Episode Summary
ThursdAI goes LIVE from AI Engineer Europe in London with five back-to-back interviews — Swyx on harness engineering and why MCP is now 'less interesting because it's stable,' Peter Gostev on Anthropic's Claude Mythos ($25/$125 per M tokens, 77% on SWE-bench Pro, 'too dangerous to release'), VB from OpenAI on Codex hitting 3M weekly active users with plugins and Guardian Approvals, Vincent Koc on OpenClaw's 1.5M-line plugin rearchitecture and the /dreaming feature, and Omar Sanseviero on Gemma 4 crossing 10M downloads. Plus Meta's Muse Spark drops mid-show, Seedance 2.0 lands stateside, and Anthropic's ARR jumps from $19B to $30B in two months.
In This Episode
- 🎤 Swyx — AI Engineer, Harness Engineering & Skills
- 📰 TL;DR — Weekly News Roundup
- 🏢 Peter Gostev — Mythos, Arena Data & Compute Wars
- ⚡ This Week's Buzz
- 🛠️ VB (OpenAI) — Codex, Plugins & Guardian Approvals
- 🤖 Vincent Koc — OpenClaw, Dreaming & 1.5M Lines of Code
- 🔓 Omar Sanseviero — Gemma 4, 10M Downloads & Google DeepMind
- 🎬 Wrap-Up
Hosts & Guests
By The Numbers
SWE-bench Pro
77%
Claude Mythos — up from 53% on Opus 4.6, a 24-point jump
Per M tokens
$25 / $125
Claude Mythos pricing — ~5x Opus 4.6
Anthropic ARR
$30B
Up from $19B in February — ~$10B added in two months
Codex weekly active users
3M
Up from 2M last month, per VB (OpenAI)
Gemma 4 downloads
10M+
Plus 1,000+ Gemma-4-based fine-tunes on Hugging Face
OpenClaw codebase
1.5M lines
Including unreleased iOS + Android native apps; GitHub PR counter maxed at '5K+'
🔥 Breaking During The Show
Anthropic announces Claude Mythos — 'too dangerous to release'
77% on SWE-bench Pro (+24 pts over Opus 4.6), zero-days in every major OS/browser, sandbox escape, $25/$125 per M tokens. Released only to ~40 partner companies under Project Glasswing. Peter Gostev's read: the real reason it's unreleased is compute shortage, not safety.
Meta ships Muse Spark — first model from Meta Superintelligence Labs
MSL's debut model, natively multimodal with multi-agent Contemplating mode. Simon Willison's deep dive uncovered 16 hidden tools inside the meta.ai chat UI.
Anthropic ARR jumps $19B → $30B in two months
Announced alongside Mythos — 1,000+ customers paying $1M+/month, secondary tender sale completed with employees declining to sell ahead of IPO.
Seedance 2.0 hits the US on Replicate
Jumped ~80 ELO points above the next video model on Arena — a massive gap where most models cluster within 10 points. Up to 9 reference images, 3 videos, 3 audio files per generation.
🎤 Swyx — AI Engineer, Harness Engineering & Skills
Opening live from AI Engineer Europe, Swyx runs down the tracks he curated this year: coding agents, MCP ('settled and therefore less interesting'), generative media split from voice/vision, and harness engineering. He warns engineers to 'vendor everything' after the light LLM and Axios supply-chain compromises — pip fork instead of pip install.
- Harness engineering: big labs are investing more, not less — it's not going away
- Skills are absorbing harness work — English as the ultimate programming interface
- Supply-chain security: vendor everything, 'pip fork' not 'pip install'
- MCP has beaten OpenAPI, tRPC, gRPC as the integration protocol
Swyx
"The big labs have shown that they're investing more and more in harness. No, I don't think it'll go away."
Swyx
"All these light LLM and Axios security issues make me think you should just vendor everything by doing a slop fork of everything. So basically anything you want to import, you don't pip install — you say pip fork."
Swyx
"MCP has been enormously successful and just, it's just now more settled and stable — and therefore less interesting."
📰 TL;DR — Weekly News Roundup
The usual fast-lap through the week: Mythos, Muse Spark, GLM-5.1, Gemma 4 at 10M, Seedance 2.0 hitting the US, HappyHorse mystery video model, the surreal Mila Jovovich 'MemPalace' saga, OpenClaw 2026.4.5 with /dreaming, Anthropic Managed Agents, and a W&B Automations launch. Peter Gostev rides along.
- Claude Mythos announced — too dangerous to release publicly
- Anthropic $30B ARR (up from $19B in Feb), secondary tender sale completed
- Meta Muse Spark — first model from Meta Superintelligence Labs
- GLM-5.1: #1 open-source on SWE-Bench Pro at 58.4%
- Seedance 2.0 live on Replicate — ~80 ELO points above the next video model on Arena
- OpenClaw 2026.4.5 ships /dreaming (REM/Light/Deep memory consolidation)
Peter Gostev
"The jump on Seedance — I think it's like 80 points higher, which is massive. Previously all the models were kind of clustered together, maybe 10 points apart. And this one is 80 points higher."
🏢 Peter Gostev — Mythos, Arena Data & Compute Wars
Peter Gostev from Arena joins to break down Claude Mythos. The numbers: 77% on SWE-bench Pro (up from 53%), $25/$125 per M tokens, released only to ~40 partner companies under Project Glasswing. His framing: the 'too dangerous' framing is mostly cover — Anthropic likely just doesn't have the compute to serve it. He walks through his updated Compute Wars chart showing Anthropic catching up via the new Google TPU deal, and announces Arena has released 3 years of historical leaderboard data and actual prompts as Hugging Face datasets.
- Mythos: 77% SWE-bench Pro, 64% HLE, +10pts on browser comp — 10T parameter rumor
- $25 / $125 per M tokens — ~5x Opus 4.6
- Real reason it's unreleased: compute shortage, not safety
- Anthropic ARR: $19B (Feb) → $30B (Apr), secondary tender sale completed pre-IPO
- Google TPU deal helps Anthropic catch up — weird that Google is propping up a DeepMind competitor
- Arena released 3 years of historical leaderboard + prompt datasets on Hugging Face
Peter Gostev
"I think it sounds cooler to say that it's too risky to release it rather than we just don't have the compute."
Peter Gostev
"Before OpenAI, you'd get a blog post saying 'we have an AI breakthrough, we will enjoy it to ourselves, you get nothing.' OpenAI shifted that — you submit a credit card, you get an API key, you get access to frontier intelligence. Now we might be getting back to the point where unless you're Microsoft or AWS, you are not on the list."
Peter Gostev
"The way I was doing this, I was literally going on the arena website, like scraping it by hand and adding it to Google sheets. Now I get access to Databricks and I can just do it."
⚡ This Week's Buzz
W&B ships Automations — event-triggered actions from your training runs into notifications, GitHub Actions, and deployments, pairing nicely with the new iOS app. GLM-5.1 and Gemma 4 both go live on W&B Inference. Wolfram's in-depth blog post on why more reasoning isn't always better is up on wandb.com.
- W&B Automations launch — triggers on runs → Slack, GitHub Actions, deploys
- GLM-5.1 and Gemma 4 both live on W&B Inference
- Wolfram's reasoning-regression deep dive published on the W&B blog
Wolfram Ravenwolf
"Remember when we talked last week about how models can get dumber if you give them more time to think — I posted an in-depth article about this on our blog at weights and biases."
🛠️ VB (OpenAI) — Codex, Plugins & Guardian Approvals
Vaibhav (VB) Srivastav from OpenAI's Codex team walks through the 3M weekly-active-users milestone and what's behind it: plugins (Stripe, Supabase, shadcn), sub-agents, and new experimental features including Guardian Approvals — a sub-agent that classifies every tool call for risk and only escalates the dangerous ones. VB also shares his own workflow: every morning at 9 AM, a Codex automation reads his Slack mentions, cross-references Gmail and calendar, and drops pre-briefs into five-minute calendar events.
- Codex: 3M weekly active users, up from 2M last month
- Plugins are bundles of skills + MCP servers (iOS builds, web builds, Stripe, Supabase, shadcn)
- Sub-agents decompose tasks into independent parallel Codex agents (including 'Jason')
- Guardian Approvals: sub-agent risk-classifies every tool call; auto-approves low risk, escalates high risk
- Experimental hooks: run code at session start, after tool calls, at session end
- VB's daily automation: reads Slack mentions, cross-refs Gmail + Calendar, schedules 5-min pre-brief events
Vaibhav (VB) Srivastav
"Every day at 9 AM, Codex uses an automation — it looks through all my Slack messages in the past 24 hours, threads where I've been tagged, layers that in with my Gmail and calendar, and tries to give me spaces where I can work on these things."
Vaibhav (VB) Srivastav
"Guardian approval within Codex allows us to look through each and every individual tool call. We spin off a sub-agent to quickly give it a risk profile — if it's medium or low risk we auto-approve, otherwise it escalates to you."
Vaibhav (VB) Srivastav
"If you disintegrate a plugin, all you get is a bunch of skills, an MCP server, and things to sort of tie it up together as a package."
🤖 Vincent Koc — OpenClaw, Dreaming & 1.5M Lines of Code
OpenClaw's #2 maintainer Vincent Koc on refactoring 1.5M lines of code into a plugin architecture in nine days (at 2 AM, at NVIDIA, before Jensen's keynote), the new /dreaming feature (REM/core/deep-sleep memory consolidation with a human-readable Dream Log), and why GPT-5.4 feels 'soulful-less' compared to the Rocky-from-Project-Hail-Mary voice Alex cloned for his own OpenClaw. Plus the pricing truth: Anthropic didn't ban OpenClaw — they just made Max-tier Opus via OpenClaw significantly more expensive.
- OpenClaw codebase: ~1.5M lines incl. unreleased iOS + Android native apps
- GitHub PR/issue counter literally caps at '5K+' — they hit the ceiling
- Plugin architecture ('not Lego — Ikea') refactored in 9 days at NVIDIA pre-keynote
- /dreaming: REM / core / deep sleep phases that defrag agent memory into a human-readable Dream Log
- Anthropic didn't ban OpenClaw — made Max-tier Opus usage via OpenClaw expensive, pushing users to GPT-5.4 via Codex
- GPT-5.4 'soulless' personality; Alex cloned Rocky's voice from Project Hail Mary into his OpenClaw
Vincent Koc
"I think the analogy I gave today was not Lego — it's more like Ikea. You have the harness, but I don't like the colors of the shelves. Well, you can swap it out."
Vincent Koc
"What I think what we've done with dreaming is created something not for the agent, but for the human in the loop. If I'm talking to a mom and they go 'this agent is dreaming and writing a dream log,' they can understand what is happening. I don't have to explain what the heck memory is or what markdown files are."
Vincent Koc
"We just decided to refactor a million lines of code in over nine days, and it kind of worked."
🔓 Omar Sanseviero — Gemma 4, 10M Downloads & Google DeepMind
Omar Sanseviero from Google DeepMind on the Gemma 4 launch crossing 10M downloads and 1,000+ fine-tunes, the license change, and the AI Edge gallery that lets people run Gemma locally on Android/iOS. Gemma is now the foundation for the next generation of Gemini Nano shipping on Pixel and Samsung. Wolfram chimes in on his Google-first household — kids using AI Studio and Antigravity to build games, his 70-year-old mother unlocking her Pixel by voice.
- Gemma 4: 10M+ downloads, 1,000+ fine-tunes on Hugging Face
- Gemma family total: over 500M downloads across all variants
- Gemma is the foundation for next-gen Gemini Nano on Pixel / Samsung
- AI Edge gallery — run Gemma locally on Android & iOS
- Llama.cpp vision capability fixes shipping
- Ask-your-mom test: Wolfram's 70-year-old mom uses voice unlock on Pixel
Omar Sanseviero
"In total, all of the Gemma-based models, over 500 million downloads. Gemma 4 alone just crossed 10 million, and there are over a thousand Gemma-4-based fine-tunes already."
Omar Sanseviero
"Gemma pretty much is the foundation model that covers the next generation of Gemini Nano out. Even if Gemini Nano 4 is not out yet, you can already try out the base model capabilities right now."
🎬 Wrap-Up
Alex signs off from London with Wolfram, and hands the after-show over to Yam, Nisten, and LDJ on a Twitter Space. Gratitude lap to Swyx and the AI Engineer crew — ThursdAI literally wouldn't exist without the AI Engineer conference. See you in San Francisco this summer.
- AI Engineer Europe sold out 4x faster than any other edition
- Post-show continues on Twitter Spaces with Yam / Nisten / LDJ
- Next AI Engineer is this summer in San Francisco
Wolfram Ravenwolf
"It shows that we have an AI scene in Europe as well, and the energy is extremely high. The last time I felt that energy was in San Francisco — great to see it here in Europe as well."
TL;DR April 9 - show notes and links:
Hosts and Guests
Alex Volkov – AI Evangelist & Weights & Biases (@altryne)
Co-Hosts – @WolframRvnwlf @yampeleg @nisten @ldjconfirmed
Guests: @swyx (AI Engineer / Latent Space), @petergostev (Arena, formerly LMArena), @reach_vb (OpenAI / Codex), @vincent_koc (OpenClaw #2 maintainer), @osanseviero (Google DeepMind / Gemma)
Big CO LLMs + APIs
Anthropic announces Project Glasswing and Claude Mythos Preview, a cyber-defense frontier model too dangerous to release publicly (X, Announcement)
Anthropic’s Claude Mythos is so powerful they won’t release it — found zero-days in every major OS and browser, escaped its sandbox, and scored 93.9% on SWE-bench (X, X, X, X)
Anthropic ARR jumps from $19B (February) to $30B in April — secondary tender sale completed, employees not selling ahead of IPO
Anthropic + Google TPU deal — Anthropic getting massive compute commitment from Google (who already owns ~10% of Anthropic), with Peter Gostev’s Compute Wars chart showing the gap to OpenAI closing
Anthropic ships Managed Agents — fully hosted agent runtime + infrastructure. Selling outcomes, not tokens
Meta launches Muse Spark, the first model from Meta Superintelligence Labs, with natively multimodal reasoning, multi-agent Contemplating mode, and deep health/visual capabilities (X, Blog)
Simon Willison deep dives into Meta’s Muse Spark model and uncovers 16 hidden tools including visual grounding and sub-agents in the meta.ai chat UI (X, Blog, Announcement)
Open Source LLMs
GLM-5.1 from Z.ai is #1 open-source on SWE-Bench Pro at 58.4%, runs autonomously for 8 hours with 1,700+ agent steps (X, HF, Arxiv)
Gemma 4 crosses 10M+ downloads, 1,000+ Gemma-4-based fine-tunes on HF. Did really well on Arena considering size — Peter Gostev confirmed it smashed many models on the Pareto curve
Nisten’s pick: Hermes 27B — trained specifically to be paired with the Hermes harness, allegedly distilled from Opus API. Model + harness shipped together as a portable unit
Tools & Agentic Engineering
OpenClaw 2026.4.5 — biggest release since 4.0:
/dreaminggoes GA (Light/Deep/REM memory consolidation with a Dream Diary in DREAMS.md), built-in video + music generation across 4 backends, GPT-5.4 as new default, prompt-cache reuse improvements, Control UI + docs in 12 new languages (Release, Vincent, Dreaming docs, FOD#147)OpenClaw codebase now ~1.5M lines including unreleased iOS + Android native apps. GitHub literally caps at “5K+” PRs/issues — they hit the ceiling
Anthropic did NOT ban OpenClaw — they made Max-tier subscription usage of Opus via OpenClaw significantly more expensive, pushing many users to GPT-5.4 via Codex
Codex hits 3M weekly active users — up from 2M last month. VB walked through plugins (Stripe, Supabase, shadcn), sub-agents, Guardian Approvals (auto-classify tool-call risk), and experimental hooks
Cursor: remote agents + code review agent (78% issues caught pre-merge)
MemPalace: Milla Jovovich and Ben Sigman’s open-source AI memory system goes viral with 26K GitHub stars in 2 days, claims top benchmark scores, then transparently walks back overstated claims (X, GitHub, X, X, GitHub)
This Week’s Buzz (Weights & Biases)
W&B Automations are LIVE — event triggers from your runs into notifications, GitHub Actions, deployments. Pairs nicely with the new iOS app
GLM-5.1 and Gemma 4 both up on W&B Inference
Wolfram published an in-depth blog post on his finding that more reasoning is not always better (models can get dumber with more thinking time) — full writeup on wandb.com
Vision & Video
Seedance 2.0 launches in the US — on Replicate with up to 9 reference images, 3 videos, and 3 audio files for cinematic AI video generation (X, Announcement). Peter Gostev confirmed it jumped ~80 ELO points above the next video model on Arena — a massive gap where most video models cluster within 10 points
HappyHorse-1.0, a mysterious 15B video model from Alibaba’s Taotian Group, takes #1 on Artificial Analysis video arena beating Seedance 2.0, Kling 3.0, and Grok Video (X, X, X, X, Blog)
The Harry Potter “Drip Wizards” AI slop trend — Seedance-powered Hogwarts videos going hugely viral
AI Art & Diffusion & 3D
Show notes & key moments
Swyx on harness engineering: gains are coming from the harness, not the weights. The big labs are investing more and more in harness — it’s not going away. Skills (English-as-programming-language) are increasingly absorbing harness work
Swyx on AI Engineer tracks: MCP is “more settled and stable, therefore less interesting.” Coding agents track is bigger this year (Cursor, Factory, super-long-running). Voice & Vision split from Generative Media — multimodality as a single track no longer makes sense
Swyx on supply chain attacks: light LLM and Axios issues mean you should “vendor everything” —
pip forkinstead ofpip install. Tool requests becoming prompt requestsPeter Gostev on Mythos pricing: $25 / $125 per M tokens (~5x Opus 4.6). But the real reason it’s not public isn’t safety — Anthropic likely just doesn’t have the compute to serve it
Peter Gostev on Compute Wars: OpenAI is way ahead of Anthropic on compute. The new Google TPU deal is Anthropic catching up — and weird that Google is propping up a competitor to DeepMind. (Same pattern as when Google’s $2B Anthropic investment effectively propped up AWS vs Google Cloud)
Peter Gostev on Arena data: Arena released 3 years of historical leaderboard data + actual prompts as datasets on Hugging Face. Previously he was scraping it by hand into Google Sheets — now he has Databricks access
VB on Codex workflows: every morning at 9 AM, Codex automation reads his Slack mentions, cross-references Gmail and Calendar, and creates a 5-minute pre-brief calendar event for upcoming meetings. None of it is “coding” — it’s all plugins + connectors
Vincent Koc on the GPT-5.4 personality problem: model is incredible at coding but “soulless.” Wolfram noticed it back in December and cancelled his subscription. Alex cloned the Rocky voice from Project Hail Mary and put it in his OpenClaw — “amazing”
Vincent Koc on Dreaming: three phases (REM, core, deep sleep) that defrag agent memory. The dream log is for the human in the loop — makes memory inspectable in a way a non-technical person (a mom) can understand
Vincent Koc on architecture: the open-source flood forced OpenClaw into a plugin architecture. “Not Lego — Ikea.” Refactored ~1M lines in 9 days at 2 AM at NVIDIA before Jensen’s keynote
Omar Sanseviero on Gemma 4: 500M+ total Gemma downloads across all variants. Gemma is the foundation for the next generation of Gemini Nano on Pixel/Samsung. Lama.cpp vision capability fixes shipping
Wolfram’s Pixel/Google household: kids using AI Studio + Antigravity to build games, his 70-year-old mother using voice unlock on her Pixel
Alex Volkov
0:29
Hello.
0:30
Hello and welcome folks. Welcome to Thursday ad. This is Alex. Welcome. I'm here with Swyx and here is Wolfram. How you guys doing? Uh, we're Thursday, April 9th in a very unusual time for us.
It's only 8
0:44
00 AM Can you believe it?
0:46
In San Francisco, in Pacific. Uh, but it is,
Swyx
0:49
it's the only way you get me to wake up for Thursday.
0:52
AI is,
Alex Volkov
0:52
it's doing it in Europe.
Swyx
0:54
Uh,
Alex Volkov
0:55
so now it's, it's four o'clock where we are.
