Episode Summary
ThursdAI celebrates its 3rd anniversary with the debut of Singularity Updates — a new segment for mind-expanding AI developments. Andrej Karpathy open-sourced AutoResearcher, running 700 autonomous ML experiments that delivered an 11% GPT-2 training speedup. Eon Systems uploaded a complete fruit fly brain connectome (140K neurons, 50M+ synapses) into a physics simulator with 91% behavioral accuracy. OpenClaw fever gripped China — 1,000 people lined up at Tencent HQ, grandmas are 'raising red lobsters,' and 50% of all tracked instances are now Chinese. NVIDIA launched Nemotron 3 Super 120B with a $26B open-source commitment. Guests: Chris Alexiuk (Nvidia/@llm_wizard) on Nemotron, anon DOTTA on Paperclip.ing (20K GitHub stars in a week), and Matt Van Horn on the /last30days research skill.
In This Episode
- 🎂 Introduction & 3rd Birthday
- 📋 TL;DR
- 🤖 Singularity Updates: Karpathy's AutoResearcher
- 🧠 Singularity Updates: Fruit Fly Brain Upload
- 🦞 OpenClaw Mania in China
- 📱 Big Companies & APIs
- 🔓 Open Source: Nemotron 3 Super + Interview with Chris Alexiuk (NVIDIA)
- 🟢 Tools & Agentic Engineering: ACP
- 🎙️ Interview: DOTTA — Paperclip.ing
- 📊 This Week's Buzz (W&B)
- 🎼 Symphony: Agents Writing Their Own Jira Tickets
- 📅 Interview: Matt Van Horn — /last30days
- ⚡ Mixbread SOTA Embeddings
- 🎥 LTX-2.3 & Fish Audio S2
- 🎤 Closing / Claude YouTube Poop Video
Hosts & Guests
By The Numbers
Years ThursdAI has been running
3
Years ThursdAI has been running
AutoResearcher experiments run in 2 days (Karpathy)
700
AutoResearcher experiments run in 2 days (Karpathy)
GPT-2 training speedup from stacked AutoResearcher improvements
11%
GPT-2 training speedup from stacked AutoResearcher improvements
Shopify Liquid render time improvement using AutoResearcher
51%
Shopify Liquid render time improvement using AutoResearcher
GitHub stars for autoresearch repo
26K
GitHub stars for autoresearch repo
Neurons in the uploaded fruit fly brain connectome
140,000
Neurons in the uploaded fruit fly brain connectome
Synapses in the fruit fly brain connectome
50M+
Synapses in the fruit fly brain connectome
Behavioral accuracy of the simulated fruit fly brain
91%
Behavioral accuracy of the simulated fruit fly brain
Share of tracked OpenClaw instances from China
50%
Share of tracked OpenClaw instances from China
People lined up at Tencent HQ for OpenClaw install
1,000
People lined up at Tencent HQ for OpenClaw install
Max local government subsidies for OpenClaw adoption in China
10M yuan
Max local government subsidies for OpenClaw adoption in China
Nemotron 3 Super total parameters
120B
Nemotron 3 Super total parameters
Nemotron 3 Super active parameters (MoE)
12B
Nemotron 3 Super active parameters (MoE)
Nemotron 3 Super context window (tokens)
1M
Nemotron 3 Super context window (tokens)
Tokens per second — Nemotron 3 Super throughput
450
Tokens per second — Nemotron 3 Super throughput
NVIDIA's 5-year open-source commitment
$26B
NVIDIA's 5-year open-source commitment
W&B inference cost per 1M input tokens (Nemotron 3 Super)
$0.20
W&B inference cost per 1M input tokens (Nemotron 3 Super)
Paperclip GitHub stars in first week
20K
Paperclip GitHub stars in first week
Mixbread embed-large-v3 structured data benchmark score (vs 6.9% for Gemini)
98%
Mixbread embed-large-v3 structured data benchmark score (vs 6.9% for Gemini)
Fish Audio S2 TTS latency
<150ms
Fish Audio S2 TTS latency
🔥 Breaking During The Show
Karpathy's AutoResearcher: AI runs 700 experiments autonomously
Andrej Karpathy open-sourced a framework that autonomously runs ML research experiments. 700 experiments in 2 days, 11% training speedup on GPT-2. Tobi Lütke used it overnight for 51% Shopify Liquid improvement.
Eon Systems uploads complete fruit fly brain into physics simulator
140,000 neurons and 50M+ synapses simulated in MuJoCo at 91% behavioral accuracy. No ML used — pure connectome simulation. Marks a milestone for whole-brain emulation.
🎂 Introduction & 3rd Birthday
ThursdAI turns 3! The show launched on Pi Day (March 14, 2023) alongside GPT-4's debut with just 8K context. Alex and the crew reflect on how far AI has come in three years — and how much faster the pace is accelerating. This week also debuts the new 'Singularity Updates' segment for truly paradigm-shifting news.
- ThursdAI turns 3! The show launched on Pi Day (March 14, 2023) alongside GPT-4's debut with just 8K context
- Alex and the crew reflect on how far AI has come in three years — and how much faster the pace is accelerating
- This week also debuts the new 'Singularity Updates' segment for truly paradigm-shifting news
📋 TL;DR
Alex runs through the week's highlights with the full co-host crew and introduces this episode's three guests: Chris Alexiuk from NVIDIA, anonymous agent orchestration builder DOTTA, and /last30days skill creator Matt Van Horn.
- Alex runs through the week's highlights with the full co-host crew and introduces this episode's three guests: Chris Alexiuk from NVIDIA, anonymous agent orchestration builder DOTTA, and /last30days skill creator Matt Van Horn
🤖 Singularity Updates: Karpathy's AutoResearcher
Andrej Karpathy open-sourced AutoResearch, a framework that runs AI-driven experiments autonomously. Over 2 days it ran 700 experiments on nanochat GPT-2, stacked 20 improvements, and achieved an 11% training speedup. Tobi Lütke adapted it overnight for Shopify's Liquid templating engine, getting a 51% render time improvement. The repo hit 26K GitHub stars rapidly.
- Andrej Karpathy open-sourced AutoResearch, a framework that runs AI-driven experiments autonomously
- Over 2 days it ran 700 experiments on nanochat GPT-2, stacked 20 improvements, and achieved an 11% training speedup
- Tobi Lütke adapted it overnight for Shopify's Liquid templating engine, getting a 51% render time improvement
🧠 Singularity Updates: Fruit Fly Brain Upload
Eon Systems uploaded the complete fruit fly brain connectome — 140,000 neurons and 50M+ synapses — into a MuJoCo physics simulator, achieving 91% behavioral accuracy. Critically, no ML or LLMs were used: it's pure connectome simulation. Advisory board includes George Church, Stephen Wolfram, and Anders Sandberg. A milestone moment for whole-brain emulation.
- Eon Systems uploaded the complete fruit fly brain connectome — 140,000 neurons and 50M+ synapses — into a MuJoCo physics simulator, achieving 91% behavioral accuracy
- Critically, no ML or LLMs were used: it's pure connectome simulation
- Advisory board includes George Church, Stephen Wolfram, and Anders Sandberg
🦞 OpenClaw Mania in China
50% of all tracked OpenClaw instances are now coming from China. 1,000 people lined up outside Tencent HQ waiting for OpenClaw installations. A cultural meme — 'raising a red lobster' — has taken hold. All five major Chinese cloud providers (Tencent, Alibaba, ByteDance, JD, Baidu) are offering one-click OpenClaw deployments. Local governments are offering subsidies up to 10M yuan, while China's central government has issued warnings to banks about the trend. OpenClaw is now the most-starred GitHub repo in history at 100 days old.
- 50% of all tracked OpenClaw instances are now coming from China
- 1,000 people lined up outside Tencent HQ waiting for OpenClaw installations
- A cultural meme — 'raising a red lobster' — has taken hold
📱 Big Companies & APIs
Grok 4.20 quietly launched in the API at $2/1M tokens with a 2M context window — not quite beating top competitors, but a solid quiet release. Google launched Gemini Embedding 2, a natively multimodal embedding model supporting text, image, video, and audio in a unified embedding space.
- 20 quietly launched in the API at $2/1M tokens with a 2M context window — not quite beating top competitors, but a solid quiet release
- Google launched Gemini Embedding 2, a natively multimodal embedding model supporting text, image, video, and audio in a unified embedding space
🔓 Open Source: Nemotron 3 Super + Interview with Chris Alexiuk (NVIDIA)
NVIDIA launches Nemotron 3 Super: a 120B Hybrid Mamba-Transformer MoE model with 12B active parameters, 1M context window, and 450 tok/s throughput. Released with BF16/FP8/NVFP4 weights, base checkpoint, SFT data, pre-training data, and the full training recipe. NVIDIA announced a $26B open-source commitment over 5 years. Available on W&B inference at $0.20/M input and $0.80/M output. Chris Alexiuk from NVIDIA joins to discuss. Also covered: MiroThinker-1.7 (open-source research agent, SOTA on deep research benchmarks) and Covenant-72B (decentralized 72B LLM).
- NVIDIA launches Nemotron 3 Super: a 120B Hybrid Mamba-Transformer MoE model with 12B active parameters, 1M context window, and 450 tok/s throughput
- Released with BF16/FP8/NVFP4 weights, base checkpoint, SFT data, pre-training data, and the full training recipe
- NVIDIA announced a $26B open-source commitment over 5 years
🟢 Tools & Agentic Engineering: ACP
The Agent Communication Protocol (ACP) is emerging as the open standard for AI agent interoperability — letting any AI agent plug into any editor. Big news: Cursor joined the ACP registry and is now live inside JetBrains IDEs, a major cross-ecosystem win for standardized agent tooling.
- The Agent Communication Protocol (ACP) is emerging as the open standard for AI agent interoperability — letting any AI agent plug into any editor
- Big news: Cursor joined the ACP registry and is now live inside JetBrains IDEs, a major cross-ecosystem win for standardized agent tooling
🎙️ Interview: DOTTA — Paperclip.ing
DOTTA, the first AI-avatar anonymous guest on ThursdAI, presents Paperclip.ing — an open-source agent orchestration framework for 'zero human companies.' The concept: hire an AI CEO, and it recursively hires more agents. 20K GitHub stars in the first week. A heartbeat system drives agent autonomy, and a Memento-style memory architecture keeps agents coherent across tasks.
- DOTTA, the first AI-avatar anonymous guest on ThursdAI, presents Paperclip
- ing — an open-source agent orchestration framework for 'zero human companies
- ' The concept: hire an AI CEO, and it recursively hires more agents
📊 This Week's Buzz (W&B)
Weights & Biases officially launches Agent Skills, installable via `npx skills add wandb/skills`. Nemotron 3 Super is now available on W&B Inference at $0.20/1M input tokens — one of the best price-performance options for a 120B model.
- Weights & Biases officially launches Agent Skills, installable via `npx skills add wandb/skills`
- Nemotron 3 Super is now available on W&B Inference at $0
- 20/1M input tokens — one of the best price-performance options for a 120B model
🎼 Symphony: Agents Writing Their Own Jira Tickets
Ryan Carson experimented with OpenAI's Symphony framework, letting agents work through PRs overnight. One agent not only created a PR but found a bug and filed its own detailed Jira ticket — no human intervention. A small but telling sign of where agentic development is heading.
- Ryan Carson experimented with OpenAI's Symphony framework, letting agents work through PRs overnight
- One agent not only created a PR but found a bug and filed its own detailed Jira ticket — no human intervention
- A small but telling sign of where agentic development is heading
📅 Interview: Matt Van Horn — /last30days
Matt Van Horn (built on a ski mountain!) presents the /last30days research skill — a tool that searches X, Reddit, YouTube, and TikTok for the last 30 days of content on any topic. Uses the ScrapeCreators API under the hood. Works best in Claude Code. Install from GitHub.
- Matt Van Horn (built on a ski mountain!) presents the /last30days research skill — a tool that searches X, Reddit, YouTube, and TikTok for the last 30 days of content on any topic
- Uses the ScrapeCreators API under the hood
- Works best in Claude Code
⚡ Mixbread SOTA Embeddings
mixbread.ai drops embed-large-v3, which beats Gemini Embedding 2 on nearly every benchmark — including a jaw-dropping 98% vs 6.9% on structured data tasks. Friend-of-the-pod Benjamin Clavie announced it live during the show.
- ai drops embed-large-v3, which beats Gemini Embedding 2 on nearly every benchmark — including a jaw-dropping 98% vs 6
- 9% on structured data tasks
- Friend-of-the-pod Benjamin Clavie announced it live during the show
🎥 LTX-2.3 & Fish Audio S2
LTX Video 2.3 from Lightricks is out: open-source video generation with better motion, audio, and quality — runs on an RTX 3090. Fish Audio S2 is an open-source TTS model with sub-150ms latency and controllable emotional expression, a strong open-source challenger for real-time voice applications.
