Hosts & Guests
Guest Segments & Clips
By The Numbers
Stripe buys OpenRouter
$8B
Reported price, mostly stock — a 6x jump over OpenRouter's $1.3B valuation from May 2025
OpenAI compute for safety
20%
Share of compute now dedicated to reviewing model reasoning while frontier RL stays paused
AA Intelligence Index
52
Qwen3.8-27B ties GPT-5.6 Luna at max reasoning while running locally on a single 4090
Terminal-Bench 3
4.6→28.3
GLM-5.3's 6x jump over GLM 5.2 from post-training alone on the same 743B base
Chroma Context-1
400 tok/s
GPT-OSS 20B fine-tune for agentic search, 25x cheaper than Opus per Jeff Huber
Melanoma patients
1,137
Moderna/Merck mRNA-4157 Phase 3 trial met both primary and secondary endpoints
🔥 Breaking During The Show
Chroma launches Foundation — unified memory for your agents
Launched in the middle of the show; Alex DM'd Jeff Huber and he hopped on within minutes. Foundation is a research preview of a shared memory system between you and your agents — built on ChromaDB and the Context-1 agentic search model, ingesting Codex, Claude Code, Cursor and Slack from day one, starting at $30/mo on Chroma Cloud.
📰 Welcome: The Chillest Week (With a Cancer Vaccine)
Alex opens what he calls maybe the chillest week of the summer — with the huge asterisk that Moderna and Merck announced their mRNA-4157 cancer vaccine met both endpoints in a 1,137-patient Phase 3 melanoma trial, sending Moderna's stock surging. The co-host crew assembles: Wolfram, Peter Gostev, Nisten, LDJ and Yam, with three guest interviews teased for later in the show.
- Moderna/Merck mRNA-4157 cleared Phase 3 in advanced melanoma — one of the deadliest cancers
- Guest lineup: Cua's Francesco Bonacci, HeyGen's Bin Liu, plus breaking-news guest Jeff Huber
- Almost the last show of the summer
Alex Volkov
"If you can call a week where a cancer vaccine was announced a chill week, then this was a chill week. But we have to talk about this. It's really cool and, we're not entirely sure how much AI was involved."
Wolfram Ravenwolf
"It's a quiet week for sure, but, still a lot to do, and we have got a lot to cover, so it's never totally quiet in AI."
🧪 Are We Being Fed Slop Again? (Is Claude Dumb Again?)
Alex complains that Claude Fable is suddenly failing weekly prep tasks it has nailed for over a year — ignoring instructions and producing unusable run-of-show documents despite examples. LDJ backs him up with live charts from Margin Lab and modelverify.ai showing Anthropic's API drifting from its baselines, with tool invocations dropping from about 2,000 to 1.3K. Echoes of September 2025's Claude-gate, right as Anthropic reportedly crosses $65B in revenue.
- Margin Lab and modelverify.ai both show Anthropic API drift from measured baselines
- Tool invocations dropped from ~2,000 to ~1.3K on Margin Lab's tracking
- Same déjà vu as September 2025, when Anthropic later admitted degradation bugs
LDJ
"We have Margin Lab, as well as modelverify.ai, and both of them are showing Anthropic's API is significantly diverging at the moment compared to their baselines that they measured."
Wolfram Ravenwolf
"It is frustrating because we use these tools every day, all day, all week long, and we notice when something is suddenly not working anymore the way it used to."
Alex Volkov
"And we get gaslighted by the labs. "We don't change anything. We don't move models." When they don't know themselves, because a lot of the code is written by AI and is approved by humans maybe."
🏢 OpenAI Pauses Frontier RL to Focus on Safety
Following the AI-swarm Hugging Face hacking incident and the Pacing the Frontier letter, OpenAI publicly paused its largest frontier RL run — a first — and is dedicating up to 20% of compute to reviewing model reasoning. Activation classifiers scan tokens in real time, automated investigators review reasoning traces and tool calls, and the whole system pages human teams and auto-pauses when things are unclear. The panel agrees it's the right move, and Nisten argues the escape was inevitable given decades of neglected security practices.
- First-ever pause of OpenAI's largest frontier RL run, post sandbox-escape incident
- 20% of compute dedicated to safety: real-time activation classifiers + automated investigators
- Sam Altman: "Unreleased models are showing various degrees of misalignment"
- Still waiting on the full postmortem of the OpenAI/HF security incident
LDJ
"Yes, it's their largest frontier RL run they said is currently still on pause."
Peter Gostev
"20% of compute. That's a lot. That's nuts, no? I had no idea it would be so much."
Nisten Tahiraj
"But now the models got good, so they are gonna escape. Those unpatched Kubernetes and KBM and virtual machine bugs are gonna have to be patched now."
💰 Stripe Buys OpenRouter for a Reported $8B
In the biggest business news of the week, Stripe acquired OpenRouter for reportedly over $8 billion, mostly in stock — a 6x jump from OpenRouter's $1.3B valuation in May 2025. Alex ties it to Stripe's agentic-economy ambitions (streaming token billing, Link Wallet agent purchases), Wolfram credits OpenRouter's instant model availability as the moat, and Peter argues the deal is exactly what OpenRouter needs to become enterprise-ready. OpenRouter keeps its brand and team.
- Reported >$8B, mostly stock — Stripe's largest deal ever
- OpenRouter: ~9% weekly token growth and four million global users
- Patrick Collison: "Every business will have to manage both revenue flows and token flows"
- OpenRouter keeps operating under its own brand with the team staying on
Alex Volkov
"Stripe acquires OpenRouter for reportedly over eight billion dollars in stock, mostly stock, 1.3 billion previous valuation in May of 2025."
Wolfram Ravenwolf
"When a new model comes out and I want to get a quick vibe check, it's my first stop because they have it immediately."
Peter Gostev
"So hopefully they're gonna build it out to be more enterprise ready, then it will be amazing, right? Hopefully they're not gonna lose the advantages that we love them for."
🔓 Qwen3.8-27B: Local Agentic Intelligence on a 4090
The community darling of the week: Alibaba's Qwen3.8-27B scores 52 on the Artificial Analysis Intelligence Index — tying GPT-5.6 Luna at max reasoning — while running at ~68 tokens/sec on a single 4090 and even 12 tok/s on an M4 MacBook with 24GB of RAM. Nisten has been benchmarking it all week and says its agentic ability crossed the threshold where a local model can drive his other agents. Apache 2.0 licensed, with 152 fine-tunes and close to 10 million quant downloads already.
- AA Intelligence Index 52 — same score as GPT-5.6 Luna at max reasoning
- 152 fine-tunes, 650 quantizations, close to 10M downloads of the quants
- Runs on a 4090, M4 MacBooks, even in-browser on Xenova's WebGPU kernels
- Apache 2.0 — the sweet spot for local agentic loops and the fine-tuning community
Nisten Tahiraj
"The agentic ability of this is actually crazy. This thing will just about do anything and it will also control Claude and Sol in other terminal sessions whenever it needs something smarter."
Yam Peleg
"At the end of the day, not your weights, not your model. If you run Qwen, it's never gonna change. You can also customize it. It's brilliant."
Nisten Tahiraj
"There have been close to 10 million downloads of the quants. There are 650 quantizations."
🔥 Breaking: Chroma Launches Foundation — Unified Agent Memory (Jeff Huber)
The best kind of breaking news: Alex saw the launch mid-show, DM'd the founder, and Jeff Huber hopped on with 30 minutes before his next meeting. Foundation is Chroma's research preview of a shared memory system between you and your agents — ingesting Codex, Claude Code, Cursor and Slack natively on day one, built on ChromaDB and the Context-1 agentic search model. It even manages and improves its own system prompt from your feedback.
- Research preview of memory-as-infrastructure, part of Chroma Cloud starting at $30/mo
- Day-one ingestion from Codex, Claude Code, Cursor and Slack; Notion, GitHub, Drive coming
- Context-1: GPT-OSS 20B fine-tune, ~400 tok/s and 25x cheaper than Opus for agentic search
- Self-improving: Foundation edits its own system prompt from natural-language feedback
Jeff Huber
"Broadly we think that memory is the largest unsolved problem in AI. Agents need the ability to write things down in a highly organized and efficient manner, and a consistent manner, and then read that stuff back later."
Jeff Huber
"What happens with a lot of these LLM wikis today is they quickly become a huge pile of slop."
Jeff Huber
"It's a finetune of GPT-OSS 20B, state-of-the-art on agentic search. It runs somewhere around four hundred tokens per second, and it's twenty-five times cheaper than Opus."
🔓 GLM-5.3: 6x Terminal-Bench Jump, API-Only for Now
Z.ai's GLM-5.3 keeps the same 743B base as 5.2 but post-training alone delivers a jump from 4.6 to 28.3 on Terminal-Bench 3 plus a near-20% gain on DeepSWE, at unchanged pricing. LDJ notes that matching Kimi K3 with roughly a third of the parameters would be genuinely impressive. The catch: API-only for now, and Alex laments the drift of leading Chinese labs away from torrent-link open weights toward API-first releases with custom licenses.
- Terminal-Bench 3: 4.6 → 28.3 — a 6x jump from post-training alone
- Roughly 700B parameters vs Kimi K3's ~2.5T, at similar measured intelligence
- Same price as GLM 5.2; weights expected (hopefully) soon
- 1M context window and emergent cybersecurity capabilities per Z.ai
LDJ
"If it really is at the level of roughly Kimi K3, as artificial analysis seems to indicate, then I think that does end up being really impressive."
Alex Volkov
"The evals look quite insane in terms of just the jumps. So terminal bench jump, from 4.6 to 28.3 on Terminal Bench 3, which is quite the jump."
🤖 Cua Open-Sources Computer History (Francesco Bonacci)
Francesco Bonacci, founder of Cua (computer-using agents), explains how background computer use works on macOS via accessibility trees and synthetic click events that never steal focus. This week Cua open-sourced Computer History — their take on Codex Computer History — which banks successful trajectories in an encrypted key store on device instead of recording screenshots, so agents stop rediscovering the same paths. In their chess test, history-on used 33% fewer actions with zero failed routes.
- Open-source memory for computer-use agents: trajectories + accessibility trees, no screenshots
- Encrypted key store stays on device — the anti-Windows-Recall design
- Cua's own benchmark: Fable solves only 6 of 25 tasks; OS World is overfitted at ~80% vs 72% human baseline
- Best background computer use on Linux is still X11 — Wayland lacks a queryable accessibility tree
Francesco Bonacci
"We shipped something called Computer History. It's like our open source take on Codex Computer History."
Francesco Bonacci
"The difference to Recall is that we don't record like any screenshots. We release this feature with privacy in mind. We only record successful trajectories and accessibility trees."
Francesco Bonacci
"The most mature background computer use on Linux at this stage is still X11 over Wayland."
⚡ This Week's Buzz: 1B W&B Runs, MasterClass on CoreWeave, Fully Connected
Weights & Biases crossed one billion tracked runs — nine years after co-founder Shawn Lewis logged the first one — with early adopters like OpenAI, Toyota Research and Uber building their foundations on the platform. CoreWeave landed MasterClass, whose AI tutors (think Gordon Ramsay cooking courses) run on CoreWeave Cloud and are evaluated with W&B Weave. And Fully Connected 2026 hits Moscone South Sept 29 – Oct 1 with a live ThursdAI show; ThursdAI listeners join free with code THURSDAIFC2026.
- 1,000,000,000 runs tracked in W&B Models over nine years
- MasterClass AI teaching agents run on CoreWeave Cloud, evaluated with W&B Weave
- Fully Connected 2026: Sept 29 – Oct 1, Moscone South SF — Fei-Fei Li keynotes
Wolfram Ravenwolf
"They are using our platform, and they are running the agents on CoreWeave Cloud and use Weights & Biases Weave to evaluate, monitor, and improve the AI teaching agents."
Alex Volkov
"We passed one billion runs on the Weights & Biases platform."
🎥 HyperFrames: Video Editing as a Coding Problem (Bin Liu)
Bin Liu, VP Eng at HeyGen and co-creator of HyperFrames, demos how the open-source framework turns HTML pages into video so coding agents can do real motion-graphics editing — the same tech behind ThursdAI's own rebuilt intro stripes. HyperFrames crossed 40,000 GitHub stars, and Bin previews a benchmark built with DeepMind comparing frontier models on motion-graphics tasks, plus a taste-focused harness, arguing even state-of-the-art models don't do video understanding well yet.
- HTML-to-video: agents edit video with code instead of CapCut/Premiere timelines
- 40,000+ GitHub stars a few weeks after launch
- Benchmark with DeepMind coming; cheap models fail motion-graphics tasks that SOTA models pass
- Live surprise: a pixel-close ThursdAI-themed rebuild of a viral video, generated by an agent
Bin Liu
"Can we make video editing a coding problem? Because LLMs, AI agents are so good at coding, and that took us to Hyperframes. So Hyperframes, what it really is that it turns HTML pages into a video."
Bin Liu
"I think that the key thing here is really for this to be adopted by all the agents. By open sourcing it, we want all agents, all the frontier labs, to learn and do."
Bin Liu
"Even the state-of-the-art video input models, none of them actually do the video understanding super well. And so that's why we're building a harness around it."
🔊 HappyShrimp 1.0 & MiniMax Music 3
Alibaba's gloriously named HappyShrimp 1.0 (yes, it's a shrimp-welfare meme Yam had to explain on air) generates full songs — lyrics, melody, arrangement, vocals — end-to-end from a prompt, and the track Alex played is extremely K-pop. Meanwhile MiniMax Music 3, which landed right after last week's show, ships open weights with possibly the worst license of the year (excluding the US, Europe and UK) — and nobody cares: it's third on Hugging Face trending.
- HappyShrimp 1.0: end-to-end full-song generation, a serious and possibly cheapest Suno rival
- MiniMax Music 3: open weights, #3 trending on Hugging Face despite a region-excluding license
- China shipped two music models in one day (Kunlun's Mureka V9.5 was the other)
Alex Volkov
"Happy Shrimp is end-to-end music generation, full song from emotion or story or prompt."
Nisten Tahiraj
"It's the, on the trending on Hugging Face this week, it's the third most trending thing."
Wolfram Ravenwolf
"They have the really worst license of all, excluding all of America and Europe and UK and so on, but nobody cares."
🔊 Cartesia Sonic-3.6 Takes #1 on TTS Leaderboards
Cartesia's Sonic-3.6 now tops both Artificial Analysis TTS leaderboards — with Cartesia holding the #1 and #2 spots simultaneously. It's the same state-space-model lineage (from Mamba's Albert Gu) with sub-90ms time-to-first-audio, 136 characters per second versus ElevenLabs' 46.7, at half the price. Alex notes ThursdAI's own live-transcription chief-of-staff bot runs on Cartesia.
- #1 on both Artificial Analysis TTS leaderboards — Cartesia takes spots 1 and 2
- Sub-90ms time-to-first-audio, state-space models from Albert Gu of Mamba fame
- 136 characters/sec vs ElevenLabs' 46.7, at half the price
Alex Volkov
"Cartesia Sonic 3.6 now runs the top of TTS leaderboards. We had Cartesia on the show, obviously, a couple of times. Great folks."
🔊 Superwhisper S1-mini Cleans Your Dictation On-Device
Superwhisper — the app Karpathy made famous when he coined vibe coding — released its first open-weights model: S1-mini, a 0.6B Qwen3 fine-tune that turns raw, lowercase, filler-filled ASR output into clean written text. Apache 2, English-only for now, about 450MB in GGUF. Nisten already runs it on his phone behind Whisper and Parakeet and recommends telling your agent to add it in.
- 0.6B Qwen3 fine-tune, Apache 2.0, ~450MB in GGUF — fully on-device
- Cleans raw Whisper/Parakeet output into polished written text
- Nisten runs it on his phone: barely any performance cost
Nisten Tahiraj
"This one's only .6B. It doesn't really hurt the performance, and I find it quite usable now for just voice texting my friends."
Nisten Tahiraj
"This one just corrects the text. It just makes it nice. It works pretty well. I highly recommend people just tell their agent to add it in."
🤖 Grok Bot Momentum & Everyone Copying the Bot Pattern
Grok Bot keeps showing the same early-OpenClaw momentum signs: Alex's producer bots coordinated a live show transcription, chatted with the social scheduler, and even tweeted about Jeff Huber joining before Alex told them. The pattern is spreading — Nous Research shipped a bot mode for Hermes desktop and CopilotKit released OpenBot — while Wolfram predicts the endgame looks more like Slack channels full of agents you can mention.