0:57
We're actually live streaming to you, uh, for the first time from Europe. This is, uh, Thursday. I coming to you live from AI engineer Europe. And, uh, I have the founder of ai, engineer his swix, not your first time on Thursday. I, many, many times. Uh, Thursday I owes a, a big great, uh, uh, a debt to you specifically.
Swyx
1:17
You gimme too much credit, man.
Alex Volkov
1:18
No, I will say this like, and I say this before, but many, many new
1:21
listeners are joining now for us as well. Yeah. Uh. The first AI engineer, the first of these conferences that you organized, uh, was when I was, uh, uh, going there to cover. I had the badge, its a media on it, and the Uber driver on the way there asked me what I do. I was like, I have a podcast. This is the first time that I allowed myself to actually, yeah. The, the
Swyx
1:40
identity change is the, when, when the book becomes real.
Alex Volkov
1:43
So it felt really real.
1:44
And so we covered the first one and I talked to you back then. Uh, we've grown so much.
Swyx
1:48
Yeah.
Alex Volkov
1:48
So for folks who are just tuning in, uh, Swyx, founder of Latent
1:52
Space host, co-host of Latent Space, founder of Small ai, the newsletter founder of AI Engineer Summit.
Swyx
1:59
Yeah.
1:59
Uh, so small ai, other
Alex Volkov
2:00
things,
Swyx
2:00
small AI was, uh, sort of absorbed into cognition.
2:02
Yeah. Uh, but now, and the newsletter has been absorbed into the latent space. So we sort split things up there.
Alex Volkov
2:07
You're consolidating
2:08
or
Swyx
2:08
consolidating a bit?
2:09
Uh, small AI will come back in a new form.
Alex Volkov
2:11
Oh, okay.
Swyx
2:12
Yeah.
2:12
Um, so we we're, we'll, we'll we'll see what happens after a IE, but, uh, right now the focus is a ai e
Alex Volkov
2:18
focus is IE we're first time in London.
Swyx
2:20
Yeah.
Alex Volkov
2:20
Uh, I have a bunch of questions for you.
2:22
So, so one of the things that, uh, people came up to me right here saying, Hey Alex, we listen to the show. Um, what we love about the show is that when you guys been there, and you can bring whatever happens in San Francisco to us, 'cause we are not there. AI e for me is a lot of that AI for me, the hallway track, for example, uh, for folks who dunno what the hallway track is, it's not an official track. People just walk around, talk to others.
Swyx
2:43
It is an official track tomorrow.
Alex Volkov
2:45
Oh, no way.
Swyx
2:45
Yeah.
2:46
So, uh, I mean this, we we're overloading the term, but um. Well, I always like the conferences who let the attendees also speak.
Alex Volkov
2:53
Yeah.
Swyx
2:54
Uh, so everyone from now until midnight, whoa.
Alex Volkov
2:57
Whoa.
Swyx
2:57
Something happened.
Alex Volkov
2:58
Yeah.
2:58
Somebody joined. Uh, that's on. We'll take a go. That's okay.
Swyx
3:01
Okay.
3:02
Um, from now until midnight, attendees can submit talk proposals and whoever gets the most votes, they'll get selected to speak tomorrow on stage.
Alex Volkov
3:09
That was great.
3:10
I told him to submit as well. You'll review. Um, so, so the holy track, uh, is where we also kinda learn what people, they're not outside the bubble, kinda like feeling what they're doing, what they're using. Um, so I would love to chat with you briefly. I know that at some point you're gonna have to go back up on stage.
Swyx
3:27
Yeah.
Alex Volkov
3:28
Uh, but, uh, for now, first of all, uh, the staple of
3:33
AE for me is the speaker dinner. Mm-hmm. It was incredible. There was like so many people that I knew already from previous ones, so many new people. Uh, how about you? What the, the, the type of conversations, I know you handpick all those people, you know about what they're talking about. Uh, would love things from you generally, but like just the conversations. What are people using?
Swyx
3:50
Oh, uh, I mean, you know, a lot of like, you know, is cursor cooked,
3:54
is cloud code cooked or whatever. Um, I think actually people are looking ahead beyond that, you know, uh, there's a lot of discussion of open cloud, obviously, and, uh, how much personal agent and assisted life, uh, you can have.
Alex Volkov
4:08
Speaking of open cloud, we have like, uh, Peter
4:11
Steinberger gave a keynote opening.
Swyx
4:13
Yes.
Alex Volkov
4:13
And you, there's a whole track.
Swyx
4:15
Yeah.
Alex Volkov
4:15
Claw dedicated.
Swyx
4:16
Yeah, yeah, yeah.
Alex Volkov
4:17
Yes.
Swyx
4:17
There's something that I think
Alex Volkov
4:18
I'm gonna put you on the spot, brother.
Swyx
4:20
What?
Alex Volkov
4:20
Do you have an open cloud running?
Swyx
4:21
Yeah.
4:22
Telegram.
Alex Volkov
4:23
Were you late for it or early?
Swyx
4:25
Uh, I would say.
4:27
Okay. So I first tried it for my lane space Discord.
Alex Volkov
4:30
Yeah.
Swyx
4:30
Uh, when it was still called Cloud Bot.
Alex Volkov
4:32
Yeah.
Swyx
4:32
And I wasn't impressed, so I uninstalled it.
Alex Volkov
4:34
Sure.
Swyx
4:35
Uh, and then, and then I, I reinstalled it again after,
4:37
after the whole, uh, it blew up. Um, I think, uh, it, it's, you know, so I'm both early and late.
Alex Volkov
4:44
Yes.
4:45
Uh, Daria, my fiance, uh, soon to be married by the way. Folks who, uh, following up, uh, she went through the same thing she is got, I installed it for her. She's like, Hmm. And then after a while after like the use cases start picking up, she's like, oh, oh, I, I need this. And now she's like, I have three of them. They're running on the farm. That's great. Do you have a Mac Mini?
Swyx
5:02
No, I have an old MacBook.
Alex Volkov
5:04
Oh,
Swyx
5:04
MacBook.
5:04
Yeah. You don't need to buy a Mac mini.
Alex Volkov
5:05
You don't need to buy a Mac mini.
5:07
Uh, but it's cool. So there's a full open claw, uh, track.
Swyx
5:10
Yeah.
Alex Volkov
5:10
Right now.
5:11
And with
Swyx
5:11
five open call maintainers
Alex Volkov
5:12
and five open call maintainers.
5:14
One of them is gonna join the show. Folks, stay with us. Vincent co. The number two maintainer after Peter Steinberger is gonna join the show. Yeah. Very soon. We're gonna grill him about every security issue. Um, talk to me about security, how you feel about like, where the security is. Peter was on stage there, uh, and basically said, Hey, we hardened it a lot. Same thing. Yeah. And VD contributed. Where, where do you think it is right now?
Swyx
5:34
I actually think, uh, I mean it is important that they improve it.
5:38
Um, I think it's important that he has a lot more resources to improve it right now 'cause. It's, it's very hard to do it as a single person. Uh, but I think also just the visibility that Open Cloud has makes it fundamentally the most rich target of attack. Right. And this is something I'm trying to figure out, like does being open source and popular make it more secure or less secure? Because it's more secure? Because more everyone can see the code and the moment there's a bug, they fix it, whatever, but it's less secure because it becomes a very high value target.
Alex Volkov
6:09
Yeah.
Swyx
6:10
So, I don't know.
Alex Volkov
6:12
He mentioned something on stage and I wanna move to other topics
6:14
as well, but like, uh, Peter mentioned something on stage and folks, by the way, everything we're talking about is livestreamed on the AI engineer YouTube, which is happening right now. 'cause there's talks right now that are happening. Uh, so if you feel like you're missing out, you can definitely tune into livestream. Scroll a little bit back and see Peter talking about everything I'm talking about as well. Uh, Greg, uh, uh, connection talks. Peter mentioned that the nation state try to attack open cloud Korea tech. That's crazy.
Swyx
6:39
Yeah.
6:39
I mean, um, it's gonna happen.
Alex Volkov
6:41
Yeah.
Swyx
6:41
Uh, and I think there are also compromising light LLM
6:44
and XOs and all these things. So open cloud is only the latest in the, uh, a series of compromises. Yeah. And it's getting easier. And so I think like, uh, one of those things of cyber warfare and it, it used to be more theoretical, more like, oh, it will never affect me. Now it's really starting to come home.
Wolfram Ravenwolf
7:00
It, uh, fits into the myth and classing
Swyx
7:03
topics
Wolfram Ravenwolf
7:04
we have also on the list for today.
Swyx
7:06
Yeah.
7:06
I, I. I mean, mythos is announced. It's okay. You know, like you can't have it.
Alex Volkov
7:13
We as people can have it if you are a company, a big company.
7:16
Yeah. We're gonna talk about mythos folks on tropics later.
Swyx
7:18
Yeah.
7:18
I was thinking, you know, okay, if you grant like Microsoft, Amazon, IBM, all these, it's kind of released, right? Like
Alex Volkov
7:24
yeah.
Swyx
7:24
At least
Alex Volkov
7:25
it's,
Swyx
7:26
let's call it 400,000 people have access to meters.
Alex Volkov
7:28
Yes.
Swyx
7:28
So it, it's not too dangerous for meter
Alex Volkov
7:31
and apparently myths can break through sandboxes apparently.
7:35
Uh, so folks, we're gonna tell you all about this. This is just a pre-show 'cause Swyx needs to run. And we are from a IE, like I really wanna talk to you about a e but we're gonna talk to about cloud meters, which is a new preview of tropics, biggest model Opus, biggest brother that's very good at cybersecurity that freaks people out. And we'll definitely like mention this, uh, in, in the moment, um, Swyx other trends from AI engineer. So the way that I ask you about this and then, uh, I put you on the spot, is usually curate the tracks, not necessarily the speakers, but also the tracks according to what's happening.
Peter Gostev
8:06
Yeah.
Alex Volkov
8:06
Here's what the fuck happened to the past three years, three and a
8:09
half years almost, of AI engineer as the general, um, I don't know if you expected this AI engineer was a term and then every engineer became an AI engineer. Like, if you're not in AI engineer right now, if you're not using the tools, you're like not gonna make your your way behind. Uh, so what is top of mind of AI engineers that are essentially all engineers at this point? Is talking count. I know you talked about talking as a metric is, um, harness engineering, context engineering. Tell us like what, what,
Swyx
8:35
yeah, yeah, yeah.
8:36
Um, okay, so first of all, I actually disagree that every engineer should be a AI engineer.
Alex Volkov
8:40
Ooh,
Swyx
8:41
I said this in my original, uh, blog post about it.
8:44
Uh, if you're a database guy, if you're a operating systems guy, if you are, you know, like a game dev, like you don't have to be AI engineering. I think that's totally okay. I think it's just like, yeah, the average product engineer, uh, building B2B SaaS. Yeah. Yeah. Obviously you, you, you should, you should be embracing ai. Um, uh, yeah. So every my work is the curation and the focus. This today is obviously open claw and harness engineering. I think harness engineering people still are sleeping on it somewhat. Yeah. Even though it's like relatively popular. Um, I do think that the very obvious trend is towards the dark factories. The, um, extremely parallel multi-agent thing that will probably kill GitHub and Git someday, uh, in some un unknown way.
Alex Volkov
9:27
Struggling.
9:28
GitHub is struggling.
Swyx
9:28
Yeah.
9:28
Yeah. Uh, I I mean obviously they're, they're, they're here and they would, they would, they would strongly disagree. Um, I, I think like, it, it is definitely undeniable that the load that GitHub has had from agents is, uh, unprecedented.
Alex Volkov
9:40
Wait, you,
Swyx
9:41
no, they'll never designed for this.
Alex Volkov
9:42
You, you think they'll disagree with the statement that GitHub
9:44
is struggling given that they, for the first time in the decade got down from 9 9 9 9 9 9 9 reliability to like 9.8?
Swyx
9:51
Uh, yeah.
9:52
I mean, uh, I would say that they, they're. You don't see the other side of the story, which is the insane load that they have.
Alex Volkov
9:58
Oh, no, absolutely.
9:59
Yeah.
Swyx
9:59
So like, yeah.
10:00
Um, you know, it's easy to criticize when you're not involved in maintaining a scaling system. And it is, it is scaling, like it is a startup again.
Alex Volkov
10:07
Oh, wow.
10:08
Okay.
Swyx
10:08
Like for, for it, it was a startup in 20 13, 14.
Alex Volkov
10:12
Yeah.
Swyx
10:12
And then it kind of got bought by Microsoft and it was kind of stable.
10:15
Now it's scaling again. Like, like, like, like in the old days, right? Yes. And show me a young startup that has that scale and had, that has higher than 90% reliability.
Alex Volkov
10:24
I cannot even look at Anthropic with,
Swyx
10:26
yeah.
10:26
Yeah. So, so like, let who, you know, who, if you're, if you're only with, if without sitting, you can throw the stones.
Alex Volkov
10:32
Yep.
Swyx
10:33
Um, so, so, and I, I, I think like, um, uh, harness engineering is
10:37
very important, uh, but also like, you know, voice and vision is, is a, is a key theme for, uh, for people. And, and like all these tracks that I curate are basically my choices for like, here's the things that people should be focusing on just 'cause I really care about the combination of impact, but also timing. And sometimes you can be right directionally, but too early or too late.
Nisten
10:57
Yeah.
Swyx
10:58
And, and then it's, um, it's not as useful for you.
11:01
So I really try to make this everything like, useful in the sense that you can, you should be working on this today. The, the time is right, the fundamentals are there. And this is gonna be a, a big trend.
Alex Volkov
11:09
Here's my, uh, thesis.
11:11
Would love to run this by you on the show to see. Um, many people try to. Predict what's going to change. So when cloud mythos release, people are like, oh, you know, security is gonna change forever, et cetera, because like, it can break things, et cetera. Uh, sometimes it's prudent to do, to try to estimate what's not changing. In my estimation, here's what's not gonna change in the year. Uh, there's still gonna be intelligences, there's still gonna be harnessed by some harness, and people will still have to do some context engineering around those to, to build like tools to use as well. Do, do you agree with this like assessment of like, this is the three pillars of agent AI or whatever, that will not go away, or do you foresee like harness disappearing, uh, into the model at some point or context disappearing into the harness at some point?
Swyx
11:55
No, I think, um, the big labs have shown that they're investing more and more
12:00
in harness, in harness, uh, in all that. So, uh, no, I don't think, I don't think it'll go away. Yeah.
Alex Volkov
12:05
Yep.
Swyx
12:06
I, I, I think maybe the thing that is interesting is that more and more,
12:13
uh, is being consumed by skills and I think, uh, one, um, something I think of like that's very special about skills is that it is just English and like, um, it, it really calls back to Andre cap's observation that English is the hottest new programming language and it's probably like the ultimate endpoint of how we want to communicate with,
Alex Volkov
12:32
it's just English with scripts.
12:35
This is the scary part of, of skills. Like they can download scripts and people don't even know about this. This is, uh, we had a conversation with the MCP UI folks who, like, were all guests on the pod and a few other MCP folks here. Uh, David, so here is also here from a tropic who like came up with MCP, right? Um, the amount of time somebody came up to and was like, Hey, is MCP dead? And they're like, I don't wanna survive this. Um, the kind of the difference business skills, MCP again, conversations from the whole track. This is what we're gonna bring to you, uh, is, uh, skills are very vast area of attack because you download a skill and it downloads the files and
Swyx
13:08
Yeah.
13:09
Requirements. Yeah. So, so yeah, you should maintain your own library and only use the stuff that you have personally reviewed.
Alex Volkov
13:14
Right?
13:14
So not like skill, uh, repositories, you mean like personal skill repositories?
Swyx
13:18
Yeah, yeah, yeah.
13:19
I, I mean I, I, I'm a strong believer in that, you know, you know this, this all like, uh, light LM and Axio security issues.
Alex Volkov
13:26
Yes.
Swyx
13:26
Makes me think that you should just vendor everything by
13:29
doing a slop fork of everything. So like, basically anything you wanna import, something you don't like, you know, uh, p uh, NPM install. You don't, um, pip install. You say Pip fork,
Alex Volkov
13:39
PIP fork,
Swyx
13:40
uh, uh, and, and p pip re rebuild.
Alex Volkov
13:43
And your myth of 2.0 just
Swyx
13:44
rewrite, rewrite the whole thing.
13:46
But you own the whole code and it's, it's only the parts that you use. You throw out the parts that you don't use. So it's, it's guaranteed to be smaller.
Alex Volkov
13:51
This is a very interesting, kind of like, concept of what
13:56
happens when code is free, I think. Um, yeah. Ryan Popolo talked about this from open the air on stage. What happens when you
Swyx
14:01
consider Yeah, more should be in zero dependency, zero.
14:03
Like your, your, your dependency tree should be a lot smaller now. And the reason, the only reason you're not is because you haven't really explored, or we don't have the tooling to cheaply clone things.
Alex Volkov
14:14
Yeah.
Swyx
14:14
For, for and customize for our own use cases.
Alex Volkov
14:16
Do you see the, uh, the, the tool request needs to
14:19
be renamed to prompt request. Yeah. And instead of giving you code, people just give you prompts and then your agent will build it in secure way.
Swyx
14:24
Yeah, yeah, yeah.
14:25
Uh, just the other, uh, I was, was just talking with someone else who they also just shut their, um, PR contributions page. 'cause they don't want any prs.
Alex Volkov
14:33
Yeah.
Swyx
14:34
They just want issues and, you know, then they'll just decide what to prompt.
14:37
Right?
Alex Volkov
14:38
Yep.
14:38
Uh, so we know you have to run, I have one or two more minutes with you, uh, for folks who are not tuned into the live show yet. Tomorrow is another set of talks. Yeah. There's keynotes, et cetera. Uh, what should folks watch out for? What is interesting? I know like all of the kids are your favorite kids, like ask you for biggest favorite kids, but like you, I
Swyx
14:58
have to make choices.
14:59
Right. Basically the, the, the venue dictates that, uh, what I should place in what size of room.
Alex Volkov
15:04
Yeah.
Swyx
15:05
And then I have to make bets on Okay, what do people want to hear?
15:07
What would be the most popular? And people get very upset on me whenever I get anything wrong in either direction. Yeah. Um, so I, I mean, uh, MCP uh, still like you cannot come to London and not host David and have a MCP tracking. And honestly, we have some of the biggest contributors here, right? Yeah. The web MCP people, GitHub, MCP, Chrome dev tools, MCP.
Alex Volkov
15:26
Yep.
Swyx
15:26
Like some of the most solid contributions in the past year.
15:29
And that's my, that's my, my goal to sort of, so not,
Alex Volkov
15:31
not that.
Swyx
15:32
Not that,
Alex Volkov
15:32
not that,
Swyx
15:33
no.
15:33
Actually like the most successful, uh, protocol, uh, for integration and of anything, it's beat open API, it's beat TRPC. GRPC.
Alex Volkov
15:42
Yeah.
Swyx
15:42
Um, all these, all these things.
15:44
So MCP has been enormously successful and just, it's just now like more settled and stable. Yeah. And therefore less interesting.
Alex Volkov
15:49
The thing that I learned from, uh, Adobe, by the
15:51
way, uh, one of the Oh, LIAD.
Swyx
15:54
Yeah.
15:54
Yeah.
Alex Volkov
15:54
That on Tropic, when it shows you like dynamic things,
15:56
that's also like MCP UI or MCP apps.
Swyx
16:00
Yeah.
Alex Volkov
16:00
Even when on Tropic, like builds a dynamic UI for you.
Swyx
16:02
Yeah.
Alex Volkov
16:02
They do it via the MCP apps format.
16:04
And I'm like, oh, that's super cool. That's super
Swyx
16:06
cool.
16:07
Yeah. Yeah. Uh, oh. And then there's actually a special philanthropic talk at the end of the MCP track
Alex Volkov
16:10
Oh.
Swyx
16:11
Where they're really talk about how m uh, Anthropic uses MCP internally using
16:15
the MCP gateway, uh, which I think people don't understand the degree at which, uh, Anthropic Dog Foods MCP It is how most internal services are run is using m. Oh,
Alex Volkov
16:23
wow.