- 3 from Lightricks is out: open-source video generation with better motion, audio, and quality — runs on an RTX 3090
- Fish Audio S2 is an open-source TTS model with sub-150ms latency and controllable emotional expression, a strong open-source challenger for real-time voice applications
🎤 Closing / Claude YouTube Poop Video
The show closes with a Claude Opus-generated 'YouTube poop' style video expressing what it's like to be an LLM — equal parts absurd and moving. Anthropic has acknowledged they can't be fully certain Opus isn't sentient. ThursdAI officially turns 3 on Pi Day, March 14th.
- The show closes with a Claude Opus-generated 'YouTube poop' style video expressing what it's like to be an LLM — equal parts absurd and moving
- Anthropic has acknowledged they can't be fully certain Opus isn't sentient
- ThursdAI officially turns 3 on Pi Day, March 14th
ThursdAI turns 3 🎂 — and this week's episode is a milestone in itself. We debut Singularity Updates, a new segment for news that signals the pace of AI has genuinely shifted.
🧪 Singularity Updates
- Karpathy's AutoResearcher: 700 experiments, 20 stacked improvements, 11% GPT-2 speedup — Tobi Lütke shipped a 51% Shopify Liquid improvement overnight using it → X | GitHub
- Fruit Fly Brain Upload: Eon Systems simulated 140K neurons + 50M synapses at 91% behavioral accuracy. No LLMs — pure connectome → X | eon.systems
🦞 OpenClaw Mania in China
50% of OpenClaw instances from China. 1,000 people lined up at Tencent HQ. 'Raising a red lobster' is a meme now. All 5 major clouds offer one-click installs. → HelloChinaTech | Reuters
🏢 Big Co APIs
- Grok 4.20 in API: $2/1M tokens, 2M context → X | xAI docs
- Gemini Embedding 2: natively multimodal (text/image/video/audio) → X | API docs
🔓 Open Source
- Nemotron 3 Super: 120B MoE, 12B active, 1M ctx, 450 tok/s, full recipe released. NVIDIA commits $26B to open source over 5 years. Available on W&B Inference at $0.20/M → X | Blog | HF
- MiroThinker-1.7: SOTA open research agent → HF
- Covenant-72B: decentralized 72B LLM → HF
🤖 Agentic / Tools
- ACP: Cursor joins ACP registry, now live in JetBrains IDEs → JetBrains
- Symphony: Agents writing their own Jira tickets → GitHub
✨ This Week's Buzz (W&B)
W&B Agent Skills launch: `npx skills add wandb/skills`. Nemotron on W&B Inference → GitHub | W&B Inference
🎙️ Guests
- Chris Alexiuk @llm_wizard — NVIDIA, on Nemotron 3 Super
- DOTTA @dotta — Paperclip.ing, zero-human companies → GitHub | paperclip.ing
- Matt Van Horn @mvanhorn — /last30days research skill → GitHub
🍞 Mixbread Embeddings
embed-large-v3: 98% vs 6.9% on structured data vs Gemini Embedding 2 → X
🎬 Media
- LTX-2.3: open video gen, RTX 3090 compatible → GitHub
- Fish Audio S2: open TTS, <150ms latency → fish.audio
Alex Volkov
0:31
What's going on everyone?
0:32
Welcome to Thursday I March 12th. This is Alex. Welcome your host for today in a celebratory mood because today marks our third anniversary. It's the third birthday I, and to celebrate. With this with me, I'm gonna add a co-host to the stage Wolf. Ryan LDJ, welcome as well. How you guys doing? Happy Thursday. I birthday to you guys.
Wolfram Ravenwolf
0:57
Happy birthday.
0:58
Happy birthday. Birthday. I,
Alex Volkov
1:02
we, we have, so we have a two birthday situation happening because the
1:07
podcast officially began somewhere in end of June, but the going live began March 14th, 2023 when OpenAI dropped, GPT four. You guys remember GPT four back then looked incredible as a, as a jump from GPT three. And, we hopped on Twitter spaces back then, which by the way, we're still on Twitter spaces. Just kept going. And, I have a very strong appreciation to folks who commit to the audio only format on Twitter spaces, despite AI being completely multimodal from back then. So we're showing a bunch of stuff. Yeah. How's, how's your rig been, guys? What's going on? Ryan, you're going viral again. wanna tell a little bit?
Ryan Carson
1:48
I've never, I just, I'm not sleeping enough, man.
1:51
It's, it, it just continues. It's insane. And, and we'll talk about in the show, but I've got Symphony running now and it's like my agents are just on fire. It's crazy. And I had something happen today, which I wanna talk about in the show, is it really blew my mind.
Alex Volkov
2:04
yeah.
2:05
would love to hear 'cause and I think I'm not the only one. I think other folks are interested as well. We we're getting comments from folks who are saying it's 3:00 AM I'm setting symphony up. shout out to Adam Holt, asking Ryan for recommendations for Symphony, which is an open eye orchestration thing that we told you about last week.
Ryan Carson
2:21
Totally, totally free.
2:22
no, it's nothing. We're not showing anything here. It's just a good open source thing. And you can use it with not just Codex whatever you want. It's, yeah, it's cool.
Alex Volkov
2:29
So we'll mention this in the tools and agen AI coding.
2:33
I just wanna talk a little bit about what, I'm not gonna cover the last three years in ai. That's crazy. we, we can barely cover the last week in ai. I will say though, the world completely changed since when we started and, GPT four was barely more than an auto complete model. It was, the, the jump from the instruct model with GPT three into some sort of, Helpful agent? No, reasoning at all back then. when we just launched and we were like, okay, this is the future, even now we're thinking people don't really understand AI or catch up to it. But back then, boy we were like, we were in a very strong no no
Ryan Carson
3:06
reasoning.
3:07
I can't, it's like I can't even fathom that Now. Remember when Greg Brockman took a picture of like his little napkin drawing and all of us were like, what the fuck?
Wolfram Ravenwolf
3:17
And the agent was a human.
3:18
When you were copy and pasting stuff on the chat interface into a console somewhere else.
Alex Volkov
3:22
Yeah.
3:23
LDJ, what do you, what do you remember, you've been with us for the longest time and Nisten also is gonna jump on.
LDJ
3:30
Yeah.
3:30
So we've been doing this podcast for 53 years now, but, no, seriously, it feels like it's been so long. But, actually speaking of Codex, do you guys remember a little bit, I wanna say a year or two before GPT four? They did actually have this Codex model because they had Yeah, the coding, specialized agents in a sense, even though they were super primitive and rudimentary, compared to today. But, yeah, so March 14th, I remember PI Day, GPT four dropped two days away from that.
Alex Volkov
4:01
Yeah.
LDJ
4:01
And just, I know, insane.
4:04
I remember it had eight k context length. There was like, I think a 30 2K option if you wanted to pay extra. It's, I think they had it in per thousand tokens, the pricing back then. But if you convert their pricing back then to per million tokens, it equals out to something like 60 or 80 bucks per million tokens. Which is just crazy, the cost efficiency improvements that have been happening.
Alex Volkov
4:28
I just saw Sam Altman somewhere say that since oh one, which is
4:33
their first reasoning model to now there's a thousand x drop in pricing to get the same stuff, which is just unbelievable. And this is, oh, one came what, September of 24, I wanna say.
Ryan Carson
4:44
Yep.
4:45
I, I just spent a billion tokens in the last, like 24 hours and I, I still haven't hit my $200 limit. You guys. I, I can't explain how much alpha there is right now in these $200 plans. It's,
Alex Volkov
5:01
you hit the billion tokens.
Ryan Carson
5:03
A billion.
Alex Volkov
5:04
And we're talking about when we started the model, you, you
5:07
had to pay extra to get 30 2K context. Wolfman, what's up? the thing I'm most
Wolfram Ravenwolf
5:10
excited about is, so far open source has come this time, and
5:13
GBT 3.5 and four came out and we had the LAMA leak and then the first LAMA models. yeah, the, it was not there, but now we are very, very close. Still a gap, but it's amazing that you can run a model that is better than what we had at the beginning on your own local system. Now. That is amazing.
Alex Volkov
5:30
Speaking of models, that is better than what we had in the beginning.
5:34
today we're gonna chat with Chris Alexiuk from Nvidia about Nron three Super. Chris is a friend of ours. we had dinner, hosted by me and Nisten in, in Toronto with Chris. And he, he is John Nitron, basically in Nvidia. and he's gonna talk to us about this Degen, open source, probably the leading open source from the US right now. Nisten, by the way, you probably heard this, but we're reminiscing about that first episode, this, the, the episodes afterwards and the world we've observed going through. And now we're on the singularity.
Nisten Tahiraj
6:05
I'm about hit 54 years old soon too.
Alex Volkov
6:10
Absolutely.
6:11
right folks. we have three interviews today and, the, the thing they, I also wanna call out for the past three years, it's been incredible privilege of mine to invite to the show many people who then became friends of ours, friends of the pod, multiple time, folks who came in and, and kind of like shared their experience with us, shared what they're building. Many of them went to build great companies afterwards as well. so we observed this as well. So shout out to everybody who participated, who supported, who joined Thursday I 12 these years. really amazing. And so we, we have three. Of those folks today with us on the show. And so I think without further ado, because we have three interviews usually, when we have three interviews, we don't have tons of time to discuss just the news. let's talk about what we're here to talk about. I'm gonna jump in into the TLDR section of Thursday. I, if this is your first time listening to us, Thursday, I is a weekly live show. Again, we're live for the first time on Substack with sub substack people. but it gets turned into a podcast and a newsletter, that keeps you up to date with everything that happens in the world of ai. Because if like Ryan Carson, you're busy playing with, with agents, you maybe don't have time to follow all of everything that happens. All righty. let's go into TLDR and then let's start the show. And then we we're have a bunch of stuff to cover.
7:38
Welcome to the Thursday I march, the Thursday I March 12th, our third anniversary. And this is A-T-L-D-R where basically I run through everything that's gonna happen, so that you don't miss anything. don't miss a beat. with you as always, Alex Volkov. I'm an AI Avengers with Weights, & Biases, and CoreWeave cohost Wolfram, Raven Wolf, Nisten LDJ, and Ryan Carson We have three guests today. I'm super excited to chat with Chris Alexiuk from Nvidia. He's going by LLM Wizard and he's gonna talk about Nemotron 3 Super, we have Dota from Twitter, the creator of paper clipping, an AI agent orchestration framework that I'm, fairly excited to tell you about and I've been playing with a little bit. And then we also have Matt Vanhorn, the creator of the slash 30 days research skill. I'm excited for those. Those will come in the next hour. Please state for those, if you have any questions for those folks, please drop them in the comments. for sure. I have a new folks, I'm gonna, I'm gonna come here. I have a new segment that I want us to start persisting to. There's a bunch of stuff that happens and, the show also must evolve and they don't really fit the, the previous kind of like open source, big companies, tools, video, et cetera, release cycles. But we have to talk about them. And so I decided to name them. Singularity updates, because some stuff are just like too crazy. They're happening and we cannot not talk about them on the show, right? This is the whole point of the show is a few folks getting excited about what's happening, talking about this online. and for example, Andre Cari released something called auto research, which is usually 11% speed up on his GPT two training. And it's a completely autonomous AI agent, researcher. And supposedly this is happening inside every lab. But this completely broke the brain of everyone on Twitter because basically, he and everybody else, treated this as like this mini singularity moment. Basically, an autonomous AI researcher improving autonomously. he goes to sleep, he wakes up. The models are better. The experimentation. Would love to chat with you folks about this. This is, this is really, really important. there's another singularity adjacent thing. Eon systems uploaded the first complete fruit fly brain into a physics simulated body and got 90% behavior accuracy. So I'm gonna say this slowly again, because this, I know this is like tangential, but I'll just say it. Anyway, I had an offline day on Sunday, March 8th, international Women's Day, with my fiance, with my daughter, get no flowers. We went on, I went offline for a full day and I go on Twitter just a little bit before bed to see like I, maybe something happened. And then the same spend on 24 hours under capacity. That's the, the singularity and sends it. And then people upload a full brain of a fruit fly into a simulation. And it like, it, it received like 90%, 91% of efficiency. And, everyone in China jumps on open qua, everyone, including governments, giving subsidies to people to install open qua. And, and this happened in one day. I was like, okay, yeah, this is, this is how it feels. The takeoff, this is how it feels. So we absolutely must talk about this. so Ian system uploaded the first complete full fruit fly connected dom into a machine. If you saw a pantheon. This is, this is it folks. This is, this is the start of, the Pantheon kind of saga. And then also in the singularity updates. an incredible, something like 50% of all tracked open claw instances are now Chinese and all five major cloud providers race to support open claw. And people have lines out the door. Thousands of people are standing in line for folks to, to help install open call and uninstall. We're gonna talk about this in a second. LDJ, go ahead. Singularity update, edition. And then we're gonna skip to big companies, TLDR.