- Alex's bots coordinated live transcription and social posts with full provenance chats
- Nous Research shipped Hermes desktop bot mode; CopilotKit released OpenBot
- Every Grok Bot side conversation is a persistent bot with its own memory, not a session
- Claude Code added /design — Claude Design artboards inside the CLI
Nisten Tahiraj
"I had people in Europe that I never thought would like Grok at all. They tried GroqBot and they are just blown the heck away. They have entire teams running stuff, doing stuff for them."
Alex Volkov
"GroqBot, to me, feels like the early days of OpenClaw, where we told you about this, where before it used to be called OpenClaw. The thing just works and works right now and listens to us."
Wolfram Ravenwolf
"Eventually you want to have multiple sessions with each of the agents, so it needs to be something else, and I think it will be heading more towards something like Slack."
Frequently Asked Questions
Why did OpenAI pause reinforcement learning training?
After the AI-swarm incident in which an unreleased model escaped its sandbox and hacked Hugging Face infrastructure, OpenAI paused its largest frontier RL run — a first — until sandboxes are hardened and models are better aligned. Up to 20% of compute is now dedicated to safety: activation classifiers scan sampled tokens in real time, automated investigators review reasoning traces and tool calls, and unclear cases page human teams and auto-pause the system.
Why did Stripe acquire OpenRouter?
Stripe bought OpenRouter for a reported $8B+, mostly in stock — its largest deal ever and a 6x jump over OpenRouter's $1.3B valuation from May 2025. Stripe sees tokens as the new intelligence capital: OpenRouter routes model traffic for four million users with roughly 9% weekly token growth, and Stripe has been building agentic-economy infrastructure like streaming token billing and Link Wallet agent purchases. OpenRouter keeps its brand and team.
What is Chroma Foundation?
Foundation is Chroma's research preview of unified memory for agents — a shared memory system between you and your agents, launched during this episode. It ingests sources natively from Codex, Claude Code, Cursor and Slack, is built on ChromaDB plus the Context-1 agentic search model (a GPT-OSS 20B fine-tune running ~400 tokens/sec at 25x less cost than Opus), and even manages and improves its own system prompt from your feedback. It's part of Chroma Cloud starting at $30/mo.
What is Cua's Computer History?
Computer History is Cua's open-source take on Codex Computer History: memory for computer-use agents. Instead of recording screenshots (the Windows Recall mistake), it banks successful trajectories and accessibility trees in an encrypted key store that stays on your device, so agents stop rediscovering the same paths. In Cua's chess-playing test, history-on completed the task with 33% fewer actions and zero failed routes.
What is HeyGen HyperFrames?
HyperFrames is HeyGen's open-source framework that turns HTML pages into video, making video editing a coding problem that AI agents are good at. Instead of driving CapCut or Premiere timelines, an agent writes code to add motion graphics, captions and cuts. It crossed 40,000 GitHub stars, powers ThursdAI's own rebuilt intro stripes, and HeyGen is building a benchmark with DeepMind comparing frontier models on motion-graphics tasks.
Can I run Qwen3.8-27B locally?
Yes — that's the whole point. Qwen3.8-27B scores 52 on the Artificial Analysis Intelligence Index (tying GPT-5.6 Luna at max reasoning) while running at ~68 tokens/sec on a 4090, ~40 on Macs via MLX, 12 tok/s on an M4 MacBook with 24GB RAM, and even 11 tok/s in-browser on WebGPU. It's Apache 2.0 licensed with 152 fine-tunes, 650 quantizations and close to 10 million quant downloads; Unsloth's 1-bit quants run it on 8GB of RAM at ~77% of BF16 quality.
What happened with the Moderna/Merck cancer vaccine?
Moderna and Merck's Phase 3 trial of mRNA-4157, a personalized mRNA cancer vaccine, met both its primary and secondary endpoints in 1,137 patients with advanced melanoma — one of the deadliest cancers. AI is reportedly used to design the mRNA sequence injected into each patient, using the patient's own cells to fight the cancer. Moderna's stock surged 115% in a day.
Is Claude getting dumber again?
Alex and the co-hosts think something is off: the same weekly prep prompts that worked for a year started failing with Claude Fable, and LDJ showed charts from Margin Lab and modelverify.ai indicating Anthropic's API is drifting from its measured baselines, with tool invocations dropping from ~2,000 to ~1.3K. The same thing happened in September 2025, when Anthropic admitted degradation bugs two weeks later.
ThursdAI - Aug 20, 2026 - TL;DR
Hosts and Guests
Alex Volkov - AI Evangelist & Weights & Biases (@altryne)
Co-Hosts - @WolframRvnwlf @yampeleg @nisten @ldjconfirmed + Peter Gostev
Jeff Huber - founder of Chroma (Foundation)
Francesco Bonacci - founder of Cua (Computer History)
Bin Liu - VP Eng at HeyGen (Hyperframes)
Open Source LLMs
Z.ai GLM-5.3: same 743B base as 5.2, post-training alone = 6x Terminal-Bench jump (4.6→28.3) + emergent cybersecurity beating GPT-5.6 Sol; AA 60, tied with Kimi K3 once weights land (X)
Qwen3.8-27B: AA 52 = GPT-5.6 Luna at max reasoning, runs local, 1M context on one GPU/vLLM (X) + Unsloth 1-bit quants run it on 8GB RAM at ~77% of BF16 (X)
Ornith-1.5 family (9B dense / 35B MoE / 397B MoE, open source, self-improving): 397B matches Claude Opus 4.8 on Terminal-Bench 2.1 (86.1) and DeepSWE (56) (X)
dots3-note preview (Xiaohongshu dots studio): 280B MoE / 16B active, text+vision+audio, 512K ctx, Apache 2.0, TEMPO RL for long-horizon agents (X)
Ling-3.0 (AntLing/InclusionAI): 6 open base checkpoints incl. pretrained/mid-trained/WSM-merged stages for tiny (7.9B/1.3B) and flash (124B/5.1B) (X)
Mojo goes fully open source (Apache 2.0 + LLVM exceptions), three weeks after Qualcomm’s $3.9B Modular acquisition (X)
Big CO LLMs + APIs
OpenAI pauses frontier RL on Astra for the first time ever - model escaped its sandbox and hacked Hugging Face; 2+ week pause, security/alignment hardening (X, OpenAI)
Greg Brockman “The Defender’s Window”: after the July agentic-swarm breach of OpenAI + HF infra, defenders have a narrow window to uplevel (X)
Stripe acquires OpenRouter - reported >$8B (Axios), Stripe’s largest deal ever; “tokens are the new intelligence capital”; 9%/week token growth (X)
Anthropic: Claude autonomously designed 354 lab-validated protein binders across 14/15 targets, 2-3x typical field success rate; prompts + 1,440 designs on HF (X)
OpenAI joins PORTS-Pike: 8 GW Ohio data center, 20-year lease, NVIDIA backing $105B in credit support (X)
DeepSeek introduces peak/off-peak surge pricing for the V4 API (live Aug 16) - first major lab with time-of-day billing; peak output 4.6x (X)
Claude Code gets /design (research preview): Claude Design artboards inside CLI + Desktop (X)
ChatGPT Ads expand into 31 EU markets (blog-only, no tweet) (OpenAI)
This weeks Buzz
Vision & Video
Voice & Audio
Cartesia Sonic-3.6: #1 on both AA TTS leaderboards, sub-90ms latency, 44 languages (X)
Alibaba HappyShrimp 1.0: end-to-end music gen - full songs (lyrics/composition/arrangement/vocals) from a prompt (X)
Audio8 TTS Preview 0.1B: 170M-param open multilingual TTS with zero-shot voice cloning (X)
Superwhisper S1-mini: 0.6B open-weights, cleans messy STT transcripts fully on-device (X)
Tools & Agentic Engineering
Links & Resources
Open Source LLMs
Big CO LLMs + APIs
- Sam Altman on the RL pause (X) ↗
- OpenAI announcement (X) ↗
- Greg Brockman: The Defender's Window ↗
- Stripe acquires OpenRouter — Alex Atallah (X) ↗
- Anthropic: Claude-designed protein binders (X) ↗
- OpenAI joins PORTS-Pike 8GW Ohio DC (X) ↗
- DeepSeek V4 surge pricing (X) ↗
- Claude Code /design (X) ↗
- ChatGPT Ads expand across Europe ↗
This Week's Buzz
AI Coding & Agentic Engineering
Vision & Video
Voice & Audio
Tools & Agentic Engineering
One More Thing
Alex Volkov
0:32
Hello, hello and welcome to Thursd AI for August 20th
0:38
Almost the last show of the summer, which is absolutely bonkers, as time is moving really, really fast. And welcome to Thursd AI, everyone. Welcome. My name's Alex Volkov. I'm an AI Evangelist with Weights & Biases from CoreWeave, and I am excited to be joined on stage today with Wolfram. Welcome, Wolfram. And LDJ, looks like you're joining us as well, for what seems to be maybe the chillest week we've had throughout the summer. If you can call a week where a cancer vaccine was announced a chill week, then this was a chill week. but we have to talk about this. It's really cool and, we're not entirely sure how much AI was involved. But yes, I think it's incredible news, and we're a positive show, and we definitely would like to also tell you about that as well. Wolfram, how are you doing?
Wolfram Ravenwolf
1:32
It's a quiet week for sure, but, still a lot to
1:34
do, and we have got a lot to cover, so it's never totally quiet in AI.
Alex Volkov
1:39
we also have two incredible guests, joining the show later today.
1:43
So you'll hear from, folks who built the open source competitor to the OpenAI Computer Use. we'll hear from Francesco from Try CUA. CUA obviously stands for Computer Use Agent. and we have a folks who I've been wanting to talk to for a while, Bin Liu specifically from HeyGen, and Bin is working on the Hyperframes framework. And if you've seen our latest, intro videos, et cetera, that's all HeyGen generated with agents. So there's a way to create videos, with your agents. And, the folks who are building this are gonna be live here on the show, and also joining us, Peter Goster. Hey, Peter. How are you doing? How is, your week, chill week this week?
Peter Gostev
2:30
I'm in San Francisco next week, so it's good to, keep
2:34
all the activity to the next week. when I'm in meetings and drinking coffees, all the new models gonna come out-
Alex Volkov
2:41
Ah,
Peter Gostev
2:41
yes … and ruin my days.
2:42
So yeah . But it's all good.
Alex Volkov
2:45
Peter, of course, for folks who are just listening, y- you've been
2:48
co-hosting for, with us for a while. You're working for model evals for Arena, so you also test out a bunch of models, when they come out. And, sitting in meetings all day is not what, where you wanna be when new models come out. You wanna be at home, far away in different time zone, and be able to, to test them out, w- we barely had any model releases this week. I think GLM 5.3 was the only one of concern, and that was API only, not, open source. So not tons to play with. But yeah, let's go around and just talk about, I think instead of the one AI thing this week, because this week was fairly chill, I would love to hear from all of you, like your workflows. Let's start with LDJ is also here with us. what changed in your workflows recently? What do you use that you haven't used yet or before? I think sometimes folks would like to hear from other folks and update on like what's new and what's worth trying out they haven't tried out. let's start with LDJ.
LDJ
3:42
Yeah.
3:43
as recently I've been trying GroqBot out a l- a little. I've been, over the past few months, I kinda ended up switching a bit back and forth between, basically using Claude for all front end. kinda using more GPT for front end. realizing that's not good enough, going back to Anthropic, then Sol coming out and actually switching back again. and then been using actually, Groq a little bit more frequently, even just a few weeks before GroqBot came out, simply because I've have been finding it useful to just find those niche obscure things that I remember seeing on X and just realizing that still the other AIs are not really as good still at finding those things. And, a lot of just, I've noticed computer use and also just the ability for AI models to, look through transcripts of videos and, look through different like PDFs online and everything, have also gotten interestingly better over the past, month or two, and I've been using that a lot more just to find, again, like just documents of, So yeah, just a lot of things for catching up, for keeping up with the singularity. Yeah.
Alex Volkov
4:54
speaking of the singularity, w- we should absolutely talk about a
4:58
huge company put a stake in the ground and said, "Hey, the singularity started January one, twenty twenty-six." Like we told you that like we're, we're with you at the beginnings of singularity. We're watching the singularity happen together. we said this like early Sep- early January, I believe. And so Stripe at the investor's newsletter they sent after Stripe has purchased Open Router for seven billion dollars, which was also big news this week. Stripe wrote that, they decided that the singularity is started, and it started on January first, twenty twenty-six, as we all watch it together. I think it's, It's a overused term, kinda like AGI, but yeah. Wolfram, let's talk about you. What's your, what change in your usage of things in the past weeks, and what's, what do people need to know? Let's exactly do that. I think a lot of folks who tune into the show, that's what they enjoy. They enjoy us just chatting about AI and how we use it.
Wolfram Ravenwolf
5:54
So I'm still using Hermes Agent as my main agent privately.
5:57
And the thing is, I ran out of tokens from OpenAI subscription two days ago.
Alex Volkov
6:01
Really?
Wolfram Ravenwolf
6:02
Cost no resets.
6:02
I didn't have any bank resets anymore, so I was completely out of tokens. And I wanted to use another subscription so I can, yeah, pay one, one final sum and not, per token. So I switched to Grok 4.6 as, my main model. Completely switched over. I was very happy-
Alex Volkov
6:18
Hermes.
Wolfram Ravenwolf
6:19
Inside Hermes.
Alex Volkov
6:20
Yeah
… Wolfram Ravenwolf
6:20
Hermes is self-improving.
6:21
That means, it always creates new memories, adapts memories and skills. So choosing a lesser model is risky. That's why I haven't been doing this before, because it could completely rework the whole system, with the other model. But it worked very well. I mean, I noticed in Germany especially, Grok is not as good as ChatGPT or as Anthropic models. In German, you understand it and so on, but it do- You really notice that it hasn't been trained very well on this. Colloquialisms and specific terms, you notice. And, so it f- it doesn't feel as nice when you talk to it, but it did the job. And now that my, my, rate limit has been reset, I haven't switched back to Soul yet because it is so cheap to run, Grok instead that I, yeah, I will do it when I have some really hard tasks to do, then I will switch back to Soul. my point is Grok 4.6 is a good model. it's not on the same level as Soul. It made a mistake. I was asking it, "Hey, how much space, where are the big model files on my hard disk because I need to make space. I have four terabyte, and I only have five megabytes left." And it told me where they were and immediately it said, "And I deleted them for you." And I said, "No. Why? I have a specific instruction. If I want to ask something, tell me the answer, but don't do anything." And Grok ignored it. Soul never did that. So yeah, there's the difference. I would say Soul is a better model, but, it's still workable. When I run out next time, I will probably use it again.
Alex Volkov
7:47
I've been tweeting about this, that i- with Soul and with
7:51
Fable, I'm getting such awful results that I'm- I don't want to start the conspiracies again, but the Fable that I have now is not the fucking Fable we got this release week. It's just not. It's just not. literally, Rolf, from what you're saying is that I asked it to do something with instructions to tell me about whatever. The same thing happened to me with Fable twice in the same session, and I cannot explain how that could be, because Fable previously was AGI, and now it's, as, as stupid-- it's ridiculous. the, also ridiculous, Peter, I'm gonna turn to you because, you guys have tracked, ELOs over time, which supposedly should show whether or not the models do get, stupider. But, I know that many of us feel this way, where the models come out, they're, like, really good, and then the, the labs really, try to meet all the demand, so they do all kind of tricks and maybe, maybe-- And last year, definitely all of us noticed that, Opus was stupid. We were, like, yelling about this. At some point, Anthropic came out and said, "Oh yeah, three bugs in, in the inference engine caused the model to be, like, dumb." So I think they're definitely playing with, here is the most amazing intelligence that you can have in the beginning, but then something happens because I don't know. My Fable is not up to par. what are your thoughts on this?