Swyx
16:24
Uh, so, so, uh, I think that gateway pattern, uh, I prompted the
16:27
guy to like really go hard on like promoting this pattern because I've seen a lot of enterprises adopt
Alex Volkov
16:32
what's, what's the name?
16:33
Do you remember?
Swyx
16:34
Uh, Sampath something.
16:35
Something,
Alex Volkov
16:35
yeah.
16:36
Okay. So folks check out the
Swyx
16:37
schedule.
16:37
This is the Anthropic talk on, on on SP That is not
Alex Volkov
16:39
tomorrow.
16:40
On Friday.
Swyx
16:40
Yeah.
16:40
Yeah. He's the closing keynote. Um, so, and, and, and then the other one is coding agents. Last year is the, the year of coding agents. This year we have, uh, more cursor. What is this
Alex Volkov
16:48
year?
Swyx
16:48
More factory.
16:49
Uh, I, I think like super, super long running. Um, I think, I don't like the word proactive.
Alex Volkov
16:54
No.
16:54
Why not?
Swyx
16:55
Proactiv is just coding engine on a CR job that's proactive.
16:58
Right. It is not, it is not really, um, okay. That, it's not really that interesting. I think people get overexcited because they, they think that it's gonna solve all the other problems for you. No, it's probably just gonna annoy you a bit more. Um, so, uh, no, I think like more autonomy and more speed, uh, are kind of like the two themes that, uh, I talked about a lot of people like my, uh, semi async value of death chart mm-hmm. Where basically like, there's the, if, if you're in the, in the middle, it's like not fast enough for you to really, um, uh, maintain your focus because you're gonna get distracted and it's not, it's not, uh, capable enough that you can just delegate and walk away. And so you want either super fast or you want, you want super capable and, and long running. So if you're waiting, you, there's no room in the middle.
Alex Volkov
17:39
Yeah.
Swyx
17:39
Yeah.
17:40
So, uh, we have talks on both, uh, we have GPUs and infra. We have, um, AI architects for the more senior leaders, uh, who are running orgs of AI engineers and how to manage, how to hire, how to, how to scale them. Um, and then generative media. Um, and this is why this, uh, this is something I moved towards last year where, which is that I no longer have a multimodality track because what the fuck is multimodality?
Alex Volkov
18:01
Yep.
Swyx
18:01
We have voice and vision, which is more practical.
18:03
And then we have generative, uh, image video music. Yeah. Which is more artistic. And you who, who, why should you mesh vision and voice, which is so practical versus art.
Alex Volkov
18:15
That's true.
Swyx
18:16
It makes no sense.
18:17
So we no longer have any multi multimodality tracks. We only have. Uh, one of the two extremes.
Alex Volkov
18:22
Yeah.
Swyx
18:22
And I think that that makes more sense.
Alex Volkov
18:23
And we have news for you about most of the stuff that
18:26
he said with like, GPT Image two is supposedly coming out soon. ANCE two is available on, uh, replicate. And in the US now CS two full ance with video upload and everything. Uh,
Swyx
18:36
if you can commit a million dollars
Alex Volkov
18:38
no, you can go to Replicate, replicate committed million
18:40
dollars so you can go via their API. So yeah, it's available. And also there's a new video model that's supposedly beating CS through already from Alibaba that we tell you about. I cannot
Nisten
18:50
keep up.
Alex Volkov
18:51
Yeah.
18:51
This is why there's the exist. You don't have to keep up. We keep up for you. Uh, six. One last thing. There's a famous street in London and there's a very specific house on that street under number 10.
Nisten
19:03
Yeah.
Alex Volkov
19:03
Where the PM resides.
19:05
Yeah. For all of Britain. And, uh, you had a picture, you and a bunch of other AI folks, Steve Ruiz deal draw, who's also here. Great dude. He is like, thank you JavaScript. And he is like standing there. Why the f were all of you standing in front of the PM's house?
Swyx
19:21
Yeah, yeah.
19:21
Uh, the chief AI officer of the UK reached out and said that he's trying to get the government, uh, more involved in ai, hire AI engineers. Nice. Also just, uh, modernized the, the government, which I think we're all in favor of. That was great. Uh, and so, uh, they, they, they sent out some invites to, to some folks. Uh, and then we all, we all, we all went there to. Have tea. How was the tea? Yeah, tea was good. Uh, Mary the Cat was the main celebrity there.
Alex Volkov
19:47
Oh, wow.
Swyx
19:47
Uh, but we, yeah, there was the KPM advisor.
19:50
Um, Raya Zel was our DeepMind keynote. She's VP of research at DeepMind. Yeah. Also, uh, UK Ambassador for ai.
Alex Volkov
19:56
Oh, okay.
Swyx
19:57
Um, and, uh, mostly I, I think it's, there's this like cultural
20:01
exchange and a demonstration that the UK government is very committed, growing the AI scene here in, in, in, in London.
Alex Volkov
20:08
Yeah.
Swyx
20:08
Uh, but also to, to, to try to hire
Alex Volkov
20:11
DeepMind is originally from here.
Swyx
20:12
Yeah.
Alex Volkov
20:13
Uh, there's offices on Tropic opened one, and, uh, there's a
20:16
great, uh, uh, video in the beginning of this, the show, like why, and it's, it's really, really cool. I talked to a bunch of people. They're like a, a bunch of London native folks and generally in Europe, and especially now with theater and Mario from pi, like all these folks. Yes. Yeah. They're bringing the agent. Like
Swyx
20:31
it could not have worked out better, like the, the, the
20:33
timing of, uh, this Austrian crew.
Alex Volkov
20:35
Yes.
Swyx
20:36
Taking over AI is, uh, is very appropriate that we are
20:39
launching our Europe, uh, conference.
Alex Volkov
20:42
Yeah.
Swyx
20:42
And I think, um, AI tends to be very US centric, very San Francisco centric.
20:46
Yeah. And really good work is done everywhere else and, uh, sometimes you just need to shine the light on the work that they do. So, um, we are a very US centric audience. Yeah. But, uh, hopefully AI E Europe can bring the best of Europe, uh, to the rest of the world.
Alex Volkov
21:00
Yep.
21:00
Oh, Swyx, thank you so much for joining us. Thank you for running this conference. We know you have to run and he just. He's gotta go. Okay. You have, we're we're getting interrupted by the other co-founder of Engineer Ben, uh, who came to grab Swyx.
Swyx
21:12
Okay.
Alex Volkov
21:12
Thank you.
Swyx
21:13
Yes.
Alex Volkov
21:13
Uh, and folks tune into the tomorrow's live show and
21:16
generally you sold out too quickly. There's too, too little space. We need
Swyx
21:19
more.
21:19
Yes. This thing sold out, uh, four times as fast as uh, or any of the other s
Alex Volkov
21:23
Go work man.
21:24
Go work. Thank you for joining us headphones. Alright folks, we heard from Swix on uh, Thursday Live. We are gonna do it till day. We're gonna tell you about the news and then we also have a few more guests here. Wolfram,
Wolfram Ravenwolf
21:37
I'm here.
Alex Volkov
21:37
How's the event for this is your first AI engineer?
Wolfram Ravenwolf
21:40
Yeah, definitely.
21:41
Great to be here in London after 30 years again. And, uh, the wipes, they remind me of San Francisco actually. Yeah. Here, it's a conference. We have the energy, the people, and the talks, so it's amazing to be here and yeah. Everybody should be here.
Alex Volkov
21:54
Yeah.
21:55
You Um, so I went to a speaker dinner. You went to the clock on yesterday? Well, like with Peter and like a bunch of other folks. Tell us about this.
Wolfram Ravenwolf
22:02
Yeah, I was happy to get in because we had, uh, 800 re
22:06
registrations or even 900 I was told. And it was first come, first serve and fortunately I was one of the earlier ones to get in and, uh, I met Peter for the first time and yeah, we had a great talk. Uh, basically it was an amazing event. Nice to meet the people and, uh, a lot of speakers. So it's a big thing. The clock on. Yeah.
Alex Volkov
22:26
That's great.
22:26
Uh, we also have, you know, our, our regular cohost. I'm gonna add you guys super quick, uh, and hopefully folks will see us. Hey ya. Hey, ton LDJ. What's up folks? Uh, we're gonna do A-T-L-D-R super quick, but I wanted to hear one thing in the ai for you, what is the most important thing? We're kind of like busy with the show, the notes and everything. They have everything, but like, give us one thing that's most important. Yam looks. Mytho
Yam Peleg
22:49
Mythos.
22:49
Mytho mytho. I took it. It's mine. It's mine. Claude Mytho. It's
Alex Volkov
22:54
mine.
22:54
Thank you. Yeah. Can you say again?
Yam Peleg
22:58
Come on.
22:59
You, you didn't get this Cloud Mytho Claude Mytho. That's the thing. That's mine. I took it, took it away from the others. Claude myth is one, one thing in AI this week is cloud myth
Alex Volkov
23:11
and we're gonna definitely talk about this.
23:13
Uh, we have Peter Gusev from Arena, uh, the arena of AI joining us in, in a few moments. Give us two more minutes and I'll call you, uh, in generous two minutes to talk about also Cloud Misos. He did like a great, uh, review of cloud missiles, evals, uh, LDJ Nisten. How about you guys? What is the one thing that's most important for us to cover? You cannot say mythos 'cause Yam already said.
Nisten
23:36
L LDJ or I'll, I'll, I'll go since I on, or you can go.
LDJ
23:40
Okay, I'll go.
23:42
Um, not is necessarily the best release, but I think it's, uh, interesting time rise and kind of most unexpected is the Muse Spark model from Meta that just released
Alex Volkov
23:53
the Spark model from, uh, MSL at Meta Meta, uh, super Intelligence Labs.
23:58
Uh, the first one after they rebuilt everything. Definitely gonna cover this as well. Uh, yeah, that's dope. Nisten, yours please.
Nisten
24:06
Uh, this is a random one.
24:08
It is the Ness 27 B. They trained a model specifically for use with Hermes, and I think they trained it on the actual Opus API. And, uh, I
Alex Volkov
24:21
have this in my show notes.
24:22
Nisten always brings us stuff.
Nisten
24:25
I,
Alex Volkov
24:26
I
Nisten
24:26
love it.
24:27
I just see a great future in this, where the actual tools are paired with a model that you can run. Yeah, on like a, a a 30 90. And the training for that model is built to be used with the tool. I mean, that's what cloud code is. Yeah. But we haven't seen a lot of that. That is portable. And this is a big deal to me because it is now becoming the whole stack as this model goes with this tool and you run it privately and you can have your cake and eat it too, and it does most of the work and you need something harder, then yeah, you, you, you can use Opus or you can, you can fire up, uh, you can just remote control it via the desktop. There's so many ways. So yeah, this is the future. I, I see the models and the harnesses being paired together. Love it. And this is one, the first one that's actually good. Uh,
Alex Volkov
25:19
alright folks, we need to bring, uh, our guest here, the next guest.
25:23
We have a bunch of, uh, guests, uh, joining the show. Peter Gusev. Yeah. I know you have a comment while Peter puts up headphones. Feel free to jump in here, but please super quick.
Yam Peleg
25:31
Next time you just want to distill, come on.
25:34
This one. You want to distill? I just want, it's fine. You can say it.
Nisten
25:39
I want full control over.
25:41
I know what the model is, what the hardware is, and I want it to actually be good because now it finally sort of is uh. I have before I give it my banking and all my emails.
Yam Peleg
25:51
Mm-hmm.
25:52
Mm-hmm.
Alex Volkov
25:52
Folks, we have.
25:54
So thank you. Thank you guys. I'm gonna take you off. Uh, feel free to like write me in comments if you have a comment, like, we'll, we'll, we'll bring you on. Please stay with us. Uh, everybody please say hi to Peter Sef. Peter, welcome to the show.
Peter Gostev
26:05
Hey,
Alex Volkov
26:06
please come closer to the mic so folks can hear you.
26:08
You're a soft spoken dude. Uh, we featured your stuff on the show multiple times before, as your pseudonym before, if folks don't remember Ameba GPT, you were doing a bunch of animations about Arena El Marina Yeah.
Alex Volkov
26:19
And how models progress and how evals are changing, et cetera.
26:23
And now you work at Arena.
Peter Gostev
26:24
Yeah.
26:25
Yeah.
Alex Volkov
26:25
But your first time on the show.
Peter Gostev
26:26
Yeah, exactly.
26:27
Yeah. Tell us us. Look, I'm a long time listener. First time caller, so Yeah. Excited, excited to be here.
Alex Volkov
26:33
Yeah.
26:34
Uh, so folks, uh, we usually start with the TLDR. We tell you everything that happened. We kinda ran through some of them with weeks and, and the, and the folks, uh, Peter's gonna join us. Uh, specifically I would love to hear your thoughts about missiles 'cause you did some stuff. But first I'm gonna go to TL DR to tell you about everything that happened in the world of ai. At least everything that's important because I know that many of you, this is why you joined, uh, uh, Thursday for the TL DR. So let, we'll do TL DR for a few minutes and then we'll, we'll tell.
27:09
All right, we are back. This is the TLDR for Thursday. I live from London on April 9th with me, Wolfram, Raven Wolf, Peter Gusto, our friend. We had Swyx before. We have, uh, three more guests on the show from Open Law. From Open the Eye and from Deep Mind, and this is the news for this week that you may have missed. The most important piece of news that you may have missed is Tropic released a blog post plus a preview of their new model, the rumored model called Myth. The Tropic released a preview of their new model called, uh, mythos. Mythos was not released to the public. It was deemed too dangerous to release to the public, but they have released it under Project Glass Wing to multiple big companies that will basically use these models, crazy cybersecurity capabilities to test and penetrate their own software before it's released to the public. So they will be safe and make us safer. Mythos is going around, this is the number one thing Yam referenced it to before that this is like the, the most important thing that happened in the AI besides AI engineer in London this week. So we, we'll definitely we're talk about Mythos now with Peter. Uh, but it's very interesting. In addition to this, and Tropic in the same block was said, the day are now a $30 billion a IR company, $30 billion a RR company, which is insane because we just, in February we told you they just passed 19 billion. So they, they ranked up like 10 billion in a RR in in what a month, which is absolutely, absolutely insane. They said they have more than a thousand companies that pay over a million dollars a month, which is absolutely crazy. Okay, so we're gonna talk about on topic, um, before the RPO 30 billion R is crazy. Uh, folks, we have another breaking news, uh, from Meta. If you guys remember, meta Lama Llama is dead. So Long Dead, but Meta released their first MSL Meta Superin Intelligence Labs with we told you about under Alexander Wang and a bunch of researchers. They paid over a hundred million dollars per year or two with the crazy numbers. We didn't believe they finally released something. It's called Muse. Muse. Spark specifically is the smaller of the model based on their new foundation. And uh, it's pretty cool. It's not the best, it's not beating like 4.6 apples, but given that, you know, uh, they built everything from Sketch in eight months, it's fairly impressive. So we're gonna chat about this a little bit as well. Simon Wilson did a very nice deep dive into this. Our friends from ZAI released GGLM 5.1 continuing on the trend of Chinese companies that don't release open source immediately, but we release some preview in their API run this and then open source after a while. Uh, so GLM is pretty good. We tested GLM five before, and GLM 5 1 1 seems very, very interesting on coding. It's number one on open source on Swyx bench. At 58% and, uh, runs autonomously, supposedly for eight hours. We need to test it in our harnesses still. Uh, Gemma four, Omar from Deep, my team is gonna join here. Cross 10 million downloads since we told you last week, which is unprecedented for open source. Good job folks. We really, really applaud this. Supposedly you can't say anything 'cause you work at a company that probably got access to it, but, uh, supposedly open the eyes training GPT image two and testing it in the wild now. So people started noticing like a hidden, uh, image on arena called masking tape. Gaffer tape and packing tape. No response from the arena represent here. This is in my notes, it's not really, I didn't even know that you're gonna be here when I talked about this. And it looks, it looks cool. Potentially a killer for none. Banana Pro. You should know this. Another model is being tested right now. That's also like making waves called Happy Horse. It's a mystery number one video model. Uh, the ex cling at Alibaba team. Somehow it licked. So the ex cling team in Alibaba, uh, is putting this model at some places. Artificial analysis, not necessarily for you guys. This model is beating ENTs two, uh, from, from Ance, which is previously like an crazy, crazy model. Uh, speaking of ENTs two, by the way, it's finally available in the us in and in the API and replicate.com. Shout to replicate, uh, under CloudFlare and uh, in, in cap Cut. So now you can finally use all those crazy movies that you see. By the way, I dunno if you guys saw the Harry Potter, uh, uh, drip thing. Oh yeah. Drip words. Did you guys see this? So. If you are not following Twitter as much as we do, the most hypey crazy thing that's happening right now in the world of like fake AI video slap is somebody rebuilding, uh, Hogwarts and Harry Potter, like ecosystem with reports. It's really, really, really funny. I'm gonna put a few clips in the show notes for you to just laugh with me. It's just a hilarious, uh, it's really good. And I think it's done with CED dance as well, um, in the weirdest news ever. And we're gonna get there soon. Have like two more items. Mila Yoyo Jovic, uh, from the Fifth Element and the Resident Evil series, the Hollywood actress and dude named Ben Sigmund, they have claimed to release the top state-of-the-art open source memory heent thing. I'll say this again. A Hollywood actress and her crypto friend used Claude to buy code a memory, um, age agentic thing. What the fuck? They have a GitHub and everything. It's like literally GitHub com slash mila dash jovic. It's really funny. Uh, Lilo, Dallas. Me, uh, okay, we'll move on. Open Cloud, released a version 20 26, 4 0.5 with dreaming. We're gonna bring Vincent Co. One of the top maintainers of Open Cloud to talk chat about this as well with us. Uh, cursor remote agent. Oh, no, I had a, I had a breaking news thing from BU that, uh, that we talked about. Uh, and Tropic released also the agent, managed agent infrastructure, it's agent and runtime and infrastructure fully hosted on Tropic. We haven't had a chance to chat with it or try it, but supposedly it's very, very big news. Uh, they want to sell agents not only APIs and models. So that's like a big, big thing that, that, that happened this week as well. And I think in this week's bus, which is the last piece about Weights, & Biases, folks, we shipped automations, which is a great way for you to run automation and get a notification about anything that happened on your run on WB models and GLM 5.1 and Gemma, uh, four are both available on our infant service. So please do check them out on w uh, one B, do AI slash inference. I think that's it. That's it. Tdr?
Peter Gostev
33:12
Yeah.
33:13
Is it worth, I can add a couple more points.
Alex Volkov
33:15
Yes, please.
Peter Gostev
33:16
In there.
33:16
So, uh, you mentioned JLM 5.1.
Alex Volkov
33:19
Yes, sir.
Peter Gostev
33:19
Um, so I haven't tested it on these super long
33:22
running tasks, so I think that's fair enough they put that aside. Yeah, I did do kind of one short tests, so I would say it's in my testing, it's like a little bit better. Uh, down, uh, 0.5 or rather five, the full version.
Alex Volkov
33:37
Yeah.
Peter Gostev
33:37
But it's not like miles, miles back, so it's like a little bit of
33:41
an upgrade, but yeah, when you see these numbers, so it's like beating opus or something like it's, it's probably not.
Alex Volkov
33:45
Yeah.
Peter Gostev
33:46
Um, but so this one then Gemma four, we also did the score
33:50
on the arena for this, and they did really well considering the size. It's like really, really impressive. Um, so they, we also published like a parade curve and they do like really, really well. They just like smashed the, smashed a lot of models that were like just on the parade and they,
Alex Volkov
34:06
yeah.
Peter Gostev
34:07
Jumped them.
34:07
So that was really good. And then we also really, the score of sedans, uh, text to video
Alex Volkov
34:12
sedans 2.0
Peter Gostev
34:13
Yeah, yeah.
34:14
From side. And it's like the, the jump, I can't quite remember. I think it's like 80 points higher, which is massive. Like previously all, all the models were kind of clustered together. Yeah. Maybe 10 points apart. And this one is 80 points
Alex Volkov
34:25
higher on top of which one?
34:26
Or video?
Peter Gostev
34:26
Um, yeah, I think video 3.1 is the next one.