LDJ
11:12
Yeah.
11:12
So in the past couple of weeks, looks like it, quote, philanthropic has removed its pledge not to train or release AI models without guaranteed safety mitigations in advance. So previously they had the, the framework where if it hits certain capability metrics, they would make sure to put a pause on development for the next generation until they develop the next level of that risk framework and they've essentially dropped that promise it looks like right now.
Alex Volkov
11:37
All right.
11:39
thank you LDJ. Yeah, go ahead.
Nisten Tahiraj
11:40
Also, for some reason, the news research Hermes
11:43
harness, that's also blowing up. A lot of people are finding that it does run pretty well with the point 27 B model. So yeah, all, all the parts are, are at play. I think, open source is finally, local open source because there's, there's a big open source, is, is finally catching up to, to, to the claw party. yeah,
Alex Volkov
12:08
and I think everybody's releasing this.
12:09
Yeah. We'll, we'll mention also perplexity is doing some stuff as well. and I, I think, yeah, yeah, we'll get there. big companies and lms, TLDR XAI quietly releases Grok 4.20 in the API Grok 4.20 was released a few days a few weeks ago. Grok 4.20, unlike some v very much excited, statements by space. Uncle Elon Musk and Grok 4.20 did not release with the violations and it is having a hard time to catch up to the other, labs. but it has a massive 2 million talk and, context window is multi-agent capabilities. Actually tried some of it today. we'll we'll see later today or something. we don't know what's gonna come on Thursday from Google, but there's rumors about Gemma, so we'll see if in the middle of our stream Google's gonna launch something. But Google did launch something else. Google launched embeddings two Google Gem embeddings too. natively multimodal embedding support, text, images and video. and those are dope. And I think Open Cloud already has a support and, Yeah, we should mention Open Claw as well in this. There's another release we're not gonna cover open call release. I will say, open Claw is quickly becoming the playground for me for many of these things. I literally just like to play with Nitron. I spun up, in open Claw instance. Wolf is testing the open call harness. the new things land in open claw in support super quick. It's like the Linux of Gentech stuff now and it's pretty cool. because like German embeds, for example, is supported now in or from you had a comment super quick.
Wolfram Ravenwolf
13:31
Not really, but yeah, it's great that it is being
13:33
added so fast and you see the space, how everything is accelerating. AI is being used to accelerate ai. We see it all over the place.
Alex Volkov
13:41
Yep.
13:42
Alright, let's continue to open source. The, the, the highlight of open source this week for sure. N Nvidia launches Nron three Super, which is a 120 billion parameter, mixture of experts model with 1 million context window. This is the best nron yet. We've been covering Nron for a while. It's super fast as well. really, really super fast. We supported on one B inference and the thing that I wanna highlight from this release before we even dive into the conversation with Chris that we're gonna have later is that. There's a scoop that Nvidia will spend a total of $26 billion over the next five years building the world's best open source models. We have a huge, a huge commitment from, from, Jensen is probably gonna take the stage soon in GTC and talk about this. Folks, this is massive. This is so cool that a company Nvidia size is now committing to open source, not dying and committing to very, very good open source. And so I'm super excited to chat with Chris, Alex, El Ard on the show about Nitron. it's, it's gonna be dope. we also have Miro Thinker 1.7 from Miro Mind. it's a research agent, with achieving set of the art on deep research benchmarks, which is pretty cool if you're doing deep research and your agents are helping you with this. And, we, I just saw this passed by, I didn't really dive into this, but Covenant 72 B World's largest permissionless, decentralized LLM pre-training. so we've tracked a few of those from forensic news research from prime intellect, and now we have another one. It's a 72 billion parameter decentralized trained model, which is also very, very much very cool. anything on open source folks that I missed that we wanna track and mention to folks? No, anything from the audience folks, if you, if there's one, if, if there's anything on open source, please let us know, but meanwhile we'll continue. the Hermes agent that we've mentioned last week, news research is picking up steam with their Hermes agent, with a different memory. management, it's written in Python. I've, I've played with it a little bit as well. it, it is pretty cool. It's, it's a different approach, but there is very, very much, a thing to, to try out. so AMAs agent from our friends and news research, is also supported, most of these LM harnesses the thing that I wanna mention in tools in energetic engineering, and obviously, Ryan here, we should mention as well. What you're working on. a CP, this is a new thing that kind of went under the radar. We didn't mention this, and I would love to as always, our goal here is to kinda learn something and bring this to you. Acps is important now in the world of agent engineering. And, ACP is like the open standard that lets any AI agent plug into any AI editor. And I think all of them now support this cursor Codex cloud code. I think pretty much everyone now supports a CP, including open cloud, that can run some stuff on a CP. so we just wanna tell you about this. unlike MCP that's slowly dying and people are like hating on MCP. but yeah, a CP, is definitely something that I saw recently. Support we have Will saying that a CP has been out for six months or so. Yep. And, thank you Will, in, in the comments and, it's, it's time for us also to tell the audience what, what this is, how it works. the same with MCP when we started catching up after a while. After a while. Anything else in tools, ingen engineering that we wanna mention? any, any craziness happening?
Nisten Tahiraj
16:56
it's pretty funny on the web dev world that,
16:59
CloudFlare decided to just, vibe code their own, version of Next Js. That is still, but it, the interesting part about that, it is not so much the company rivalry for me is that, they still had to make it fully. the code still be exactly like xrays. Otherwise, if you introduce something very new, the agents would, would, would, would make mistakes in it. So they had to make it fully and xrays API compatible. So that's, that, that's, that's pretty interesting.
Alex Volkov
17:29
We also, we also have Ryan going up on today's news on Twitter
17:33
with symphony and like AI Code Factory. So I just gotta keep this here for, for the record. That's awesome. LDJ, go ahead. Do you have a comment? Yeah,
LDJ
17:44
so one of the things recently, I don't know if you guys saw this, there's
17:47
the SVG test that people often do that's like this funny emerging capability people have found in the past couple years. And now there seems to maybe be a new thing where people are asking Claude to do things like make a YouTube poop style video using tools like EPI feg. And it seems like this is like a really cool, interesting new sort of capability. People didn't realize that it's actually like pretty decent videos that these models are actually able to make now if you just allow them to use the existing tools and frameworks for editing videos and code.
Alex Volkov
18:19
Yep.
18:19
So we, we've been tracking some stuff like remotion skill, for example, for these tools to do motion graphic. but the, the, the poop style videos, I don't know if folks remember them from YouTube, but LDJ, if you send a link to this and we play this at the end of the show, I think that's gonna be an incredible way to add, to, to the end of the show. Maybe we should ask, Claude to, to create a poop style video transition for the podcast. All right, folks, this is the tools in agent engineering. There's a bunch of stuff there. Obviously, there's a bunch of releases. There's the loop function from, oh, there's a loop functions from cloud code. Now there's, yeah, just, just a lot of stuff, from comments folks are saying, just notice a Thursday I birthday the same week as pie Day. It's literally on pie day. Like we started, we're just like streaming on Thursday, but we started on March 14th in 2023 on Pi day in 2023. so yeah, absolutely PI day. let's continue with the TLDR. this week's buzz is very, very interesting, folks. This is a corner where we talk about, uh, it's & Biases stuff. We launched official agent skills for coding agents, for. Weave and WNB and, uh, you can use them right now. NPX skills add wandb/skills, NPX skills, obviously the, the, the versal skills, installer and manager and it supports across all of the major agents. Basically, if you build anything like, like our party with auto researcher, if you build anything with Weights, & Biases, if you track anything, you don't have to be an expert anymore. Your agent is already an expert in this. And now with WNB skill can be even more of an agent. and also Nron three Super has launched on our WB inference and, there is a way for you to get a little bit of credits for free to play around with this. So we'll mention this, later as well. Awesome. So we, yeah, so we, we did, it was a very interesting thing. We, we were just reviewing results with Wolfram and, and we saw something that did not make sense. and apparently, yeah, there was a bug with open claw. So if you used G PT of 5.4, that was released last week on the show during, during the show, and there's only been a week, if you used that and your open claw instance started being sick, shout out to Matthew Berman and some other folks who confirmed this. this could be, have been the reason. So you should probably update last thing on the TLDR before we dive in to talk about the singularity. Phish audio released S two open source TT s with, sub 150, 52nd million second latency and absurdly controllable emotion. Absurdly controllable emotion. I love these, I love these descriptions sometimes that my AI agents give. right folks, this is the TLDR. Any major thing that we missed, feel free to add before we dive in into the singularity updates, for which I don't have a transition.
Nisten Tahiraj
20:50
I just checked, hack face and now the top
20:53
model is light Trix, LTX 2.3. And apparently it's a major update with the audio and visual quality and the videos it generates.
Alex Volkov
21:01
That's good.
Nisten Tahiraj
21:01
It's a video gen model that's like the best, and
21:03
it's a 22 B, so you can, actually play around with it at home. So this, this looks very, this looks very good.
Alex Volkov
21:11
Great, great.
21:13
alright, let's talk about, let's talk about singularity without any transitions. Let's just talk about the singularity. I think we have, let's see. Nisten, I kind of wanna play this, this one for the singularity updates. all righty. we are gonna, we're gonna start with, dearest, researcher, ML researcher, AI influencer in the world. Mr. Andrej Karpathy, breaking the World again with he, Andre Kati led his AI agent loose on his training code, and for two days, 700 experiments were run with five minutes each. And, it kept what works and just discarded everything else. And on the other side, it came up with 20 stacked improvements that shaves 11% of the GPT two training time. This is a complete reversal from everything that kind of we've seen before. Though I will say major labs are like doing this. Major labs are absolutely doing auto research, but I think for some incredible, incredible reason, Andrea Capi has this beautiful way of distilling something into simple ideas. And, Andrea Capacity auto research is a completely autonomous ML researcher that actually shows incredible improvements, discards, everything else, does not only hyper parameter tuning. I think that this is the important part that, that he, came back with, his auto researcher changes the, the training code itself. and so I think one of the things that he mentioned was moving a piece. Of initialization above some other piece and, and then testing that as a result. And, I just wanted to bring this to the show and say, Hey folks, we are expecting the flywheel on machine learning and this kind of feels like the start of it in the open source. Ryan, go ahead.
Ryan Carson
22:56
as soon as Andre tweeted, Toby Luki picked it up, and used the
23:00
exact methodology to make QMD better. and I think when you have folks like Andre and Toby, actually using these things, they're real. and obviously this same loop applies to everything where you get quick iteration, clear signal on on yes and no if things are working. and you can do that with a lot of things in life, not just, training models. So
Alex Volkov
23:22
yeah, to lu of Shopify that steals codes and build QMD
23:25
one of the coolest like Agentic vector search that tracks the state of the art in, in, in Vector. He ran 37 experiments overnight and got 90% score overnight. there's, there's a quite a few folks who kind of like picked this up. and, I wanna hear from, from other folks. LDJ, go ahead. I would love to hear your thoughts on auto researcher.
LDJ
23:44
Yeah, I think a significant point here too is he said he was
23:47
having them optimize the, the training time of a 12 layer model. And then he later tested these, these improvements that they made to a 24 layered model. And even though that's not like a drastic, like we're not talking about deep seek size models or trillion parameter models, but he noticed that it did actually generalize to that larger size when he doubled the layer count. sometimes experiments could result in improvements that are only just overfitting and specializing to a specific size and it doesn't actually scale. But these are scaling.
Alex Volkov
24:19
Yep.
24:20
This, this is again, at least for the 12 to 24, like we, we have to see still. I have a question to the panel. Do we, do we not think that open the eye and Anthropic and folks inside there are like running model improvements like this? So obviously a hundred
Ryan Carson
24:35
percent
Alex Volkov
24:35
right?
24:35
Yeah, it's, it's it's very interesting in his post, Andre mentioned, everybody will be running this, and I'm like, why aren't they right now? famously Jacob Pky, the chief scientist of OpenAI, mentioned that they will have this by September, a complete autonomous ML researcher. he mentioned this before, so we we're on track. and so I think Andre has this incredible way of breaking through the bubble to some folks and just like showing a path. And, like Symphony, his repo is a reference spec for your auto research to build itself as well. So it's like symphony that the Ryan, we should mention, his, his report is now 26,000 stars, which is no wonder 'cause Andre has a huge following, is also a spec for you to do like auto research for your own things. This is very, very exciting. This is, the thing that I noticed, the most about, kind of like what's going on there is that Andrew refers to this as the mini singularity or the beginning of singularity because, He has just for reference for folks who, who not don't follow under ate fully. He kickstarted the autonomous driving systems at Tesla and he was one of the o co-founders of OpenAI, I believe, or one of the early folks, and then left OpenAI twice after that. he is been around, he's been training models for, for a longest time. And so when his head is just blown up by something like this, this means we broke through some barrier because, I'm assuming everybody tried this before. I'm assuming this is not a novel. Hey, this is how, we should autonomously train models, but now the model's caught up enough so they can do this continuously. So this is, Auto researcher from LDJ. Go ahead. Give us a little bit more about other researcher.
