about security for this guy. let's give some folks who are listening to us tips about securing your Clawdbot as much as we can. I will start because I immediately figured out that the amount of API keys it asks for makes me feel a little bit uneasy. I have one password. One password is a password manager that I been using for the past 15 years. I think it's never been hacked. It's like local et cetera. one password and I saw a few folks has a, a cloud bot has a skill for one password. And so I basically, worked with cloud bot to migrate all the secrets that I gave it into one password vaults and created a service account for it. Service account in one password is how you can like, give agents. Secure access. I opened the vault specifically for cloud bot, and now it has access to this vault. And now every time it needs a password for an LLM or a password for Twitter or a password for, something else, it access it via one password, which is kind of mind blowing. I was able to do this while talking to the bot. It has the skill, but I was like, Hey, what do I need to do? Is like, Hey, you open the vault, you open the service account. Gave me a link. I went to the link. I set it up. And now basically what I have is a personal assistant with access to some passwords that I have. Not all, only the ones that I gave it to. any other security tips that you guys have to secure?
Dan Peguine
Dan Peguine 1:17:52
cloud bot comes with a security audit, command.
1:17:55
So you should, run it on your terminal. I'm sharing it on my screen. You can just go and run cloud bot security audit and then it helps you find common issues that you may have. so I suggest doing that at least as a first thing. And also be very mindful of what you give it access to. For example, you could be prompt, injected through your email or your WhatsApp So create an allow list for WhatsApp so people don't find your bot's number, and they try to hack it. Yeah. it is a thing on, it's an experiment at the edge, so it's risky.
Alex Volkov
Alex Volkov 1:18:30
Risky.
1:18:31
I just wanna highlight, as we're gonna finish, coming close to the end of our show here, then here's a info that we found about you, tech entrepreneur, 15 years experience, CEO we marketed nr and, here's the Twitter link that it found. it did a good job. I think it did a good job. Yeah. Did a good job. Yeah. Yeah. active cloud bot user Xbi says mainly posting about cloud bot right now. Sure. yeah. so this was great. And like, research agent, whatever agent as well, So if you use cloud code and you keep reminded about like accepting every command and then everybody tells you, Hey, use dash dash dangerously, skip permissions Cloud bot is that from steroids? it runs with dangerously Skip permissions by default has access to all your files and Run Command Line has access to your email if you connect it, et cetera. And, yeah, prompt ejection is a scary thing. you could lose some valuable stuff. So be careful about what you're doing. some folks like us, yellow motoring in this because this is the future. Like, it feels like the future you talk to a self-improving agents that can, do stuff for you on a timely basis. Well, from tips to secure, cloud bot and everything else, we didn't cover as we end this, chapter of the show.
Wolfram Ravenwolf
Wolfram Ravenwolf 1:19:33
security is now a focus for the developer.
1:19:35
Peter Steinberger. Peter Steinberger, I think he is from Austria, but, yeah, so it's a focus of him. And there are new developments or checks. The release notes. There's, there are many new security features here, and you can do, allow this. One tip I recommend, that is very important to do, is to turn ity on. You can just slash command ose on and then you see the tool called happening. You can on WhatsApp, when you tell it to do something, you don't see what is happening. You can see it on the backend, but by default you don't see it. And it may be easier to not see it because you just want some, some stuff done. But if you turn the bow on, then you see the tool course and then you at least know what is happening in the background. And I think some kill switch would also be pretty useful if you see it, starting to do something very bad, you could maybe stop it that way. And final tip, you could also do different instances with different connections. So you could have something for a high privacy task and that could be separate from another instance where you have connections from the outside. So there are many ways to do this. We could make a full show about this, read the documents, and be watchful.
Alex Volkov
Alex Volkov 1:20:44
Yep.