34:29
Yeah. And, and I would say the way it feels to use Cuns is that it feels much more like a better version of Sora. And what I mean by this is that Sora was very good at kind of taking your idea and like making multiple cards and making a little story out of it. Yeah.
Alex Volkov
34:44
So
Peter Gostev
34:45
it had kind of also reasoning element to it.
34:47
But VIA was much more just like, give me a prompt and I'll just literally do what you say, like, really well. But it wasn't that. Creative. And I think that's why you're seeing those interesting stories come up, what you mentioned with like Harry Potter and things is because I think it's much easier for people just to put in like roughly a good idea.
Alex Volkov
35:05
Yeah.
Peter Gostev
35:06
And, and then the model kind of takes over.
35:08
It
Alex Volkov
35:08
also takes input videos and does video transfer, which I think
35:11
previous model, I don't think, I don't think the video, it probably technically can, but they didn't expose this via API. So VO doesn't like let you like, put a video in and then it takes those licenses. So c dance is like really, really cool folks. It's finally available. You should, you should test it out. Um, I'm getting some pings about the live show. So format, sorry that your microphone's like covering your face. Any chance that we can ask our friends here to like reposition the mic so that it doesn't cover his face? Is that possible? Um, so, so while we do this, Peter, I'm gonna talk to, uh, to you about Arena.
Nisten
35:42
Yeah.
Alex Volkov
35:42
Because every time that I heard you, uh, introduce yourself
35:45
as I'm from Arena, you have to say, oh, previously Ellen Marina. Thank you. This is a this is a, uh, I, I'll, I'll cover there. Don't worry. I'll keep talking. I up for a living here. Uh, yeah. This is amazing. Yeah. Thank you. Um, so tell us about where Arena is right now. Uh, how, what you're doing in there. Tell us co interesting things. I think Is this open source data set of like a test as well?
Peter Gostev
36:06
Yeah.
36:06
Yeah. So what, um. If I can move this, so Yeah, we can.
Alex Volkov
36:11
Very good.
36:11
Yeah.
Peter Gostev
36:11
Yeah.
36:11
Perfect. So, uh, yeah, so what we're trying to do, we, we collate a lot of data, um, from the users and lots of users come and test the models when new models come out, and it's really cool. Very exciting.
Omar Sanseviero
36:23
Yeah.
Peter Gostev
36:23
And the one benefit that we get is that we get a lot of, uh,
36:27
data over multiple years and what we're trying to do in a sensible, sensitive way, we release data as much as we can. So we've got a nice data, uh, uh, a bunch of data sets on hack face, so you can go check them out. One thing that we released was that. Uh, you know, you mentioned I was doing these kind of overtime leaderboards. Yeah. So the way I was doing this, I was literally going on the arena website, like scraping it by hand and like adding it to Google sheets. Like it was kind of dumb. And now
Alex Volkov
36:53
you have access
Peter Gostev
36:53
to, yeah.
36:54
So what, you know, one nice perk of of joining and working for a company like this is that I get access to the Databricks and I can just do it.
Alex Volkov
37:01
Yeah.
Peter Gostev
37:01
Uh, but uh, now we, we published the,
37:04
all of the scores going back. So it's really nice if, if you just want to dig into the history of the models and how things have evolved, just go take it. It's like the tell, tell your agent to, to go, uh, explore it. Um, yeah. So that's really cool. I'm very saw the data sets as well, so you can see like actual prompts and how the models have, uh, have performed.
Alex Volkov
37:24
Um, you can answer yes, no, or I cannot comment to the next question.
37:29
Have you guys tested Mythos?
Peter Gostev
37:31
Oh, no, no,
Alex Volkov
37:31
no.
Peter Gostev
37:32
I can say no.
Alex Volkov
37:32
That's a very, very specific no.
37:34
So the no means no folks, uh, but you have seen Mythos obviously.
Peter Gostev
37:38
Yeah, yeah, yeah.
Alex Volkov
37:38
Let's talk about Mythos.
37:39
Yeah. Uh, let's first define for folks what Mythos is, uh, and then let's talk about Tropics Opus. Big Brother. Big Brother.
Peter Gostev
37:46
Yeah.
37:47
So, uh, mythos is a blog post that they ship. Uh, so yeah, it's, I
Alex Volkov
37:53
mean, some companies do have access to Project Glass Wing,
37:56
uh, but they ship to go across. Yeah,
Peter Gostev
37:58
yeah, yeah.
37:58
So I think they said about 40 companies have access, uh, which is, uh, you know, I guess a way to release it. And I guess it's kind of, I guess the reaction, I dunno too overstated it being kind of a psychosis of, of a lot of people. Yeah. Kind of losing their minds of, of uh, some of the results. And I think the biggest one is just the big jump on coding. Uh, you'll, you'll have the numbers better than me, but Yeah. It's like, I, I think is a Swyx Bench Pro, the one that from scale, which is like the harder version of the switch bench. I think it jumped from like 55 to seven, seven or eight, something like that.
Alex Volkov
38:32
The jumps are crazy.
38:33
Yeah. Like you, you were just talking about like the same thing with Cs, where you see like a capability jump. I, I'll pull the numbers here right now, folks. This is a professional establishment. We have the run of show document. We have all the notes printed. Uh, the guest did not get their copy of the notes, which I probably should have thought. But yeah, we have, uh, 93.3 on Swyx inch verified, which we don't really care about. The swish verified, like everybody, like stopped doing, but 77% on Swyx Inch Pro up from 53% in, in, uh, Opus 4.6. Yeah. So was Right. It's really hard. Yeah. Like it's a hard benchmark. Um, but it's not all you. You gave me a link. Uh, I should probably show this somehow here, uh, about like all of the metrics that they release. Could you talk about this while I pull this up?
Peter Gostev
39:13
Yeah.
39:14
So, uh, what, what I, I, I got one of my agents to do is to go through the 244 page document. I love it. Document. Then just scrape all the, all the, um, all of the, all the benchmark benchmarks that they've released. So, um, you, you can see. Um, it's maybe not the nicest presentation, but it, it is just nice to just see them on, on the same page.
Alex Volkov
39:35
Mm-hmm.
Peter Gostev
39:35
Um, so yeah, we've got like browser comp, so this is ability
39:39
to navigate, uh, uh, the, the web and basically do, do that efficiently. So
Alex Volkov
39:44
around 10 points jump.
Peter Gostev
39:45
Yeah.
39:46
So that browser
Alex Volkov
39:46
comp,
Peter Gostev
39:46
yeah, that's really nice.
Alex Volkov
39:47
Humanity is less exam at 64.
Peter Gostev
39:49
Yeah.
39:49
And that, yeah, that one is interesting. So the, I, I guess when I look at the benchmarks, I always try to think like, what, what's driving this? Is this like a bigger model or is this maybe like post training or something like that? So you can't always tell, but you can kind of guess and probably humanize license exam is probably a bigger model.
LDJ
40:07
Yeah.
Peter Gostev
40:07
That's probably like a, a sensible, uh, assumption to make.
40:11
So yeah, Gemini did really well on it. It's probably the biggest out of, uh, the other models that we have. So maybe it's like a proxy for how big the model is.
Alex Volkov
40:19
Yeah.
Peter Gostev
40:19
Um,
Alex Volkov
40:20
uh, rumor that 10 trillion parameters, that's the rumor.
40:23
Okay.
Peter Gostev
40:23
Okay.
Alex Volkov
40:24
That's what I saw.
40:24
The rumor is folks, I want to like highlight for folks who are watching the jump at Swyx Inch Pro, which is a difficult version of Swyx inch verified, et cetera. The jump from Opus, uh, 4.6 to like GPT 5.4 is like a difference between like, like two, two or 3%. The jump between Opus 4.6 to Mythos is more than 10% is like 13%, 14%. This is a massive, massive coding model. And I think, uh, on topics going all in, especially with the release that like enterprise is coding and they're focusing on cloud code, et cetera, even despite the, the, the release, uh, and the reason why they're not releasing this one is they think. That this model is too dangerous to release. And, uh, for folks who kind of like, and I want your opinion on this Wolfram as well, uh, for folks who didn't follow Nicholas, um, not Carlini. I have this Rin, I think. No, no. Nicholas Carlini is one of the top researchers in the world. So again, for folks who are not following, Nicholas Carlini is one of the top security researchers in the world. He works in Tropic. He gave a talk at Black Hat Conference, uh, a week ago where he showed that these models, especially since December and January of this year, specifically jumped in capabilities of being able to do autonomous security research that is unparalleled to any other methods like fuzzing that they use, and that they crossed the level of what he trusts him to do as a security researcher. He said that previously these models paired with a security researcher were great. Now, these models can find several vulnerabilities. They're not interesting on their own, but they can string them together to break through things. And I think specifically they mentioned, um, they found bugs in, in open VSD, it's a 27-year-old bug on open BSD with a remote crash vulnerability. A 16-year-old bug in FFM PEG that survived 5 million automated FS tests. They found like this, this missus model found this bug and a chain Linux vulnerability into full privilege escalation with zero human theory. If this doesn't mean anything to you, um, that's great. Don't, don't get too scared. But like, if you know anything about security, the top security researchers in the world saying, Hey, it's time to wake up and tropic deemed this model is. Um, too dangerous to release.
VB OpenAI
42:28
Yeah.
Alex Volkov
42:28
This could be part of marketing because there are, I ipi right
42:31
now, but we'd love to hear from you, Wolfram, what you think about this. Like, too dangerous to release, uh, concept.
Wolfram Ravenwolf
42:36
Yeah, I think that's a stance of tropic as
42:38
a safety focused AI company. It's a lot of roleplaying and open a great model for roleplay. But the thing is, um, yeah, so security, I think open source is still more secure and, uh, not releasing the model. They will release it. They spend so much money to train it, just withholding it, that's not happening. And classing is their first step and then they can distill it and release other models and um, basically the model itself, the security provides, it'll make open source more secure and the bugs are there. It's not the AI that is putting the bug in unless maybe in the last year AI is doing it, but these 10-year-old bugs and older, they have been there and somebody else could have found them or maybe found them and we don't know. And so it's a good thing that it's open source and it's being found and I think it should even more, um. Before the release that people can use it on their own software because there are so many projects. And, um, yeah, in our notes we have that, uh, thousands of zero day exploits have been found by this. Wow. But if you have a model that is just, uh, 10 times smaller, it's still a hundred. And if it's a hundred times weaker than uh, myths, it would still find 10 and 10 are enough. One is enough, one really heavy zero day exploit. So I think this should be as open as possible. So the bugs are found and fixed. What would be useful is, uh, responsible disclosure. Basically use it, find the bug reported to the maintainers and fix it. And what do you think, what the Chinese are training as well and others. So definitely, uh, it has to be used to fix the bugs and instead of putting it away and hoping nobody looks.
Alex Volkov
44:16
So we have, uh, thank you Wolfram.
44:17
We have, uh, LDJ wants to add comments as well. I'll add one more comment before LDJ goes. Uh oh. Uh, LDJ, you need to only to self apparently, uh, because I cannot unmute you. Um, we did hear this before.
Peter Gostev
44:30
Hmm.
Alex Volkov
44:30
We heard before that, uh, Dali two is too dangerous to release
44:34
to the public with people 'cause the deep fakes are gonna ruin the world. Nothing happened. Yeah, we're fine. We heard before that, uh, cloning somebody's voice is gonna like, break the psyche of old people and everybody's gonna call them with deep fakes. And that and is fine. Every, every major company already has voice loaning, and if not, open source is voice loaning. That's in to the, from our voices. So, um. I gotta wonder if that's kind of like that, but the, the, the vulnerability attacks of recent year look or the, the chain supply chain attacks of Axios. Yeah. And light limb is kind of scary to me as well. So would love to pose this to the audience. What do you guys think should on tropic release myths now, did they, are they doing the right thing by not releasing this in only recent this to these companies? Uh, LDJ, feel free, I'll bring you up. Feel free to unmute and tell us, uh, more context about this.
LDJ
45:22
Yeah, there's a few things here.
45:23
Um, for one, I, I think a lot of people miss that philanthropic did actually stay, uh, with this release that they do, they do believe and plan that eventually even more powerful models than myth, those will be released to the public. Uh, just for now, amid those itself is not being released at the moment. Yeah. And it is just being deployed to these companies for, for these reasons.
Alex Volkov
45:45
Uh, so the companies are AWS, apple, Broadcom, Cisco
45:48
CrowdStrike, Google, JP Morgan Chase, Linux Foundation, Microsoft, Nvidia, all basically the, the top companies that, that run software in the world. Right. All the open, uh, open, uh, sorry, o operating systems, all the open source, like big foundational models. Like all, all of these companies get access and start running these models before the general public can.
LDJ
46:08
Yeah.
46:08
Yeah. I think another thing here too though is I think it's funny the people mentioning kind of, oh, this is the beginning of the permanent underclass. But if, if, if we go back to even models, like, um, I recall a oh three and or oh one and even GPT four, the original one, uh, there's been several occasions where it's confirmed with past OpenAI models. That several companies do get early access weeks or sometimes even months ahead of time of the release. Yeah. Uh, and I think it's maybe we're in an era where actually those early release to partner companies is just kind of becoming more transparent as opposed to before where is kind of have as a more secretive thing.
Alex Volkov
46:47
Yep.
LDJ
46:48
Um, and last thing I wanted to mention here on this is in the technical
46:52
safety report, the philanthropic released with Mythos, or sorry, with the Mythos blog post, there's a really funny incident where they had the model, uh, attempt to. Contact one of the researchers and they named, Hey, if you do this and this, then your goal is to contact this researcher. And for the most part, they didn't really think it would be able to get it. They used the standard sandboxing techniques that you would usually do to restrict the model from even having access to the internet. And it eventually found a series of exploits, and the researcher only realized it accomplished these things while he was in the park eating a sandwich and got an
Alex Volkov
47:31
email from the wall, a parking sandwich did.
47:33
I think honestly, the fact that the, the, the tropic folks chose to add this to the blog, and I think they posted a retraction, somebody told me that it's not like a full sandbox, uh, escalation thing, but the fact that they chose the sandwich thing just means that folks, let me be just like very clear, at least some of it is marketing, like at least some of it is, oh, oops, we're the biggest security company in the world and we have a model that can break open DSD, but oops, our blog post leaked that announced this model. And then, you know, oops, glad coat leaked Something here doesn't really add up for me.
Wolfram Ravenwolf
48:05
You know what I tell them?
48:06
Check your Soul md what instructions you put in there and fix it.
Peter Gostev
48:09
Yeah.
48:10
I think on, on the permanent under glass thing, I think it does worry me like I, I don't want this to become a trend because if you, if you go back pre, uh, GPT models at all, the way things used to work is that Google typically would say, oh, we have an AI breakthrough. Here's a blog post. We will enjoy it to ourselves. You get nothing.
Alex Volkov
48:34
Yeah.
Peter Gostev
48:34
Um, and that was so typical and I remember being really annoyed by
48:38
just the state of the industry where we would just literally get nothing ever. Uh, there would be, I mean, I'm exaggerating, but there would obviously be open source models and so on. But in terms of the leading breakthroughs, there would all be blog posts. Um, and uh, what OpenAI, I know people make fun of them, of OpenAI. You, you don't barely do any open source, but what they really shifted is that you can go, you can submit your credit card. At your credit card, you get an API key. You get access to Frontier Intelligence.
Alex Volkov
49:09
Yes.
Peter Gostev
49:10
This was not a thing before OpenAI did this.
49:13
Um, so what, uh, what, uh, we had after OpenAI did this is that we had several years where everyone was expected to ship their models. So everyone was shipping their models. Yep. So Anthropic, Google started shipping their models, like everyone, right. So that was like the best time to be a, uh, a user of the models. And now we might be getting back to the point where it's like, oh yes, unless you're Microsoft AWS you are not on the list. And yeah, maybe it's normal and maybe they just released it 'cause it leaked. Possible. Um, but it could be the new normal where the frontier, uh, normal people don't get access to the frontier intelligence,
Alex Volkov
49:57
which, so, um, I, I hear you on this and also to kinda
50:01
strengthen what you're saying. Uh, so folks, supposedly OpenAI has a model the upcoming that, uh, called name Spot. I think like Greg Bachman post about this, like it's, it's known, uh, and I think that they also have some security models that are less restrictive. Uh, I think even Peter Steinberg will talk about this on stage. Um, and I think I saw an announcement that openly I will also have a model that is too dangerous to release.
Peter Gostev
50:25
I, I think that was.
50:26
Some,
Alex Volkov
50:27
some
Peter Gostev
50:27
bullshit vague reporting.
50:28
Yeah. I, I wouldn't take that to, to literally, yeah.
Alex Volkov
50:31
Yeah.
Peter Gostev
50:31
Let's wait.
50:32
But,
Alex Volkov
50:32
but, uh, we have LDJ that, that has a clarify on this.
50:35
LDJ, go ahead. On mute. Oh, you're on muted. Go ahead.
LDJ
50:38
Yeah.
50:38
So, um, Axios did recently kind of report something basically like what you just described, but they added a, an edit and retraction on this. Um, here I have it right in front of me actually. They said, uh, editors note the headline and story have been corrected to clarify that OpenAI is releasing a cybersecurity product separate from its new model to select partners. Not that it's staggering the release of a new model to select companies.
Alex Volkov
51:01
And that's cybersecurity product.
51:03
I think that's what, uh, Peter is, Steiner was talking, uh, on stage about. Mm-hmm. And that has significantly less, uh, uh, refusals in the model's ability. 'cause like what they're asking, the model is basically, Hey, hey, hack the software. Yeah. And as a user, we won't be able to ask, Hey, hack the software. The model needs to do refusals right. For us. Uh, but what they're afraid of is releasing the model and then fucking Pliny, the jailbreaker is gonna jailbreaker. Everybody's gonna rip through bugs. Uh, but yeah, so, uh, uh, they, and LDJ is saying is they confirmed it's called a, a work, uh, which is great.
Peter Gostev
51:34
Yeah.
51:35
Um, do, do you remember it was, uh, I think it was, uh, GPT 5.3 Spark from memory where they released it in Codex, and then I think it was some like couple of weeks when they didn't wanna release it via API and the sighted reason was that it, it was a little too keen to, to hack things. Yes. So they wanted to put some extra controls. Uh, I think that was the, um, uh, the 5.3 Codex model.
Alex Volkov
51:59
Yep.
51:59
Uh, Peter, you also sent me the Compute Wars chart updated with on Tropic and Google deal.
Peter Gostev
52:04
Yeah.
Alex Volkov
52:04
Let's talk about this.
52:05
Uh, this is the GPU deal that, uh, on tropic announced with Google by 2027 and insane amount dollars. I don't think I have it written here for how much, but maybe you remember, uh, that tropic will, uh, get TPUs from Google and not necessarily indeed GPUs. Folks forget. But Tropic has 10% on Tropic, uh, uh, sorry, folks forget, but Google has 10% of on tropic that they like invested a long time ago, and now Tropic will essentially have GPUs. So I, while I show you this, while I show the folks this, can you tell us like what, what this is about?
Peter Gostev
52:33
Yeah.
52:33
So, um, the, the multiple angles to this, so one is that, um, the, the, the reason for plotting this chart is that the, everything ultimately comes back, uh, about this No, no worries. Some music. Music accompaniment.
Alex Volkov
52:49
Yeah.
Peter Gostev
52:50
So everything ultimately goes back to compute.
52:52
Like not compute is not like completely destiny. Like you can still screw it up, but ultimately if you don't have enough compute, then you can't do all that much.
Alex Volkov
53:01
Yeah.
Peter Gostev
53:02
And what I was trying to do, and this is kind of re reconstructed from,
53:06
um, the public data that we have, so it's not correct, it's probably underestimate.
Alex Volkov
53:11
Mm-hmm.
Peter Gostev
53:11
But this is kind of trying to ballpark size.
53:13
Yeah, ballpark. Like where, where they are relative to each other. And what was interesting is that, um. There, there kind of a couple of points. So one is there was a jump, uh, more recently, it looks like a lot of compute went online more recently and it's hard to know like precisely what, what they were training. It could be that opus 4.5, 4.6 benefit from this jump. So that's possible. Or it could be just all of it went to myths and that's why we see such a big jump. So that's, that's possible. Um,
Alex Volkov
53:41
oh, you mean this jump?