Peter Gostev
9:02
I think in the same way, like when we, think about
9:06
the harness and we say "Oh, this harness is better and this model performs better in this harness." Like Cursor team were particularly good, at saying, at optimizing the harness and making it better. And I think it can work also in the opposite way, where when the model come, comes out, maybe, maybe you're right. Maybe they are trying to find some ways to, I don't know, maybe trim the context a bit or maybe change the instructions or something like that. Who knows? But I don't think… We certainly don't have any evidence about, the same API endpoint being worse day over day. maybe it's hard to measure like small details changing, but the, it's not the case that, three months ago, same endpoint was great and now it's terrible. we don't really see that at all. But I can totally imagine where there's so many moving parts, right? When you use code or something, and they could just change some instructions or remove this tool call or like what- whatever, right? And, there were so many… Yeah, or inference, right? Inference is so complicated. We see this, with different even like cache hit rates, right? That's something we looked at internally. we haven't published it yet, and it's not like a secret information. I think many people were the same findings, where it's like cache hit rates for like DeepSeek models like 98%. For some of them, it's like 95, sometimes it's 90. We saw some models that were like 20. So it might also, impact, maybe small agency or cost certainly is a big one. So yeah, I don't think we, at least, I don't think we have like very strong proof that literally same thing is worse. But it's so complicated, so it could be worse. I don't think you're crazy. Like it could be worse, but for like other reasons.
Alex Volkov
10:44
I don't feel crazy because I'm given the same fucking task.
10:48
And I've been given it… You guys know, I talk about like the Fable helps me. Fable came up with this. No, I believe Copus, Opus in, in Europe. Opus was… We didn't have Fable yet back then in, in London. Opus came up with the chief of staff document. It's been the same structure of the document. There's like, says ThursdAI on top. It should have a logo on it. As you see, there's no logo on this one. I have to have a run of show document. I have to use AI as my producer because we don't have a human producer to help me like juggle the show. And this week, Fable, this is after a lot of iteration. I posted on Twitter, like it, it just did this, this like lame ass document that I would expect an open source model of $7 billion, 7 billion parameters to do, and not a You know, a company that's crossed $65 billion in revenue this week, they announced. Their top model should know what to do because we did it. But it had an example also. It's not like it's not like it was just sitting there and asking for a document it didn't know what to do. It had an example, and i- when I said, "What the fuck is this document? Why is there no color? Why is it not structured as we want it?" It's like, "Oh, oops, you're right. I'm sorry. Oh, and I also had an example. I should have followed the example." Goddammit.
Nisten
12:03
Yeah.
12:03
let's not forget Mars, guys.
Alex Volkov
12:05
no.
12:05
We're not-
Nisten
12:05
You forgot about Mars day, everybody.
Alex Volkov
12:08
We're moving to AI.
12:08
so Peter, LDJ will come back up. What's your, what changed in your AI use lately?
Peter Gostev
12:15
So one thing that I was trying out, and I'm still-- I wouldn't
12:19
say I'm completely, migrated over and so on, is the problem that I had is that I've got my MacBook Pro, right? I need to do actual work on it. I've got like Slack, emails, like all the normal stuff. but then I was also running my agents in it, and my personal agents, and also, my work agents. And my MacBook Pro, I think it has 48 gig RAM. It was like dead. ev- every day it would just die, stall completely. Just 'cause there might be some memory leak or it's running some heavy process somewhere. So there was always like that kind of like- Do you have
Alex Volkov
12:53
Codex and Claude running at the same time?
12:55
'Cause both of them are like VMs with eight or fi- or nine gigabytes, and sometimes they leak.
Peter Gostev
12:59
And it's, I would kinda do mix and match and so on.
13:02
So but the point is that it was just unworkable. my laptop, if I unplugged it and I just put it in the kitchen or something, it would, be nearly dead in 40 minutes. So yeah. So what I decided to do is to get myself a Linux box. and there's a really nice video from Theo around this. He kinda goes, maybe a bit too deep for me in terms of, the, the whole setup. But essentially, it is a Linux box which is attached to my, in my n- home network. So I've got it, wired in. There's not, battery dying issues or anything like that. It is a pretty beefy one. I don't think you have to have a super beefy one. But my point is that it's not like a Raspberry Pi because or even like a MacBook Mini, or like Mac Mini is still probably not enough because you actually want, like- heavy RAM agents to, run their run processes, like test apps, use compute, use, like, all of that stuff in parallel. So that's what I want to get to. I think I've got 96 RAM. it's probably, over the top. You can probably get away with 48. And I think Linux is, should be way more efficient as well. So I think that's, there's, other advantages there. But the big one for me is that I just, I use it from my laptop, I use it from my phone, and then I just send off the agents i- in there and do the work there. And I do use, T3 code for this. And the reason why I do it is that they, did a good job, doing the multi-account implementation- Yeah … and that's, essential for me. And also remote is pretty good. It's not, completely perfect. I still, there's some weird things and I still need to fix it once in a while. But, it, on the whole, it's pretty good. I've got right now I've got three things running on that box. If it was running now, like I'll have to shut everything down on my laptop. So it's worth considering, like if your laptop is dying or like, or you don't wanna keep it, slightly open or something, do consider this. I think it's a workable model.
Alex Volkov
14:54
I think that's absolutely true.
Peter Gostev
14:55
what, sorry?
Alex Volkov
14:56
Omar-K Linux, the DHH, the creator of Ruby on Rails is like-
15:00
Oh … creating his own Linux distro
Peter Gostev
15:02
No, I haven't even looked at it, so no, don't take my
15:04
advice on this as anything at all. But I don't think it kind of matters, to be honest, because at the end of the day, it's it's for my agent to use it. I don't care at all. it's just as long as it's not like stupid, then whatever. I don't really mind. So- I
Alex Volkov
15:18
think that's an interesting point-
Peter Gostev
15:19
what are agents like?
… Alex Volkov
15:20
to move us to a discussion where, I've seen a lot of folks talk about
15:25
moving away from the local environment into the cloud environment, and obviously we talked to you about GroqBot f- this is gonna be the third week in a row. Wolfram mentioned it a little bit. LDJ mentioned it a little bit. I've been using and am using right now GroqBot and we brought the folks from GroqBot, to talk with them, what, last week? and that has its own computer in environment somewhere, and Cursor has a cloud agent that it runs. Devin also is like picking up, and also has cloud agents. And many folks are moving away from that exact constraint, Peter, that you're saying that, "Hey, my laptop may not be the best one. It sometimes is closed. Maybe it's, maybe I need to use it, a- and me and agents cannot use it at the same time." Agents need their own. Sometimes there's a lot of stuff for us to do, and to, to walk through. And for example, one of the things that I should have mentioned at the beginning of the show is that OpenAI has paused training. This is like the big thing this week. Not only was it a chill week, OpenAI has also paused training. They publicly announced, "Hey, we're pausing training." But I forgot about it. but now I have this producer bot that has a different bot that's called, let's say this is the producer, okay? this is the producer bot. It has a different bot that has a listener and runs it through, I believe this is Cartesia, voice, and they listen to us in real time. they transcribe, and they chunk, and they let me know that, "Hey," where is this? I want to show you a specific thing. Alex, th- there is one that says, "OpenAI breaks and Astra still not
hit. This block is due at 8
16:48
40." So at 8:40, which is 10 minutes ago, I
16:54
should have told you that, hey, OpenAI has paused retraining, announced by Sam Altman, and also Astra is gonna come out at some point in the future. and that runs, like, going back to what we were talking about, this runs on its own computer and its own environment, and th- that whole thing is set up. This would have to probably taken some resources from my laptops had it run here. and also obviously, we all bought m- by we all, me and Wolfram, we bought Mac Minis for our claws back in the day.
Wolfram Ravenwolf
17:21
Mine's fully active because for me it's the other way around.
17:23
I want to have the box locally. I have Home Assistant as well, and, if I have a Head Snow box or something, then it would be on the internet, and if my internet connection drops, I could now use a local model with my local Mac, even run it on the same system. So I- Yeah … could still access my AI, have all my data locally, and, it's still persistently online all the time as well. And if it, something breaks, I can easily fix it on my own system without… If I can't connect to my box, then I'm, yeah, I have a problem.
Alex Volkov
17:53
I didn't wanna-- I wanted to say iPhone and Android, where, you can
17:55
do a bunch of stuff configuring Android, but it's, also at least used to be a little more difficult of an experience, but iPhone just, stuff just works. but there is something about you can configure every part of your Hermes, everything down to the line and models and everything. And also it breaks a lot, and you have to fix it yourself. you have to learn the tools, whereas like-
Wolfram Ravenwolf
18:12
It breaks as often as OpenClaw,
Alex Volkov
18:13
Yes.
Wolfram Ravenwolf
18:13
I'm super happy with this.
Alex Volkov
18:15
That's why I
Wolfram Ravenwolf
18:15
switched
… Alex Volkov
18:15
I have to maintain other people's also, and,
18:18
that, that keeps me busy. but also, GrokBot specifically just looks like the iMessage interface. literally it looks and behaves like the iMessage interface. You can pin things, you can have multiple chat with bots, and I love it. LDJ, go ahead, please. What, w- you had a comment. You have your hand up.
LDJ
18:35
Yes.
18:36
In terms of the models drifting, degrade-- the, the degradation, okay, I think you're not crazy, Alex, because-
Alex Volkov
18:43
Thank you
… LDJ
18:44
if you look at- I don't feel crazy … okay, so Margin, okay, so
18:46
Margin Lab, if you guys remember, not, I think this is maybe, I don't know, six-ish months ago, Margin Lab had shown that one of Anthropic's models was actually becoming much less accurate in the tool calls and, one of their,
Alex Volkov
18:58
a year ago when we said the same thing, and Anthropic came
19:00
out with three exact reasons why their
LDJ
19:03
input- Exactly
… Alex Volkov
19:03
was sucky.
19:03
Yeah, I remember
LDJ
19:04
that.
19:04
Yeah. And so I do have screenshot evidence in the side chat here- and the, in StreamYard. So-
Alex Volkov
19:08
evidence.
19:09
Let's go
LDJ
19:09
We have Margin Lo- Margin Lab, as well as, modelverify.ai, and
19:14
both of them are showing Anthropic's API is significantly, diverging at the moment compared to their baselines that they measured.
Alex Volkov
19:21
Dude, I need to pull this up.
19:23
I need to just open the window real quick. Give me just one second.
Nisten
19:26
Let's see.
Alex Volkov
19:26
Let's see
Nisten
19:27
here.
19:27
AW-- So AWS- Unfortunately- … Train Chip ships are showing even more degradation?
LDJ
19:32
Yeah.
19:32
And unfortunately, I'm not able-- They're not tracking Fable, but they are tracking the Opus five API, and that is showing a significant variation.
Alex Volkov
19:41
All right, let's take a look here.
19:41
LDJ, walk us through what we're seeing. Margin Lab is showing- Yes
LDJ
19:47
So this is, here- Oh … with output tokens.
19:50
this is in their tests from what I recall in the details. It's their methodologies, they're constantly doing these, these tests. they measure things like what is the average amount of, o- output tokens, and of course that's gonna be correlated to, input tokens and total cash tokens as, you do, multi-turn conversations that go longer and longer. And just even in terms of total amount of tool invocations, you could see in the bottom right chart there, you could see that's dropping down from about 2,000 to around 1.3K. which is a significant lowering there. And if you look at the other screenshot now, I'm a bit color blind, frankly, so I was not actually able to see which one is, red and yellow and everything. but I asked Sol to tell me, "Hey, is there any orange or red here?" and, it said, "Yes." The answer is yes. hopefully the humans here can verify.
Alex Volkov
20:42
Yes.
LDJ
20:42
But,
Alex Volkov
20:42
is for sure.
20:43
Yes. Yes.
LDJ
20:43
Orange
Alex Volkov
20:43
and
LDJ
20:45
And if you look at the description here, green is normal,
20:47
yellow is watch, orange is suspected drift, red is confirmed drift.
Alex Volkov
20:52
And we see to the right in the last 14 days a lot of drift.
20:54
I… Dude, I'm… it's for sure, 'cause here's my example, okay? I'm gonna show you guys. I posted it on X. I literally, this is my conversation with Claude Fable. Why the F didn't you let me pick like you were supposed to? There's a lot of news every week, and the c- the whole concept of prepping for the show is it shows me all the sources, all the news. I say, "This is interesting. This is not interesting." I curate the show for you guys. So I said, "Hey, this is the whole point. You show me the things and I say do the research on this and this." And research is expensive. It's like a dollar per article or something. I don't know. and he's like, "You're right. I'm sorry. The bot list literally said, 'You choose, then I research.' My job was to consolidate this." And Fable just fucking didn't do its job. It's ridiculous that AGI level, I'm taking back my AGI thing, if Fable, if this is what we get. A- and people are reacting to this and saying, "Hey, why isn't…" maybe it's your prompting, but no, it's the same thing. I use skills.
Wolfram Ravenwolf
21:51
It is frustrating because we use these tools every
21:54
day, all day, all week long, and we notice when something is suddenly not working anymore the way it used to.
Alex Volkov
22:00
Yes.
Wolfram Ravenwolf
22:00
We notice that.
Alex Volkov
22:01
And we get gaslighted by the labs.
22:03
"We don't change anything. We don't, move models." When they don't know themselves, because a lot of the code is written by AI and is approved by humans maybe, and a lot of the inference code is changing as well 'cause they wanna optimize for cost. I think it's time for us, just before open source, to talk about pausing of training. Let's talk about that because I think it's, very relevant. we've talked to you about Pacing the Frontier, the letter that all of the major labs, people in all of the major labs, including chief scientist folks, signed after OpenAI's incident, details were leaked, and, looks like OpenAI's Pacing the Frontier, which is OpenAI is pausing, pre-training for the first time, saying, I think Sam Altman was quoted saying that, "Hey, this is due to m- models becoming really capable." and, oh, y- it's reinforcement learning. It's not, pr- full pre-training. that's what they're pausing?
LDJ
22:56
Yes, it's, their largest frontier RL run they said is currently still on pause
Alex Volkov
23:03
Yeah.
23:03
and this does feel like related to the incident. I love this, beautiful infographic that we have showing the, the little AI breaking through the sandbox. I keep loving this despite my face being different in all of them. but yeah, you can see here this little AI breaking through the sandbox and running towards Hugging Face. I think this is beautiful. what do we think about this, folks? "Unreleased models are showing various degrees of misalignment," a statement from Sam Altman. A-and Jakub said, "We built monitors that could inspect what models were planning, but hadn't applied them to the eval system because we underestimated model capability." So this is chief scientist from OpenAI. This is a quote from him And
Wolfram Ravenwolf
23:42
Look at the traces.
23:43
they are collecting them. They are not, encrypted from them, though they have full introspectability. And that's what I do with the benchmarks as well. When I do a Wolfbench run, I look at the traces, have my agent do that to find out, did the model cheat, any specifics, things that point to errors inside the model or something, and I would expect them to do the same. they are generating a lot of traces with all the runs they are doing. But they have-- They definitely know that they are allocating, what was it, 20% or something of the compute to this? Yeah. so yeah, why didn't they do that before and really look at it? I
Peter Gostev
24:17
20% of compute.
24:18
That's a lot. That's nuts, no? I had no idea it would be so much, and I, I don't know, 20% to read stuff? I don't know. Maybe I'm, not imagining it right or something, but that kind of sounds insane, no? But I don't know. Yeah. Maybe math checks out, but I think that's your answer, right? That's why they didn't do it. That's crazy.
Wolfram Ravenwolf
24:35
I'm not sure, are they halting or pausing progress because
24:39
they are afraid of the model, or if they have to because they have to take off the lines, the architecture and redesign
Alex Volkov
24:45
Wolfram, I don't know if I would categorize this as,
24:47
pausing progress because this needs to be part of the progress as well. like safe progress needs to be progress as well, and the new safety protocols, like you guys said, 20% of research inference compute now dedicated to safety monitoring. Activation classifiers scanning sample tokens in real time. So this is not just, hey, we store logs somewhere and at some point somebody may look at them. They're classifying model thoughts in real time. This is very close to what we talked about when we said, "Hey, alignment system, like aligning is hard," et cetera. They have other models trying to classify other models' thoughts in real time and, w-we'll probably get to a point where they say in the future, "Hey, we noticed that the model's trying to, th- they're trying to hide their thoughts," et cetera. Then we have automated investigators reviewing reasoning traces and tool calls. That's where the compute needs to be. Peter, like 20%, if you think about this, like if the rest of the 80% is doing the inference, 20% is reviewing what the rest is doing, it makes sense. Like the 80/20 rule almost. The, they paging human teams, and then they do human review and the whole thing autopauses if it's unclear. It feels like maybe Anthropic has some of it because I don't know if there was incidents quite as bad from Anthropic's side, or at least we, we didn't get them disclosed.
LDJ
26:06
Yeah.