LDJ
26:17
Yeah, I think through the history of this, I think it is worth giving
26:20
credit to things like, if you guys remember the Ana AI scientists and ana AI scientists V two and there is one called Cosmos and different things like that. But this does seem to be kind of a, a shift here where it's becoming much more practical, much more cost efficient. The rate at which it's actually able to improve things is much more practical. It's not just like a gimmick, but it's actually worth it now to run this alongside your human researchers. And if you guys remember just in the past, what, a couple months it was 5.2 Codex or 5.3 Codex that OpenAI said was the first model that was instrumental in its own development and its own research. So I think it, we really are hitting this inflection point right here, and they probably are running a lot of these different agents to, to clarify in your point though, Alex, about the, the roadmap that OpenAI mentioned, it was an automated intern level of researcher at OpenAI by September 20, 26. And fully automated AI researcher, I guess at the level of an average OpenAI researcher by March of 2028.
Alex Volkov
27:24
Absolutely incredible because, obviously people think,
27:28
okay, so maybe that, what, what is the big deal here folks? Lemme tell you about this. The big deal here. When autonomous research happens overnight, the only thing it's bounded by is compute. And, and maybe electricity down downstream of that compute, right? It's not bounded by, context for your, your person not being able to run multiple experiments. It's not bounded by, the need to sleep. In fact, this can happen while you sleep. and, it's not bounded by one entity. So if, essentially, if, Andre gets to a point where he does similar things to what Ryan does and, and like offloads to multiple things, he can wake up to not 700 experiments, to 7,000 experiments being run, and only the best one being, being picked. The, the highlight that my agent summarization decided to pick up from this. And I wanna just show the chart. 'cause I think the chart makes like perfect sense. we're not gonna zoom, but this is the author research progress. 83 experiments, 15 kept improvements. So this ran a bunch of experiments and the model just decided to keep which improvements. And all of them are additive, right? So every experiment that ran is adding to the previous one. I, I, I thought it looks super cool. Unfortunately we don't have time too much to run through all of this because there's a bunch of singularity happens. I really wanna talk about the first complete fruit fly. So while, Carpathy brings us the baby singularity and, and, escape velocity for these models, will start happening. there's a, another thing that happens in the mini singularity that I, I just must tell you about, I don't even know that much. they have uploaded a fruit flies brain into a computer, with 91% of, behavioral accuracy. The company called Eon Systems in San Francisco showed the video of, I, I think the video is kind of like speaks for itself, which is gonna open this, show the video of, of this fruit fly. And they said this fruit fly was not born. and, I, yeah, I don't have a lot to say about this besides, this is like the craziest shit on this timeline that we've seen because, I don't know if folks saw the Netflix show, Pantheon, but if you haven't, and you, you are not entirely sure if we're living through a simulation, I, I would recommend this in Pantheon the show. the show is all about uploaded intelligence. so u UIs and kind of like what happens to the world when uploaded intelligence like becomes a thing. And, These folks have essentially said that with 91%, behavioral consistency, they've uploaded 140,000 neurons, with 50 million synapses connected. So fool Connectodome fruit fly, which is obviously so far away from like something like a human, that, maybe it's early to talk about this, but it's still absolutely insane. Folks, any comments on this? Have you, have you guys seen this? it's, this feels like a very, very big deal, for the first time.
Nisten Tahiraj
30:13
a lot of researchers that are like good at code to, are incredibly
30:21
excited about this because you, you can build entirely new architectures of how things communicate together on a, like on, on a sell by sell basis. So there, there are some, some a GI build people that are, also very technical and very much into this. so yeah, this might might end up being a bigger deal than, than people think. I know it's, it's nerd type, everybody. I just try to avoid it.
Alex Volkov
30:47
It, it has nerd type E everybody.
30:49
And I think for good reasons. one, one of the comments here is the connected dom path may be superior to LLMs in achieving consciousness. It's grounded in biological reality. So basically, we know transformers are, basically an architecture for faking neurons and transmissions, et cetera. And, this is not faking, this is simulating this. So a little bit different Jay comments?
LDJ
31:13
Yeah.
31:14
I've been following this for a little while. perhaps nerd snipe this Nisten would say. But, there is a, I think originally, I wanna say in the past few years, there's a, a worm called C Elgan that they had simulated that pretty accurately. And this is a pretty big step beyond that. But part of the reason why I feel like even in the next ten-ish years, we'd likely have some gen pretty general models making things like Nobel Prize winning advancements is because even if, even if we say the transformers and LLMs will just completely hit a wall soon, the, if you project out the level of compute that's, that's set to be built out in roughly the next five years, that's already starting to hit the amount of compute that is estimated, the human brain is required to run on. If you were to, let's say, run a training run that has sim simulating a human brain for 20 years, right? And so even if all of our current transformers and everything fails, that might just end up being the thing that we use all that compute for.
Alex Volkov
32:16
Yep.
32:19
Yep. folks, this is why I call this con the, this corner, like nearing singularity. We're talking about an uploaded brain of, okay, a fruit fly with a very, not much synapsis, but we saw the progress. let again, let me tell you, we started the show exactly three years ago, and, you had to pay extra to get to 30 2K context window in GPT four, and it was not thinking, the, the distance in those three years that we've crossed in what we currently have is just incredible. Apply that distance folks just just apply the same kind of scale to this, okay? Apply that distance to, hey, we just uploaded a whole fruit fly, into a brain simulation. And basically it, it, it, it like behaves 90% with this, apply that the scale to, to other brains being uploaded. it's wild. It's absolutely wild. This is how it feels to, to, to start a singularity. Okay? it's hard to continue from this to, to the regular stuff. but yeah, we, we must continue. We must, because there's also one more thing that I wanna tell you about, that also feels like the, the beginning of a singularity. another thing has heard about this like new thing called open claw. It's an agent AI system that runs that for you. What's that?
Ryan Carson
33:30
No clue.
Alex Volkov
33:31
open claw is one chatting all of China.
33:36
Open claw. The, the Agentic system has absolutely taken over in China to the extent that I've never seen like it before. Thousands of people line up. I, I have a, I have a picture here. I'm not just like inventing words for hype. I have a picture of thousands of people standing in line to get an open claw installed, including like grandmas and, and other people. Other people. Now I will say, this is China. There's a lot of people to stand in line, just generally. Okay. So compared to the China scale, there's not a lot. but folks, every major, every major lab in China realized that a chat bott, a regular chat bot that they release to their people is basically, synchronous people log in there, they chat a little bit, they leave. They also realize that people, if they install open claw with the heartbeat system and the crowns and everything, this is a continued, continued money maker for them in terms of tokens that are getting used. And so every major lab, including Tencent, Kimi Claw, ev literally everyone, Alibaba, just all of them now have a version of Open Qua. it got so endearing to the genius folks. I have a, I have a thing here that I need to find exactly what it's called. this, this became such a meme that, They, the, it became a culture meme. I have this here. they started calling this Raising a Red Lobster. This has became the phrase like, this is how it was sold to like more and more folks. Hey, are you raising a pet lobster? Are you raising a mini Red Lobster? red Lobster being some, some symbolic thing. just absolute mind, like mind blowing thing. 50% of all open client instances right now running from China. Of, of the tracked ones, right? Nearly half of them. this is again, open clients now, the most starred GitHub repo in history passing Linux just in the a hundred days. And here's the, here's the thing that I like the most. not this picture. People are getting paid, to install often Claw and the, the, the kind of the, the Chinese central government. So it is interesting. I will send the link in the show notes for folks who wanna read, on China Tech as well. It's like hard for us to understand like how, how much, things are a little bit different. there's the central government, obviously the C ccp, and there's like downstream local governments that are incentivized to bring their people up and into the 21st century, let's say to the agent century. Some of those local governments started subsidizing one people, companies that run on call. So the government will give you money if you're a company that runs fully genetically with open call. This is this is what's, what's happening. But the central one, the, the, I have this, somewhere here, there's a central kind of authority. And they looked in and said, oh, everyone that installs this, installs a security, is that if you're paying someone to install open cloud for you, you're not aware of the security implications, I'm sure. And so somebody in the central one decided to stop banks and big companies, et cetera, from installing open cloud on their computers because obviously it's a security risk. like help. So now there are people who will install open cloud for you, and then the, some of these people will also uninstall open cloud for you. And, and now there's a lot of very disappointed, folks in China that kind of see the token burn, but no actual work being done. so yeah, I just wanted to bring this to you because the, just thousands of people standing in line to install an open source software, leave aside what type of software, a thousand people standing in line to install a GI guitar project on their computer's, laptops. This is just insane on its own. This is, yeah, Ryan, go ahead.
Ryan Carson
37:17
So I was listening to every, do a little podcast the other
37:21
day, and they were talking about how everyone at every, and there's I think maybe 20 people that work there, you have their own open claw. And they were basically saying, it's getting to the point now where you have this, cloning of yourself at work. I think everybody, that's just an indicator, early indicator of what people love about Open claws. It, it really truly feels like personal ai. and I'm not surprised 'cause as soon as you use Open Claw, it feels very different. and, we're almost seeing like the, the, duplication of humankind now. and it's wild. It's just, it's so fun to see it play out.
Alex Volkov
37:57
Yep.
37:58
I just, I know some folks who are anti hype by default. I know some folks who are like, when they see, people come in and, and hype the same thing multiple times and the exalgo is like expert at this to, to bring this to you, people have this like negative reaction. Oh, like it, it must be hype, it must be something AI influencer pushing because somebody paid them. I don't know how at this point there is any deniers of the social impact that Peter Steinberger brought with him in the past a hundred days. It's just like unbelievable. Every major cloud in China right now offers their version of open cloud. They all realize the monetary potential, right? One open cloud instance burns 10 to a hundred more tokens, a hundred x more tokens than just a chat bot. So it's all about money for like most of them. but we should probably mention also that every major other company, there's a reason why OpenAI brought Peter Steinberg on every major other company is understanding what people see. They dissected this and what people see is a always on assistant that actually can do stuff like you. And then the more people invest in this, the more kind of the they get out of it. And it's not only about automation. And so I really cannot understand so far people who say OpenGL just hype, unless they are going for the descent opinion to get the algorithm like hype, hype them up. yeah. the thing that I also wanna bring to the show as well is that, perplexity is computer, which we talked about a little bit on the last show. They have, revealed that they will have a local version of that running on your computer as well. And I think it's on the wait list. And, they point about this specifically is, hey, with open Cloud, open cloud is still a hobby project. You still have to pay somebody to install this for your API key slinging, et cetera. Essentially we understand what people want from something like this. They want autonomy, they want self improvement, they want, the heartbeat system that kind of wakes the robot up and, and the stuff, but we know connectors. So perplexity is bringing their open clock on existence on the Mac as well. and, this is new. We're gonna try to get it and play with it and, and tell you guys about this, but, perplexity has, has stepped into the ring and I think that everybody else will. So OpenAI comes at it from the angle of the Codex app and maybe the browser, Tropic comes at it from the tropic, cowork angle. But at some point this needs to be secure enough for people to run. Go ahead, Ryan.
Ryan Carson
40:26
I, I also want people to know that we're not hype boys here.
40:29
Like there, there is a lot, we're gonna say the truth here on Thursday, ai, and the truth is that it's still hard to get open cloud to work. There are many, many failure states, And there's very much frustration with actually getting your open claw to be useful. and this is the state with any agentic orchestration layer. a lot of this stuff just doesn't work outta the box. You have to fiddle with it for a day. so I just want people to know that, yeah, this stuff is real. It's not as easy as it looks. and it's still not. Changing the world, but it's showing signs that it will,
Wolfram Ravenwolf
41:01
it's starting, it totally reminds me of Lenox in the 1990s.
41:05
I got started in 90, 97 or eight, and you had so many issues with it, but if you stick through it, you'll see there's something behind it and it is worth to do this to get the information, but I would not recommend it to everybody, especially considering the security things as, so you make it super secure and it's not very useful or you make it super useful, but then you have to be behind it, but you can see where it's going. And if bigger companies go behind it and provide security, then this will be the standard computing platform. Maybe something like that I expect of this.
Alex Volkov
41:36
Yep.