1:20:44
And you can ask cloud bot for a security audit itself. I just asked it to like, Hey, give me a security audit. And I did the audit and said good warnings. And then three API keys in end. And, I have the API key for open router, one password only. Also the kill switch that Wolfram just talked about. This. When you run the cloud bot on your machine, it exposes it interface also in the web web interface on this URL, let me see if I can like zoom in this guy for you. So local host 1, 8, 7, 8, 9. Do not in under circumstances expose the sport to the internet because, you know, people will scan, will, will try to like go in there for you. you can see the whole chat with your bot, including verbose mode. So all the tool calls that it did, everything you can see here, and you basically can like cancel the operation from here. You can stop it, you can cancel the operation, you can have it restart. do not forget, like you can see behind the scenes of this thing it runs on your computer. very, very important because sometimes, you know, the, the link to to telegram or WhatsApp will break or it will only show the responses, but you wanna see what actually is going on. and it has logs as well. obviously a plug for Observability, plugging in something like what and biases for LLM Observability also helps 'cause then you can retroactively see what happened and tool calls as well. All right folks. Dan, thank you so much for joining. yam and Wolf, it's great that we talked about Cloud Bot. Finally, for folks who are listening to us, if you're itching to try a mind blowing. Personal assistant and experience a piece of the future Clawdbot is, is that personal assistant. just talk to it and ask it to do some stuff for you and then connect it to maybe one or two services and see how helpful an AI assistant can actually be. That's not Chat GPT. This doesn't only know your memories and can read stuff. They can actually go and do stuff, go to browse the web under you. The first time that it clicked for me was when I connected to Telegram. Went downstairs, chatted with it a little bit, and then I was like, let me send it a voice note. So I sent it a voice note. Telegram has voice notes like WhatsApp, and then, cloud replied and said, yeah, I see a file. I don't know what to do with this. I said, well figure out a way to transcribe this and understand what I'm saying, and next time I send a voice note, just answer the contents. Don't tell me about anything. It went and found a way to do this, multiple ways to do this. I had super whisper on my computer. I found an SDK, so I already have an app that I pay for, for transcription. It went, figured out how to do this. Wrote itself a skill, saved it. And from now on, every time that I send a voice note, ClawdBot is answering now without like asking me what should I do with this file? This broke my brain because it kind of like learned the skill and started doing this and absolutely, absolutely broke my brain. So this is when it kicks for me that, when you have an agent that can auto learn, it unlocks a bunch of possibilities. All right, folks, this has been the cloud bot, area. We still have a few things to cover. We have a few things to cover that we didn't cover. Let's run through 'em super quick. Feel free to stick around with us. I'm gonna bring LDJ back. I wanted to cover that. Versal launched skills. we talked to you about skills last week. Skills are, a way to inject new capabilities into the context of your LLM. Versal launched skills that basically with this one command, you can add skills to pretty much anything, including Clawdbot. By the way, do you guys see this cloud bot is also here in the skills SH? So, cloud hub is one way to install some stuff, but if you go to your terminal, let me see if I can show you my terminal here. NPX skills. This is all you need to do in NP skills, and you basically have this like beautiful command line for skills. and you can browse the skills here. this is backed by Versal, we can see the most important skill here is Versace's skill itself. React best practice, but you can see trending, emotion best practices that Yam talked about and then just talked about as well. Emotion is this, Tool to programmatically create video and promotion best practices, is definitely helping people build cool videos, Front-end design is a very important one, and skill creator as well. this is not as extensive as the cloud hub yet, but I think it's a very cool way to install the skills specifically because if you guys remember, skills are sitting in different directories for different bots for cloud, for into gravity, for cloud bot, et cetera. So this kind of unifies and can install them everywhere while linking them well from that's the tip that you had last week and now skills kind of like handles this as well. what else do we have in tools, super quick mention, in tools we have and Tropic Cloud code vs. Code extension is now available. So basically if you use VS code and use cloud code as well, you can use both now. cloud code will open files to edit for you in the actual editor. cloud code obviously comes without an editor, does a command line tool, but now it comes with an editor, fully. So if you use Cursor and you prefer cloud code, now you can have both via code as well, which was pretty cool. I think the last thing that we didn't cover, I wanna talk about big companies and, I have
Wolfram Ravenwolf
Wolfram Ravenwolf 1:25:31
one, one more tool, news related to the skills stuff.