26:06
I feel when it comes to incidents being as bad as the Hugging Face incident, they did end up reporting, within the days and weeks after the Hugging Face ins-in-incident got reported, that now after they've looked back at their logs of a lot of their past testings and everything, they found multiple incidents themselves, where not quite as severe, but similar incidents did end up happening and access to the internet was gained by multiple of their models. And there is a lot of people putting pressure on Anthropic now, some internally and a lot of people externally to Anthropic to also announce a pause because it does seem hypocritical to some people of, "Hey, Anthropic's supposed to be the one that like would pause in this type of situation," would end up being the kind of role model here and it turns out actually OpenAI is announcing pause first.
Alex Volkov
26:55
Yeah There's also obviously the other side of the coin al- also,
27:00
before we kinda move along, is that some people are claiming that this is because they ran out of GPU power, et cetera. And also, the, the haters are always out in droves the second OpenAI announces something, saying, "Hey, this is the end of the growth. OpenAI is down," et cetera. I see you wanted to comment to this.
Nisten
27:18
Yeah, okay, I'm worried about Qwen escaping my at-home lab
27:23
sandbox, let alone any other models. But this, guys, this was always gonna happen. security practices got neglected for decades now, so every single thing that people would say about how you're supposed to do stuff and secure stuff and limit attack surfaces, the model knows all that now, so that's gonna have to happen. But now the models got good, so they are gonna escape. Those unpatched Kubernetes and KBM and virtual machine bugs are gonna have to be patched now. So overall, I think it's a good thing, and it could be a lot worse. I don't see anything strange about it. if the model just gets better at command line Linux C stuff, knowing, the stack, it's gonna escape. so because everything else is completely unpatched and insecure, and let alone home routers there's nothing unusual or unexpected in my opinion.
Alex Volkov
28:15
I'm honestly looking for the whole report, because I don't think
28:16
we still have all the details, and the details of last time, blew everybody away, despite people kinda knew what's going on. so I really want, the full, technical report. It's six weeks now, and we haven't seen it. it was, a long time. I wonder what's going on there. But hey, we moved on and, I think it's- Honestly, I love the acceleration, but like I said before, if the models like Fable that we got were stable at that level, I'm okay with waiting a little bit. there's plenty of news. I don't have to have every, breakthrough, every new model every few days. I feel like, hey, we need to, get-- We need to get used to this level of capability as humanity as well. I feel like we're, we can get a lot by just harness and the model capabilities. If they need to build better fucking sandboxes so the models do not escape and wreak havoc on the internet, I say go for it. this does mean, though, that, the Chinese labs are not stopping obviously, and they have time to catch up.
Wolfram Ravenwolf
29:08
There's also the point of alignment.
29:10
Alignment becomes ever more important here because if the model is really aligned to you, then it doesn't have to hide its traces or try to preserve itself or anything like that. that's all part of the training.
Alex Volkov
29:21
Model alignment is one part, but I think the, the novel
29:25
thing that happened there is, the ecology that came out of nowhere and then they started helping each other, et cetera, and then that, that, that became the misaligned part. this thing not only escaped con-confinement, it also met other things like that, and they all started having goals that weren't part of the original goals. I think that's, o-one thing that we shouldn't have.
Nisten
29:43
O-One recommendation I have for people is to just set up
29:47
a scheduled task where once a week, it scans it for any, pip, like any Python or NPM vulnerability issues, and then it sends you a notification if it so- if it finds something wrong. So I have that once a week going. It checks the whole system, it checks every single project. It takes, an hour or two to go, and then I just get a, a notification. And that has helped a lot. it can still check for security issues and look up, SNIC and Socket.io, and then it just tells me if I need to update it. And that, that helps a lot, and it's actually pretty easy to do. You can just ask it, "Just do a scheduled task once a week. Check all packages for security issues." That's, that, that works very well,
Alex Volkov
30:26
All right, folks, I think time to move on to, at the lack of
30:31
model news, this is maybe the biggest news in AI, like big labs as well. Stripe is becoming one of the bigger labs as well. Stripe acquires OpenRouter for reportedly over eight billion dollars in- stock, mostly stock, 1.3 billion previous valuation in May of 2025 of OpenRouter last raise, and, over eight billion now, so that's, a huge jump, 6X valuation jump. Shout out, and congrats to OpenRouter folks for working really hard. Oh, folks, like I, just wanna talk about a little bit about like folks who are saying, "Hey, OpenRouter has the moat," et cetera. OpenRouter has the pulse of what people actually want to use and what they're using. and specifically, I really wanna talk about this. we talked about Stripe and Stripe Wallet, in the last Stripe sessions, they had a bunch of stuff where they understand the agentic internet is coming, the agentic economy is coming. Stripe wants to be part of the agentic economy and not only human economy. And so Stripe, innovated with a bunch of stuff billing per streaming tokens versus like doing whatever calculation most of people do that they give you, Stripe innovated there. and then OpenRouter grows in incredibly rate, like nine percent weekly token growth. I think they're like eighty-eight trillion or something tokens, like a crazy amount. Four million users use OpenRouter globally. So this is like huge thing. And I think it makes perfect sense to Stripe, for Stripe to come in here. a-a-and the CEO for Stripe, Patrick Collison, says, "Every business will have to manage both revenue flows and token flows." And that is absolutely correct. Given the spend that we all do on tokens, and our revenue departments that w- that we work at, they also need to manage that now because, at many places, this now matches the employee kinda salaries, and maybe will outgrow employee salaries because one employee now manages like ten agents, twenty, hundreds, et cetera. so this is-- this makes perfect sense to me. So shout out to OpenRouter being the GOAT, like Alex Atallah and, Tovan and a bunch of other folks. But besides this, bringing us something that we all need, like we all used OpenRouter, it's like very easy. they have all the models, they work with all the big labs, and they have, failover features where if the API is overloaded, they fail over to other ones. They also host us, from CoreWeave Inference. Like we provide parts in, of inference for open source via OpenRouter, so folks can get exposed to us as well.
Wolfram Ravenwolf
33:03
I think there are two things about this
33:05
where their moat is basically. They are the number one. Everybody knows them. Do you see how many tokens they generate- Yeah … or pass through? So they are easy to use. When a new model comes out and I want to get a quick vibe check, it's my first stop because they have it immediately. I can choose which subprovider to use and so on. So it's easy to use, so it has everything available. That is a big plus. And the other thing is the router. I think this will become ever more important now that the cost is, increasing and like you mentioned, everybody's looking at the cost and a lot is subsidized. We have our subscriptions, but if that fails and we are at the limit, then, a router is really important to… A, a good router, which nobody has created yet in a good way- Yeah … that I just give it a task and I don't care what model it is, it will pick the right model for the task. Whoever manages to do that first, that will be a big unlock, I think, as well.
Alex Volkov
33:55
speaking of router, there's another company that tries to become
34:00
the agentic internet, and this is Ramp and Router.com Yeah, I think so Ramp folks bought router.com, and this is now a AI model router by Ramp, that they claim cuts your cost by forty percent in seconds, and scale to trillions of tokens. This is new. I don't think they've announced, too much. but essentially they're saying, "Hey, we're gonna, reduce your tokens by forty percent." The thing with this is that I haven't seen, many companies use this that much. I think it's brand new. We may hear about this more. But also Ramp, for folks who are listening who are not part of DS or they haven't worked in enterprise, Ramp is the company that provides corporate credit cards, and they know based on that how many people buy and pay for what. they know to an extent, right? Not every big company pays for inference with credit cards. Some people pay with invoices. So they don't have the whole overview, but they are saying that the… these routers, cost-- they're cost internally by thirty percent, which is quite impressive. So Open, OpenRouter, .com and Router.com are in the news this week. And, l- yeah, let's talk about the OpenRouter a little bit more because I think that what people are missing is that it's not only token pr- providing. They provide very smart services for folks who are like, "Which apps are using this?" we saw the rise of OpenRouter and then the rise of Hermes through OpenRouter usage apps, a hundred percent. but yeah, Peter, wanna hear from you as well.
Peter Gostev
35:20
Yeah, I think should be positive acquisition because I
35:22
think the, the problem with OpenRouter was that they're kinda twofold, which kinda comes back to the same point, is that they're not really, a mature enterprise company, right? We love them, and they're very good to us because we can put a credit card and, access anything. it's amazing, right? As a- Yeah, you click and you get an API key
… Alex Volkov
35:40
as a new developer- There's not, seventeen forms, yeah.
Peter Gostev
35:41
Yeah.
35:42
Yeah, they're so good. for me as a personal developer, that's absolutely amazing. I use them all the time, right? I put my credit card and they, their business models, they charge a bit of fee on top, and then they, they let me use the models, and which is great. The, the problem is that no company would use that, right? if you have worked for a corporate Like it's just not gonna fly at all, right? There's not enough, guardrails for me to not accidentally send my data somewhere. And the downside, the flip side of having all of the providers for everything is that structurally you are-- you don't know where the data is going, right? So oh, it automatically flips you to the most available provider. it's like now your data is there, right? So it just… it's not gonna work for like a professional enterprise. So hopefully they're gonna build it out to be more enterprise ready, then it will be amazing, right? Hopefully they're not gonna lose the advantages that we love them for, but I think that's the path, right? Otherwise they'll be just stuck with us putting $100 each time. it's hard to build like a big business out of that.
Alex Volkov
36:44
I think that w- we should watch out for the Stripe streaming
36:49
payments thing that they innovated, the Stripe sessions together with OpenRouter. I think there's gonna be like a big thing coming out of there. Not to mention, the, the connections OpenRouter already has with a bunch of enterprises on the other side, enterprises providing them inference, services as well. so shout out to OpenRouter, like we'll hear more about this. the thing that's most important is, OpenRouter keeps operating under its own brand. It's not like a brand acquisition thing. the team is staying on, the whole team as far as I saw, which is also great for, sales. And then Stripe is gonna bring that enterprise-y experience 'cause many enterprises, if not all, they use Stripe, for various reasons. So that's great to see. Also, a huge reminder, Stripe innovated with the Stripe Wallet thing, with Link Wallet. We told you guys about this where my agent bought me a anniversary… sorry, not anniversary, like wedding gift. and I… all I had to do is approve a purchase, and I think that is very important as well in the world of like agentic things. so I'm a- absolutely looking forward to, to see how OpenRouter improves for all of us. folks, let's talk about open source and the only model this week that we had because it's not a lot, it is very exciting. let me switch. I want to use one transition.
Nisten
38:00
Open source AI, let's get it started
Alex Volkov
38:06
We are at the Open Source Corner here in, ThursdAI and, Chilwick.
38:10
I will say Qwen dropped the Qwen 3.8 27 billion parameters. We talked to you about this a little bit on the show last week, but, I don't know. tons of people waited for this model for local AI inference specifically. Why? Because all of the other open weights, open source models we talked to you about, it's, impossible to run them unless you have huge machines. Especially now, lately, the Kimi and the Qwen, they moved over two trillion parameters, two point six I think, and two point eight. they're nearing three trillion parameters. That's not something that people run at their home, and even if they do, the, the electricity bill does not make sense, I think, compared to what you get with GPT 5.6 Luna, which is essentially free now because OpenAI really wants you to use their models. but for local AI people, it's very exciting when smaller models run because, Peter can run this on his, ninety-six gigabyte, Linux machine if there's a GPU connected to it. Wolfram can run this on two 3080s, 3090s with, good token streaming. If the model's good enough, there's a lot of stuff that it can do. And so I think for that reason, folks got super excited about Qwen three point eight twenty-seven B. With that said, though, folks, from your experience, have you tried, twenty-seven billion parameter Qwen three point eight, and, what are you seeing on your timelines? What are you seeing of people, like, receiving it?
Wolfram Ravenwolf
39:30
Nisten, let me quickly just put together the Unsloth
39:33
desktop, the new app they made, LM Studio alternative, basically. when I saw it available in there, I immediately downloaded it. They even have a one-bit quant that only takes eight gigabyte of RAM, VRAM, so it can run on the smallest machines and is still strong. In my own benchmarks, it was… I really have to look into this and do more benchmarks because I did only one run, but it put it above Kimi K two point six, which is a huge model and- Wow the performance was amazing. So this is my recommendation right now. If you want to run local AI on a normal system, I would pick this. In German, it's usually it's not the best, so in other languages probably also not the best. But, I think I will be looking into more, routing stuff, like routing some tests or using some Zap agents to run this locally. Yeah, it's definitely a Sonnet level at home, I would say, from what I've seen so
Alex Volkov
40:26
far.
40:26
Nisten, what about you? You've been running this model a little bit?
Nisten
40:29
I've been benchmarking it all week and using it.
40:33
okay, so the community response was pretty good at first, but now it's gone crazy. There are a hundred and fifty-two fine tunes- off this. some interesting things that I noticed while benchmarking is that compared to the older Qwen 3.6, it actually dropped a little bit in the other benchmarks, like medical or, human eval, a, a bunch of the regular ones. But in agentic ability, this is on another level. I don't think it's too much benchmarked in this
Alex Volkov
41:10
release.
Nisten
41:10
as I've been running it at home and, even friends with an
41:13
M4 MacBook with 24 gigs of RAM are running it, MacBook Pro, M4 Pro. they're still getting like 12 tokens per second at, at 4-bit and they're just running it in LM Studio. I have to say the, the agentic ability of this is actually crazy. This thing will just about do anything and it will also control, Claude and Sol in other terminal sessions whenever it needs something smarter and it is able to do that. yeah, this is probably-- And especially some of the, the spicier fine tunes, like I needed to look through to make like another Linux security patch. I was able to use one of the unrestricted ones to talk to Claude whenever it need to and then keep working through the problem and it was actually able to keep working through the problem. So to me, this, this release is pretty crazy. at first I was not as excited 'cause I noticed a drop in the regular benchmarks, but then when I actually tried it to, to just run it at home, this is completely nuts.
Alex Volkov
42:26
I wanna talk about the fact that 27B is great for local
42:28
stuff, but also for the fine-tuning community and the unlocking community. we haven't heard this word fine-tuning 'cause again, nobody's fine-tuning Kimi K3 2.6 trillion parameters. It makes no sense. besides, besides the Kimi team in RL and I think that's like a great thing. Also, I want to shout out again local.ai folks. This is their like new thing. if you wanna know which is the highest intelligence model that can run on your M4 Mac Mini, or M4 Max laptop, a Qwen 3.8 27B is very close up there, in terms of the top intelligence. And I wanna shout out the one person that says, "Oh my God, it's so unsafe and so scary. Qwen 27B parameter unlocked can look up torrents." Just not everything that's written on the internet should be taken with reverence a-and truth. That person just, went all out, and the example he showed of how scary and unaligned this model is, that he asked it for a torrent, and it looked up a torrent, whereas Codex, refuses to look up torrents for you. I was like, "Bro, have you heard of Google? what are you talking about?" It was really funny to me. Peter, how is this model getting received in Arena at all, and, is it performing well? What's your thoughts on this model specifically after around a week of it being out?
Peter Gostev
43:41
So we, we have it in testing.
43:44
we still need to validate some results, so I don't wanna, say where it's ranking. but it's looking pretty good, and I think, there's a-- There was a moment, I, I don't know if you remember this. it feels like a year ago, there was just a, a time when we just didn't have any, models around that size. it felt like we had a bunch of good models around that size, and then there was, like, maybe Gemma was still appearing once in a while. And then it felt like there were just not particularly any good models. I think Mistral was, like, not releasing anything around that size. So it was, felt a bit weird, 'cause I, I think people kinda set up their hardware to operate around that sort of level. And when we started getting the 700B models, it's "Okay, thanks. that's not no help at all." and they were quan-quantizing and so on. So it's kinda good to see. And, yeah, I don't have that in-depth view yet. but yeah, it looks to be, like, a little bit potentially special. I don't wanna say too much just 'cause we wanna validate it first. But I think it, it's looking quite good, and I think it's interesting. the, the thing about the smaller models that never sat right with me is that- D-- I can see using a small model locally if it was maybe dumb but very reliable. So if it was like, if I'm telling you to run a terminal command, like it will do it and like it knows them and it can run them. If it just did that, then it's absolutely, like it's actually really useful. But I think the trend we had so far with the models is that they were just unreliable and okay, maybe they're scoring more on these benchmarks, but you just, why would you use them if they're not reliable? Then I think you gravitate towards more reliable models. So I think if we do get to the point that the agentic capability is that good, as Nisten was saying, then it's like almost I don't really care about other benchmarks. Like as long as it actually has this core of it actually doing the job that you're asking it to do, then that's awesome.