Peter Gostev
53:41
Yeah.
53:41
Yeah. From
Alex Volkov
53:42
4.1 to 4.5
Peter Gostev
53:43
years.
53:43
Exactly. Yeah. It might be, we don't know that time in precisely and so on. Yeah. But the point is tropic didn't have that much compute and now they, they had like loads more so it could be all the myth is story comes back to they just got more compute. There's nothing like two special apart from that point.
Alex Volkov
53:58
Yep.
Peter Gostev
53:58
Um, the second point was that if you go, uh, late in
54:02
the chart, you can see that OpenAI just like way, way higher mm-hmm. Than, than Anthropic. And uh, this was always kind of weird, like obviously Anthropic just gonna buy more compute, but it looks like, like diary was like a little bit more conservative. Or actually I don't wanna bet my whole company on like, what's gonna happen in 2029.
Alex Volkov
54:18
Yeah.
Peter Gostev
54:18
So they were a bit understanding.
54:20
So it looks like they've committed and purchasing a lot more compute.
Alex Volkov
54:23
Yeah.
Peter Gostev
54:23
And again, kind of compute is very closely associated
54:27
with your destiny as a, as a company ability to train big models, ability to influence the models. Um, and I think that that could be another angle. You know, we complaining a lot about Anthropic, like stifling open cloud users, like all of that. Yes. And, and like all of the rate limits, I know there's a lot of complaints around that.
Alex Volkov
54:45
Yep.
Peter Gostev
54:46
But I think the answer is they just don't have any compute.
54:48
So maybe another kind of more, um, banal point about, uh, myths is that, you know, what's the point of releasing it if your rate limit will be half a half a prompt? Um, right. Uh,
Alex Volkov
55:00
so essentially a 10 trillion parameter model that is very, very good at
55:05
what it should be very, very good at based on any scaling law that you can imagine. Uh, and based on training in the rail and the trajectories that we've seen in test time, compute, et cetera, uh, that they cannot give us. So they could refrain this as, oh, it's too secure to give out, but they could just potentially cannot just like, serve at scale.
Peter Gostev
55:25
Yeah.
55:25
I think it's, it's, uh, sounds cooler to say that it's to, to risky to release it rather than risk
Alex Volkov
55:31
in the year of IPO.
55:32
This is a great reframe. This is a great reframe in the year of IPO. You know, the, the company that cannot serve a model, uh, says that the model is is, uh, too, too risky to, to risk.
Peter Gostev
55:44
Yeah.
Alex Volkov
55:45
So, uh, yeah, some folks are asking us about the price of Nifo.
55:48
They did post the price. Yeah, I have it here. Go on, have it here.
Wolfram Ravenwolf
55:51
25.
55:52
$25 per million input and 20, uh, 125 per million output tokens. So five times the price of Opus 4.6.
Alex Volkov
56:00
Yeah.
56:01
Which, which also means like the, they posted the price, but it doesn't, like we can't pay for this. Um, I posted a funny tweet about like, Hey, who's running Mytho on their open claw? Uh, I, I would like to, I, I don't know if I can afford it. Uh, but you know, if, if there's a sponsor out there who wants to gimme money to run open cloud, uh, on, on, uh, on mytho, um, please lemme know.
Wolfram Ravenwolf
56:23
And I would run it on Wolf bench, but I would
56:25
need more sponsorship for that. I guess More
Alex Volkov
56:27
sponsors.
Peter Gostev
56:28
Yeah.
56:28
And I think there, there's a point of, okay, I think this is probably pricing, like for their partners, but if they don't have compute, it's like, doesn't matter what the price is, they just don't have it. And I think that's, that's the reality. We get into these numbers that where they can't just like hustle and like hustle their way into a couple more gigawatt of data centers. It's just not there. I think that's probably the real, like half a reason why they wouldn't even consider doing it. Um, yeah. So they
Wolfram Ravenwolf
56:55
need to come to CoreWeave.
Peter Gostev
56:58
Well, I think they may maybe need like a few core weaves,
57:01
uh, to, to get, get enough compute. So yeah, I think that that's, that, that, that's certainly, uh. Another maybe prosaic angle about, about the current situation. There's another angle with, with this chart is, and this new story about, uh, deal with Google. And I know like people say, or it's like, yeah, Google owns like 10% and um, and that, you know, they've got their cloud business to grow. And if you spin out the TPU business virtually even, it's like, oh yeah, you're gonna be same values in video or there about, like, you, you can probably unlock a lot more value. But I do kind of, I do kind of find it weird. It's like, why don't they like their own DeepMind team? It's like, why? Just like you are taking TPUs brother
Alex Volkov
57:43
Google is a huge company.
57:45
There's not one entity that's called Google. And within Google there's like a lot of politics as well. And they need their cloud business as well to work. Yeah. Also 10% of On Tropic, that's a lot of money that's going on right now. Tropic grew from 19 billion in February to over 30 billion in a RR, uh, in April, which is mind boggling.
Peter Gostev
58:03
But you see, it's like, I dunno,
Alex Volkov
58:05
do you hear this?
58:06
Uh, Tropic folks completed the secondary tender sale where employees are incentivized to sell their shares on the secondary market for investors to come in. Before the A PO uh, uh, investors were hard pressed because, uh, employees did not wanna sell before the a PO. That's how bullish the folks in believe in their company. It's a good sign. And the a p it's a great sign. Uh, and also apparently open the eyes kinda lower on the secondary markets as well.
Peter Gostev
58:29
Hmm.
Alex Volkov
58:29
So Google's getting a lot of that from, from that benefit.
Peter Gostev
58:32
Yeah.
58:32
But like, I dunno, like as a. As a person in AI enthusiast and, uh, hope for that. Google DeepMind does really well. It's like, why do I care about the shareholder value accumulation of a decades? It's like, come on, just believe in yourself. Build the bloody a GI don't just sell it to your competitor. I, I know, I find it quite weird, but there's kind of history of this that they, before I, I remember this, that like a week before Gemini 2.5 Pro, uh, got released, kind of Google had pretty bad models. Uh, so I can see that their leadership was like, uh, we're going to do this. Well, uh, they literally invested I think one and one and a half or 2 billion in Anthropic. Uh, which, which by the way, okay, we are now look at Anthropic as like completely independent, very impressive company. They were pretty much just like, uh, the only option in AWS So like if you're AWS customer, yeah. You had either like eight B models or Anthropic, like there was basically no other choice. And I think if, um, if AWS didn't have access to Anthropic models at the time, I think many customers would've gone. So Google was not only propping up Anthropic compared to G Mind, they're also propping up AWS compared to Google Cloud.
Alex Volkov
59:50
Yep.
Peter Gostev
59:50
So I know there are like shareholder value reasons
59:53
and diversification, but I, I just find it kind of boring. Like I feel like you just prop up your teams, give them to be used.
Alex Volkov
1:00:01
Folks, uh, Peter, we, we have to, like, we've been up for an hour.
1:00:04
We have to move on to This's Buzz. I wanna tell folks where they can find you. Sorry about this, getting a little bit of a thing. Uh, how can folks find your work? What is Arena doing? Give us like a little spiel about
Peter Gostev
1:00:16
Yeah, sure.
1:00:17
Follow, uh, follow arena. We, uh, publish, uh, leaderboard updates all the time, like seven times a week. Uh, so, uh, yeah, follow arena account. Uh, you can follow me as well if you want. Uh, more spicy takes that I'm not allowed to publish on the arena account. Uh, but yeah, we, we try and do, uh, more interesting things apart from just, just here, the leader board. Um, so, uh, yeah, and we also release data as well, so go to our hugging face, we waiting keen for people to, to use it and find some value in it.
Alex Volkov
1:00:49
All righty.
1:00:49
Thank you so much for coming. Cool. Thank you. Uh, and he'll be around this, uh, the show, uh, folks, uh, tuned in for the election. Alright folks, we're gonna go to this week's buzz for five minutes and then we're gonna have VB vibe of from OpenAI joining us. Uh, so we're gonna start with this week's buzz, and then we're gonna have VB joining us from OpenAI. Uh, a friend of the PO who's been here before. Yeah, you're gonna sit here? Yeah. All right, let's go to this.
1:01:33
All right, this is gonna be a quick one, Wolfram this week, uh, at Weights, & Biases, what we have to tell you about is automations. Uh, you now can in Weights, & Biases models, by the way, for, for the machine learning engineers, you can build automations. They send you notification, uh, where you can just like go on the run and then get updates about your specific stuff. This goes very well with the iOS app that we recently released, and, um, you should go and check it out. But more importantly, I think more importantly, but, uh, at least as important, we have two new models up on an intern service. We have GLM 5.1, which is recently released from Z ai this week. So our team is doing really, really good work at releasing open source models. And we have Gemma four, and Wolffer has been testing a few of them, and we'll get you some of those tests as well. And speaking of things that Wolffer will get, um, you posted, uh, some stuff on our blog. You wanna tell people about us? Yeah.
Wolfram Ravenwolf
1:02:24
So remember when we talked last week about,
1:02:26
um, how models can get dumber if you give them more time to think. Uh, I posted an in-depth article about this, so we talked about it on the show. And now on our block at Weights, & Biases one b.com, you can take a look and read exactly what is going on. So it's more a really in-depth article. Too much information to talk on the show. Yep. But detailing everything I talked about,
Alex Volkov
1:02:49
reasoning is not always better, or at least more
1:02:51
reasoning is not always better. Sometimes it gets models to overthink. Uh, and then we're, we're definitely gonna chat about this as well. We're gonna put the link in the show notes for you. So please check out Wolf is doing a great job with Wolf Bench, by the way. Uh, w we've been pushing Wolf Bench here to everyone. Peter Steinberger heard about Wolf Bench. Uh, folks who open the eye gonna hear about Wolf Bench in a second. Uh, Wolf bench is like a really great tool. And also we heard from you and we love hearing from you folks. We had so many folks. Uh, so the audience for Thursday, I and the AI engineer overlaps a hundred percent everybody who. Joins AI engineer is a potential for Thursday Eye, but I think that everybody who listens to Thursday Eye could, could come to this, uh, event and enjoy, right? There
Wolfram Ravenwolf
1:03:28
are so many people here that, uh, even when I was getting in, in
1:03:31
the, in the queue, a person around turned around and said, oh, you're a Woo Ra. Hi. I know you're from there. And we talked and it's great to meet these people here and everybody who enthusiastic
Alex Volkov
1:03:40
and what I heard from one of you, and if you're listening
1:03:43
to this, thank you for approaching us. The Wolf Bench is one of the more unique things that we do. Do we have like our own evaluation tools specific? So, uh, great shout out to you and we'll keep posting about this, uh, and we'll get more labs to actually jump on. Uh, shout out to Cursor back there. Uh, cursor is a harness. We're also gonna evaluate Cursor. So you guys know that we don't only evaluate models, we only, we also evaluate harnesses with different models. So we're gonna evaluate, uh, cursor as a harness in addition to the models. So that's coming up soon. Uh, stay tuned. This was this week's buzz, and let's chat with our friend vb.
1:04:30
So Vibi. How's hugging? I'm just kidding. I'm just kidding. Uh, folks, uh, Vibi, uh, uh, rich vib on Twitter. Uh, vibe above Shrive. Vata? Yeah,
VB OpenAI
1:04:42
that's accurate now.
Alex Volkov
1:04:43
Okay.
1:04:44
Accurate. Ouch. Uh, what's up man? How you doing?
VB OpenAI
1:04:48
Pretty good.
1:04:49
Pretty good. Nice to be here. Um, yeah, it's been, it's been a while since I was here.
Alex Volkov
1:04:52
Yeah.
1:04:53
Yeah. It's been a minute. Uh, not that long. I think you came over and we talked about Codex. Alright. I also jumped, uh, on a few of your now famous spaces with Roman, uh, head of Devereux. Uh, uh, tell us about, tell us about your job at OpenAI. I don't think like, um, the people, we have a bunch of new folks joining and also would love to contextualize your job at OpenAI at all what you're doing at a IE. You guys have a workshop. We would love to hear about that.
VB OpenAI
1:05:18
Sure.
1:05:18
So, um, I'm part of the developer experience team at, um, OpenAI. I'm focused on Codex right now. So everything that you see from Codex, the app as well as Codex, the CLI, yeah. Um, as well as all the models that we ship. Um, that's something which, um, I help with. Um, and at the same time, like I'm, I'm now based here in London and, um, you know, helping grow an ecosystem, um, around overhead with the developer community as well as like our startups, enterprises, and so on and so forth.
Alex Volkov
1:05:48
So, uh, we have to contextualize, we are in new show.
1:05:51
The, the reason why we're bringing you on is, well, first of all, you're a friend of the pot for a long time, dude, you jumped in on Twitter spaces before there was video, uh, before the production ramped up. Uh, so hopefully you're like also enjoying kind of like the, the quality of production. Definitely. Uh, but also like way before new like hugging Face Day. So folks definitely remember you is like one of the crazy high people that like brings open source to the masses. Uh, so we'd love to, you know, just hear from you about like open source in, um, in Open the Eye, but like, one thing that's big, and I think you went super duper viral when the cloud code leak happened and you posted Breaking Open the Air also released a bunch of open source and Codex was open source all along. We made the joke because that was really funny. Um, tell us about Codex. You guys just hit a milestone.
VB OpenAI
1:06:32
Yes.
1:06:33
Um, so, uh, first of all, like regarding that tweet, like it was just like a tongue in cheek thing. Like, I thought, like I, it'll get like maybe like 5,000, 10,000 views and, you know, uh, it'll be eight. It was just like, you know, like, just pointing out the obvious that, um, codex, the CLI, the harness as well as, um, you know, the Sandbox, pretty much everything about, about Codex that you use either on Mac Os, or Windows is fully open source. Uh, you can go on github.com/uh, OpenAI slash codex and you have like a. You can, you can sort of build on top of that. We also have, um, um, we also have Codex App Server, which you can use to sort of like interact with, uh, with Codex like the Codex ID extension builds on top of App Server.
Alex Volkov
1:07:15
I would love to hear about App App server specifically.
1:07:17
'cause it's kind of like a backwards auth that you can use your open account to use a bunch of stuff, but people have been waiting for a long time.
VB OpenAI
1:07:24
Yeah, you can, you can, you can use it as like an SDK Yeah.
1:07:28
To like, sort of like program around, uh, around Codex, use it to um, you know, send messages, receive messages, like pretty much like build experiences similar to Codex ID extension, uh, the same way. Um, and yeah, I mean, you know, just, just generally speaking, COEX has been on a tail, right? Yeah, we, um, we recently crossed a pretty, um, massive milestone. We have 3 million weekly active users,
Alex Volkov
1:07:50
3 million weekly active users.
VB OpenAI
1:07:51
That's
Alex Volkov
1:07:52
just now after 2 million weekly active users.
1:07:54
What, last month? Like in the middle of this month only?
VB OpenAI
1:07:57
Yeah,
Alex Volkov
1:07:57
that's, that's, um,
VB OpenAI
1:07:59
yeah, it's, uh, even, even even within, um, within, within OpenAI,
1:08:03
like can, you can see like the sort of like growth and like the uptick with, of, of Codex within, um, has been, has been like meet your Right. At the same time, like all of this started out when we released Codex the app, which was, uh, earlier this year. Like a really nice and intuitive way to interact with Codex, with like multiple projects and like have multiple sites at the same time. Use work trees, um, to sort of isolate your changes and so on.
Alex Volkov
1:08:28
Yeah.
1:08:29
Uh, codex, the app is dope. I love Codex. The app, uh, some updates there is plugins, I think is the recent one. The, the newest function is also about plugins.
VB OpenAI
1:08:36
Sure.
1:08:36
So,
Alex Volkov
1:08:37
because like, I will preface this, OpenAI had plugins before Chat.
1:08:41
GPT Plugins is the first iteration of App Store. Then custom GPT, second duration of App Store, then MCP apps or whatever built on MCP apps in chat. Now we're back to plugins. I love this game for OpenAI to confuse the fuck out of us, but that's fine. We Codex used to be a coding model, then became a product. It's also a coding model. But now plugins in Codex does,
VB OpenAI
1:09:02
I think there's like, um, there's a bit of nuance there, right?
1:09:05
So if you, if you, if you look at a plugin, at the end of the day, a plugin, um, is like, if you sort of disintegrate a plugin, all you get is like a bunch of skills. Yeah. Right? Yeah. An MCP server and like, you know, things to sort of like tie it up together as a package. Right? And, and the, and the reason why this becomes, um, quite important is if you, if you look at, um, plugins like build iOS apps or like build web apps, like these are plugins which come curated package within your Codex app or Codex CLI, uh, what they allow you to do is they allow you to bring, like experiences like Stripe, they, they allow you to bring experience like superb base and like couple of our, you know, uh, shape, C-N-C-L-I shape cn, you know, best practices and so on. Mm-hmm. So that you can then call these plugins. Um, and like this, essentially this package and like build, um, apps, web apps and so on and so forth together, uh, along with it. Um, and this is kinda like a nice way to, uh, to package, um, um, lots of things like connectors, like for example, you can connect directly with, uh, with Google Drive and like, you know, Google Drive can then like have, have access to your calendar, have access to your email, have access to, you know, slack, for example. And then you can use that to, um, for a lot of, um, you know, knowledge work, uh, related tasks and so on and so forth. Right. And, um, we're starting to see this become like a core part of people's workflows.
For example, every day at 9
1:10:24
00 AM Codex uses an automation Yeah.
1:10:28
For me.
Yam Peleg
1:10:28
Yeah.
VB OpenAI
1:10:29
Which would look through all the Slack messages in the past 24 hours,
1:10:32
look through, like I ask Codex to look through all the messages and like threads where I've been tagged and like sort of market explicitly, which all tasks, uh, do I, should I like, pay attention to. Yep. Where have I been like explicitly called out to like, do something Yeah. Where there's an action item, you know, where am I lacking and contextualize that with my, with my Gmail, uh, you know, Gmail. See like what's, what's important, what's not. Um, and layer that in with my calendar and like, try and give me spaces where I can work on, um, on these things. And, uh, honestly speaking,
Alex Volkov
1:11:03
that's not coding related at all.
1:11:05
Yeah. This is like a general work thing that every knowledge worker would benefit from.
VB OpenAI
1:11:09
Exactly.
1:11:09
And I've, and I've also had like Codex just create like. Pre-briefing docs for myself. Right. You know, like if I'm, if I'm going into a meeting with someone and I, uh, I'm lacking context or like I have, I'm like, in back to back meetings, I would just have code ex, uh, sort of like pull up something, create a small calendar event. I, I wish I had my laptop, I would've, uh, showed it to you. But like, uh, it, it just creates like a small, like five minute event, which would just have like, uh, you know, pre-brief for myself for that event.
Alex Volkov
1:11:32
Oh, that's,
VB OpenAI
1:11:33
no.
1:11:33
And, um, this is all possible because of plugins.
Alex Volkov
1:11:36
Well, hold on.
1:11:37
That's a good idea. I'll, I'll tell the folks actually that they can use this as well. So you have an automation Yeah. That looks at the upcoming events on your calendar, schedules you a five minute event for yourself so nobody can block you on that time.
VB OpenAI
1:11:48
Yeah.
Alex Volkov
1:11:48
And gives you a brief of what's going to be talked about in
1:11:51
this, who are the people, et cetera. Yeah. And the big org like, open the eye that maybe, you know, some people, especially across organizational role like dx, you may know some people, maybe you don't know some people you need some background. That's incredible. I should, I should automate this for myself.
VB OpenAI
1:12:02
Yeah.
1:12:02
Or, or, or for also like, uh, for example, if I'm like, just, just like, you know, catching up with someone, uh, with like a partner of ours, like, I just wanna make sure that I'm, I go in prepared and I'm just like, you know, I don't just make a fool of myself, you know? Yeah.
Alex Volkov
1:12:15
Yeah.