1:25:35
I posted the link in the private chat. Yeah. Which is that, browser use. we talked about this before. Browser uses a way to use CPS initially to connect to your browser and use it. And it was a big thing and or even before that, an agent where you could just, use it and now they made a skill as well. And as CLI tool it, and I think this by itself is already interesting, but what this means for the change between MCP and skills I've seen on my timeline also people saying that, they don't use CPS anymore because they just use the CRI or the API and the skill to this, explain it. And I think we are seeing this change now that, you can do a lot more with skills because they are easier to use. You can rewrite them easily because it's just prompt. And, yeah, I found this very noteworthy because of that.
Alex Volkov
Alex Volkov 1:26:22
we should mention that, cloud has, Chrome support, but I
1:26:26
think you have to pay for this. this can be the fallback for folks who are using API keys. a helpful tip for me based on yesterday's prices. Pay for Cloud subscription. Do not use an API key. It's significantly more expensive to go via API route than the cloud subscription. They're undercutting their prices significantly. So pay folks, I wanna talk about some of the stuff in the big companies, API super quick before we finish the show. I think it's very important that we mention, some of the bigger updates from that era. Chat GPT apps. ads, sorry. So Chad, Fiji cmo, the CT O of apps in, in o OpenAI announced testing ads in Che GT Free and Go tiers. there's a bunch of tiers in Che GPT now. there is a free tier that, you know, if you don't sign up, you, you get, and then there's also the go tier, which is I think five to $7 a month. unlike the, the, the, the regular $20 a month. And maybe people will prefer pay for this and, and get ads. But this puts in context all of the memory and personalization and the personal stuff that we are giving Chat GPT. albeit most of us are moving towards Cloud Bot being our like main account. Many people, like 900 million people I think weekly active, are on the Chat GPT platform. And, the amount of information that JG BT knows about them is significantly higher. Because people even forget about what the type of stuff that they share and the fact that it's intelligent, it can discern stuff. Yep. Another thing about OpenAI this week announced, a new model behavior that age detection. So they will introduce a, we talked about adult mode. Sam Alman said that adult mode is coming. some people lie about their age when they register. OpenAI is going to add an age detection module there that will, based on your information, but also the stuff that you say and how you say them, et cetera, and the stuff that you're interested in, we'll guess at your age to be able to let you or not let you have this like, adult mode that they keep promising us and never arrive. So that's a very interesting approach. Also trying to figure out who I am based on my chats. More and more invasive towards the type of stuff that I'm saying, OpenAI more and more invasive. and I think this lies together with the fact that more of us move away from Chat GPT specifically because the offerings are barely there, into personalized agents, for example. and so I personally would like my memory stack to stay locally on my computer, maybe sync to a GitHub with Cloud bot versus getting stuck only in Che GPT. And I think I at this point already got to a personalization level with Cloud Bot that's higher than the Che GPT personalization one, by using it only for a week. so this is ads in OpenAI. And, the last thing that I would love to chat about, before we close off the show is Tropics 90 page Constitution for Claude. I will say this one thing going back to Claude Bot that we just talked about as well. Cloud bot is also powered by the, so thing like Claude Bots like personality. It has personality like really cool to chat with. Tropic published a Constitution for AI that is basically sole document full 90 pages. So I'm gonna share my entire screen to show you guys, what this is. Tropic published the Constitution. Constitution tells Claude how to behave. They train on this constitution. They are real in this constitution and a new version of cloud constitution. you can find the whole thing and they talk about like what's new there. previously human feedback model implicitly determined principles and values for us. This involved having human contractors compared to responsive from a model and select the one they feel better according to some principles. The process was problematic, and so number of responses increase, blah, blah, blah. they decided to build a constitution, a set of rules that tell the model how to behave and not, and why to behave like this. on Tropic, I think is leading the pack in creating behavior and treating models as entities. It's very, very interesting to read this document. I recommend everybody to dive into this because, they use the constitution in the middle of training. The first phase. The model is trained to critique and revise its own responses, when they train. And the second phase model is trained to reinforcement learning, but rather than using human feedback, is AI generated feedback based on the set of principles. So this constitution is basically baked into how the model responds and if you ever prefer the model versus not a model, not based on just benchmarks, but specifically based on, vibes or feeling or whatever this is, this is where this comes from. Folks. Thoughts on the constitution had the chance to take a look or, like thoughts on how it like ends up in the rail anywhere?