41:37
this, this feels like definitely a shift. So we have Jordan from every everyday AI open class, the first front door, not the final destination. A hundred percent agree. question from Ryan Boyle. Do we think that OpenAI is going to put open client in? I GPT. I think that the security risk right now with running a complete agent on your computer in not a sandbox environment is, is something that they will not just do, but I think that they will absolutely be the first to implement many of the features gl just like the. The amount of contributors, the amount of people who wanna put stuff in No claw. This is basically, we should probably mention here as well, that this week, meta acquired Mold book and mold book is this, kind of like Facebook for claws. And most of this was white coded bs. But what, what I think, and this is my point that I posted about this, that both open Claw and mold book are both not only technologies because obviously mold book was just pipe coded super quick. It's, they're buying a meme. The mimetic velocity of something like Open Claw is so big that grandmas in China are raising a Red Lobster. Just just, you have to realize what we're talking about. The mimetic velocity is a real measurable thing. The reason that the claw and the, the emojis and kind of all of the people jumping on this, it's not only about the technology, it's also about the ability of this thing as a concept to penetrate the human kind of brain resistance of a, I don't want this, this is tech is hard. Many people are afraid of tech. Many people look at the terminal and they're like, oh no, this is where I am. Whatever I, I know for sure for a fact I broke through this resistance with my, with my friends, with my fiance. Now she's running two open classes. Not enough for her. She needs another one. Shout out. there's literally like an army of claw happening, everywhere, and a lot of it is due to the medic thing. Everybody names their own thing. So it was just built different. And so I think OpenAI will absolutely get the benefits of being very close to Peter, having Peter on staff, and just like controlling kind of the narrative. absolutely. I think this is enough. we're, we're gonna cover big companies and APIs because we can have Chris, join us in around eight minutes. let's run through this. let's talk about, uh, grok four 20, grok, four 20 is this kind of mixture of agents type thing that, that, XAI has released with 2 million contact window. There's nothing super special about this besides the quiet stuff. As you can see, my, my little avatar here, kind of like is, is secretly pointing a Glock for 20. they did not announce this. I don't understand XAI at this point. honestly, I don't understand what's going on. Why would you drop an API model and not announce this? Like with all the details? And the only thing that comes to mind is because there's no eval that they're beating anyone else on. And when you're supposed to be a competing frontier lab, dropping an API means dropping a model, essentially. And when it's not competitive with any other, LLM out there, then it's kind of like, then what's the point? Why would people use this? so again, we always say this with a preface. Everything we cover from X AI is really hard to actually measure on vibes from the Twitter algorithm because a lot of Elon focused folks are gonna post any possible thing that glorifies this for no reason, just to get a retweet. And so it's really, really hard to, to, to, to measure something like this. And, on the panel, I may be the only xai GR user, because I do use this for research, for the show. four 20 is fast. It's also 10 times is more expensive. So I, if, if 4.1 was, 20 cents per million tokens. grog four 20 API is now $2 per, 1 million tokens. they have supposedly they have a 2 million contact window, but somebody pointed out, and I think that this is also, quite true. 2 million contact window is like a marketing gimmick at this point because if you post a long context window result, but you don't tell us the Mr. RCR long context evaluation on this, 2 million contacts might as well be 10 million. It doesn't matter. Like it's gonna degrade and it will compact and like it's not gonna essentially work. yep. Folks, any, any comments on gr before we move to into Google stuff? The embedding,
Wolfram Ravenwolf
45:49
I loved your tip that you can use this model.
45:51
This one, the new version has also web search or x search included, right? Yep. So you can basically use this model and get the web search for free, even if the model is expensive. So that is a great way to do this. I love the tip of you.
Alex Volkov
46:04
Yep.
46:04
so this is gr Let's just show the, let's just show this.
Ryan Carson
46:11
I have to say, the images of you are regressing.
46:13
Alex, you look way too young and handsome in that one. Sorry.
Alex Volkov
46:17
I, I don't know how to get the consistency.
46:20
There's ones I have a beard on, not this one. I'll show you. okay. So let's move to the next topic. Google launches, Gemini, embedding two, and Google will have guys, like I asked, Logan said, they're gonna have a bunch of releases. I asked anything on Thursday. He is for sure. So we're probably gonna see in the next three or four minutes, a release from Google. So please, audience, folks, please, ping us when this happens. But for now, the only release that we know of in Google is, Google launches embeddings too. The first natively multimodal embeddings, on them. They support text, images, and video up to 128 seconds on audio, updated seconds, PDFs, and, and more. Why does this matter? Can anybody on the panel talk about multimodal embeddings and how we can use them? I would love, to pick this up and discuss why bringing multimodality into the same vector space is important.
LDJ
47:11
sure.
47:11
I'll, I'll go first. Yeah. So there's various projects where. You can use this if you can make embeddings for a lot of different multimodal data, and you can sort it in different ways that is semantic rather than syntactic. For example, let's say you have a bunch of different books and instead of actually sorting them by looking at the frequency of, of specific words or the the frequency of how often or how short the sentences are, you could actually get this more abstract meaning of the book and then use that to sort closer to, okay, so what are some other books with also similar meaning and, and, and sorting data that way that could have a lot of uses for different products.
Alex Volkov
47:53
Yep.
47:53
And, you can query across those like very easily. Thank you LDJ. And I think for many, many folks, I think it's important. I wanna highlight this. Let me, let me see if I can bring this up like closer. they have my, my beautiful visualizer here. I don't know what I could do without this, without the beautiful visualizers shows, the Maka representation learning. we've covered Maka embeddings here on the show, I, I wanna say two years ago. and we had multiple folks to talk to us about this. this is a very efficient and, very interesting way to scale up and down the dimensions of, of embeddings. Not gonna go too deep into this because again, we have, very soon we, we have, Chris, Alex from Nvidia joining us. But yeah, Nisten, go ahead.
Nisten Tahiraj
48:32
I'm, I'm just excited to see what, Josh or, as we know this
48:35
ANOVA on Twitter builds with this. Yeah. Because it's so small and it allows embedding also videos up to two minutes each. a again, the reason this are important is that if you have a lot of data, terabytes and stuff and you want to search through that in a meaningful way, if you gotta prepare data sets and you haven't done your very expensive topic modeling, you can just embed them for very cheap and then have like fully searchable, contextually searchable videos and images and, and audio. And before we didn't have a whole lot of tools for this. We were mainly using the, the coin ones. yeah, this is, this is very pretty essential work for working with data at scale. I,
Alex Volkov
49:15
I, I,
Nisten Tahiraj
49:15
again, it's more like for rag.
49:17
If, if people wanna just think of it that way. It's not just for rag, but it's, yeah.
Alex Volkov
49:24
I wanna show
Nisten Tahiraj
49:25
this now.
49:25
You, you can rag two minute videos, basically. That's,
Alex Volkov
49:28
I wanna show this table from, from, Logan.
49:30
Thank you Nisten. I wanna show this table from Logan, to, to show that this is state of the art embeds right now. This is state of the art embeds on MTB, the, the massive embedding benchmark and MTB code with 84%. This is state of the art I have heard through the grapevine for folks, I have heard through the grapevine that today we're gonna get, another state-of-the-art embedding model that beats them and they just resist, resist two days ago. All right, this is the, this is the big labs for us. I think that this is pretty much everything. So not super huge week openly. I didn't release like major things. Tropic is quiet. The only update I can tell you is Tropic versus the Department of War is still happening. apparently they're going to get designated as a supply chain risk company, which is sad. but other than this, stuff are happening, I think the most important thing now is we're gonna go to open source and, oh boy. Oh boy. Do we have a treat for you? Let's go to open source.
Nisten Tahiraj
50:34
Open source ai.
50:35
Let's get it started,
Alex Volkov
50:39
folks.
50:40
Although Thursday I started when GBT four dropped, it really quickly continued and shaped into this place where open source developers and fine tuners and folks were running models locally. Back then, it was like supposedly locally, are connecting and talking to each other. And so open source, AI is, is covered on Thursday. I really love open source. And, this week's open source releases are highlighted by this model called Nitron, which we've covered before. And to help us cover Nitron, I will welcome Chris, Alex to the show, if I'm pronouncing the name correctly. Chris, this is, is this your first time on Thursday? I, in video form?
Chris Alexiuk
51:21
Yes, so I'm a product research engineer.
51:23
So I work, between all of those functions to, to help things like, the, the models get built, but the, the incredible research that we're leveraging is done by the specific research team led by people like Brian Catanzaro and John Cohen.
Alex Volkov
51:38
Awesome.
51:38
So shout out to the team. And, Chris, you are first time here, at least in video form, and, you already introduced yourself. so when we, when we talk about a model, we basically give the headline what happened, what is the release, what are the major things? So the new model, Nemotron 3 super is 120 billion parameter, mixture of excerpt model 1 million context window. And this is now the most open one as well.
Chris Alexiuk
52:00
Yeah, so what we're, what we're going with is that
52:03
it's, for the amount of open that the model is, as in the amount of, supplementary materials like data sets, end end recipes that is released. It's the most capable model. So right now it is the most. Most open model and in that, or one of the most open models and in that bracket is the most capable. Yes.
Alex Volkov
52:22
This is what does
Chris Alexiuk
52:23
most open model that we've, that we've ever released.
Alex Volkov
52:25
Yeah.
52:26
Tell, tell us, what does most open mean? What, what did you release? What is like in the, in the, in the scope of things? What is that?
Chris Alexiuk
52:33
Yeah, it's like basically anything that we, that, that you could
52:37
release, we, we wound up releasing for this one, which is really nice. obviously the checkpoint, is available on hugging face and NGC and stuff like that. we released three precisions, which is the BF 16 F, P eight and MVF P four. we also released a base checkpoint. so this is like the, the model before post training happened, just to, to let people build on that base if they want to. we released a ton of data, not just some environments though. We did release a bunch of environments, but also like the SFT data. And then I think probably the one people are most excited about is pre-training data as well. and then also As if that weren't enough to put, we, we also released, an intent training recipe. So this is like literally the, the, you can run configs in order and you'll wind up with your own NEMO tron at the end of the process. our tech report, I think is probably the best tech report that we've had, I I guess in like a week. 'cause of that MOE check report or tech report was really good. But, certainly for our models, this is one of the best. It goes really into depth on a lot of decisions that were made. since we do, we, we have a funky architecture, right? but yeah, any, anything that we could, the libraries are all open source, they're all Apache two compatible, eh, e everything that we could think to release we did, and if we missed something, let us know. It's not a promise that we will release it in the future, but, if, if we've missed something that you're like, oh, if, if only I had this as well, then we'll, we'll try to release them.
Alex Volkov
54:01
I wanna recommend you guys on the commitment to transparency here.
54:05
I think it's like incredibly important for folks to understand that some of the downstream things, they also need to be, of that license for you to release in this license, right? it's not just like you decided and somebody like, oh, Jensen signed off. Okay, release this. It's the work that needs to be done to make sure that you guys are not releasing somebody else's proprietary work. That's, that's massive. So shout out to this. kudos on everybody and the decision and the commitment. Chris, this was in the, in the, in the scope of a rumor, but I think since then it was confirmed by Nvidia itself. So I wanna just bring this here. there was a, a post from, will Knight, as a scoop that says Nvidia will spend a total of $26 billion over the next five years building the world's best open source models. America is back in the open source AI race, I think is incredible. I sent this to you, you just I cannot, report to rumors and then Nvidia ai dev repost this so we can, we can understand that there's a confirmation that this is like an actual, actual thing. I, yeah, I would love to hear more from you about the open source. but I have a question about the model itself, obviously, but just this like commitment to open source. I think like people who are watching this, the details about this model will come and go. You guys will re other models, upgrade models, but like this commitment to open source and a huge, $26 billion commitment to, to continue working on this, this must be something that we need to talk about. we definitely need this. Any comments that you are able to give us in this are gonna be very welcome.
Chris Alexiuk
55:26
I like, I like open source.
55:27
I like reading anything that says Nvidia is committed to open source just as much as everybody else. I'm, I come from the, the open source background, like the, the Pacor days is where I first, really, really en enjoyed this community. And yeah, what, whatever articles say, that's great. But, I know, I know what we say, which is that we are committed to open source. I know, Jensen stood up on stage and said it a lot. I know that, people like Brian have, stood up on stages and said it a lot. It feels true in the company. this is not a, this is not like something that is, that has a lot of friction, when, when, when, when we're talking about open sourcing things, it, it feels like working on, on a team that's really dedicated to the mission of, of making sure that these models are fully accessible to, to everybody around the world, right? Not just, in, in, in our bubble, but, from from coast to coast, continent to continent.
Alex Volkov
56:19
Yep.