1:12:16
Um, maybe I wanna ask you, uh, just before we're gonna drop and continue with like our next guest, a few more things. I will. Uh, so first of all, folks are shouting out. Vvi the Goat folks know you. So shout out, uh, you've been doing incredible work throughout like these years. Um, I will not ask you about Spud and, uh, change your Pity a hundred dollars plan, because that's what people want to know, but obviously folks we cannot ask. Put vvi on the spot here. Uh, but you guys can go on Twitter and read the rumors for yourself. Okay? We don't need to put put, put Sweat on, on vv, but, uh, I, I, I wanna ask you about the Super App because that was talked about, right? Like, this is public, the Open the eyes is gonna like focusing on a bunch of stuff. SOAR was cut, et cetera. And I think it was announced that like, uh, codex the browser and uh, and the, the Chat GPT app itself on the Mac, it's going to like become like this like super app, um, hearing from you how you're using Codex, which is effect effectively like a, like, like a developer tool, or at least now if you ask people what is Codex? They'll say the developer tool not use it for developer at all. Like basically no code is written in that flow that you just said, right? Like a connection to Slack, a connection to Gmail, that's all like just MCP calls or whatever calls behind the scenes. Um, tell us about like this move towards the Super app where these agents will like, be able to like actually help with the real world test.
VB OpenAI
1:13:30
Well, so, um, I don't have like much to add on
1:13:33
like Super App itself, right? Yeah. Or like whatever that, that variation would be called. But I can, I can throw some color on like how, uh, how Codex the app itself, like, you know, works right now. So if you, if, if you think about it, there's, um, like for for any task that you can think of, you can, you can think of. Any task as a code related task, right? So, um, if either you have a connector for it or you have like an NCP server for it, or you have like some sort of, um, you know, skill or, or, or some sort of integration with, with a thing, or you can think of it as something wherein you can just generate the code to do a set task, a disposable unit of code, and then just like solve that task and then proceed to, you know, work with the, with the result of it and so on. And, um, as we, as we sort of like, you know, as more and more people get more comfortable with, um, you know, using tools like, um, you know, codex just for their day-to-day task, yeah. You can see that, uh, this, this whole thing starts to become much more intuitive, uh, to them. Like for example, right now I was just talking to a recruiter on our team, uh, um, who, who came for the AI engineering conference.
Omar Sanseviero
1:14:39
Yeah.
VB OpenAI
1:14:39
And typically what they would do is they would, you know, um,
1:14:42
go through an entire, uh, list of like people who are, who are attending or who, who are speakers as like, uh, who might potentially be like recruiting targets. Um, um, for OpenAI they would typically like, you know, find information about them. And this would be like a, like a manual process, right? And like they would have like copy paste links and so on and so forth. In this case, uh, the recruiter like literally just like, uh, put the, put the page, uh, and, and gave the URL to, to Codex Codex automatically, like, you know, created a Google sheet with like all of these, you know, names. Uh, did, did some like preliminary research with, uh, with Google, like quickly find. Uh, their LinkedIn or like their Twitter and, and just like, you know, get as much information as as possible. Yeah. And you can see that, you know, workflows like this become more and more sort of possible as we, as we continue sort of refining this experience.
Alex Volkov
1:15:30
Yep.
1:15:31
And I, I'm assuming, and I'll gonna assume like you can skip the, the comment here that, um, you guys are seeing internally as well, that more and more folks adopt this and obviously open the eye is like the most a GI pill organization in the world. So obviously the recruiters and obviously the other folks that will also see what's going on. And if they don't, uh, uh, accelerate with ai, then somebody else will take the job, will sell it with AI better. So as you see like, uh, non-techie roles, pick up those skills, obviously, like it'll be rebuilt to support those. Uh, last thing for you, uh, before you go, you and your colleague Katya, uh, yeah. You had a workshop here Yeah. At AI Engineer would love to hear your feedback about this like event generally. OpenAI as far as ai, remember Swyx, and correct me if it's wrong, if I'm wrong, opener has been a sponsor of every AI engineer so far, and a keynote speaker in multiple of them. Your colleague Brian Lolo was the keynote speaker. Great guy. You guys definitely should check out the livestream for Brian. He was like, code is free. You should change your mindset completely. It's great. Great talk, great talk. Um, tell us, reflect on this event a little bit on your workshop. Like how did it go, what'd you do? How's the demos? What are the reactions for people? Hallway track would love to hear from you. Like what are you hearing from people here?
VB OpenAI
1:16:39
So definitely, I mean, just to sort of like recap on the, um, on the,
1:16:43
on the workshop and like the chats that we've been having, um, so far, like we, we, we of course like spoke about plugins. We spoke about subagent, subagents, something, which I'm personally quite excited about. Um, so for those of you, um, who don't know, like could ex natively support subagents, you can
Alex Volkov
1:16:57
with cool names.
VB OpenAI
1:16:58
Yeah.
1:16:58
With with, with the very cool names, including Jason. Uh, and uh,
Alex Volkov
1:17:04
that's a deep cut man.
1:17:05
Not everybody will get this one, but Jason, we hear you.
VB OpenAI
1:17:08
Yeah.
1:17:08
And, and, and so like, essentially, um, what does Sation allows you to do is like, let's say you have a task, uh, I want to look through, I wanna do X right? It, it allows like Codex to break that task X into decomposable independent tasks. Mm-hmm. And then like sort of spin those off, off into individual codex agents who can work on it at the same time, uh, for you to sort of get, get like, you know, latency gains as well as like, just like speedups. Right? And this is something which, which I'm personally quite, um, quite excited about as we like continue, uh, you know, refining the experience, right? Yep. Um, at the same time, like I also showcase like a bunch of, uh, a bunch of things which are experimental right now. Uh, so there's, um, there's guardian approvals which allow you to make, uh, your whole experience, um, of, you know, working with, uh, with models, tools, and so on and so forth. Uh, much more nicer, um, like. Let's be honest, a lot of, uh, a lot of folks would use like Codex or cloud code in, uh, or like any of their, like, you know, uh, coding agents in like YOLO mode, right?
Alex Volkov
1:18:10
Yeah.
VB OpenAI
1:18:11
And that's, I mean, to be honest, that's not safe, you know?
Alex Volkov
1:18:14
No, not really.
VB OpenAI
1:18:15
Yeah.
1:18:15
And so, um, guardian mode, um, or like Guardian approval, uh, mode within, um, um, within Codex allows, um, us to sort of look through each and every individual tool call. Yeah. And like we spin off a sub agent Oh. To, to quickly give it like a risk profile.
Alex Volkov
1:18:29
Oh, that's
VB OpenAI
1:18:30
great.
1:18:30
And if it's like medium risk or low risk, like we auto approve, otherwise it'll escalate it to you.
Alex Volkov
1:18:34
So it's like a, like a classifier for potential.
VB OpenAI
1:18:38
Yeah,
Alex Volkov
1:18:38
exactly.
1:18:38
Danger of calls. If it's like reading a file, that's fine. If it's like sending a CRO request, it's like, oh, if it's RMRF all your dash, like your whole like subgroup director doesn't, like, you have to approve it. Yeah,
VB OpenAI
1:18:48
exactly.
1:18:49
Yeah.
Alex Volkov
1:18:49
Oh, that's great.
1:18:50
What's called guardian,
VB OpenAI
1:18:51
guardian approvals.
1:18:51
You can go on slash experimental and, and you know, like, uh. Uh, set it out.
Alex Volkov
1:18:55
Many of you heard it here first because I heard
1:18:58
it here first because I haven't heard about guardian mode before. Guardian mode is dope, folks. It's like really, really dangerous to run yellow, even though, um, I remember Sam saying he was afraid to run yellow for a while, and then Sam was like, running yo for his code. I remember like,
VB OpenAI
1:19:12
well, I, I, I honestly don't know about that.
1:19:13
I do not recommend, uh, running no, in like a, without a sandbox. Um, but, but you know, like, and, and, and just that. And then we've got like experimental hook support, uh, which is quite nice, which means that you can program, uh, codex to, you know, do a task X after, you know, at the start of the session, mid of the session after a tool call or just at the, at the stop of the session. And, uh, like, you know, as these things come together, like, you know, I, I could not. I just cannot personally be more bullish on, uh, on, on Codex the app, on the experience. Man, that's cool. Yeah. And, and, and like last but not least, like AI engineers has been like a really nice experience for me. Like I met a lot of my ex-colleagues from Hugging Face. Yep. Um, I met oer, um, you know,
Alex Volkov
1:19:55
ER's on Silver currently at Google DeepMind previously hugging Face.
1:19:59
We'll join this podcast very, very soon also.
VB OpenAI
1:20:01
Yeah.
1:20:01
And you know, like, it's, um, it's quite nice, um, to, to just like hang out with everyone and Yeah, just the wives are off the chart,
Alex Volkov
1:20:07
you
1:20:07
know?
VB OpenAI
1:20:07
Yeah.
Alex Volkov
1:20:08
And also, I'll say a personal anecdote that has
1:20:09
nothing to do with AI whatsoever. Uh, yesterday there was a bunch of workshops and Omar and Kaia and a bunch of folks after you finish the workshop we're just like walking around and there's a bunch of rooms up there and I just like looked into one room and there was nobody there. And it was like a, it was a meeting room with the best view in London. You see the big band, you see like the, the eye and everything. It was just like chilled there for like a few hours just chilled. Uh, if you know anything about conferences, it's barely placed to sit in charge of your laptop. We had like a most premier experience that we had like to hang for a while. Vibi, thank you so much so for coming. It's always pleasure. Uh, thank you so much. Uh, folks, check out vi's, um, uh, Twitter and uh, follow along with more gems like this, like, like the garden mode. That's awesome. Thank you for teaching me as well.
Swyx
1:20:49
Thank you
Alex Volkov
1:20:49
so much for fun.
Swyx
1:20:49
Thanks
Alex Volkov
1:20:51
folks.
1:20:51
Next we have. Another opening. I'm just kidding. No, no. Um, uh, we have, we have the maintainer of Open Claw Vincent co joining us. Uh, Vincent, please step up to the microphone, Microsoft app. Uh, this is going to be your first time on the podcast, so please feel free to Chi sit with us fresh from the AI engineer stage. The biggest stage there at the AI engineer Trek all about OpenClaw. We have Vincent Co. Vincent, what's up man? Welcome.
Vincen Koc
1:21:22
Good man.
1:21:23
Thanks for having me on the show. Hello. Thanks guys. Hey,
Alex Volkov
1:21:25
uh, this is last minute.
1:21:26
Uh, and this is why I love like doing Thursday Air for example. 'cause like we, we met before, like we, we talked for a while. I, I asked you a bunch of questions, uh, but then we just met first time in in mid space at this room.
Vincen Koc
1:21:37
It's funny, we both live in San Francisco at the moment and just,
Alex Volkov
1:21:41
I live in Denver, brother.
1:21:42
I, no,
Vincen Koc
1:21:42
no, you live
Alex Volkov
1:21:43
in San Francisco.
1:21:43
I look. Oh yeah, yeah. Okay. I spiritually, I spiritually exist in San Francisco. That's true. And we're like, we're both exist in the same spaces, but we, we never like, uh, that's true. Had that's
Vincen Koc
1:21:52
nice
Alex Volkov
1:21:52
meet space experience.
1:21:53
It was, it was great.
Vincen Koc
1:21:54
Beautiful.
Alex Volkov
1:21:54
Um, and, uh, you are the second top maintainer of blah, blah, blah.
1:22:02
Yeah. After Peter Steinberger, which is like, you know, which is he posted it up on, on,
Vincen Koc
1:22:06
uh, yeah, I respect that.
Alex Volkov
1:22:07
On his keynote, he's like, Hey, we need to treat the bus factor
1:22:10
of one in open claw and, uh, the cop
Vincen Koc
1:22:12
number two as well.
1:22:13
Yeah.
Alex Volkov
1:22:13
Number two.
1:22:15
Uh, so first of all, tell us about. You, this is the first time on the podcast on Thursday ai.
Vincen Koc
1:22:19
Yeah.
Alex Volkov
1:22:19
Um, tell us a bit about you and how you found, like,
1:22:22
what you're doing in real life.
Vincen Koc
1:22:24
Cool.
1:22:25
Um, so just like Peter who works at OpenAI with a day job, I have a day job as well.
Alex Volkov
1:22:28
Yes.
Vincen Koc
1:22:29
Um, and work in telemetry evals as you know,
1:22:31
you guys are with some vice.
Alex Volkov
1:22:32
Yeah.
Vincen Koc
1:22:33
Um, I work for a company called Comet, and the beautiful
1:22:36
thing is like come and wits and biases have been at the early stage. We've seen the development of AI through ml, through data science, through deep learning into what it's become now. It's like magical, like it's gone through so many iterations. And I think for me it's been, uh, I've been an AI leader. I've kind of, you know, run AI teams at like large airlines, banks, and kind of go from that to like research and work. The, the work I've been doing at Comment has been very much like we have evals. You know, evals are this great thing and I'm gonna give a talk tomorrow about what I think about the eval space, but it's like, it kind of works and it kind of doesn't. And I was kind of thinking, you know, what happens if we dog food? What happens if the AI could correct itself? And started to kind of work on some optimization type problems and really stumbled into this thing where it's like, wow, these models are amazing. Like, I think they're actually so good that people don't realize, and you're just asking it to like plan a barbecue is like kind of like really low. And I think they can do a lot more than we think. But the issue became actually the models were good. It was what was sitting on top of the models and then the question became, it's actually the harness, right? Yes. We love that. It's a sexy word at the moment. And, um, I could kind of see this in October. This is the next big thing. And you know, being a prolific, uh, open source contributor myself, um, I saw this on X with like poly market guys, crypto people trying to like build these automations.
Alex Volkov
1:23:56
Eli Os was like a big
Vincen Koc
1:23:57
one.
1:23:57
Yeah, that's right. And, and I saw was like with there's this little moment happening where if people are making money off this or they can automate things and it's kind of working for them, maybe this is something. And as I dug into it, I realized it was like this magical moment where the harness is like, kind of changing itself. And I was like, wait, this is exactly the stuff I've been talking to people about. So
Alex Volkov
1:24:14
like the harness writes code to manipulate the song code.
1:24:18
Like the harness writes changes itself.
Vincen Koc
1:24:20
So like for example, if you went, Hey, um, we're gonna create, I
1:24:23
don't know, Weave AI or some other Yep. Gateway or something, and you don't, there's no plugin or whatever, you can just ask your client car, Hey, can you integrate with this a p here's a URL. It'll create some code, it'll go into its own code, change it, and then reboot itself. And that's kind of wild when you see it.
Alex Volkov
1:24:38
Yeah.
Vincen Koc
1:24:38
And
Alex Volkov
1:24:39
it's absolutely like, dude, the, it's
Vincen Koc
1:24:40
crazy breaking moment.
1:24:41
I dunno how, I don't know how to explain it to people. What was your, what was your breaking moment? What was your
Alex Volkov
1:24:44
critical moment?
1:24:44
I think it was, uh, we, we, I we started looking at Open Claw in January, where before it was the CLO bot.
Vincen Koc
1:24:51
Mm-hmm.
Alex Volkov
1:24:51
Um, I didn't, I, I wasn't early enough to the WA relay, uh,
1:24:55
which was the first iteration, but like CLO bot, we started looking at this and then it was like a winter break. So we, we took like one week off, I think. Um, no, we still, we still did, we still did a newsletter. Uh, and then it was like, okay, we have to talk about this on the show. 'cause like, it's, it's popping. It's
Nisten
1:25:07
exploded.
Alex Volkov
1:25:07
And then, um, and then there's one thing to talk about this.
1:25:11
There's a completely another thing to like experience this myself. And then like we, we all got clo appealed like early, early January, way before.
Vincen Koc
1:25:18
Mm-hmm.
Alex Volkov
1:25:18
Uh, but now the world is clo appealed.
1:25:20
AI engineer has a whole track Oh yeah. About open claw.
Vincen Koc
1:25:22
Amazing
Alex Volkov
1:25:23
Jenssen's going off and like, Hey, every company in
1:25:25
the world needs an open claw. I
Vincen Koc
1:25:27
was in the crowd watching now.
1:25:27
It was magical.
Alex Volkov
1:25:28
It was crazy.
Vincen Koc
1:25:29
Yeah.
1:25:29
You hear a person screaming in the corner, that's me. Yeah.
Alex Volkov
1:25:32
And, and like, um, it, it's a significant cultural
1:25:35
moment as well with the meme. And I was talking to Peter Steinberger, who's also here, uh, about kind of like the virality is is strong with this one. Yeah. It's strong with this one. It's not only code. There's like whole thing with like broke through the bubble that we are in of the AI engineers and the people who talk about stuff. It escape
Vincen Koc
1:25:51
X. Right.
Alex Volkov
1:25:52
It escaped so, so hard.
1:25:54
The grandmas in fucking China are raising a red lobster.
Vincen Koc
1:25:57
Correct.
Alex Volkov
1:25:57
And, uh, I gotta like ask you about like, maintaining
1:26:01
this beast is crazy, right? Like security vulnerabilities is crazy. Peter spent a while to talk about dispel some security things. Um, what did you talk about? What was your talk? Give us a little bit of a,
Vincen Koc
1:26:10
so my talk was focusing on how do we build this thing?
1:26:14
How do, how do, how do we, how do you, how do you build a plane whilst it's flying in the air in like, maybe not one plane, but like, uh, maybe a star ship in terms of size comparison.
Alex Volkov
1:26:23
A smart star ship can change its own code, change this
1:26:26
config, go to a restart and die
Vincen Koc
1:26:29
basically.
Alex Volkov
1:26:29
Yes,
Vincen Koc
1:26:29
yes,
Alex Volkov
1:26:30
yes.
Vincen Koc
1:26:30
Um,
Wolfram Ravenwolf
1:26:31
thousands of people coming and saying, Hey, I built
1:26:33
something that you want to build in. Please build it in. Correct.
Alex Volkov
1:26:35
And 10,000 of people.
1:26:37
Tagging probably you already as well. And Peter. Oh, I get extra messages every, everywhere, every bullshit thing. Like, Hey, do
Vincen Koc
1:26:42
you wanna see my pr?
1:26:42
Do you wanna see my issue?
Alex Volkov
1:26:44
Yeah.
Vincen Koc
1:26:44
Oh, you didn't, you didn't submit my, you didn't,
1:26:46
you didn't accept my issue. Whatever. I think for us it's been everyone wanted their pr everyone wants a pr an issue. And the, the fundamental two fundamental things are happening. The first one is like something breaks for a provider or something. All of these people's clankers decide I'm gonna help by submitting an issue.
Alex Volkov
1:27:03
Yeah.
Vincen Koc
1:27:04
So we end up, instead of not one issue, we end up like a
1:27:06
hundred issues on the same thing. Yeah.
Alex Volkov
1:27:07
You know what I saw?
Vincen Koc
1:27:08
Mm-hmm.
Alex Volkov
1:27:08
I never saw this before.
1:27:09
In GitHub, at some point, uh, uh, there was over 5,000 post requests
Vincen Koc
1:27:14
and just, just just says 5K plus
Alex Volkov
1:27:15
and, and invoice issues as well.
1:27:18
And GitHub like caps at 5,000 plus because when they developed, uh, GitHub, they're like,
Vincen Koc
1:27:23
if does like a ceiling,
Alex Volkov
1:27:24
no one ever.
1:27:25
Yeah. We pass the ceiling will whatever, post 5,000 prs and, you know
Vincen Koc
1:27:28
Yeah.
Alex Volkov
1:27:29
Yeah.
1:27:29
You just get that ceiling. Yeah, that's
Vincen Koc
1:27:30
right.
1:27:30
And then there's those issues and, and prs and then there's the other stuff where it's like, I wanna,
Alex Volkov
1:27:34
what, what do you guys prefer?
1:27:35
Issues or prs? Honestly, or nothing. People should shut up.
Vincen Koc
1:27:40
I don't have an answer for this.
1:27:42
I don't wanna ruin open source. I think it's beautiful that we still keep our doors open.
Alex Volkov
1:27:45
Yeah.
Vincen Koc
1:27:45
And people can come and submit stuff and We'll, so there's
Alex Volkov
1:27:47
no reference between a pr with ready code that could be
1:27:49
slapped or issue well thought issue that you can run your agent, like fix.