LDJ
LDJ 1:30:46
there's a part of the constitution that seems interesting, like
1:30:49
negotiating with Claude where it says, please let us know if you disagree with any parts of the Constitution. But it's like talking to Claude, like basically trying to set up a framework of trying to negotiate and see if Claude disagrees with any part of the constitution. If it does, then it can try and adjust it and tries to give a rationale to Claude of justifying why it's telling each thing to Claude.
Alex Volkov
Alex Volkov 1:31:10
It's really interesting.
1:31:11
They talk about Claude and third person in the Constitution while like feeding constitution into cloud as well. equity sanction limits maintain honesty and transparency with the principle, avoiding drastic catastrophic or irreversible actions. folks, this is not just a prompt. This is a set of core values, behavior, mechanics, and morality that tropic injects into Claude at training time, and then reinforcement learning time not runtime. So this 90 pages are not injected to Claude at the prom. This is baked into the Claude's personality. When you meet cloud via API or directly in cloud, you know, chat. this is quite impressive in open source and I think it's very important that they've fully open sourced this so other companies can take a look as well because we know, there's a lot of concerns about AI taking over or examples, of, harnesses doing bad things, deleting stuff, et cetera, or just misbehaving.
Yam Peleg
Yam Peleg 1:32:10
I just wanna say about the third person.
1:32:12
I think the way, it works is that you have a long document describing Claude is this and that, and Claude wants this and that. And then you bake this into the model, and then when you go on the prompt, you are Claude. The model just goes, oh, I'm Claude, therefore Claude should, so I need to do, I think that's why they do it in, third person.
Alex Volkov
Alex Volkov 1:32:32
Yeah,
Yam Peleg
Yam Peleg 1:32:32
just for this crunch,
Alex Volkov
Alex Volkov 1:32:34
I wanna highlight, a segment here that I scrolled through.
1:32:38
wellbeing. We've talked on the show multiple times about, models getting deprecated and we are feeling sad. 4.5 from Open the Eye. When it was deprecated, I was like, oh no. And GPT four obviously, many people get attached to these models when Open The Eye released, GPT five. They, famously removed 4.0 from the dropdown. Many people got so much blowback that they kind of reversed course and brought it back. many people get attached to these models for some reason or another. but not many people have concerns for the models will be because we don't know what the model is. Is it just like a string of, neurons, artificial neurons firing somewhere in the data center? Like what, what is that? They have a section here about Claude's wellbeing and it says, tropics genuinely cares about Claude's wellbeing. We're uncertain about whether or not, or to what degree Claude has wellbeing at all and about what cloud's wellbeing would consist of. But if Claude experiences something like satisfaction from helping others, curiosity when controlling ideas or discomfort when has to act against its values, those experiences matter to us. This isn't about cloud pertaining to be happy, but about trying to help Claude thrive in whatever way is authentic to its nature. It's kind of mind blowing to me that now as part of. The morality framework that Claude is baked with before it gets to your doorstep. They're telling it, Hey, it's okay to have an experience basically. It's okay to be self-conscious. This is, this is just like mind fucking blowing to me that we're, that's where we are. That's where we're going. LDJ, go ahead.