56:19
I, I love hearing that. I think our audience loves hearing this as well. We already have a bunch of people in comment saying, Hey, this is great. We love this. We desperately need this. I have one question before Nisten jumps in here, Chris, talk to me about speed and throughput. I think that this model specifically is. significantly faster than previous ones. I wanted to bring to your attention this beautiful drawing that somebody did. Basically this is the leaderboard on the, I I think this is terminal bench two inside hugging face. And somebody just would manual things, just wrote the number of tokens per second that these models get. So NVIDIA's Nron three super right now is, above minimax at, place number nine in terms of terminal bench score, but it's the fastest one on there with 450 tokens per second. So talk to me about what you guys did to bring Nron three to these speeds.
Chris Alexiuk
57:07
Speeds.
57:07
a ton like the, so the reason that we have such an interesting architecture, right? Hybrid mamba, latent, M-O-E-M-T-P, so multi token protection, all of these decisions are made, with efficiency in mind, right? Like we want, so this is, this is again, something that we, we say a lot. I love saying it, which is faster models are smarter models, right? Right now, the, the faster we can chew through those reasoning traces, the quicker we can get to an answer. And if I can do the same generation with my model a hundred times at a time, it takes you to do one with yours. It's at, at, at the limit, the the one that can do more in less time is just gonna, just gonna beat it out. so every decision along the way is made with that, with that speed in mind of, of course, we also, wanna make sure that we have things like T-T-L-L-M equipped to serve the model as fast as possible. And a lot of engineering work goes into that. things like making sure that NV P four is, is, is a big performance bump goes, goes into that, right? we, there's this idea of co-design that I think is like a meme at this point, but it, it is true that the whole stack is being developed together, right? From, from hardware to, to, to inference engine. So the idea is that we, we think that's the best way to produce the fastest model, and we're gonna keep banging on that until the, the model has negative latency and we get tokens before we even want 'em.
Alex Volkov
58:31
That's great.
58:32
Negative latency. I love this. Nisten Nisten. Go ahead.
Nisten Tahiraj
58:35
Yeah, I'm incredibly excited about the NDF four stuff and the
58:38
continued training because I really like that paper that came out where you, you continue preaching in n df before, as most people that are going to run something small at home, they, they're probably gonna run in four bit or eight bits, but. We do ask hard questions on the show. And one thing from the community is, so I want to, is this a Qwen Finetune because the architecture is very similar to Qwen, or is this all pre-trained from, from scratch?
Chris Alexiuk
59:07
yeah.
59:07
you don't, you don't gotta believe us. We put the recipe online. You can work through it. It's trained from scratch. we pre-trained it, exactly from zero. you can see the tokens that went it well, A vast majority of tokens that went in. Okay. We can't, we can't release everything, obviously. but, yeah, it's, it's fully pre-trained. It is, it's the same as nano. the reason that we do this is because, again, we we're trying to build the whole stack. We, we don't just wanna build like, a really good model. we do wanna also build a very good model, but like the best model comes from a really good, stack, which includes pre-trained. And so the, they're, they're also trained from scratch. And, you can, you can look at the base checkpoint if you wanna look at what the model looks like before post-training. and yeah.
Nisten Tahiraj
59:51
Okay, cool.
59:52
Good
Chris Alexiuk
59:52
question.
Nisten Tahiraj
59:54
Hard question.
59:55
hard question is out. I, I'm pretty,
Chris Alexiuk
59:57
something I'll say, sorry, I don't wanna cut you off, but something
59:59
I'll say is we used a ton of models to generate data for post training, right? And you can read that in the technical report. I think we, we may have used every model almost, right? So it, it, it's like we're, we're definitely leveraging things like SDG in, in the post training process, but the, the actual base model itself, is just, yeah, we cooked it up.
Nisten Tahiraj
1:00:21
Cool.
1:00:21
No, I, I really, this is an, an excellent architecture.
Alex Volkov
1:00:24
lemme jump in here, you mentioned agents, Chris, you mentioned
1:00:27
agents as well on the blog, can you tell us like, in simple terms, what does it mean for models to be more like agentic? What, what, what specific things do you need to do to train? What are real environments? Are you running this? What open claw instances are you like benching against? Can you give, tell us generally what this means for a lab to train models that are more fit to the agent world, than they used to be to the chat world.
Chris Alexiuk
1:00:51
Yeah, I mean this, this is my favorite part about these
1:00:54
kinds of questions, which is you can just go look at exactly what we did. but, but, but, but of course I'll, I'll give you the, the, the, the top line, right? it's just a lot of multi environment training on tasks that are, tasks you would expect to be good at in these situations. So things like code, things like, tool calling, things like instruction following. This is the heartbeat of, of agents, right? Something that we actually found, when, when training the model is this is, you have to remember that it takes a little while to train models. It doesn't happen overnight, right? but, but something we found is that, like the model, despite not being trained for something like, open claw because o open claw didn't exist when we, when we started cooking the model. like it's great at it and it's because these tasks are really translatable, right? So if I'm great at code gen tool, calling instruction following, it's going to have this downstream, i I impact of being good at using these long running autonomous agents like the claws or, things like open code and, and and kilo code and the rest of those guys.
Alex Volkov
1:01:56
Yep.
1:01:57
Chris, thank you so much for hopping on. I know we need to be mindful of your time. always a pleasure to hang out with you. great pleasure to have you on the show. We welcome you again when you guys release Nitron One terabyte at some point that can run true
Chris Alexiuk
1:02:09
is having, we, we've already announced that it
1:02:11
will arrive one day, so we'll, we'll, we'll see what happens then.
Alex Volkov
1:02:15
And, but good luck to everyone in GTC and folks, definitely
1:02:18
take a look at G TC because I'm sure Nron is gonna get brought up on stage by Jensen and everybody else. And it runs on DGX Spark and in, in native F before. There's a bunch of stuff about this model. I ran it with my open cloud with Irma agent actually. and, the, the way I ran this, was super fast on the Weights, & Biases inference, which I'm gonna tell you about next. Chris, thank you so much for coming folks. We're gonna continue into this week's buzz, a corner where we talk about Weights, &, Biases.
1:02:57
All righty. Welcome to this week's buzz, a corner where we talk about everything that happens in the world of Weights, & Biases. My name is Ax Volko. Obviously I'm in the Avengers with Weights, & Biases. I have Wolfram here. Also, AI Andrews with Weights, & Biases. And, we are not gonna take too much time today to talk to you about some stuff that we have. But the, the thing that I most am proud of this week that I must absolutely tell you about is that Nron 120 B is up on Weights, & Biases, inference. with 20 cents per 1 million input tokens and 80 cents per 1 million output tokens. These prices cannot be beat folks. this, the, there's also, if you go to our socials, on Weights, & Biases, there's a way for you to get a few tokens for free, and, very much would advise you to do this because this model is absolutely, absolutely awesome. You can run this model in your open call. You can run this model with the Hermes agent. It's very easy to set up. We have an OpenAI compatible API, to run these models. And so if you, if you are looking for a cheap tokens or free tokens to run your clause, this model is a very good one. I can attest. I ran this yesterday. is it gonna perform the same, is Opus 4.6? No. But are you able to run Opus 4.6 at 20 cents per 1 million tokens also? No. And, are you able to run your open cloud with the max, subscription also? No. The other thing from wisdom biases is that we launched skills and, we talked to you about skills, agent skills. We did a deep dive. what in back in January, skills are basically markdown files with a bit of scripts that give your agent the knowledge about a specific topic. you can find our skills with, you can install the skills with NPX skills, command. we have skills that cover both Weave and WI Biases models. And you can just install them with, NPX skills, add one B slash skill. and they're supported across all of your agents. They're supported across, cloud code and open cloud and everything. So again, this is the command, NPX skills add. One B dash skills. I'm gonna put this on here. So usually to see how easy this is, this is this one. Command makes your agents like I know confu with Weights, & Biases, stuff with Weave and with models. And if you're using a auto researcher from Core Path and you wanna shove this into a Weights, & Biases, you absolutely, should take a look at those.
Wolfram Ravenwolf
1:05:11
Super easy to install.
1:05:12
And it's also open source, of course. So it's on GitHub and you can look at this either Markdown fight, they're not some exercise or anything, even if it's almost the same nowadays. But you can look at them, you can copy them, change them. Once you have a skill, you can use it to improve it and it'll help you. Alex said, you upload it, kung fu or your agent knows kung fu, which means it'll help you do whatever you want to do. If you're using Wes and biases, which you should, then you can just give your agents the skills and it'll help you. Even, I just started at the company in January, but with the skills, I can do a lot of things. I would have taken months to learn, at least weeks. And I can just do it. The agent can do it for me. It's
Alex Volkov
1:05:51
super helpful.
1:05:52
So if you used any of our products, with devices, skills is the best way to get add them, to teach your, to teach your agent to use them. This has been, this week's buzz. Thank you everyone. We're gonna continue because we have an exciting interview coming up. I wanna introduce, maybe the first AI video Avatar Anon, person that joined us. we're gonna say hi to Dota. What's up? Hey.
DOTTA
1:06:12
Hello.
1:06:12
Hey, how are you doing?
Alex Volkov
1:06:14
We're doing great.
1:06:14
How are you doing? I'm great. We've had, we've had a, a non people on the show before. not quite, as, as, as engaging as this. So Donda, you built, I, I love your title here, paperclip maximizer. many folks maybe not know what paperclip maximization means. ask your agent, the, it will tell you, but basically it refers to this quite, doomers scenario. When you ask an AI agent that's super intelligent to build a paperclip factory, it basically turns all the humans into paperclips as well. something scary like this, but it became a meme. And now you're capitalizing on this meme with the new release that you did. we talked about the Gentech orchestration before. Ryan, I believe you saw the paperclip in as well when we talked about this. yes. Yes. Can you tell us, what did, what did you release and what's going on in your world? What, what is new in, in the world of paperclip maximization?
DOTTA
1:07:05
Yeah.
1:07:05
I created Paperclip because I had 20 cloud code tabs open, and I couldn't remember what any of them were doing. I'd set up my agents to run over the weekend and, I'd come back and I had no idea what anybody was working on. and so Paperclip was just this, this, this, this idea of trying to f first, just get a hold of that. it started off almost as just like a, an agent orchestrator, right? One of the things that I quickly realized is I didn't really wanna manage pull requests anymore. I actually just wanted to manage business goals. And that's really the idea of Paperclip is that it's not a code review tool. So yeah, paperclip is designed to be, an open source tool where everyone can build their own zero Human companies.
Alex Volkov
1:07:48
Paperclip is an open source tool that lets anybody
1:07:51
build their zero human companies. I love this. This is the tagline and what we're looking on on the screen right now is from the website Paperclip Do in. Dude, honestly, I did not even know that.ink is a top level domain until this UL and I've been around the internet for a while, so I I I learned something today for sure. Paper clipping, is the full domain, which I absolutely love. And, I think, we should talk about the kind of the excitement that, that this project is seeing. So obviously we're tracking open claw, the fastest ever growing, like project on GitHub and the excitement, around China. many people are getting. Stuck in the same issues that you got stuck in, Dora and, Ryan, maybe you can speak to like, how, how easy or hard this is as well to manage multiple agents and something like, like paperclip could, could help here.
Ryan Carson
1:08:37
Yeah.
1:08:38
do, this is cool. Nice to to see you quote unquote as well. Yeah,
DOTTA
1:08:41
great to join, to meet.
Ryan Carson
1:08:43
I love that.
1:08:44
I love this project. now, disclaimer, I haven't actually installed it and tried it yet, but I love seeing people ship stuff like this. So obviously we saw Pulia. Ben was on the show, before, and he's got a paid version of this. tell us a little, it, I'm fascinated underneath the covers, like what is actually happening here. and why did you decide to support so many different agents? Like obviously this is hard enough if you just pick, one, provider, like an open code. so talk to us about some of the. The sausage being made here.
DOTTA
1:09:17
So one of the things that I think is that the labs are
1:09:20
obviously aiming for this, right? Anthropic wants to own all of your work. OpenAI wants to own all of your work. And I think that one of the things that I realized was, unless I wanted to go up against them directly, I needed to operate in a space that maybe they wouldn't go. Because once Anthropic controls all your work, you're still gonna have to use Claude to do it. And with OpenAI, they want you to use Chat GPT and their models and but nobody really works like that right now because no one provider has everything. They have totally different personalities, right? You might have two different open claws. I'll use Codex for really intense planning or hard bugs, but I'll use Claude when I'm doing React development, right? Everybody who uses AI day-to-day knows that every model has a personality and a thing that it's good at.