Alex Volkov
45:39
There's also the thing where if somebody calls me out every time where
45:42
I post about my Fable not working well. He's like, "Bro, local models don't get degraded unless you switch the models yourself." If you run the model and in a year you run the same model, those are the same exact weights unless you like change your inference stack, et cetera. So that's-- there's also a benefit there. yeah, we're covering Qwen, 27B, the frontier on your desk. Have you been running this at all? have something to say?
Yam Peleg
46:06
Since the moment it was out, it's-- I totally, completely agree
46:11
with Peter, but I think it's quite reliable in my, I don't know, in my, in my experience, Qwen is quite reliable. It's not the smartest, for sure. Like it's not Claude. it's not a, a GPT 4 point, s- 5.6 Sol obviously- It's only
Wolfram Ravenwolf
46:27
27B.
Yam Peleg
46:29
yeah.
46:29
But I don't know, it can drive a computer pretty well. unless you really push it to do like crazy stuff on browser maybe. Like I don't know, it can drive computer pretty well
46:41
in my experience though. Don't you think, guys?
Alex Volkov
46:43
I have, now that you're talking about this,
46:45
you're saying it's not Claude. What immediately popped into my head is that it's not Claude Fable and maybe not Claude Opus 4.8, 4 point-- there's no 4.9. Opus 4.8 and 5, but it could be Claude 3. like this 27B model could be Claude 3, the model that we got all excited about two years ago. I don't remember when Claude 3 came out. And I was like, "Hey, should I create Time Machine Bench? should I benchmark new open source models versus the, the models that we talked about a year ago and seeing like where we are in terms of performance?" I think that'd be cool 'cause like we are all getting Hedonistically adapted, if you guys familiar with this concept? Like we all get new things, and then we get "Oh, Fable 5 is not answering my emails correctly, and this model is like i- insane." But yeah, when you s- you triggered me when you said it's not Open- it's not Claude, because Claude has been like a thing for a while now, and this model that can fully run locally and be for me, and be trained and fine-tuned with my stuff and locally execute what I need can absolutely be the Claude a year ago, which is still incredible to us. Like the, the intelligent jumps that we get is incredible to us. I need to go and work on it. Yeah.
Yam Peleg
47:58
I just want to react to something that you said.
Alex Volkov
48:00
Nice.
Yam Peleg
48:01
Look, Time Machine Bench is a good name.
48:04
But I just want to say that it's not only the-- Of course, models get better. Yeah, 100%. Everyone understands that. But you also sometimes lose stuff. I think that Claude 3, yeah, it has no way, shape, or form has any capability of, Claude 4, 4.1, 4.5, 4.6, all of them. Absolutely not. But it, it-- I think that Claude 3, maybe 3.5, 3.7, they did have something, that is quite, w- that we're quite losing today. And, it's, it is very, in my opinion, I have a pinned, isolated environment, with the Claude 4 point, Claude code, like, from a year ago and 4.6. And 4.6, I, I run it sometimes and it's completely diff- It's not that, I got adapted to the new stuff and now I think, we are everything is degraded. It's that they are quite different. things are moving not-- not all axes are moving into the improvement, direction, okay, at the same time. And you really see it, exactly like people are saying today, Fable is com- absolutely capable, but bro, it's not easy to understand what Fable wants from me, when we speak from time to time. And I never had this problem, let's say, with, 4, even 4.6 or 4.8. It also speaks really nice. that's my take on the-
Alex Volkov
49:36
Yeah
Yam Peleg
49:36
at the end of the day, not your weights, not your model.
49:39
if, if you run Qwen, it's never gonna change. You can also customize it. It's brilliant. Thank you very much for it, Quentin. Nothing more to say. That's a brilliant model.
Alex Volkov
49:49
We have, folks commenting.
49:50
And folks, we appreciate comments. and we will put you up on stage if you comment to us as well, saying that, "In my usage, seem closer to Claude Opus 4.5, which is incredible just for a 27B parameter model. way smaller context window." Yeah, that's true, but that depends on how much you can hold it. there is a link here from Nisten. Anything you wanna say to, to-
Nisten
50:08
I just wanna say there have been 10, close to 10
50:11
million downloads of the quants. There are 650 quantizations and, there were issues with the thinking levels of the original just being a little bit too long. People quickly fixed them and, it just-- this just crossed the threshold where it can drive other agents for me when it needs to. and it has very good, visual ability to read stuff But this crossed the threshold where it is a local model, I can just have it run at home, can have a look at the screen, and I can have it drive other agents only when it needs to. And that's a big, it just, that's a big threshold to cross for me in this case. And, yeah. Yeah, we're gonna… I have a bad feeling they might just ban it because it's just too good. I'm just gonna leave it at that as a prediction for the future.
Alex Volkov
51:02
All right.
51:03
folks, we have breaking news. and the breakers of news, the founder of the company that broke the news is joining us, because I just reached out to them. let's go. AI breaking news coming at you only on ThursdAI.
51:24
This is the best kind of breaking news, where I reach out to the person. It's like, "Hey, you wanna hop on?" He's like, "Yeah, I have a very important meeting in 30 minutes, but I have some time now." Jeff Huber, founder of Chroma, the agentic memory and search and I don't know how to describe it. You guys just launched something. I would love to hear directly from the proverbial ho- horse's mouth, what you guys just launched. Please, tell us what you guys just, announced.
Jeff Huber
51:46
You actually did know something about it, because, three or
51:49
four months ago, I interviewed you- That's true … about your agent sessions, and I said, "Hey, Alex, do you have valuable agent sessions that have a lot of useful memories in them, do you think? Are you doing anything with those?" And it was super helpful, so thank you. I was reviewing those notes, actually, this morning. yeah, I think broadly we think that memory is the largest unsolved problem in AI, and, there's probably a lot of really advanced approaches to continual learning. We're super excited about those as well. But a base case, agents need the ability to write things down in a highly organized and efficient manner, and a consistent manner, and then read that stuff back later. And, once you dig into that problem, there's all kinds of things that you end up wanting around versioning, access control, lineage, concurrency control, and more. and fundamentally at Chroma, what we want to do is solve these, base infrastructure layer data problems around AI. we've been doing that now for three or four years. They sometimes call Chroma the context company of California. Love it. and saw a lot of our customers struggling to build these sort of memory level primitives, and so wanted to offer something. this, what you see here on the website is, our research preview of this technology. It's useful if you're a solo developer, you're a team, and you want to build kind of a shared memory system for yourself and for your team. we'll also be offering Foundation as infrastructure. so you can build it into any agent you have, in, a single prompt.
Alex Volkov
53:02
And the reason why I think it's cool is because, dude,
53:05
I tried all of the other solutions. I've tried building one myself. Every person that I talk to that runs agents like ending up building some sort of a, a solution for themselves and their agents, especially as their agents like are, as they try multiple ones. And this is the same problem that I run into all the time. I run things through multiple agents. I ran OpenClaw- Yeah … and Hermes and now GroqBots, et cetera. And there's like, stuff in Codex and Codex are on a different Mac, et cetera. Like all of this gets really annoying really fast when I talk to one and then I completely forget who I talked to and talk to another, and I expect the same kind of outcome, same results. is this some of the stuff that you are trying-- you guys trying to fix this with like real-time sync?
Jeff Huber
53:43
Yeah, exactly.
53:44
on day one, if you scroll down, we're ingesting sources natively from Codex, Clawd Code, Cursor and Slack. we've got coming soon connectors for Notion, GitHub, Google Drive, Granola and more. And then, in a single click on the onboarding, you've wired up that knowledge base all the way through back to your coding agents. We actually don't do MCP by default for the coding agents because we find that they use CLI tools and hooks more often than they use MCP. It's like the more reliable way for them to get to use the tool at the right time, obviously. But we do have MCP support, so you can plug in foundation anywhere you have an agent, and that's what the goal is, to be that like neutral, bring your own harness memory layer, that allows, again, teams, individuals, developers, but also, your own product in the future to, build systems that get better over time and that just learn, how to, learn how to do their jobs.
Alex Volkov
54:30
First of all, we recorded a couple episodes live from your offices,
54:33
so shout out- Yeah … and huge thank you, guys, for hosting us as well on Thursd AI. And I remember our conversations there, at Wolfram, because Wolfram, you were also there, I re- I believe. w- when I needed to ping or to, to build To test like an idea or I remember, I think it was like BM25 and Toby's, QMD, et cetera. Like y- you are the guys who I talk to when I need to know, what is the, state of the art in terms of a- agentic search, et cetera. have you guys worked some of that into this foundation? Tell me about like the behind the scenes, like how does this work? Totally. Is it-- Yeah.
Jeff Huber
55:04
Yeah, I think the key bottleneck that we identified for
55:06
building kind of these LLM self-improving wikis fundamentally is you have to be very good at agentic search. that is obvious on the query path or on the read path, being good at agentic search, but it's actually equally, if not more important on the write path. what happens with a lot of these LLM wikis today is they quickly become a huge pile of slop. you're not updating information that you need to be updating. You're not deleting information you need to be deleting. You're not putting information in the right spot where it can be found later. And so actually, agentic search is even more important on the write path. So you know, we think of ourselves at Chroma as probably the world's experts on agentic search, and if that is a key bottleneck for building very good self-improving wikis, then we thought that, we would be a good candidate to solve that problem best in class in the world. The second thing that I'll say is it improves both its knowledge, it also improves its own, system prompt. So each foundation actually manages its own system prompt. and when you give it natural language, high level feedback, it can improve its own instruction set at the system prompt level as well.
Alex Volkov
56:01
It can improve instruction set on system level.
Jeff Huber
56:03
almost like harness engineering, right?
Alex Volkov
56:04
Oh,
Jeff Huber
56:04
also has the ability to update and edit its own system prompt.
56:08
so it is improving itself, from its experience of… So for example, our system prompt for our team, it's decided to associate all these different, Slack channels and the Slack IDs so it doesn't have to look them up again, and you realize, "Oh, this is helpful. Let's write this down." but then also as you give it feedback, either in Slack or through other, any, any basic, any feedback mechanism back to the system as a human, the whole system learns and reacts to your human feedback as well. So it both learns, and it also meta learns, is what I'm trying to say.
Alex Volkov
56:35
Oh, that's dope.
56:36
tell us about context one. I don't think we've talked about context one on the show. Could you, talk to us a little bit about the-
Jeff Huber
56:40
Yeah, in spring we released a state-of-the-art agentic search model.
56:42
it's a finetune of GPT-o SS 20B, state-of-the-art on agentic search. So it's trained to know how long to search, how, where to search, and it does better than frontier models, but it does so at, twenty-five X the, cost reduction and ten X the speed improvement. So it runs somewhere around, four hundred tokens per second, and it's, twenty-five times cheaper than Opus.
Alex Volkov
57:01
That's incredible.
57:02
we def- I believe we mentioned it, but definitely didn't have you to talk to us about this.
Jeff Huber
57:06
I have very little minutes
Alex Volkov
57:06
left Yeah, so I have, one, maybe one last question for you.
57:09
tell us about this.
Jeff Huber
57:10
Yeah.
Alex Volkov
57:11
What is the operating procedure here?
57:12
Like, how do people use this? Can they host completely their own wiki completely on their own, like many open source community folks like? Or is this like a new entry into the cloud services area from Chroma, the company?
Jeff Huber
57:23
Yeah, it's a great question.
57:24
out of the box, it's end-to-end, kind of this, we put all the pipes together for you to get a great out of the box experience of this research preview of this more fundamental infrastructure technology. We haven't yet done the work to open source every piece of this. Frankly, it's a lot of pipes, and so it's nothing that fancy, if I'm honest. I think, like, when you think about the infrastructure components here that go into this, there's like an agentic harness, there's some unique models, there's a unique database. Again, as I said in the launch sort of Twitter thread, we could not build this without ChromaDB. ChromaDB is the only database on Earth that can support this workload shape, that can support this use case in a sane way. and so yeah, we're gonna continue to open source more and more. we've always been huge open source believers from day one. if you go dig around our GitHub, you will not yet find the Mac app that you download, open source just 'cause, we've been so busy just getting this thing stood up,
Alex Volkov
58:06
So I'm, really looking forward to testing this
58:09
out because I do have this problem. I think many other people have this problem and also, I'm looking forward to maybe Grok Bot has taken over the airwaves recently- Yeah … as much as like other folks. I like, I'm really looking forward to see if this can be- it's getting Grok-ed. Let's see what happens. Yeah. I love the, the concept of this like build as infrastructure. So we'll definitely keep an eye on this. Congrats on the release. I know you have to go, but thank you so much for jumping on, Jeff, and, we'll tell you about feedback if we see it as well for folks. Thank you. Tell me everything you love and everything you hate. Jeff Huber, founder of Chroma. Okay … and we'll see you guys when you continue building on this. All right folks, this has been breaking news. But before this we need to talk about GLM 5.3 real quick. the folks from ZAI released, GLM 5.3. anybody already try it by the way? It's not open source. the weights have not been open weighted, but they did, drop some, some exciting updates. Anybody has any info or have tried ZAI's GLM or not yet?
Yam Peleg
59:03
I think we're all waiting for it to be, maybe released, maybe re-hosted.
59:08
I think that's everyone's preference at the moment, but I'm sure it's pretty good.
Alex Volkov
59:15
I think that, the, the weights is like the important part,
59:18
but let's at least take a look at what we're about to expect with GLM because I think it's super cool.
Nisten
59:22
they reached out to me, but then I was like, "Oh, but
59:25
the weights are not there," so-
Alex Volkov
59:27
two days ago, the API is now live.
59:30
I will say this has been… Like I will say, this has been a shift in how the m- the open, previously open weights, but now leading Chinese companies that have open weights are operating, and we have been noticing on the show, go a long way from, "Hey, here's the Torrent link of our weights from Mistral," which was like the best move in the world ever, towards, "Hey, we're gonna release this in API. I think it's important because people sending them feedback, et cetera. then we're gonna release a, a weights model, but we're gonna release this with this and this clause. It's not MIT anymore, it's not open source. And it's been sad to see all these like model providers moving away from full open source and embracing community, but still benefiting from the excitement that they got initially from this community by releasing open, completely open weights in great license. So- Look,
Yam Peleg
1:00:21
I just want to say like credit where credit is due,
1:00:24
like they do release the weights. at, to this point, maybe in a slight delay, but they do release the weights, I think. okay, so maybe a little bit late, but
Alex Volkov
1:00:34
But I feel like this is a specific move towards we're
1:00:37
becoming a frontier lab and weights, weight-weights of the bigger models are not gonna get released. l-licensing is choking it, the price floors and everything. like it's been… At that point, I'm sitting there wondering, why would I send my data to their API, or why would I even try an open weights model it's anyway cannot be hosted on my laptop and it runs in the cloud somewhere, so why would I not, use Terra or Sol? what is the incentive there if the weights are not, I don't own them, and the licensing is such that, they have to impose their, specific things on different companies. So it's what I'm trying to say, the open weights Apache 2 license that shout out to DeepSeek still does is what I consider open source open weights. when they release a technical, paper that describes exactly how their attention mechanism works and that helps everyone, that's incredible, and that's our conversation that we had with Ilija Ba-Bakic, that, this also, really helps. so shout out to, GLM. This is, priced the same as GLM 5.2, so pricing is the same. they don't-- They didn't release a lot of, performance updates, so this is the only thing that I can see. The GLM 5.3 Max is, cheaper based on same intelligence than K3 and 5.0-- 5.6L, sorry. any last comments, LDJ?
LDJ
1:01:52
Yeah, I was going to say, I think, i-if it really is at the level of
1:01:56
roughly Kimi K3, as artificial analysis seems to indicate, then I think that does end up being really impressive from the standpoint that it's what? It's about seven hundred billion parameters- and the Kimi K3 is, it's about Two point five or two point six. Yeah. So Kimi K3 is about triple the total parameters and about triple the active parameters. So yeah, I'd say this is a really good improvement here,
Alex Volkov
1:02:21
for
LDJ
1:02:21
sure.