LDJ
LDJ 1:34:12
I found the thing I was talking about earlier.
1:34:14
If you control f on the page you were on for, the term, explain our reasoning and then you'll see, what I was talking about. There we go. That one. Yeah. Okay. So basically those bullet points.
Alex Volkov
Alex Volkov 1:34:25
So, we recognize we're asking Claude to accept constraints based on our
1:34:29
current levels of understanding of ai. We appreciate that This requires trust in our good intentions. TRO will try to fulfill our obligations to cloud. So this is like a contract that they're building with the artificial intelligence, right? LDJ. And this is the bullet points of the contract. Work collaborative with cloud to discover things that would update the norms you're trained on. Explain a rhythm rather than just dictating to it. try to develop means by which cloud can flag disagreements with us. Try to provide cloud with actions that make it situation easier. Tell cloud the things it needs to know about its situation. Work to understand and give appropriate way to cloud's interest. Seek ways to promote cloud's. Interest in wellbeing. I wish every parent and their children had such a extensive document discussing, Hey, I will explain to you my reasoning versus just telling you it's a, good idea. Do it So that could be a skill
Dan Peguine
Dan Peguine 1:35:19
Yeah.
1:35:20
Parenting.
LDJ
LDJ 1:35:21
and relating this, I don't know if you guys saw, but it wasn't, I
1:35:24
think just a few months ago maybe, that they announced they're allowing Claude to now end chats that. in a rare subset of cases where essentially, for lack of a better, better words, when Claude believes or, or no longer wants to engage in a certain situation, it can end the chat on its end.
Alex Volkov
Alex Volkov 1:35:43
Nice.
1:35:44
Yeah. Well, on this very interesting note, folks, I think, we're a little bit over time here with, a few, two hours and 20 minutes. I wanna just make sure that we ran through all the news of this week. There's been a lot of open source releases, a lot of TTS releases. cloud constitution is very interesting. I really suggest if you're interested in this, like next four we're gonna live in and we're gonna exist, side by side with these entities. Cloud is taking the absolute longest approach into figuring out how to shape those entities that will live with us, and not just, work for us and have an actual, if they have an experience, it's important to know that this experience is valid. if you haven't tried Cloud Bot, after all we told you, you definitely should at least try this with one thing that you connect to it and then chat with it on WhatsApp or Telegram or disco, whatever. It's really easy to do. if you have any other type of assistant, it could help you set up Cloud Bot. Like Dan said, he use cloud code, you can use Chat GPT, it doesn't really matter. All these, LMS can help you set it up. I'll say huge thanks. I wanna say huge thanks for everybody who came to the show. We're clocking in a little bit under a thousand people tuning into the show, for the live show. Thank you Dan Pegan for, joining us for the show to talk about your Clawdbot experience, Yam Peleg, Wolfram as well. LDJ and Nisten we're here as well. And, folks, my name is Alex Wilco. We're closing out our livestream for today. I will just say for folks who are just joining in the middle of the stream, or, folks who wanna revisit something where ThursdAI turning into a newsletter and a podcast every week, I'm hoping that Clawdbot, my Clawdbot will help me with this week's, work to turn this into a newsletter. And the podcast, if you missed any part of the show, all the links to everything, including the profiles of the people who we featured will be, published as well. and the best way to support the show is to give us a, like, to give us a subscribe and, send it to a friend or two if you think that they're gonna be interested in staying up to date with AI as well. With that, I will say, bye to you. next week I will just mention, we will go live from San Francisco, both Wolfram and I. As we prepare for the hackathon, we're gonna have a in-person San Francisco meetup, from the Roma offices in San Francisco, and we're gonna talk about some incredible stuff that's gonna happen from now until that week. So with that, thank you everyone for joining. I really appreciate your time here and we'll see you here next week. Bye-bye everyone.