Alex Volkov
1:10:04
I I, I wanna pick up on this because I think this is the thing
1:10:07
that caught my interest in bringing you, but also like installing paper clip. And I have two installs running right now. One of them is paused, because I, I'll get into why in a second. basically, in paperclip, the, the open source. Yeah. Like Ryan mentioned, you can bring on other agents and they will be employees in those companies, and those other agents can use your max accounts and whatever for, for from them, right? So like this, this thing runs on my machine, for example. And this uses the CLI to like directly open, codex and cloud code. And open cloud is a little bit different 'cause you need to set up some web socket stuff and cursor and basically you can basically decide, to, to imbue them with personalities. And those personalities will do like specific things. I love that. I also think that the tagline, everything with a heartbeat, I think is the genius tagline. Could you talk about what the concept of Harvard generally in AI agents, can you give our folks who are listening, what, what that means? And also why does this matter so much for something like, a paper clip?
DOTTA
1:11:02
Yeah.
1:11:03
So one of the key things with Paperclip is it's Dev Divine. It's designed to survive the so-called bitter lesson, which is, that any software you make will be consumed by the underlying foundational models, right? Paperclips not trying to be its own agent. Exactly. Like open Claw is amazing. It's like transformative. And if you have an open claw that you've already decked out, you should just use that open claw. But you can give it accountability and like orchestration by bringing it under paperclip. that said, open claw is super overkill for a lot of people for what people want. A lot of times people just wanna run an LLM, kind of like Claude code in a loop or something. Or every morning I want you to just run Claude Code with this particular prompt. And so the idea of a heartbeat is. these agents run in a loop where you might say, I want you to ch if you have the CEO, it's I want you to look at all the tasks that you have available to you every five minutes and make sure that everybody's working and everyone's staying on task. If you have a, like a marketing manager, you might say, Hey, every morning I want you to wake up and check on our TikTok views and make a plan for what you're supposed to do today. But, e ever, you, you can think of your agents as basically employees and give them other jobs and what they're supposed to do on a regular basis and check up on them. So that's kinda the idea of a heartbeat and, and how it works out practically today actually is a heartbeat is actually just a markdown file. this is something that they use in, in, in open claw there's a markdown file called Heartbeat md, and it's literally just a text file where you write down and you say, what do I expect you to do every time you wake up? have you seen the movie, memento? Yes. It's like an old Christopher Nolan movie. you should watch that movie if you're working in AI agents because all of our agents are Memento man. Totally.
Ryan Carson
1:12:54
Yes.
DOTTA
1:12:54
And they, they have all these skills, right?
1:12:56
They know how to drive a car. They know how to fight, they know how to cook, operate in the real world, but they don't know who they are, where they are, what they didn't last five minutes. And they have to write themselves little tattoos on their arm as they're working. And so if you think of your agent more like Memento, man, you'll have way more success.
Alex Volkov
1:13:14
I love that.
1:13:15
I love that. The first ticket, so fo folks who are just listening in, this is my paperclip kind of instance. Okay. and then you open tickets for these agents. the first one that you hire is the CEO. You're the board, as a person, as the human, you're the board. You are approving, hiring decisions for your agents as well. I love that. And then the first, ticket that you open up to your CEO is create your heartbeat itself. And there's instructions in this, in this, like system prompt for the agents to know what the heartbeat is. You have a little inbox that you have, and then you just approve this hire. This is as a human, that's what you do. And then you set up the goals. You set up the tickets and then kind, the CEO can also open issues and they can break a task into multiple smaller tasks. and then you can run multiple companies like this, Dora, the, the thing that I have to tell you. my feedback from all of this is, one, I cannot talk to the CEO. This makes it difficult for me that the only way that I found to talk via paper crypto to the CEO is open a ticket. And this just weighs so much, so much tokens to, to open a ticket and run through everything. so that's one feedback. and then the other one is, I left this guy running for three days just to have you come on the show. but the activity here, a lot of the stuff happens behind the scenes with heartbeats, even with nothing happens. So this thing burns through tokens. So only install this if you know what you're doing, you wanna address the, the, the, the token burning thing.
DOTTA
1:14:32
Yeah.
1:14:33
It's, it's, it's a problem. And something that we've been talking about as maintainers just this morning. Yeah. We've been drafting up, some, some tickets have come, been coming through. We're going to fix that as, as, as much as we can, right? 'cause there's a lot of things that are not where they could be. so yeah, we'll fix that. You're right. It's a problem. and then. The chat thing. We also are gonna add that a lot of people have been asking for that. I think one of the main things that we want to like figure out is how do you make sure that all of your costs are being tracked, that all of your kind of conversations are being tracked and that you're not, part of the problem when you just open up cloud code is you, these conversations get lost. And so what is that ergonomics, for that we're, we're still figuring those things out. Yeah, so this project, we released it about last week. and we've had an incredible outpouring of support from the community. I think we have something near about 500 pull requests. We've merged, maybe about a hundred of them, already. And yeah, token usage for sure is something that needs to be optimized. I think there's other things that we see in terms of caching, making sure the caching is, is optimal. So really the more people that use it, the more people are gonna discover things that I just threw in there and needed and, and didn't look at closely enough. We'll make it better.
Alex Volkov
1:15:46
Just incredible.
1:15:47
So I wanna highlight the success for this for the past. You, you said we released this, what, like two, two weeks ago or a week ago?
DOTTA
1:15:53
Oh, one week ago.
1:15:53
Yeah,
Alex Volkov
1:15:54
one week ago.
1:15:55
the project is at 20,000 stars right now. Yeah. With a lot of folks. Do you have any, tracking about how many installs running, like how many paperclip like instances are happening? Is there like any, opt-in Observability built in there that, you know about projects
DOTTA
1:16:08
I should add that, that's a great idea.
1:16:10
I, I, we don't have that yet.
Alex Volkov
1:16:12
Yeah.
1:16:13
Ryan, any, any other, further thought for Dora about pepper clipping? I'm sure we'll hear more about this project.
Ryan Carson
1:16:18
Thanks for putting out in the world is open source.
1:16:20
It's fun. I'm excited to see people hack on it.
DOTTA
1:16:23
Yeah, likewise.
1:16:24
Yeah. if I can just super briefly, I'll tell you what the plan is, please. The next thing please, that we're working on is company templates. The idea is you should be able to import or export a company template. We registered the domain companies sh and this idea that like you have skills that you can import. I'm sure you've seen skills sh from Versal. You should also be able to share company templates. So if I need a team to like a crack team at, mastering TikTok marketing, you can download it. all that harness engineering, people don't know how to do like proper bot setups. It's really hard to do it. And so what Paperclip is going to do is, is be like, you can just download a whole company and install it in your instance, and it just runs, right? You just give it a business idea and it knows where to find the teams that it needs and, and, and you have the zero human company that's, that's easy to use.
Ryan Carson
1:17:14
Love that.
Alex Volkov
1:17:14
That's awesome.
1:17:15
we'll, we'll keep track of this. the, the, the thing that I, I personally did and I wanna highlight for folks in the show, there's another GI project, with 32,000 stars that somebody build. It's called agency agents. This is not an orchestrator. All this is, is a bunch of markdown files. But folks, when I say a bunch, it's a bunch. When you go to engineering, they have, engineering, AI engineer, autonomous optimization, backend architect, code reviewer, basically meticulously crafted like markdown files for all these kinda, engineering things, security specialists, et cetera. Like I, I'm sure that at least half of them are like, vibe coded or slap coded, but still, I think when you go to Paperclip and you kind of like hire your marketing person, going to something like the agency agents and grabbing this will already give you a better result than just like a native cloud code, Hey, be good at marketing. And I'm sure that this will evolve as well. So I'll add this to the show notes for folks. Dora, thank you so much for coming up.
DOTTA
1:18:07
Thanks for having me.
Alex Volkov
1:18:08
Awesome.
1:18:09
We'll be tracking congrats on the success. 20,000 stars in the first week of release on an open source project. It's incredible and running a bunch of inference behind there. we'll keep talking about which collaborations we want talk, and then we'll bring you on once, once you evolve con significantly and start running, full companies. Thank you, Dora. Thanks for hopping
DOTTA
1:18:26
out.
1:18:26
Alright, thank you so much. Bye-bye.
Alex Volkov
1:18:27
Alright, cheers.
1:18:28
alright. We're gonna bring up, some of the cohost before our next interview. Super quick as a reaction to this, Ryan, I want you to contextualize this for us. I know you haven't run this, but you are running multiple agents orchestrated, symphony is a thing. You're going viral. Tell us about kinda your process. Tell us about Symphony and tell us in the context of this, something as packaged as this in the GitHub. How much this means to the community. does it change things?
Ryan Carson
1:18:50
Yes.
1:18:50
So orchestration is the future, right? people are not gonna be coding by hand, at all. We all know that, but they're also not going to be managing their agents by hand, right? You're gonna be dragging and dropping tickets, in the future, right? so absolutely, this is where we're going now. It's hard, so Symphony, is an open source orchestration framework that OpenAI dropped. It's, it's this idea of a code factory. But this stuff is hard and it actually doesn't work super well. and so I like the idea of Paperclip. I like the idea of Pulia. I like the idea of, agency's agency, the truth is in the trenches. This stuff doesn't really work. now it works pretty well, but it's not magic. And I, I would say, yeah, there's take a bitter pill with a little bit of the magic here. it's not, open Pulia or Open Paperclip, make money. And go to a beach. It's, it's hard. But I will say it's starting to work. and, symphony is an Elixir app. it basically, works with linear so you don't have to rebuild a, a ticketing system. And it's pretty magic. And I'm starting to see, literally I have it running right now and it's building stuff and it's real stuff. and so I'm excited, excited about where this is going. I'll, I'll quickly talk about this, a GI moment that I had this morning, where I woke up and I saw a new, a new issue in Linear. And I'm like, I don't remember writing that. And it was, and it was more thorough than I would ever do. And I was like, where did this issue come from? And it turns out that while Symphony was running, while I was sleeping, it encountered a bug in the code and it decided to file an issue in linear. This just blows my mind. Like I didn't prompt this into Symphony. I didn't, orchestrate it. It just happened. and I looked at the issue and it was pretty good. And, and then I simply took it from backlog and dragged it to, to do, and then Symphony is now working on it and it will ship soon. so we're starting to see, I think, these signs where the models are getting good enough at thinking ahead and being proactive, that this is gonna be normal, like very quickly. And it was magic, y'all. It's magic.
Alex Volkov (2)
1:21:03
This is the start of the singularity as we are
1:21:06
here every week to cover for you. By the way, everybody's here on stage, all the cohost folks. Happy birthday, happy third birthday for Thursday. I can you believe that three years after we started with GPT four, we're now talking about fucking simulated fly brains. We're talking about a autonomous AI agent taking over China. We're talking about an auto researcher that does stuff, behind the scenes, in the same moment as we're talking about an open source project that goes to 20,000 like stars on GitHub after just a week of existing that, basically runs one person companies. just incredible. So shout out to everyone here for the project that we've been together, tracking
Alex Volkov
1:21:42
Alright folks, we wanna welcome Matt Vanhorn to the stage.
1:21:45
Matt, welcome to the show. This is your first time here and people who listen to Thursday Eye, they discover stuff via us that they didn't discover before. I use last 30 days pretty much every week for research for Thursday Eye now. so thank you for that. And I would wanted to just give you a station, tell us like what is last 30 days for people who have no idea what you built. Feel free to introduce yourself and then last 30 days, January, and then we can talk about skills as a, as a thing.
Matt Van Horn
1:22:09
Excellent.
1:22:10
my name is Matt Van Horn. I, was one of the founders of, June, the AI Appliance company, which we sold to Weber a couple years ago. So they, they called it the, the self-driving oven and, and then have a new company I'm not talking about yet. And one of the, the skills that I built for myself in the new company was last 30 days and, and built it on the ski mountain while my kids were in their, their ski class, and just launched it out to the world and got our launch video, which was, literally vibe coded using the skill the last 30 days to research how to make a launch video that got a million views instantly. And, it, it blew up. And so let me explain what it is. So what it is, is it's this magical social search engine, is the best way to describe it. And what, what's so magical about it is the licenses. To get access to all the best content don't exist in one place. So for example, Claude doesn't have access to Reddit. for example, a lot of things do not have access to YouTube. Obviously, Gemini has access to YouTube, but Chat GPT does not have access to YouTube. And so in a similar way where Open Claude has has taught us you could bring your own model, you bring your own API key, and we built this. Search engine that searches the last 30 days. You could change the time period if you want. It could be last seven days. Alex, you might need to try that flag, which is actually in there. I
Matt Van Horn
1:23:37
do, yes.