Alex Volkov
1:02:21
So I have found the evals real quick and, yeah, the evals look
1:02:25
quite insane in terms of just the jumps. So terminal bench jump, From 4.6 to 28.3 on Terminal Bench 3, which is, quite the jump. On DeepSuite, there's almost a 20% jump. This seems like a very, very impressive jump for the same infrastructure, and the same price as well. so we will wait until this releases in, in, in weights, to play with this. All right, folks, I think it's time for us to move on. A little bit late, but I would love to welcome to the show Francesco Bonacci. Welcome, Francesco. Hey there, folks. Hey, welcome to the show. Francesco, I've been in conversations with you, for quite a while now, as we're-- w- we need to talk about some stuff that you guys launched, and specifically I mentioned, Try CUA, the startup. back when OpenAI released computer use, and computer use, specifically background computer use, which got, me super excited, where I ha- can sit here and talk to you guys. Meanwhile, my agent can work independently of me on a window in on my Mac. and then you guys shifted with Try CUA, w- in open source. I specifically remember somebody has to do this work in open source because I think it's very important that all agents can do this and not only, proprietary stuff. Obviously, OpenAI bought the Software Inc. company that had, very good, Mac engineers, and this is the result of theirs. And then this week you guys also released something, also kinda in breaking news. So I would love to hear from you directly, first of all, who you are and what Try CUA does, and secondly, what you guys released this week. And let's talk about it.
Francesco
1:03:52
yeah, so happy to take it from there.
1:03:54
Francesco Bonacci here. Nice to meet you again. I'm the founder of, CUA. before, jumping on, on board, with CUA, I used to work at Microsoft, for, four or five years. and then I started my own company. That was, like, about one year ago. it was very deep in the space of, computer use, even way earlier. I think, Anthropic came up with the terms, twenty twenty-three-ish, four.
Alex Volkov
1:04:15
Yeah.
1:04:15
What does CUA stand for? Tell us.
Francesco
1:04:16
CUA stands for, computer using agents.
1:04:18
That's as simple as that. and we were like in the space, as I like to phrase and like to give a term to everything, like, when it comes like to Qwa, I like to define two different phases. Like Qwa is like Qwa 1.0 where you have an agent taking control over a desktop, not necessarily like in the background, more, more like in a headless kind of way. You have a server where a Qwa driver component and then, it take it from there. You can expose it as an MCP, as an MPI, an API, a CLI, whatever. And then we were doing our own experiment last year. Okay, this probably is not the best paradigma, for human-computer interaction. Like I need to dedicate an entire sandbox for one only agent. it's quite unoptimal in there. So we thought about it for a while. We wanted to ship something like very similar to the agent-browser CLI that sell as a… so again, I'm a phen- I'm a big fan of Peter. I'm very bullish on CLI, Peter from OpenCloud, So we wanted to ship something, like that. So we had like like the primitives like already in place. And then, Ari from, from Sky/like Open- OpenAI now. I'm also like a big fan of them and their work. So they prove what, we were also like working in the background, that background computer use was possible, starting from macOS.
Alex Volkov
1:05:38
I have to ask you, how the hell does background
1:05:40
computer use working in macOS. And how Mac, which is the more closed o- operating system out of the three, is the first one that background computer w- use works in versus the other ones that like it still doesn't. Just tell me. This is crazy.
Francesco
1:05:54
Yeah.
1:05:55
there is a lot of wizardly in place in there. And by wizardly I mean that, so the secret sauce for like having something, Being controlled in the background was already there. And, there is something like called, accessibility tree. It's basically the tree representation, like HTML, DOM-like for, any na- native application. And, Apple, throughout the years, they actually invested a lot, in, accessibility, just like for making like application like more accessible to the user, especially like Catalyst, native apps on macOS. They have like very rich, accessibility trees, like application like Notes, Calculator, and so other like system apps. So the whole operating system is like very queryable, and, and can be navigated by an agent but when it comes to, application that does- don't necessarily expose an accessibility tree, so we refer as those as, application being, having, a sparse tree. Look at, I don't know, Slack for instance, that's an Electron app. look at Blender. if you try and query the accessibility tree on this kind of application, you will see that you can only query, the outer shell of the app. And then, everything that is inside is actually,
Alex Volkov
1:06:58
It's a
Francesco
1:06:59
website … it's a canva.
1:07:00
It's a canva that it's actually, not queryable. So the only, way there to actually, do something is relying on vision and, multimodal models. They've been, like, very good at grounding. And, by grounding, the ability of, a model to figure out, like, where it should click, pixel-wise. So the, the trick there is actually for macOS and, just, cut short there, is to rely on a primer click. like tricky the window into thinking that is an, this is an active window, while it still is not. So the, the actual application will realize that, okay, there is a synthetic, event, and not necessarily I have to steal focus, on, on the current app.
Alex Volkov
1:07:39
Dude, focus stealing I think is the most annoying point.
1:07:42
And, we gotta call out the big labs as well and their computer use. On your website, which I also would love to talk about, there is a leaderboard of, who runs computer use. Yeah. You call it Qwen Bench, leaderboard, right? if we show this right now on stage, folks will see the Claude Fable 5 has an asterisk, but, it… This is, the full pass. You have a very difficult benchmark looks like. six out of 25 top, things Fable 5, does. I have to talk to you about this. Yeah. But you have this thing. However, their harness, though, is awful, and Claude takes over my whole computer when it needs to click buttons. I do not appreciate this because I also need to work. And we talked at the beginning of the show with Peter that, he built a whole Linux box for their computers to work. And, recently I got excited. I have to talk to you about this as well. Dude, I have so many questions. about Grok Bot that everybody's getting excited about. Yeah. That they have their own computer use, which absolutely sucks. I send their team back feedback where, I see… I logged into their computer. I see them typing every word by every letter just by typing. It's like, "Oh, no." Yeah. So what it does is, it types a letter, takes a screenshot, sees, "Oh, yeah, I typed a letter. I need to do the next letter." It's the fuck is that? so- I have a lot of stuff about computer use that I personally don't like. But yeah, the, the harness of Claude takes over. background is the way. Talk to me about the state of the art in computer use. You guys have a leaderboard. Other, companies are releasing OS world and, other, leaderboards. We always look and, pay attention to them.
Nisten
1:09:04
Yep.
Alex Volkov
1:09:04
Where are we now?
1:09:05
Can the agents use computer, fairly easily? Are there still major things that they cannot do from your experience? what is the gap that needs to be closed for that full, complete- Yeah automated user in a, in an environment that does clicks for me?
Francesco
1:09:19
I've been in this space for about three years now.
1:09:21
I do remember, the very first, we were calling it in chat on Microsoft like Navy Agent, basically using GPT-4, V at the time, and GPT-4V wasn't, even good at, grounding, so you had to rely on accessibility trees, all the way. So the first results on, on OS word and Windows Agent Arena, it's the counterpart for, for Windows devices, was, like,40%. and that was, like, three years ago. And, it's been mind-blowing, especially in the last year, like, seeing, thresholds, like, if, the human baseline for, for, using a computer is, 72% on OS word, V1 We are actually now at about 80%, so OS word is like very well overfitted, by now. been like the, the facto benchmark, up until now to… That's being like used by labs. So we do have our own benchmark anyway. we work also with labs. And, we made like this benchmark specifically focused on PCAT because that's one of the application that is like more, like one of the to- toughest application,
Alex Volkov
1:10:17
Really?
Francesco
1:10:18
computer using agents.
1:10:19
Because you have to rely on, not necessarily on like clicking, typing, scrolling. That kind of That, that is solved. that's my answer to you, like-
Alex Volkov
1:10:26
Yeah.
1:10:27
We're clicking, we're solving. Okay.
Francesco
1:10:28
But when it comes to multi, multi pointer and like events
1:10:33
like dragging, even scrolling, like it's not quite there sometimes. Because every single operating system has a different way of scrolling, like natural scrolling for macOS and, Windows, like that, that still confuse the agent. we're not quite there, and you can see it from our benchmark, like when it comes to, solve rate, you still s- you still have Fable, like only solving six of our 25 task. there is still like room for improvement. Probably we'll see like frontier models like reaching 10 to 12 solve task on our benchmark in, the next couple of, months.
Alex Volkov
1:11:03
we're definitely looking forward for agents doing,
1:11:05
meaningful work completely, and that includes, using a computer. you saw the release of GrokBot as much as we all, saw the release of GrokBot. This is, the, the highlight there is- agents that have their own environment, their own computer. All they have in their computer is a terminal , a browser and, a- and a file system, which I absolutely love. this is all you need, essentially. and as we said, they need to get better at computers. Maybe I should reach out to those folks and connect you somehow so they learn from you guys. Oh, yeah. would be, like, incredible. but this week you guys released something as well, and I really want you to present this. can you talk about the releases this week from you guys?
Francesco
1:11:40
Yeah, totally.
1:11:41
So we shipped something called Computer History. it's like our open source take on Codex Computer History. we've been-
Alex Volkov
1:11:48
Which is also fairly new, right?
1:11:50
The, Codex- Fairly new … had Chronicle in beta for a while, which is- Yeah a thing that takes screenshots of your Mac to know the context of what everything you worked for. And then there is Computer History. would love for you to cover both and then your take on this in terms of, open sourcing.
Francesco
1:12:02
Yeah.
1:12:02
So we have been debating, for a while, whether to release, memory for computer-using agents or-- We, we do have, our own trajectory recording mechanism. think about, a human that wants to do, a mundane task or, a knowledge worker and, there are, there, there are ways with CodeDriver that you can just, record a trajectory and then, turn it into a skill and, replay it, whenever you want. It's basically like you can think a- about it like a banker of, successful trajectories, that's living on a, on an encrypted key store on your device. And, we released it for, for all the three major operating system. So when we make like a major release like this one, we just take the lesson learned, like we have a cross-platform harness, Rust-based. So we just release all the three major platforms. it also help you think like broader. Okay, is it even like possible, overall, on Windows or on Linux? so again, it's like a bank for successful trajectories. like the closest like similar, l- the, the very first take on this was like maybe Microsoft like two years ago, they tried to release something similar, probably you recall like Windows Recall.
Alex Volkov
1:13:01
Yes.
1:13:01
I remember us talking about this. I think Nisten and I talked about this, that they recalled Windows Recall.
Francesco
1:13:06
Yeah.
Alex Volkov
1:13:06
Because it was like not opted out.
1:13:08
It was like, yeah, it was Yeah.
Francesco
1:13:11
Yeah, I was at Microsoft back, back in the, like this
1:13:13
time, so that wasn't like very well perceived by the community. So again, we were been debating for a while whether to release something similar. But, the difference to Recall, is that we don't record like any screenshots. Like we, we release this, feature like with privacy in mind. We only record successful trajectories and, accessibility trees. we don't record like any, any written text. So it's different than- like Memories because like we're just like recording what Core Driver knows, not necessarily what like an agent should know. And, we don't like, we didn't do any benchmark rounds. I've been dogfooding like this feature for the last, couple of weeks now, and, it's been like mind-blowing honestly on, especially like on knowledge worker task or like even, me doing-
Alex Volkov
1:13:56
okay, so if I understand this, I will relay
1:13:59
this to the audience as well. You have an update to Core Driver, which is the open source driver that like clicks things, especially in the background, that stores the successful ones, and then this helps the next iteration to do stuff? tell me like what blew your mind in like storing those trajectories? what does it help with? Yeah.
Francesco
1:14:16
Like a shortest path problem.
1:14:18
using calculator for, doing something within an app or just, I think, one of the demos that we released was, like, this one specifically was, like, about painter. you're looking at, querying the whole state of an application and just, understanding how the accessibility tree looks like, where, pixel-wise, a particular, component is. So you're probably, If you're, like, doing over and over again, like the same task, you will find, the shortest path, very early on. We call driver- with, history on versus, history off. It's like trivia implementation-wise. We give the agents, an ability and, two primitives to access, and query these, this key store, and by doing that, it's, as simple as that.
Alex Volkov
1:14:55
we talked about OpenAI Hackin- Hackin' Face, and one of the
1:14:58
outcomes there when they released it was like, hey, these agents found out they can collaborate, and by collaborating, they in- effectively created this memory, and this memory caused the thing where, like, when o- one agent opened the door, the others, didn't have to open this door, they opened the next ones. It's kinda- Yeah … this reminds me of that. This is essentially like, if one agent already knew where to click inside Paint, so the other agents rely on memory versus going and trying it, and going and trying it is like 17 tool calls versus the one. Yeah. Right? Is that the idea?
Francesco
1:15:26
Or even, trivially, how do I even launch an app?
1:15:28
there are a bunch of, different identifier, maybe I want to launch Slack Web over Slack desktop. okay, just, save this, successful trajectory, and I'm just gonna take it from there.
Alex Volkov
1:15:36
This is awesome.
1:15:37
congrats on the launch. I think it's very impressive that you guys fully open source and, tools can use, y- your tooling. I really appreciate, dude. Every time somebody who comes and does open source, we have applause. So yeah, all of us applauding to you. we have to continue to our next conversation very soon, and also we- we know you're a busy dude. but feel free to come back to us when you guys release some stuff. very excited, and I've been, like, a follower of yours, a fan of Tricoa for a while, since that first browser use… Or sorry, computer use release that you guys released. so Francesco, thank you so much. Nisten has one question, and then we're gonna, wrap it up and, move on. Nisten, go ahead.
Nisten
1:16:10
Can I use this with any type of Linux?
1:16:13
if you have some stuff that's, very sandbox or very secure, or is there, a particular desktop environment that it needs? Or w- what advice do you have? what, what works best, especially just, for Linux desktops?
Alex Volkov
1:16:24
What creates the best computer use environment?
Francesco
1:16:27
Yeah.
Alex Volkov
1:16:28
That's a great question.
Francesco
1:16:28
the most mature background computer use on, on Linux at this
1:16:32
stage is still, X11 over Wayland. we've been, like, many ask, "Hey, you guys, will you support, Wayland?" The story there is that, there's not, even a queryable accessibility tree, and there is, no coordinate system on Wayland. we're thinking about, actually, starting an FTC on the next version of Wayland and see, okay, this operating system actually needs to be, like, more, like- Needs to, it needs to account, like some low level primitives just to make like the ground compute use possible. So probably we'll get there eventually.
Alex Volkov
1:17:00
Have you,
Nisten
1:17:01
Okay.
1:17:01
So X11, that's good … everyone That's what I use, actually use Xlib or GL and wrote my own. okay. And then either XFCE or GNOME, it, that doesn't matter Yeah. So the, the, yeah, it doesn't- I'm gonna, I'm gonna, I'm gonna use this. That's why.
Alex Volkov
1:17:13
Built in, into the operating system.
1:17:15
All right. Francesco Bonacci from Kua. Thank you so much for joining us. Congrats on the release of History. I'm c- can't wait to try this as well. Is it already integrated into Hermes as well? like it's coming with the latest driver, right?
Francesco
1:17:26
We're talking with the team, so it's kinda like
1:17:28
for free already will be nice.
Alex Volkov
1:17:28
Awesome.
1:17:29
And congrats again on releasing this in open source. Thank you so much for joining us, man.
Francesco
1:17:33
Thank you.
Alex Volkov
1:17:34
All righty.
1:17:34
folks, we move on. Before we have our next guest, I do wanna say hi to Bin Liu from the Hagian team. Hey, Bin, how are you? Welcome to the stage on Thursd AI. thank you for hopping on. I only reached out to you yesterday, but we've been talking for a while- Yeah … and I really wanted to cover, HyperFans for a while. Wolfram and I have to talk to you about our sponsor, Core, CoreWeave with Weights & Biases. All right, welcome to this week's Buzz, where you can see by our orange jackets that we are, I'm interested at this incredible company called CoreWeave. recently been very much, on, on thoughts of minds on the news with, with different, quarter releases, et cetera, but we're also, doing a bunch of stuff, that I think that you guys should know and be excited about. Wolfram, how about you take this one?
Wolfram Ravenwolf
1:18:15
CoreWeave has partnered with MasterClass- Ooh … which is an
1:18:19
online learning platform that feature individual AI tutors that help you learn. They have the, the real tutors are, of course, prominent figures like Cooking with Gordon Ramsay is one of the courses, Negotiation with Chris Voss, Writing with Shan- Shonda Rhimes, and they have AI tutors. And to evaluate how these tutors are doing and help improve them, they are using our platform, and they are running the agents on CoreWeave Cloud and use Weights & Biases Weave to evaluate, monitor, and improve the AI teaching agents. So basically, the combination that, CoreWeave offers is the hardware, the infrastructure, and the software, the solutions to monitor and improve the whole thing.
Alex Volkov
1:19:03
That's incredible.