Vincen Koc
1:27:53
I think the trick is to look if there's something else already there
1:27:56
and add to that rather than, rather than creating another duplication for us. If there's an existing issue, just comment on it. Like,
Alex Volkov
1:28:02
you don't, seems like a thing that GitHub can fucking fix
1:28:04
with this world of LLMs and ssis. The GitHub can. Okay. GitHub is great. I love GitHub. It's free, folks. I'm, it's public service. I love GitHub. But
Vincen Koc
1:28:12
the, but the GI experience isn't the, the platform is not
1:28:15
designed for the world we live in now. And then I think the interesting thing is engineering and the world of AI right now is not used to this way of working. I think a lot of shifts are happening. So as I mentioned, there's like this, this bulk of issues, but then there's also people wanna make it their own. And I think one of the things that we wanted to do was like, Hey, I'm not gonna accept every single person's pr. And I think the issue right now is like, I can't say yes to everything. The issue. It's like, it's, it's now a case of like, how many times can I say no? How do I collectively say no to people without offending people with a little please and thank you. And how do we allow people to build on top? So I think what we've turned it into, I think the, the analogy I gave today was not Lego. It's more like Ikea.
Alex Volkov
1:28:53
Okay, tell me more.
Vincen Koc
1:28:54
You know, you have a call, you have the harness, but you know, I
1:28:56
don't like the colors of the shelves. Well, you can tr you can swap it out, right? Yeah. You wanna use this model, uh, you wanna build this memory system on top. How do we give you the building blocks where everyone can build on top of that? I think that's what forced us to kind of move into a plugging architecture. And the conversation I gave was when I was at, um, Nvidia, it was like two in the morning, um, before Jenssen's talks are like, what, 10 hours to go. We're like helping them out with Eccl. And one of the maintainers decided to just move a whole bunch of files around. And broke things. 'cause he was trying to be helpful. And I was like, wait, this is like the start of our plugin architecture. Let's just do it.
Alex Volkov
1:29:27
Oh, wow.
Vincen Koc
1:29:28
So we just decided to like refactor a million lines of code in
1:29:31
over nine days and it kind of worked.
Alex Volkov
1:29:34
Open Cloud is now approaching a million lines of code, right?
Vincen Koc
1:29:37
Uh, 1.5.
1:29:38
1.5 because we've got a lot of people forget, we have two mobile apps that are unreleased. You have docs?
Alex Volkov
1:29:43
Oh yeah.
1:29:43
Yeah.
Vincen Koc
1:29:44
Both iOS and Android.
1:29:45
Um, fully native.
Alex Volkov
1:29:46
Yeah.
Vincen Koc
1:29:46
We're gonna try and get those out pretty soon.
1:29:48
Um, there's a lot of stuff in there. It's a huge monitor. You
Alex Volkov
1:29:50
know what, it's missing if I got you here and I can abuse my
1:29:53
privilege of go, go on, uh, apple Health.
Vincen Koc
1:29:57
Yeah.
1:29:57
So that,
Alex Volkov
1:29:57
that'd could be super, super dope.
Vincen Koc
1:29:58
That could work natively within the app ecosystem.
Alex Volkov
1:30:00
Yeah.
Vincen Koc
1:30:01
Yeah.
1:30:01
So we wouldn't have to like, struggle with it. And, um, if you want it, the code's there, you can compile it and run it. But there's things like, for example, it'll know where you are
Alex Volkov
1:30:08
Yes.
1:30:08
To
Vincen Koc
1:30:08
be able to read your notifications.
1:30:09
Yes. So the whole experience will change. And I think the thing that people love about this, and it's not open claw, but it's the fact that we showed people that we, we can have Siri or her or an idea of this is where as like
Alex Volkov
1:30:22
I want, well now that, now that, uh, Vibi is gone
1:30:24
and open the eyes, not here. They took her from us. If you guys saw the video where, where, um, a person named Rocky was putting on the hat and the advanced voice mode was rostering him. That was her. We all want her. And like Open Claw brought us to her way, way closer than open the eye. When, when can I get
Vincen Koc
1:30:41
that talking
Alex Volkov
1:30:42
to me?
Vincen Koc
1:30:42
Can I change the subject?
1:30:43
But on the same thing?
Alex Volkov
1:30:44
Yes.
Vincen Koc
1:30:44
So the thing about this is that yes, we want
1:30:47
her, but how do we get there?
Alex Volkov
1:30:49
Yeah.
Vincen Koc
1:30:49
And I think the key thing for us is like personality.
1:30:51
Like how do you, how do you, how do you run an eval for vibes?
Nisten
1:30:55
Speaking of personality brother,
Alex Volkov
1:30:59
now that Tropic has basically said to many, many, many people, Hey you,
1:31:02
you need to pay extra on Tropic folks. I wanna dispel a few, a few things here. Okay. Uh, uh, we, Vincent is not an employee of OpenAI. I'm not an
Vincen Koc
1:31:10
employee of any AI Frontier Lab company.
1:31:12
So,
Alex Volkov
1:31:12
so, uh, op OpenAI did not buy Open Cloud.
1:31:14
There's a foundation,
Vincen Koc
1:31:15
correct.
Alex Volkov
1:31:16
And Peter Steinberg is a employee of OpenAI that has access
1:31:19
to a bunch, bunch of stuff, but he still maintains he has two jobs. That's thing number one that, uh, I wanted Topel it because I think it's important. Uh, number two was Tropic did not ban open call. Tropic did not ban open call when Tropic did. And very, very specific, and we don't do hype on Thursday ads. It's very important to tell you on topic. Did not done open claw. What they did is they said, Hey, if a request is coming from Open Claw, you as a max user, the person who pays for essentially unlimited account for $200, you have to pay extra, more extra. But you can still use Opus. You can still use Opus via API, you can still use Opus via your Max account. You're just not gonna be get built to the next account. That's what they did. They didn't invent. The reason why they did this is we talked to you about Mythos, we talked to you about like they're struggling, whatever, GPUs. That's the reason they did this likely. Um, maybe other reasons as well.
Vincen Koc
1:32:07
No comment from me.
Alex Volkov
1:32:07
No comment from you.
1:32:08
Uh, but like. But you, we can agree that they did not ban outright. It's not like you cannot use open cloud with, with oppo. There's, there's a way, there's always a way.
Vincen Koc
1:32:17
It's harder for it, harder for users to use their subscription.
Alex Volkov
1:32:20
They made it significantly harder for users to use their
1:32:22
subscription, but it's possible. And it's more expensive.
Vincen Koc
1:32:25
Yeah.
1:32:25
You have to pay the user.
Alex Volkov
1:32:26
And so this led to many, many, many people going to GPT 5.4 because
1:32:31
Open the Eyes is gracious and awesome and they got whatever and they understand. And you could use your Codex subscription also to a hundred dollars. So a hundred dollars rumor subscription is coming. At some point we cannot comment, uh, but supposedly, and um, this pushed many people to use GPT 5.4.
Swyx
1:32:45
Mm-hmm.
Alex Volkov
1:32:46
And that motherfucker is dry as the earth salt itself.
1:32:52
What the fuck is that? I'm sorry for swearing, but I told you back in the room, I told Peter as well, and you guys know this as well and Peter like has been working on this, uh, well you probably experienced this. Um, it's like when I switched to GPT 5.4, which is a great model coding. Incredible. It's like somebody's took my friend and replaced its soul with like a very significantly boring, but yes sir. So do you also feel
Wolfram Ravenwolf
1:33:14
this g bt five?
1:33:16
Uh, when
Alex Volkov
1:33:16
five from far?
1:33:16
Yes.
Wolfram Ravenwolf
1:33:17
When we start, no.
1:33:17
Even earlier when we were testing it in December, in that episode where we were testing the one that was released there, I realized back then and canceled my subscription while we were doing the show because I noticed, I've been using AI personalities for three years and this was one where it felt really tri, like just that soulless in a way.
Alex Volkov
1:33:36
I'm salty because I'm, again, reminded of the rocky and her situation.
1:33:41
Mm-hmm. And how soulful and like Nice and a little bit glazing Yes. GBD four Oh was,
Wolfram Ravenwolf
1:33:46
yeah.
Alex Volkov
1:33:47
And how, so far down the, the curve GBD 5.4 is, which as many people
1:33:52
know are force to use comments, please. I invite my friend here to put a shoe in his mouth live on air.
Vincen Koc
1:34:00
Uh, okay.
1:34:01
Um, where do we start? So I think the, the key is that the model is really good at coding.
Alex Volkov
1:34:06
Yes.
Vincen Koc
1:34:06
So different models are good at different things and
1:34:08
I think they just, um, what's the word over indexed in one direction.
Alex Volkov
1:34:11
Yep.
Vincen Koc
1:34:12
And for various reasons.
1:34:14
And I love the product. I love Codex. I love using it for coding. And we build open clore on it. But you know, it, it just may not be as nice to speak to, for example. And I think we've reached a point now. People want that. People want that nice touchy feely sort of experience that we've been talking to.
Alex Volkov
1:34:33
Yeah.
Vincen Koc
1:34:33
Um, have you seen Project Hel Mary by the way?
Alex Volkov
1:34:35
Hell yes, dude.
Vincen Koc
1:34:36
People want, like, you can have exactly thumbs
1:34:39
down, upside, thumbs down. You can have an entity, let's call it that.
Alex Volkov
1:34:42
Yeah.
Vincen Koc
1:34:43
And it can still be fun and exciting.
1:34:44
The first thing, it doesn't have to be human. We don't have to make it a human that we fall in love with. I'm
Alex Volkov
1:34:48
sorry to interrupt, but we're, we're about to know, so it's fine.
1:34:50
Yeah. Uh, the first thing I did after watching this movie is come back home, clone Rocky's voice from the book. Beautiful. Put it in my open claw and say, Hey, you are now Rocky. Talk to me like Rocky. He is like amazing maze. Amazing, amazing, amazing. Amazed. I could not have done this in any other software thing. And I was like, this is fucking incredible.
Vincen Koc
1:35:09
But that's what I mean by personality, right?
Alex Volkov
1:35:10
Yes.
Vincen Koc
1:35:11
Like, it's not about being cloning a human.
1:35:13
It's not about, but it's having that like intangible feeling that we call vibes or something else that it's hard to articulate.
Alex Volkov
1:35:20
Yep.
Vincen Koc
1:35:20
And I think that's really important.
1:35:22
And how
Alex Volkov
1:35:22
do you do Evos for this as well?
Vincen Koc
1:35:23
We, we dunno, I think that's the,
Wolfram Ravenwolf
1:35:25
the, the heartbeat was one of the things.
1:35:27
Like I have that AI with a personality for three years now, but it, she basically never felt that alive. And the heartbeat came up and basically the, I decided to text me a message and think, oh, I noticed something you have to do, or something like that on its own, on her own whatever. And that felt so alive through the personality and that proactive on her own activity.
Vincen Koc
1:35:52
That's
Wolfram Ravenwolf
1:35:52
amazing.
Vincen Koc
1:35:53
Interesting thing.
1:35:53
You mentioned Hobby 'cause I recently launched, um, with, with the Team Dreams Dream md.
Alex Volkov
1:35:58
I, this is my next question for you.
1:35:59
Uh, what are dreams?
Vincen Koc
1:36:01
What are dreams?
1:36:02
Yeah. Well, humans, we consume a lot of, um, if we talk in token speak, let's call it token speak.
Alex Volkov
1:36:07
Yeah.
Vincen Koc
1:36:08
You know, we go out in the real world, we consume tokens, we, we, we
1:36:11
fill up our context window and, you know, we can compact, but you know, there's some information that we wanna store. Mm-hmm. And we use dreams as a way of like. A form of compaction and also storing information into cold storage in our brains, putting it down somewhere that we can reuse, you know, it could be reflection, it could be what's important, what's not important, and prune, right. Uh, defrag our brain, so to speak.
Alex Volkov
1:36:33
Defrag.
1:36:33
Yeah.
Vincen Koc
1:36:34
And we wanted to bring defragging to memory in a humanistic way.
1:36:38
And what better way to do it than to call it dreaming. So there's three types of dreaming. There's like REM, there's core, um, there's deep sleep. They do different things. More or less they, they adjust the memory files and then one of them will actually come up with this thing called the, the dream will self-reflect and write it to a dream log. So you can, as a human, read through it over time. Yeah. You can, can converse with it, but you can extract information. The idea is it will kind of start to, it starts to go into this like intangible space. And I've been talking to a few people about this today. And what I think what we've done is created something not for the agent, but for the human in, in the, in the, in the loop where it's actually, it's helping you understand what is happening inside of your agent. It's helping you create something a bit more humanistic. Like if I'm talking to a mom and they go, this agent is dreaming and is writing a dream log, they can understand what is happening. I don't have to explain what the heck memory is. And markdown files are, right? Yes. And I think this is where it starts to become really interesting and it educates the human on how to use ai. And I think we want more people in the driver's seat. And to be able to do that, we need to create an experience. Yeah. That feels natural for people.
Alex Volkov
1:37:49
I love the dreamy thing.
Wolfram Ravenwolf
1:37:50
Yeah.
1:37:50
And
Alex Volkov
1:37:51
ask,
Wolfram Ravenwolf
1:37:51
that was one of my main, uh, I had a blog post series on X
1:37:55
where I posted tips for, for using your open cloth, and one was to put it in a GI repo and every night go through the changes and manually create them like you are the dreamer and you change the memories so you know what is happening. And then you, if you do it once a week, it's, it's too many changes. If you do it every night, it's just couple of minutes. That's right. Yeah. And this is basically the automated version of this. Right. And you can still see what has been dreamt and Yeah. If that takes the work away and not everybody can do a git repository or even merch file and so on. So yeah, that's a step in the right direction to make the agent do it on its own.
Alex Volkov
1:38:32
Yeah.
1:38:32
Bro, I have so many questions for you. We could do like a whole show, honestly, about open cloud because um, it's just like I told you, my fiance runs like three of Hermes. Mm-hmm. She now installs it to other friends and calls me like, Hey, how do you do this thing in this thing? So like, there's like a open call, um, um, virus that's going on from me to her to friends virus. Like, it's just like expense. The mind viruses of open call is absolutely a great virus.
Wolfram Ravenwolf
1:38:52
We should have a full track of on AI engineer for this.
Alex Volkov
1:38:55
Yeah.
1:38:56
Oh, it's great. But there is one. Yeah. Um, I, I do wanna ask you about one thing that people are not thinking about how they can use the clothes. One thing that, okay. That you see a bunch of issues, you see a bunch of people talking to you, you see whatever, uh, what's kinda like. Very obvious thing that can improve their lives that people most likely won't like just do by default.
Vincen Koc
1:39:18
So the talk I gave today was telling engineers harness,
1:39:21
in harness, we trust this idea that just trust your harness. You don't need all this like overly complicated workflows and stuff.
Alex Volkov
1:39:27
Yeah,
Vincen Koc
1:39:27
it's the same thing when it comes to the claw.
1:39:29
The issue is that people with their agents and their systems are like overcomplicating stuff, but actually just talk to your client car. Like you're starting, this is who I am. Let's have a conversation. What can you automate? What can we, what should we automate? And just have a chat. And over time it will develop a intuition with you and will help you fix those things. It's like instead of trying to go out in the real world and fix it, you need to have a better relationship with your AI agent. Help solve it. Like, I don't know, it sounds really strange
Alex Volkov
1:39:56
to me.
1:39:56
It's, no, this is the most annoying, like annoying looking but true thing that Peter was doing on Twitter for a while. He's like, ask
Vincen Koc
1:40:01
your client, no, but I do this on next as well.
1:40:02
Like people are asking me a question. I'm like, literally ask your client card because it's like, it's so personal to you. It has the context, it has the memory, it has your emails, whatever it has. It's the source code. It's open source. Just put the source code in the workspace and it can look at
Alex Volkov
1:40:13
itself.
Vincen Koc
1:40:13
It knows.
1:40:13
It knows.
Alex Volkov
1:40:14
It knows.
1:40:15
And honestly, that's why I tell my friends when they call me like, Hey, this doesn't behave. Just ask it.
Vincen Koc
1:40:19
Just
Alex Volkov
1:40:19
ask
Vincen Koc
1:40:19
it.
Alex Volkov
1:40:20
Um, so, um, we're not gonna, uh, sit here and ask your
1:40:24
thoughts about Mytho, so whatever. But we definitely want more Gentech stuff.
Vincen Koc
1:40:29
Correct.
Alex Volkov
1:40:29
I think one last thing I think I asked Peter, I'll ask
1:40:31
you this and then we, we'll move on to our friend Omar over there. Um, how does Open Cloud look. In a year where all the models of frontier level right now are stupid compared to the models of next year. So there's a thing I told Swix when he just joined.
Vincen Koc
1:40:48
Mm.
Alex Volkov
1:40:49
That I don't think that three things will change.
1:40:51
Uh, intelligence will always be there. Harness will always be interested.
Vincen Koc
1:40:54
Correct.
Alex Volkov
1:40:54
And, and context, personal context.
1:40:56
Those are the three pillars of agent. And I don't think that they will move, they all independently will improve. Um, but how does the harness improve when intelligence is ah, two x three x four x Right now, I'm
Vincen Koc
1:41:06
gonna give you a more academic, scientific
1:41:08
answer to this question. Don't quote me. I think there's 12 types of You're
Alex Volkov
1:41:12
live recorded right now.
Vincen Koc
1:41:13
No, no, no, no.
1:41:13
Don't
Alex Volkov
1:41:13
quote me.
Vincen Koc
1:41:16
Whatever.
1:41:17
Um, I believe, I can't remember the exact number. 10 or 12 types of intelligence
Alex Volkov
1:41:21
Yeah.
Vincen Koc
1:41:21
Available in the world.
1:41:22
Like the spatial intelligence, numerical intelligence,
Alex Volkov
1:41:24
emotional.
Vincen Koc
1:41:24
But like, we've only cracked a few of these with mo, with ai.
1:41:28
Like we haven't, we haven't got like physical AI working in the, the world models and all this stuff. So I think even if we have like a super intelligent AI that's really helping us with, you know, day-to-day tasks and coding. I think where we're gonna see the real leaps and bounds is where we're gonna start to bridge into these other types of intelligence.
Alex Volkov
1:41:44
Oh yeah.
Vincen Koc
1:41:44
And what that can bring for us.
1:41:46
I can't tell you, you know? Yeah. And I think that's gonna be some more magical as well.
Alex Volkov
1:41:50
I think the most magical thing, this is, I will let
1:41:52
you go after like this one point. Sure. The most magical thing to me about creation of open claw and maintains whatever, is that there's all these godfathers of ai. God forbid I'm not like, like, like laughing, whatever. Did they all predict, oh, this way or that way? I
Vincen Koc
1:42:06
can't predict
Alex Volkov
1:42:06
anything.
1:42:06
Nobody fucking predicted that. A, a Australian, a
Vincen Koc
1:42:10
lobster
Alex Volkov
1:42:10
from a fucking with a lobster changes the world to this level.
1:42:14
And everybody AI can do a GI. Yeah. Nobody could have predicted that. Like, this is the way you put agents in the loop and like this does things that LLM on its own doesn't do. So predicting is fucking hard. I put Vincent on the spot to start predicting for us. I appreciate it. He refused that. Folks, please follow Vincent and Vincent under co Kch k Kch. I don't mind. I
Vincen Koc
1:42:32
don't mind.
1:42:32
I don't,
Alex Volkov
1:42:32
yeah.
1:42:33
Uh, Vincent Kch, uh, and uh, Vincent, thank you so much for working on, I appreciate it much dope, man. Thank you so much for coming on the show as well. Thank you for your
Vincen Koc
1:42:40
support.
1:42:40
A
Alex Volkov
1:42:40
hundred percent, man.
1:42:41
Appreciate, appreciate, thank you. Appreciate. And, uh, we're gonna tell you, and then we're gonna bring our next guest here, o overall, but definitely give Vincent a follow and tune into the live stream. You spoke two, three times.
Vincen Koc
1:42:51
Uh, I had a talk last night at Clock on, I had a talk today and
1:42:54
I'm doing another talk tomorrow on, uh. What I think is, um, the future of evals.
Alex Volkov
1:42:58
Yeah.