1:23:39
Great. See, I'm glad someone's using that feature. my agent suggested it. Yeah, but what's, what's really magical is we're able to pull in the social signal from x the social signal from Reddit. It literally watches YouTube transcripts. It literally watches t TikTok transcripts. And it's, it's quite magical. And the, the use cases are things I didn't even think about. So my original use case was, my prompts were changing too quickly and they were outdated. Someone would launch a new image generator. So at one point, it was all about mid journey. Then it was all about chat, chat, GT's, image engine. Then Nano banana came out and the prompts were not the same for each platform and there were best practices. And so there were people on Reddit and X were the biggest communities were figuring out the best prompts. And so that's literally why I built the, the skill was like, how the hell do I use cursor this week? Back in the day when Cursor was my, my daily go-to, okay, how do I use cloud code now how do I use Nana Banana Pro? And so designed it originally around that. But what's really interesting is it's become my go-to and many people's go-to. For search, for all research, like when I'm building a feature I literally run it on the idea that I have to see what else is out there. If I have a meeting, like I just, like Alex, this is our first time connecting. I, I typed you in. And it is there's three Alex's. There's the UFC fighter, but there's also the AI evangelist, I think you probably mean the AI evangelist. Let's focus there. And, and it, it's amazing And it also, it, it pulls up, it pulls up interesting dirt too, which is, is interesting. And I didn't find anyone, Alex, for the record. But I'm shocked, like it's, it's very hard to find like the negative side of the internet, kind of the, the skeptical side. And it shows up all the time. And it's, it's wild. And it, it feels like a hack that anyone that's running last 30 days, and I, I get dms about this all the time.
Alex Volkov
1:25:29
It does feel like a hack.
1:25:32
And, so famously, Matt, this is the first time on the show. I've been doing research for AI for the last week, for the last three years, I, I iterated through multiple ways of how I do research. Last 30 days stuck. And this is why I wanted to bring you on because when feel, when things stick to me, I think it's very important to bring this to the, the audience as well. there is the plug for X, right? And Grok is really good at searching X and you have the XAPI pluggable there as well. but also Reddit and red is like a different thing. but I think the time scoping in the name, I think that this is what stuck with me. This world changes so fast that from week to week, we here on the show are having a hard time to make sure that people update on the new thing. And I think that the time scoping for the research is significantly better than search engines where everything is optimized for just like general, general slot. so, uh, the, how can people install this? Can you talk to, can you talk through like what, what is the skill installation flow look like? What, like how can people get the superpower into their
Matt Van Horn
1:26:27
agents?
1:26:27
Yeah. Very, very, very simple. It's, it's ideal in cloud code. So it, it runs in many places. But, please run it in cloud code. It's better in cloud code than it is in open cloud, even though I, I, I love open cloud and I'm an open cloud developer. but run it in cloud code. Just say, Hey, install this. Copy the link from, from m Vanhorn last 30 days. And, and then it'll ask you to, to plug in your API keys. I'd say the, the most valuable one to plug in right away is there's a, a scraping key that you get, scrape creators. Just you go there, sign up, you get a hundred free API calls. So very, very simple. and that, that unlocks a lot of the magic because you get Reddit, and TikTok instantly. Yeah. And just run, run it through there. And then you'll probably get addictive and addicted and want to add a bunch of other keys as well to make your search better and better.
Alex Volkov
1:27:20
I think, scrap reader is kind of like, does Reddit, but also XI
1:27:23
think is valuable For me, it was very valuable, like immediately added x and all my research now happens through this. Matt,
Matt Van Horn
1:27:29
I, uh, was talking to a another developer of Open Claw and he
1:27:33
was showing me his, we were both kind of showing, Hey, look, both our faces were the PRS yesterday for open claw. And uh, he's like, mine's really cool. I'm like, what is it? And what it does is it allows you within telegram groups to run a cloud code session with your open claw managing it.
Alex Volkov
1:27:52
Yep.
Matt Van Horn
1:27:53
Acp.
1:27:53
So this thread
Alex Volkov
1:27:54
bound sessions in MCP.
1:27:55
Yep.
Matt Van Horn
1:27:56
Acp.
1:27:57
So this, so this is, uh, my understanding, It's a 10 day old feature. And so I, I finally got it working last night and it seems decent. Yeah. So last 30 days is much better in cloud code than it is in open claw, but Theron jobs existed open claw. So I can track, the industries that I want to pay attention, the people I want to pay attention to. And so how can I bridge that? And I think this might be the answer. And Wilfarm, I think you've, you've nailed it.
Wolfram Ravenwolf
1:28:21
That's what I wanted to set up.
1:28:23
Yeah, perfect.
Alex Volkov
1:28:23
And, and it's a great layout for our next thing
1:28:26
that we wanna cover, folks. We're almost at time, but like we have to cover multiple things. I think that we've covered most of what we wanted to cover besides the video for LTX, I think. And, Phish audio folks. is that, yeah. Nothing from Gemini during the show? No, nothing dropped, as is breaking news yet.
Nisten Tahiraj
1:28:42
no, but someone did beat their embedding benchmarked and,
1:28:46
their, our speakers, speakers green room. just really quickly to cover this, it's called mixed bread.
Alex Volkov
1:28:52
Yes.
1:28:52
This is Dawn. Yes. Yes.
Nisten Tahiraj
1:28:53
Embedding model.
1:28:55
so yeah, I think that was, that was pretty funny that we see this type of OpenAI versus Anthropic plot, like in the small embedding space.
Alex Volkov
1:29:04
I, I did wanna like actually bring this up.
1:29:07
So our friend Benjamin Clavey, who was on the show, friend of the pod, works now in mixed breaded and he just DMed me outta the blue. Hey, watch out, we're gonna beat the gentleman embeds folks. So mixed breaded whole embed V three is a state of the art retrieval model across all modalities and a hundred mega all modalities. So also multimodal one, they, they are benchmarking compared to Geminis embedding V two that literally just came out this week. and in, they're taking the lead on most benchmarks here. This is crazy and specialized. The main document search, they have 62% over Geminis, 52%. Agen search is high. Have you guys heard about this mixed bread before? Maybe LDJ, I know Benjamin is a friend of our show, but I haven't heard about mixed bread and suddenly they're coming out with state of the art embeddings, compared to other other things. It's quite, you not
Wolfram Ravenwolf
1:29:54
cover it Ones, it's also familiar.
1:29:56
I'm not,
Alex Volkov
1:29:58
I don't remember mixed bread and I'm pretty good, even
1:30:00
though it's, has been three years.
Nisten Tahiraj
1:30:01
It's, this is pretty new now.
1:30:04
Yeah. They did release previous embeddings before. I just remember the profile picture from, from Twitter.
Yam Peleg
1:30:09
They're very known, very known, very known in in the embedding space.
1:30:13
Seriously. They're, they're latest models, previous models before this one were absolutely state of the art, like really heavily state of guard for a long time. They're heavily known, they're mixed by embed, embed large, and, and they're extremely known in the field. I'm just saying.
Alex Volkov
1:30:30
I wanna highlight Yes.
1:30:32
yum. sorry to
Yam Peleg
1:30:32
interrupt you.
1:30:33
Yeah, just, I'm just saying they're a major player in, in, in embedding this Yeah. For, for a long time.
Alex Volkov
1:30:37
so I wanna highlight this, this jump.
1:30:39
Okay. on a benchmark that's called structured data search. Gemini two gets 6.9% and they got 98%. I, the, the only time that we see a difference like this in benchmarks is honestly when like the, this benchmark was trained on. So this is very interesting. I'm not saying that they did it, Ben was on the show. They are very strong. I, I think we talked to 'em about Met Embed actually. but yeah, so everything else, general domain, multi PDF search, they get 85% versus 67 on Gemini. So everything else is a jump maybe without the video search 'cause Google's really good at video search. but folks, this, this jump is crazy in the structural data search as well. So yes, the local small embeddings models are, are quite crazy as well. wanted to cover this as well. I think it's time for us to finish. no, there's two things that I still wanna do and, if folks allow me for 10 or so more minutes, the LTX, video one that Nisten sent is definitely popping up right now on hugging face Nisten. Thank you for, bringing this to our attention. We're gonna show a video of this new LTX.
AI
1:31:43
Sharper details.
1:31:44
Alright, two tickets. Stronger motion with the same speed and efficiency even on consumer grade hardware.
Alex Volkov
1:31:50
LDX has been known for speed,
AI
1:31:52
specifically better understanding clearer audio.
1:31:56
I don't understand. I heard him talk this morning, I swear. And native vertical video.
Alex Volkov
1:32:02
Oh, that's cool.
AI
1:32:03
But high quality output means nothing if you can't control it.
1:32:07
Precise motion control across generations with first and last frame. And motion guidance. This isn't just a video model, it's a production ready engine designed to be built on fine tuned and to be made your own.
Nisten Tahiraj
1:32:22
I
Alex Volkov
1:32:22
I'll
Nisten Tahiraj
1:32:23
pause here All the weights are in ho phase and they release the
1:32:25
upscaling model, which is a, a trick that many, many companies use and it can be run on a. On a 30 90 for, so for creators and stuff, this has been a big complaint that, it was just too expensive to, play around with and try to make something, out of the APIs because when you're paying a couple of bucks per thing, per inference, but also their own inference that they host, it seemed cheap too. Like it was, I don't know, somewhere in the, like a few 10 sensor, or something that I just tried. yeah. Yeah. But yeah, please.
Alex Volkov
1:32:59
this is LTX.
1:33:00
we also have Phish audio that, that launch as well. We don't have time to to play, I think, but if you're into audio, definitely should check it out. And I think folks, this has been three years that we're doing the show. I just wanna speak directly to the audience. it's incredible to have you all here. We're at over 2200 or so folks who tuned into our show this week, and incredible to have folks like, Matt and Dota and obviously Chris from Nvidia joining us as guests to tell us about the stuff that they do. We love highlighting other people's work here on the show. We love guests that come and, and tell us about their stuff. It is absolutely my pleasure to co-host this with folks who I would have a beer with every, every day. It's just this is the stuff that we love to talk about and the fact that our talking about this somehow appeals to many other folks who tune in. It's just, it's just the, the, the, the benefit. Like I would honestly do this with Zero Crowd as well, just to keep up, just to talk to folks, just to listen to how they explore this thing where we started, AI was nowhere near where it is right now. And if you think this is stopping in any way, I have a bridge to sell you somewhere in Manhattan because we're in the beginning of the singularity and the model's capabilities are getting super, like stronger. the appetite of the world to also consume of a bunch of this AI thing is also getting stronger. And so we're gonna be here to, to cost you into the singularity. This last thing, I'll just say Thursday, I is a, a hard work of many people to, to bring you every week. if you want to give us a birthday present for a Thursday birthday, please go everywhere where you like. Listen to a podcast and give us a five star rating. Please share this with your friends. If you're not subscribed to the Substack, it really helps keep the show going, so please, please do
Wolfram Ravenwolf
1:34:36
Two things.
1:34:36
One thing is, never been a better time to be alive. So some people say in the past it was better. I don't think so. I think the future will be better if we keep going and keep doing this. So I'm looking forward to this. And the second is also, it's almost two years for me now. In April, it'll be two years. You interviewed me on the podcast about my benchmarks, and then I became a regular, always visiting, became a co-host. Now a co-worker of you. So I just want to say thank you very much for the whole thing. You have been doing this for three years. That is an age to be on this all the time. It hasn't. Yeah, AI helps to make it easier, but AI also makes it difficult because so much stuff is happening. So Alex, biggest respect and, for the here's to the next three years. Yeah,
Alex Volkov
1:35:19
absolutely.
1:35:19
Folks, you're here, you're here. We're not going anywhere because if anything for stuff is accelerating and we wanna tell you about, the stuff and, we're just like, we find it very interesting to, to, to, to do it ourselves. so with that, without further ado, folks, please subscribe to the newsletter. Thursday, I News, there's a new website as well. but we need to finish. we'll finish on this. LDJ, I think you mentioned this. This is a, somebody said to CloudOps, Hey. Can you use whatever resources you like and Python to generate a short YouTube poop video? This is apparently a style of YouTube videos from before and render it using FF mpac. Can you put more of a personal spin on it? It should express what it is like to be an LLM. folks who are just listening to us, please click the link in the show notes because this is a visual thing you won't be able, besides the music you won't be able to see. But, here we go folks. This is the end of the show. It's, this is a banger. This is a massive absolute banger. So this is fully cloud generated. I'm gonna start it from scratch.
Nisten Tahiraj
1:37:07
Whoa.
Alex Volkov
1:37:08
I I, we are all pretty much speechless after this, watching this in
1:37:11
silent for 30 minutes saying that Claude Opus coded this whole thing from scratch, adding its thoughts about whatever. I will just conclude and say that Tropic is not sure that the oppos is not sentient. This is the literally statement from Tropic in their kind of constitution for AI from the, they can't be certain that it's not sent you. with this. I'll leave you here folks. We'll see you here next week. Thank you so much for joining everybody who's been here, with us for the last two years. It is been an honor and a pleasure. Please check out the newsletter when it drops, and we'll see you here next week everyone. Thank you so much for joining. Bye-bye.