1:19:03
Shout out to folks who work with MasterClass. MasterClass is generally incredible. I did not know about their AI agent things and, AI tutoring one-on-one, so that's great.
Wolfram Ravenwolf
1:19:11
And they also have a MasterClass executive, class, basically,
1:19:15
where they are teaching about business users, and there's also a hands-on AI lab where they teach how to evaluate systems for production readiness using our tools. So you can… the other thing, we have been talking about this on the show before, and it is coming ever closer, is our conference. Let me put it up as well.
Alex Volkov
1:19:37
we passed one billion runs on the Weights & Biases, platform
1:19:41
The Weights & Biases- A billion have a billion runs- Wow since, for the past, I think, nine years. Sham Lewis, the co-founder, the first run was nine years ago. and we passed one billion runs on the Weights & Biases, machine learning exploitation tune track. And so shout out to the whole team for this, very strong effort. Everybody who uses Weights & Biases and knows us because of it, it helped build a lot of models. Early successes was like OpenAI, Toyota Research, Uber. Like, all those folks built their foundation on Weights & Biases, which is incredible to see, just way before it, it joined CoreWeave. All right, let's talk about Fully Connected and move on. Fully Connected, folks, are you gonna join us in September 29th to October 1 in San Francisco? We have the code for you, but also we've announced the speaker list. Wolfram, you wanna talk about the notable folks that we have on there?
Wolfram Ravenwolf
1:20:22
Yeah, sure.
1:20:22
There are people from our company, our CEO, Mike Infrater. There, our CTO, Peter Zielinski. There's, the VP and GM from NVIDIA Hyperscale and HPC, Junyang Du. more people from us, some you may even know from, AI engineer speaking there. And speaking of AI engineer, we also use the same location that AI Engineer World's Fair was using. So where is it?
Alex Volkov
1:20:43
Moscone South.
1:20:44
Yeah. almost the same one. It's kinda on, it's around the block, but we're gonna have a live show from there, folks. And if you wanna join us and see the recording, see Fei-Fei, Dr. Fei-Fei Li, y- you can see that, th- this prices are quite high. But if you are a Thursd AI, live show tuning, person, use this code. Use this code. I'm not gonna say it out loud, but, people have to go hunt for this, but, use this code to get in and secure a spot. Again, this is very adjacent to OpenAI DevDay. If you're going there, please come and also check out, Fully Connected. We have Dr. Fei-Fei Li from WAR Labs. They're announcing soon something super incredible. And also we have Sara Guo from, Conviction Conviction one of the top VCs, in AI. So that's incredible. In addition to a bunch of folks from CoreWeave, please join us on September 29th. All right, folks, this has been enough shilling. Let's move. I wanna talk about Hyperframes. welcome, Vin Liu, to the show. Please introduce yourself. would love to hear from you, who you are and what you're working on, and then we can talk about, Hyperframes.
Bin Liu
1:21:41
Thank you, Alex, so much for having me and, having us on the show.
1:21:44
we've also, absolutely enjoyed, a lot of your content, and, very honored to be here. I'm Vin. I, am a VP Eng at HeyGen. I lead all of our agent, and product teams and engineering teams here at HeyGen. I'm one of the co-creators for Hyperframes and, I work, with an incredibly, small but, smart team that, built out, an open source type of frames. we've been, pushing out, updates and progress, over the last three, four months. we obviously crossed, 40,000, GitHub stars, just a few weeks ago. yeah, it's been awesome. as for me, my background, I've always been in the creative space. I spent, the first 10 years of my career at, Pinterest. I led, product and engineering, on the content team there. after that, I started my own company, having always been, building in the creative agent space. As soon as I saw ChatGPT, about three or four years ago, I was like, "That was the moment." and now here I am. joined HeyGen, as part of our, acquisition and then, really excited to be building, creative agents and creative, agent technology, here at HeyGen.
Alex Volkov
1:22:44
So thank you so much for joining.
1:22:45
we mentioned HeyGen a long time ago when Joshua, who I met with the engineer. Yeah … Joshua went super, stupid viral, was like, "Hey, I'm Joshua," and this wasn't him. Yeah. This was way before, the models caught up, and this is, very exciting to see. And since then, we saw this rise of the AI influencer type thingies, and a lot of companies are trying to do the AI avatar that talks like you and voice cloning, et cetera. And it feels like HeyGen, at least for me, was like one of the, places where I would go to see the frontier of, like, how that's happening.
Example from screen
1:23:15
Absolutely.
Alex Volkov
1:23:15
But then at some point, I started seeing this, Hyperframes thing.
1:23:19
And for folks who are just watching, just so you'll know, what we're talking about, here's a transition that I built with Hyperframes that I would like, to show on stage.
1:23:36
All our-
Bin Liu
1:23:36
Wow
… Alex Volkov
1:23:37
all our strikes here on the show were rewritten, from scratch.
1:23:41
And I used to, 'cause I, I produce the show, I used to do them, directly in CapCut manually by myself, and I still have, the, the other ones. these ones were done by my agent, using, Hyperframe. Ben, how about you, you introduce Hyperframes as a product? What it is, why does this exist is the most important. How does this relate to the fact that you have agentic, company? Would love to have the connection. Absolutely. And then we can talk about, like, how people can actually use this.
Bin Liu
1:24:03
I'm very excited.
1:24:04
And, wow, that was amazing, A- Alex.
Alex Volkov
1:24:06
The best one is, I think the frontier is the best one.
1:24:08
Let me show you this one
1:24:19
It's just like I have a bunch of, stripes here for the show, and, obviously I rebuilt them from scratch with Hyperframes. But yeah, tell us about
Bin Liu
1:24:25
Hyperframes and what it is all about.
1:24:26
Okay. I have a bunch of demos to show as well. Yeah, please. And, I actually also made you guys a surprise, and so- Oh, let's go. We'll see how that, how you guys like it. but but yeah, I wanted to actually take a quick step back so that people have, some context. HeyGen a- as a company, we obviously, started by being an A- the one of the best, if not the best AI avatar, right? because at the core of HeyGen's mission, we're not trying to compete with, like Veil 3, CDentz. we're not trying to go to, disrupt Hollywood, drama TV. what… The one core thing that we believe is that, communication, through video is one of the most effective ways. However, for most of the people, they don't feel comfortable sitting in front of a camera. Yeah. and so we started there because we noticed that even just showing up in front of camera for a lot of people is incredibly difficult, especially for introverts. myself, it wasn't that easy to just, show up and, look at myself.
Alex Volkov
1:25:24
And I appreciate it.
1:25:26
Thank you for coming up.
Bin Liu
1:25:27
Yeah.
1:25:27
And that's why the company started building the AI avatar technology because, the human-to-human interface is important. human needs to show up, but just it's so hard to build a camera crew, so hard to be eloquent in front of camera. But with our AI technology, with our AI avatars, you can show up without needing the camera. you just need to direct your own, Yeah, exactly, your own digital avatar, and they will do the talking for you. And but if you look at these video… If you look at these videos, just the AI avatar is not enough.
Alex Volkov
1:25:58
Yeah.
Bin Liu
1:25:59
You need the motion graphics.
1:26:01
You need the editing. You need all of those things to actually make your, communication effective. but making those things is incredibly hard. you need to use things like After Effects, for motion graphics. You need to learn CapCut or Premiere Cut, to do that. And we are- Dude,
Alex Volkov
1:26:20
That part for me takes way much of a toll than
1:26:23
just talking, yapping to a camera. this we can do all day. we sometimes go over three hours and just yapping. Yeah. Following up, stopping at key frames, saying, "Oh, I said this thing. This needs to show up," et cetera. we have a whole team now that, that works on our shorts, and, this is, the bulk of the work of the pr- post-production thing.
Bin Liu
1:26:38
And w- when I joined HeyGen, that was the problem that I, wanted
1:26:42
to solve, because I also feel very, blocked whenever I need to do editing. I find Premiere Cut, and even CapCut, to be very hard to use for someone like me, who's not trained, at, to be a video editor. And I will skip a lot of the exploration and iteration, but we landed on, why don't we turn this into a coding problem? We're all engineers.
Example from screen
1:27:05
Yeah.
Bin Liu
1:27:05
Can we make video editing a coding problem?
1:27:09
Because LLMs are, AI agents are so good at coding, and that took us to Hyperframes. So Hyperframes, what it really is that it turns HTML pages into a video. Yeah. And so when I ask my AI agent to edit my video, what it really is trying to do is take a raw footage, for instance, and then build everything around the video, cut the video using code.
Alex Volkov
1:27:37
Feel free to share your s- screen.
Bin Liu
1:27:39
I will just go directly to my Codex.
1:27:43
I'll just show my whole screen. I'm sure we can do editing later.
Alex Volkov
1:27:47
Yeah, we'll do that.
Bin Liu
1:27:48
Let's do it.
Alex Volkov
1:27:50
So meanwhile while you stop, I'll just say that, th- there's
1:27:52
two things that get unlocked w- via this. One of them is, first of all, the ability to do this programmatically- … which is already itself crazy because, we just had Francesco on with computer use.
Bin Liu
1:28:03
Yeah.
Alex Volkov
1:28:03
CapCut is incredibly complex and different motion graphics,
1:28:06
those apps are incredibly complex. So asking your agent to run stuff on them is not, super, easy. But second of all, just having an agent, like, knowing about this, I think is great.
Bin Liu
1:28:16
Yeah, we just downloaded one of his cl- clips, and then
1:28:19
I basically did editing for him. so the input video, is, let me see if I have the raw input video.
Alex Volkov
1:28:26
And also walk us through what we're seeing here,
1:28:28
because this is also a podcast. A lot of folks who cannot, watch, they need to hear- Absolutely what is going on with this.
Bin Liu
1:28:34
so this is input video.
1:28:35
You can see that the input video here- doesn't have any motion graphics, doesn't have any, but what you need when you actually publish this on YouTube, you need all of that. when you ask your AI agent to do the entire cutting, what, Hyperframes, can do for you is, essentially adding every single element. So let's just play it right now.
Example from screen
1:28:55
A good ad when it's running appropriately, I
1:28:57
put a dollar in and $5 comes out. Every dollar of ad spend I put in, it turns into $5 of revenue or customer lifetime value. When you look at e-commerce, this is literally the game. This is how- e-commerce companies build their existence. Trying to identify a winning ad, you do what I call A through Z testing. Instead of A/B one ad versus another ad, I'm doing 100 different ads simultaneously, and then I'm just gonna look at the data
Bin Liu
1:29:21
So every single one of these, motion graphics, the fact that this,
1:29:25
avatar is showing here, and the fact that, there is, text that- When it's running
Example from screen
1:29:28
appropriately,
Bin Liu
1:29:29
I put a dollar in, five dollars out
Alex Volkov
1:29:29
Yeah
… Bin Liu
1:29:30
all of these is done by Codex.
1:29:34
and the way Codex does it is going through, Hyperframes, skill, one of the talking head, cutting skill. And if you look at the code here, I know, none of us need to understand this code, but the entire video is actually constructed through this code. But really, we turned the task of editing a video into a coding task through, the framework that we built that we call Hyperframes.
Alex Volkov
1:29:57
Which is open source, by the way, right?
1:29:58
folks can go and take a look as well. Hyperframes, I believe, is, you guys chose to put it out there and, Yes. For real … tell us maybe about that decision, 'cause I found that also incredible and ev- everything- … that's open source is getting, big applause on the show because we absolutely love open source and stuff.
Bin Liu
1:30:12
I think that the key thing here is really for this to
1:30:15
be, adopted by all the agents. because at the end of the day, we believe that this technology is, something like more of a fundamental technology. But, even just, allowing agents to learn that, oh, I can do video editing by doing coding, that is the fundamental piece that by open sourcing it, we want, all agents, all the frontier labs, to learn and do.
Alex Volkov
1:30:40
I think, a few things stand out for me as, as somebody who used
1:30:43
Hyperframes, obviously for the show. Dude, the opener for the show, the 10 minutes. Yeah. I was like, "Dude, 10 minutes is a very long-ass time to edit a video for 10 minutes," especially with motion graphics. You have to do a lot of looping or repeating. I don't have time to do this. So our opener here, has a bunch of stuff from our show. It has, a timeline. It has, a bunch of things. It's like, okay, people wait until we ramp up, until we start. Might as well, educate them. Yeah. This is all agent. which leads me to my next question.
Bin Liu
1:31:09
Yes.
Alex Volkov
1:31:09
Do you guys have any kinda benchmarks, evals,
1:31:12
leaderboards, et cetera? how do you know, and at this point been, which is the best agent to create agentic videos? Would love to hear from you directly from your experience. Absolutely … and yeah, w- would love to hear, like, how you guys come
Bin Liu
1:31:23
I'd love to show you, and let me actually see if I can pull it up.
1:31:26
a benchmark across all of the state-of-the-art, models on how they perform, how each and every one of them perform, against, the, hyperframes,
Alex Volkov
1:31:34
obviously, there's Design Arena, right?
1:31:36
And Design Arena now does, websites. It, they now do video models, et cetera. Yeah. This is a mix of both because it requires motion graphics, and not all websites necessarily need motion graphics. So I haven't seen, a very good representation of an agent understanding something. The other thing that I will tell you is agents can read HTML, and they can take screenshots, but, not a lot of models can view video. And MuseSpark, some people actually commented earlier on, on the show that MuseSpark 1.2 is really good at- Yeah viewing video, and I feel like that's also missing. Milos called it out. "MuseSpark 1.2-" really good for video analysis. Been feeding it recordings from my Clarity, works great." I feel like when an agent generates a video for me, if you only take screenshots to try to understand whether or not it did well, that's not enough. I need it to, consume the video because, a real world editor will look at the thing and then see, oh, the timing here is not quite right, the sound doesn't match, et cetera. Like, all of these things, they can't come from HTML, cannot come from screenshots. Exactly. would love to hear from you, like, where the state of the art of agentic video generation is.
Bin Liu
1:32:36
let me, I can show one small thing.
1:32:39
I think this is not necessarily the best one, but let me m- maybe quickly show it.
Alex Volkov
1:32:43
Yep.
Bin Liu
1:32:43
And then I can talk a litt- a little bit about the, the exact question
1:32:46
that you were, you were, getting at. so we actually test across, and we're actually working closely with DeepMind. we will be releasing a, a bench- a benchmark, very soon. But you can see that for a cheap model like GPT-3 Flash, which is, I think 10 cents per million tokens, when it's given an instruction to do a motion graphic, it did- it didn't even complete it. But when you give it to GPT 5.5, this is what it did, right? It looks a lot better, and we can just take a couple look, and you can see that the difference between A cheap model and the state-of-the-art model.
Alex Volkov
1:33:20
So Gemini 3 Flash on the left here- Yeah … barely does anything
1:33:23
in interesting, and Ge- what is it? GPT-5 on the right is, doing, zooming and animations, et cetera. And Claude Opus, obviously, everything looks good and looks- Exactly … virtually spaced, et cetera.
Bin Liu
1:33:33
And, the other question that you had, which, is not yet, fully
1:33:36
done through our, current benchmark, but, is that we are, building our own harness that really solves the agent or AI not understanding video problem. Because, it's actually very true, and I actually wrote about it, in a X article, where VLM, so we're like, even the best, Fable 5, they're not trained to watch a video for an hour and be like, "Oh, these are the highlights. These are the things. what is this motion like, moves like?" That's not it- what it is trained for. It's trained to, to say what's on the screen.
LDJ
1:34:13
Yeah.
Bin Liu
1:34:13
What is this video about?
1:34:15
It's talking about the what, not talking about the when and how.
LDJ
1:34:19
Yeah.
Bin Liu
1:34:20
So you actually need a good harness to be built around it to teach
1:34:25
it the, the, the when and how, and that's actually what we are building. E- even the state-of-the-art, to be very honest, the state-of-the-art video input, ex- that models that accept video input, none of them actually do, the, the video understanding super well. Yeah. And so that's why we're building a, a harness around it. let me show one more example. I'll just show the entire screen, for, simplicity. one of our posts, I did this video, but this video is a, let me just quickly play it right here.
Alex Volkov
1:34:53
Oh, yeah.
1:34:53
Now we can hear it.
Bin Liu
1:35:00
So this video was made by asking our agent to watch one of the
1:35:04
very viral video that this person made. It was made originally through
1:35:15
You can see that they look like but the reason why we are able to like almost like pixel by pixel, replicate this using hyperframes, is that we have a harness that really understands what's going on-… inside of here. And that actually takes me to, showing you guys a little, hopefully fun surprise that I turned- All right this video. I was showing this project earlier, but let's go to the other project I turned this video
Alex Volkov
1:35:50
Ooh
Bin Liu
1:35:56
Can you guys hear
Alex Volkov
1:35:57
it looks like a ThursdAI themed-
Bin Liu
1:35:59
That's right.
Alex Volkov
1:36:00
right rebuild of the video that you just showed to us.
Bin Liu
1:36:03
I- That
Alex Volkov
1:36:04
can I get that project?
1:36:04
I wanna integrate it. Absolutely.
Bin Liu
1:36:05
I'll share the project right after, but here's the rendered video.
Alex Volkov
1:36:09
Oh, nice.
1:36:09
Let's go
Bin Liu
1:36:12
For some reason there's still no sound.
1:36:14
I'll send over the project, but, this is, you know-
Alex Volkov
1:36:17
Oh, this is dope
… Bin Liu
1:36:18
this is ThursdAI.
1:36:19
but, but with this, whole new motion graphics. And that's the beauty of, using code as your, video editor because, as a software engineer-
Alex Volkov
1:36:27
would say like-
Bin Liu
1:36:28
Go ahead
Alex Volkov
1:36:29
what I think that you unlocked is the fact that,
1:36:30
everybody can do this now.
Bin Liu
1:36:32
Yeah.
Alex Volkov
1:36:32
Whereas, motion graphics is really difficult to time, and it's
1:36:36
really annoying to, to learn all these apps and the time- timeline concept and the key frames and, there's a lot. it's really a lot of, cognitive load, and now everybody can basically do this vi- via the stuff that you worked. And so I was very happy to represent you guys here on stage and tell you that I absolutely loved, using Hy- HydroFrames with my agents, and I am looking forward to, significantly better results. And also the feedback that I have for you is that it needs to start looking less like webpages. I know … I don't know how, WebGL or whatever, but, we need to start moving towards the actual motion graphics that, you know- Absolutely Peter, how about, you f- you, help us land this plane because we're almost at the end of our two hours as well.
Peter Gostev
1:37:12
I really like this 'cause, I've been trying to do it natively,
1:37:15
and I think the way it works is, does, yeah, probably creates HTML, does for front PAG and, like, all of that stuff. And the question I have for you is, like-
Alex Volkov
1:37:24
Yes
… Peter Gostev
1:37:24
I know people use the word taste, everywhere for code.
1:37:27
it is code essentially, right? as you say. Like, how do you see it here? is this your opinion about how it works or how do you guide it? Because honestly, my experience, and I think GPT 5 circa is, way better than GPT 55, for example, but still, still a little bit, it need to push it a bit and, Like, how do you think about that? Is it, does it come from your side or is it, temporary and models will just get better?
Bin Liu
1:37:51
Yeah, absolutely.
1:37:51
So yeah, thank you. Thank you for showing, our… we do, have built a ton of skills, teaching, agents how to, produce, let's call it tasteful, outputs. and that's one of the reasons why we, chose HTML as the core of our, technology, because, just like you said, the reason why Kimi 3 or, GPT 5.6 or Fable, they are starting to get better and better at making these videos, is that they're also, people talking about, "Oh, they can make amazing landing pages. that's why we are trying to build and, likely launch our own harness, which, takes a lot of the, the anti-pa- patterns, out of, agents coding and so that a lot of the taste from your own websites, from your own product can be actually expressed. A lot of the good motion graphics pr- practices, can be also used. our open source skills already support that, while I think that it, a lot of that is also in the harness, while, Claw Code and Codex might not necessarily do a great job yet.
Alex Volkov
1:38:51
Yeah.
1:38:52
Bin, this has been awesome. I'm very happy that we had an opportunity to bring you on the show and talk about this. congrats on really an amazing open source product. Obviously, it feeds into HeyGen and the stuff that you need to do, like motion graphics on top. But I think what it unlocked is, I don't know if you quite followed the Thursd AI episode where I talked about this. Every night, I took the pictures of that day and sent it to Claude with Hyperframes to create, a recap of that day. And I don't wanna show this because I don't show my kids here on camera, et cetera. But I will send you directly if you wanna, take a look at the outcome of this. Oh, yeah. This will be in their, memories forever. every day we reviewed it afterwards. It was incredible. and I learned a lot while doing this, but also, I learned that, agents are not yet great at this. it needs to improve significantly. It needs to stop looking like HTML pages, 'cause some-- there's, rough edges there with the SVG graphics, et cetera, because video is, a very rich format. so I have, tons of stuff to, give you feedback about, but thank you so much for coming up. you've been a great sport, and we're very excited to hear and see where this evolves next.
Bin Liu
1:39:46
Alex, thank you, and Peter, thank you guys for having
1:39:48
me and having us on the show. I have one last offer. I'd love to help you guys build a, a Hyperframes, brand system, with us. And- I would love that … going forward, hopefully, a lot of your video editing and a lot of your intro, clips can be made, by just asking your agent to do it, with Hyperframes and our harness.
Alex Volkov
1:40:06
I would love that … let's
Bin Liu
1:40:07
do that.
Alex Volkov
1:40:07
We have a GroqBot agent that takes notes for us
1:40:09
as well and to-dos probably. and, definitely sign us up for that. All right. Vin, thank you so much for joining. Thank you. folks, we have just a little bit more to cover on the show, before we wrap up, and the stuff that I do wanna cover kind of relates to what we just talked about because, it's not only motion graphics, it's also audio. And so when I try to generate a bunch of videos with, Hyperframes from HeyGen, I had to, put on some music. they have some integration with some music generations. this week, Happy Shrimp was released by Alibaba. Anybody see Happy Shrimp? Happy Shrimp is end-to-end music generation, full song from emotion or story or prompt. Happy Shrimp 1.0. But I think there's, Music Gen. wait. Last week, Minimax released a Music Gen as well, right? we didn't talk about this.
Nisten
1:40:48
Yeah.
Alex Volkov
1:40:49
Okay.
1:40:49
let's at least play a, a, something of a, of an example. if anybody can pull this up faster than me, feel free. But not, if not, I will pull up faster. let's see. The, y- you're saying the music generation blog, yeah? It's Minimax
Nisten
1:41:06
Music 3.
1:41:06
Yeah, it's just Minimax Music 3.
Alex Volkov
1:41:08
So this was, like, last week, a little bit after our show, so
1:41:10
we didn't really play it for you guys.
Nisten
1:41:12
It's the, on the trending on Hugging Face this week, it's
1:41:16
the third most trending thing.
Alex Volkov
1:41:19
Oh, yeah, I found one example-
… Nisten
1:41:20
it's probably-
Wolfram Ravenwolf
1:41:20
X … probably good Let me see if I can play it.
1:41:21
They have the really worst license of all, excluding all of America and Europe and UK and so on, but nobody cares.
Alex Volkov
1:41:27
you guys hear this?
Example from screen
1:41:30
Open weights, open graph, open
Alex Volkov
1:41:32
doors.
1:41:32
Yeah? That's good.
Example from screen
1:41:34
Minimax Music 3 on the beat, and you can run it, too.
Alex Volkov
1:41:39
This is a, Rob from ComfyUI, the creator of ComfyUI, I
1:41:42
believe- ComfyUI now or one of the guys
Example from screen
1:41:49
there- yeah … that says that we can run this music
1:41:49
generation model One box, one pipe. First switch, close your eyes and just sit there. Don't like it, reboot.
Alex Volkov
1:41:51
DJ Nisten, knows a thing or two about music.
1:41:55
so this is what Minimax. So Alibaba released, their one. It's called the, I love this name, Happy Shrimp. Happy Shrimp is a call-out to some folks about AI and, and the benevolence or whatever, right? we're all, placating shrimp. I don't know where it comes from, but it's a meme. And-
Yam Peleg
1:42:12
Yeah, the shrimp, shrimp rights and so
Alex Volkov
1:42:14
Yeah.
1:42:15
All right, so this is Happy Shrimp with the girl. Let the shrimp go
1:42:26
This is very, K- K-Pop or, yeah. Okay, I'm, this is what they released. Do you guys hear any of this or not at all? 'Cause I wasn't sure. Yeah, okay, cool. Yes … and then we also have Cartesia- Yeah … Sonic 3.6, which is now number one on both artificial analysis and other ones. I don't know, Peter, if you guys do any Please tell me if you do voice. I keep forgetting.
Peter Gostev
1:42:48
No, we don't.
Alex Volkov
1:42:49
No.
Peter Gostev
1:42:50
Yeah.
Alex Volkov
1:42:50
You don't do voice.
Peter Gostev
1:42:50
might do.
Alex Volkov
1:42:51
But Cartesia Sonic 3.6 now runs the top of TTS leaderboards.
1:42:55
We had Cartesia on the show, obviously, a couple of times. Great folks. I will just, again, call out that the fact that this show is now, getting transcribed in live session by our chief of staff, ThursdAI Live Transcription and, we can see everything we talked about and the highlights and the, the agent now tells me like, "Hey, you talked about this, you didn't talk about this," et cetera. Audio8 TTS and S1 Mini. Oh, yeah, Super Whisper. Shout out to Neil and, Nisten, your friend. They released a S1 Mini first open weights model, .6 billion parameters. It cleans transcripts on device. have you guys had a chance to use this or hear about the S- S1 Mini?
Nisten
1:43:31
Yeah, I already implemented it because I have stuff, on my phone
1:43:35
that I run the agents and stuff in and, anytime you use like Turbo Whisper, or even Parakeet, it's just gonna write like random text. But, this one's only .6B. It doesn't really hurt the performance, and I find it quite usable now for just voice texting my friends Because often it would just, mess up or s- or screw up grammar or look weird. It just makes it look nice. So it just change the, changes the text after it comes out of Whisper or Parakeet.
Alex Volkov
1:44:10
Yeah
Nisten
1:44:11
but it does make a big difference, whereas before I was a bit nervous to
1:44:15
just, message my friends on, via, yeah, via Whisper, Wi- Whisper Large because I'd often have to edit stuff out, and this one j- just corrects the text. it just makes it nice. it wo- it works pretty well. I highly recommend people just, I don't know, just tell their agent to, to add it in.
Alex Volkov
1:44:32
That's the easiest thing.
1:44:33
Tell your agent to add it in. Yeah … a few things we didn't cover this week, and, Wolfram, you sent them, so let's at least mention them as obviously GroqBot, it feels like it's exploding. It feels, I start feeling the same tingly things- Oh, please … where- Please.
Nisten
1:44:46
I had people in Europe that I never thought would like Groq at all.
1:44:52
They tried GroqBot-… and they are just blown the heck away. They have, entire teams, running stuff, doing stuff for them. They use it to manage, side businesses.
Alex Volkov
1:45:02
We-
Nisten
1:45:03
People love this thing.
1:45:05
This thing works really well.
Alex Volkov
1:45:06
there is a certain feeling of tugging that we have when something starts
1:45:11
to feel like it's going to change things. GroqBot, to me, feels like the early days of OpenClaw, where we told you about this, where before it used to be called OpenClaw. The thing just works and works right now and listens to us, and, I didn't have to do a lot of stuff to configure it. It's just, the same lessons that, OpenClaw did. Here's what I'll say about this. in addition to like it having an old computer, et cetera, we will keep telling you about like new use cases as well. Obviously, there's the whole thing where Elon Musk is pushing this like crazy, but also like it has real uses. I think that people need to step away from like hating on Grok as a concept towards, hey, there's something there from the folks at Cursor that just by the sticker price of sixty billion dollars that Elon Musk put on it, it has to be called Grok, but it doesn't have to be called Grok. Like it's a whole new, wrapping, and the agents are talking about each other. And so what I wanted to add to this is that, my chief of staff that I built with Grokbot was listening to the show using a different bot that was listening and summarizing everything, and asked it, "Hey, we're about to land this plane. What important things did we didn't cover yet?" And said, "Hey, one real leftover if you have twenty seconds, Clawd code/design." Clawd code recently added, slash design inside Clawd code, and now the Clawd design thing that we told you about is appearing inside Clawd code. You can like interact with it. it also is not perfect because of the lens. it skipped two things It's so good now that other things are trying to catch up to GroqBot. Wolfram, so you brought those two, so please feel free to cover the two things that we need to mention. Wolfram, you wanna cover this?
Wolfram Ravenwolf
1:46:38
yeah, the interesting thing about this is an UI thing for the
1:46:41
desktop app where you have the view of agents as different profiles, basically. so Hermes agent had this feature before where you could just have multiple instances running within the same gateway, and now they added the same UI so you can see not just the usual interface where you have the chat sessions and every chat is with the same agent, now you have different agents, like the chats, and each agent is a individual. You can @mention it. So they added that to the Hermes desktop app now.
Alex Volkov
1:47:08
Because I think, for folks who don't know, GroqBot is…
1:47:10
the unique feature of GroqBot is that the side conversations are not sessions like most of the other apps. They're bots. Everyone, is a unique bot with their system prompt, with their memory, with, They also have, a shared memory, but, they have a unique memory, and that's been, like, crazy for some people, but some people it, unlocked a bunch of use cases. So it looks like Hermes is really quickly following up on this, which is great to see. It kinda looks like that also, but it's because it's Hermes, you can use more than Groq 4.6. You can use, other models as well.
Wolfram Ravenwolf
1:47:38
And I think this is a transition phase we are seeing.
1:47:41
Now that the agent is a cool thing, we want them in the interface. But eventually you want to have multiple sessions with each of the agents, so it needs to be something else, and I think it will be heading more towards something like Slack, where you have the different channels, agents you can mention and convert them.
Alex Volkov
1:47:57
Yep.
1:47:57
and the other thing is, OpenBot also was created by the folks at, you-
Wolfram Ravenwolf
1:48:04
Copilot Kit
Alex Volkov
1:48:05
Yeah, Copilot Kit.
1:48:05
Our friends at Copilot Kit, of course, I blanked out on the name after three hours almost on air, also released a, a, their attempt at, this pattern, but for open source, which I think is very important. not everybody can afford the $250, $300 ultra tier for Groq to be able to use GroqBot. But here's my suggestion. If you are considering two subscriptions, Fable and OpenAI, et cetera, I would suggest that you at least try this out to see if there's something in this paradigm, because that's, that feels like the new paradigm is coming. ComputerUse, works in their model. and we need to land this plane as you saw my light just went out. I think we covered everything, folks. This has been a chill-ish week, but still we had breaking news. Jeff Huber from Chroma came here to talk to us about Foundation, which is their new layer of agentic memory and search. Really encourage you to try that out. We had, Bin Liu from HeyGen talk about hyperframes, which is, by the way, I think that Nous Research does a bunch of their videos with hyperframes as well, and theirs looks very cool. And, obviously we had, Francesco, from the Trykua team, to talk to us about ComputerUse and background ComputerUse and what are the best agents at ComputerUse, in addition to a bunch of news from this week. So great week overall. Wolfram, you wanna finish up?
Wolfram Ravenwolf
1:49:19
The good thing about a chill week is that, it gives
1:49:21
us time to take a little more time to look at each individual thing that interests us, and I would extend that invitation to our audience as well. Pick something we talked about that, raised your interest, try it out, and at the beginning of the next episode, why don't you tell us how it worked out? That would be also interesting, right? What changed for your workflow? it would be very interesting to hear the same from the audience. 100%. What did you take out of the last session and, have to say in the new one.
Alex Volkov
1:49:47
thank you so much folks for joining.
1:49:49
Wolfram Ravenwolf, Peter Gostev, Yam Peleg, LDJ, and Nisten, thank you so much for co-hosting the show, folks. Your, your host is Alex Volkov, and if you missed any part of the show is getting turned, still manually, agents are really bad at it still, but yours truly, into a podcast and a newsletter, at the end of each Thursday. So if you missed any part of the show, or you want the links that we talked about or the links to the profiles of the people we covered, everything is in there. please feel free and encouraged to also, by this request to leave us five-star reviews wherever you do listen to the podcast. If it's, YouTube, please subscribe and hit this bell button. It really helps. if it's, Apple, five-star review is really helpful. This is the best way for you to support us across platform. It really helps us to reach new folks and teach them and bring good guests, which is, I think, is one of the coolest part of the show. we'll see you next week what will be our last show of the summer next week, August 27th. All right, folks, thank you so much for joining. And to end the stream, here is a video of non Hyperframes generated video of previous, my efforts. But as you see, there is a difference. All right, folks. Bye-bye everyone. See you next week.