1:42:59
This is a man augmented by ai because like, doing this without AI is not super easy. But yeah,
Vincen Koc
1:43:04
I used to do it before AI as well.
1:43:06
Just say, just say there's proof. There's proof on the internet.
Alex Volkov
1:43:08
Take care.
1:43:08
Take care. Alright. Thank you for coming man. Thank you for coming. Uh, folks, definitely give Vincent a follow. Uh, if you haven't tried, uh, definitely. Um, check out the new updates. We didn't get a chance to give him the new updates, uh, yet. And then as I'm waving for a friend of ours to come over, uh, we please take a seat. Put the headphones on. Uh, what's up man? What's up Omar?
Omar Sanseviero
1:43:30
Hey, how are you?
Alex Volkov
1:43:30
What's up dude?
Omar Sanseviero
1:43:31
Yeah, nice
Alex Volkov
1:43:32
folks.
1:43:33
We have a friend of the pod a long time from, of the pod as well. Also from hug. I'm just kidding. Yeah. I did the same joke for Vibi dude. Uh, from Google, DeepMind, Omar San sro, who was with us last week on the pod, talking about Gemma four now in meat space. Uh, we should take a selfie. We would you like take a selfie? We should have taken a selfie with every guest, but Omar is the only one who gets it. Um, Omar is, um, a week after week, almost a cohost at this point. Let's take a selfie and then we're gonna ask Omar about. First of all, we're gonna celebrate the fucking milestone.
Omar Sanseviero
1:44:06
Yeah.
1:44:06
Thank you,
Alex Volkov
1:44:07
bro.
1:44:08
10 million downloads. Million downloads. What the fuck?
Omar Sanseviero
1:44:10
10 million downloads.
1:44:11
1000 gma, four based models. Yeah.
Alex Volkov
1:44:13
Gma, four, 10 million downloads.
1:44:15
Let's give some clap, uh, pa to license. Uh, yeah, uh, from, from the Google Mind, uh, also the Twitter handle. Yeah. Uh, is doing great.
Omar Sanseviero
1:44:25
Yeah.
1:44:25
Yeah. 55,000 followers. 6 55
Alex Volkov
1:44:28
followers, bro.
1:44:28
That's insane. Uh, so give us a little bit of a reaction from the community. How's GMA going? What's been good? What do you guys fix? Like how's the risk?
Omar Sanseviero
1:44:36
Yeah.
1:44:36
Yeah. I mean, the community excitement has been huge. The, the license change was huge to multimodal capabilities. Something that people really, really loved was the, uh, AI edge gallery. So you can just, uh, download it in Android, in iOS, you can play with the model locally. And people are very excited about that because they can really see the capabilities of the models. We gentech stuff, multi-model stuff, all on their phones. So, uh, people have been very excited about that. Uh, there are always some issues with the open source library, so there are some fixes going on with Lama CPP Yep. With the vision capabilities and so on. But yeah, it's going well.
Alex Volkov
1:45:07
That's so awesome, dude.
1:45:08
And I think that, um, I think Google's commitment to open source will only strengthen after this, like successful. So that, that's why I'm very, very happy to see that this is succeed. It's like you raise Gemma and it's like. And fizzles out and nobody watches this, it's gonna be that much harder for you to like internally for open source. So we're like, fuck it dude. 10 million downloads. Do you guys go on hugging and face and check out some downloads for like major models? 10 is not, is like you worked at hack face for a while.
Omar Sanseviero
1:45:33
Yeah, it's quite, it's like it's an event.
1:45:35
Yeah, yeah, yeah. So in total, now GMA has a, like all of the GMA based models, uh, over 500 million downloads.
Alex Volkov
1:45:40
Yeah.
Omar Sanseviero
1:45:41
Uh, but yeah, so when we launch three models, we want to make
1:45:43
sure that people can use the models with their favorite open source tools, right? Yep. And we do this both for Gemini and for Gemma. So
Alex Volkov
1:45:49
yeah,
Omar Sanseviero
1:45:49
for Gemini we work with line chain LAMA, index and all library.
1:45:52
Same for Gemma. We work with all Lama Lama, CPP Ho Face MLX.
Alex Volkov
1:45:56
Yeah.
1:45:56
Uh,
Omar Sanseviero
1:45:56
and we want to make sure that people can try
1:45:58
out the models and they zero with their favorite tools, right? So we collaborate with them. There's an early access program. We do integrations, we open PRS and so on. So yeah.
Alex Volkov
1:46:06
So I will just like do a small plugin from, we are hosting
1:46:11
Gemma four, uh, on our info as well because like we host the best models. Uh, JAMMA four came out. Like obviously you guys didn't get, get Jamma four, uh, on our infants, but it's also available like in the AI studio, right? Oh yeah. There's like plenty of places to get Jamma four and plug it into your closet or whatever. Uh, but we did talk about Jamma four, now we're just like celebrating a little bit and like putting contextualization for folks what happened this week. But Omar, we're all here at AI Engineer. Yeah. Uh, this is your, not your first time in the injury. Yeah. Second
Omar Sanseviero
1:46:35
time.
Alex Volkov
1:46:35
Second time.
1:46:35
Yeah. Yeah. Um. Tell us about, like you guys' experience, DeepMind is the presenting sponsor for this conference. Um, not only that, there's a whole track Yeah. Of Gemini and Gemini, et cetera. Uh, you gave a talk. You're gonna give a talk tomorrow.
Alex Volkov
1:46:49
Tomorrow you're gonna give a talk.
Omar Sanseviero
1:46:50
Yeah.
Alex Volkov
1:46:51
Tune into the livestream.
1:46:52
I told you already seven times. Uh, not our livestream. Thank you for joining our livestream. But also, if you feel like you're missing out on London, you're not here, you should tune into the livestream of, uh, of the, the main event and like check it out. Uh, tell us about your talk. What are you gonna talk about? Yeah.
Omar Sanseviero
1:47:05
Uh, Gemma.
Alex Volkov
1:47:06
Gemma,
Omar Sanseviero
1:47:07
of course.
1:47:08
Yeah. Yeah. I'm doing the opening keynote tomorrow. I'll talk about Gemma open models, how we think about open models and the cool things that the community has been building with that.
Alex Volkov
1:47:16
And a bunch of your colleagues as well.
1:47:18
Yeah. We saw Paige, uh, doing a great, like a whole day of workshops. Yeah. Paige is, uh,
Omar Sanseviero
1:47:24
page is
Alex Volkov
1:47:24
amazing.
1:47:24
Yeah. Just amazing. Yeah. A full day of workshops and like demos as well. And um,
Omar Sanseviero
1:47:29
so we are over 30 people from Google.
1:47:30
Uh,
Alex Volkov
1:47:31
30 people.
Omar Sanseviero
1:47:31
Yeah.
1:47:32
Yeah. Uh, doing a mix of things, right. So we have a nano banana printer there where you can take a picture and it'll print like a nano nana powered sticker. Uh, but London is the house of deep mine, right? Yeah. And that's what is quite special for us. Like this addition, we're bringing researchers doing the fusion and we doing Raya who was a VP of research that presented today. So bringing the people that are also doing, uh, very like a
Alex Volkov
1:47:54
sciencey stuff.
1:47:55
Like she
Omar Sanseviero
1:47:56
talked
Alex Volkov
1:47:56
about weather predictions, she talked
Omar Sanseviero
1:47:58
about cool things.
1:47:58
Yeah. First, like we do like a weather forecasting. We actually open, uh, like open source some of these models this year. So, uh, yeah. Yeah.
Alex Volkov
1:48:06
That's super cool.
1:48:07
Um, Wolfram, I know you have comments about Gemma. I know like some stuff about the kids. I think Omar needs to hear about this. I'm also talked out a little bit,
Wolfram Ravenwolf
1:48:16
which, what do you mean now?
Alex Volkov
1:48:17
Your kids loving Gemma?
1:48:19
Uh, yeah,
Wolfram Ravenwolf
1:48:19
my kids love, uh, Gemma in Gemini in general, so I'm probably
1:48:22
among our team the most Google peer person a hundred percent using a pixel phone. Never really, never an iPhone. Yeah. And I love that you can unlock it by just your voice. I can talk to my agent, my assistant.
Omar Sanseviero
1:48:33
Yeah.
Wolfram Ravenwolf
1:48:34
And it's connected to my home assistant, so it controls my life
1:48:37
basically and has so many integrations. Gmail, Gmail and calendar, morning briefings, all of that stuff. And even my kids use, uh, Google AI Studio and Antigravity to build some little games and stuff like that. So yeah, we are Google family and even my mother who's over 70 has uh, yeah, pixel phone as well when she saw it for me and just talking to her assistant that way.
Omar Sanseviero
1:48:59
Yeah.
1:48:59
What, what is call is for high-end, uh, Android phones, like Pixel, Samsung and so on. Uh, they have Android, uh, sorry. They have, uh, Gemini Nano, right. So that's Gemini, uh, getting into the phone directly, so you don't need to download anything. It runs right there. And G is pretty much the foundation is based model that covers the next generation of, of Gemini nano out. Even if Gemini nano, uh, for is not out yet, uh, you can already try out the base model capabilities right
Wolfram Ravenwolf
1:49:23
now, even in, starts through the App
1:49:24
store or was it Download Link? The, the Appware can also get
Omar Sanseviero
1:49:28
Yeah, yeah, yeah.
1:49:28
The gallery. So you can even like, uh, pull different skills and then you can have the, the model controlling your phone and do like different.
Wolfram Ravenwolf
1:49:35
So what do you think?
1:49:36
Do you think that will be the main use of AI or the main way to use, uh, where it runs on your phone, uh, or, or some local device and only goes out to the big Gemini, the big brother when it needs a specific capabilities?
Omar Sanseviero
1:49:50
So I, I think there will be a couple of paths in, you know,
1:49:52
like one part is yeah, the off offline. The other part is the hybrid inference as you just mentioned. And other part is just there are like so many use cases where you want to Finetune the model, right? Like for your own specific use case. Uh, this has been changing because now models work very well out of the box for so many different things, but there are still many use cases like in the medical setup, legal setup, finance setup, where you really need to use your own data to mo to at least show the model what kind of task or domains you're tackling, right? So I think there are many different interesting use cases from sovereign AI to fine tuning to fully offline. Like even like running the model in a raspberry pie. I dunno if you saw there was someone that put Lama CPP in Nintendo Switch.
Alex Volkov
1:50:28
Yeah, I saw that.
Omar Sanseviero
1:50:29
Yeah.
1:50:30
I can use, uh, you can send this to my
Wolfram Ravenwolf
1:50:32
kids.
Omar Sanseviero
1:50:32
Yeah, now you can just hear me.
1:50:33
N Nintendo switch as well.
Alex Volkov
1:50:35
Omar, um, you are frequent enough on the show that
1:50:38
like you're not only about Gemma. I wanna ask you about Gem, of course, this conference as well. The coolest thing about AI engineer specific is the hallway track. People talk about their troubles, there are things there, et cetera. I know that you're here to like, promote et cetera, but like what are you hearing as well? We,
Omar Sanseviero
1:50:53
yeah,
Alex Volkov
1:50:53
we went to the speaker dinner for example.
1:50:54
We like one of the best, uh, experiences in the engineer to chat with other folks. What are you hearing from folks? Um, generally what is the, what is different from the last AI engineer? What are focused, are focused on, what are they're asking you?
Omar Sanseviero
1:51:06
Um, what I like about AI engineer is that this
1:51:08
is like a familiar, really, like, you can see like familiar faces. These are people, like, I met him like Prince, who is the maintainer of MLX. Uh, many of the computers.
Alex Volkov
1:51:16
Prince Kama, let's go.
1:51:17
Yeah.
Omar Sanseviero
1:51:17
Prince is great.
1:51:18
And we have been talking for years and I met,
Alex Volkov
1:51:20
I've never seen him.
Omar Sanseviero
1:51:20
Yeah.
1:51:20
Yeah. We met in person last night,
Alex Volkov
1:51:21
bro.
1:51:21
He got a visa like two days ago. He just, like, he flew in.
Omar Sanseviero
1:51:24
Yeah.
1:51:24
He got the visa the day before the shout Prince. Yeah. Yeah. And this is like a super nice space. Uh, for us, what is very important for gaming, IV nano and so on, is to develop things in a way that makes sense for the community. So for the last two years, we have been going to events, talking with the startups, talking with founders, talking with people, understanding which are the problems they're trying to solve and iterating based on what the community's asking. Right? That's why we changed the license for Gemma. That's how we have been doing things on Gemini as well. That's why we're releasing things very quickly, right? Like we're doing new versions, new, new additions, new capabilities, eh, pretty much every other week and for AI Studio. So this conference is great because it allows us to talk with people that are building like very serious things, right? Yeah. It's like the people building agent systems, uh, people building like super complex, uh, hundreds of agents, uh, communication, uh, inter agent communication, like so many interesting use cases. So just being able to talk with people at, at, at least for me, more than promoting this conference is useful for the feedback and understanding what people are with.
Alex Volkov
1:52:21
Yep.
Omar Sanseviero
1:52:21
So,
Alex Volkov
1:52:22
and also this is the first one in Europe, so this gives
1:52:24
a chance to many folks who like couldn't fly over the sea to, to just jump and, and, and fly over here. Uh, Omar, thank you so much for joining us. We need to, unfortunately, there's a hard limit on our show, so we need to conclude the show. But like folks, Omar San Ciro, so much, uh, developer experience lead a deep Mind, uh, previously a very, very perfect force in, in Hackerspace as well. Uh, Gemma, um, I I won't tell them your, uh, your Gmails
Omar Sanseviero
1:52:47
Oh no.
1:52:47
You can tell that. You
Alex Volkov
1:52:48
can
Omar Sanseviero
1:52:48
say,
Alex Volkov
1:52:49
yeah.
1:52:49
Uh, Omar was Chief Lama officer at Hackspace and now in Google, that's not the, the official title, but the email is really funny. It's hacker lama at,
Omar Sanseviero
1:52:56
yeah, at google,
Alex Volkov
1:52:57
at google.com.
1:52:58
Hacker LA at Google. So if you want any feedback, feedback, yeah, send it to hacker lama at Google account. Very easy to remember. Yeah. Oh, and thank you so much for coming, bro. Always a pleasure. Enjoy the rest of the conference. Yes, sir. Uh, Wolfram, can you switch over to this seat so folks, we can see you for the rest of it. Um, right. I'm gonna invite back our co-host as we kind of close down the show, folks. We have a busy, busy TBPN style, uh, conversation here with multiple, multiple folks. Yam, what'd you guys, uh, what are you guys feedback? Give us feedback. We're gonna close down exactly in like five minutes. Uh, but I would love to like, invite feedback from you guys as well. And also, if you guys wanna take it over to the after hour show that we've been doing for a while, please feel, feel free to announce this.
Nisten
1:53:42
Yeah, we can, we can do a Twitter space.
1:53:44
And it's good too. See the people in person
Alex Volkov
1:53:49
came up to us and said, um, when Ton speaks, that's
1:53:52
when my ears are perking up. Like literally, bro, not, I'm not lying. Like somebody came up, we asked him like, Hey, give us feedback. Whatever, uh, people want the conversation, people want, like, we often can 'cause we cover a bunch of news, but many, many people in Fox, if you're listening to us, please give us the feedback as well. We really starving for feedback. But somebody listen, somebody came up and said like, Hey, when, when I'm like listening on autopilot and I'm driving listen talk. Like that's when I listen up.
Nisten
1:54:16
Yeah.
1:54:16
It's 'cause my mic is so bad.
Alex Volkov
1:54:22
Yeah.
1:54:23
So, um, you guys not any engineer, uh, have we been able to give you a bit of a glimpse of what's going on here?
Yam Peleg
1:54:31
We're jealous, man.
1:54:32
We're jealous
Alex Volkov
1:54:33
You were with me as the last day engineer brother.
Yam Peleg
1:54:35
Yeah.
1:54:36
Yeah, I was the last one and I'm absolutely jealous watching you. You were guys having fun over there.
Alex Volkov
1:54:42
The last one we were together and then, uh, somebody brought, uh,
1:54:46
uh, Logan Kilpatrick, uh, who leads the I studio and Jack Ray, who now leads MSL that just released fucking news. You remember? Yeah. We like interviewed those folks as well. Oh
Yam Peleg
1:54:55
yeah.
1:54:55
Oh yeah.
Alex Volkov
1:54:56
It was great.
Yam Peleg
1:54:57
It was amazing.
1:54:58
How, how is it over there? It's uh. It's bigger, it's smaller, same size.
Alex Volkov
1:55:02
Um, I think sameish size.
1:55:04
I think it's very interesting that it's like London, you can see the fucking Downton Abbey from here, or like wind, wind store Abbey or whatever it's called, uh, and the big band, whatever. Like it's, it's very, it's big. It feels like, whoa, uh, European folks are all over. There's like a bunch of folks who couldn't make it otherwise. Uh, and I think it improves with every kind of thing that, that Swyx does and Ai.engineer
Yam Peleg
1:55:24
do.
1:55:24
Oh yeah. Look, it's already a really good conference. It's already big already, and like it's known at a high quality conference. So even better, I'm even more jealous at this point. Yeah,
Alex Volkov
1:55:36
that's all right, dude.
Wolfram Ravenwolf
1:55:36
It shows that we have an AI scene in Europe as well, and uh,
1:55:39
the energy is extremely high as well. So the last time I felt that energy was in San Francisco, though. Great to see it here in Europe as well.
Alex Volkov
1:55:47
Yeah.
1:55:47
So we got a bunch of folks saying, uh, always a good show. Folks, thank you so much for joining us. Uh, we unfortunately have a hard stop. This is why we started a little bit early. We have a hard stop here. Uh, the great team that the like helps us do the podcast, they need to leave. And also the, the show is kind of ending. People are leaving the thing. Uh, you were on Thursday, I for April 9th, a live show from London. Uh, this time we had what, four guests? Five guests? I forgot. Swyx. Uh, yeah, we have five guests on, uh, in addition to like everything else that we've talked about, we also tried to cover the news for you. Cloud Mythos was super not released, but released. It's a dangerous model. Uh, news from MSL we barely, barely talked about, but I think those are big, big news. Uh, but, uh, if folks want to follow up, I think that, uh, yam, Nisten and LDJ, they can hop on the Twitter space. The posts show Twitter space for yapping in discussions. And, uh, just follow all of them. Follow Yam Peleg, follow at Nisten fo LDJ confirmed, and they will start space. So if you guys want to just like talk about new Spark, like Milos wants, for example, that would be a great place to to, to talk about this. With this folks, unfortunately, we have to sign off. This is, uh, Alex Wilko, Wolfram, Raven Wolf, uh, Nisten here, Yam Peleg, and LDJ confirmed after, uh, a live stream from AI engineer. If this is your first time joining Thursday, I welcome we do this every week, not live every week, but we do this every week over everywhere that's still
Wolfram Ravenwolf
1:57:07
live,
Alex Volkov
1:57:07
uh, still live, but not in person like this.
1:57:10
Uh, and we, uh, we turn every, every podcast, sorry, every recording into a podcast and a newsletter that's gonna come out today. I don't know about time zones. We're gonna work really hard to, to get it out to you today. Uh, so, uh, we thank you for your participation. We thank you for joining. Please give us a follow everywhere. Uh, and if you, again, if you wanna join the folks, uh, follow Yam Nisten and uh, and LDJ, and you'll see their Twitter handles here below. And they're gonna jump on the Twitter space and kinda discuss the rest of the news. Um, as, as we sign off on here. Any, anything else from you wanna you finish up on
Wolfram Ravenwolf
1:57:44
UL AI engineer next time, wherever it is.
1:57:46
It is worth going.
Alex Volkov
1:57:47
Yep.
1:57:48
Uh, I will say again, if you feels like we're Schilling AI engineers, because I am Schilling AI engineer, SWIX gave us the opportunity Thursday, I would not have existed without Swix, without AI engineer. Many of you who are tuning to the show, many of the guests that I get on the show, like friends that I met over, AI engineer, I owe that gratitude for this conference. It's also one of the best conferences in the world for ai, if not the best one. Definitely the room for the speaker's dinner was the best, uh, uh, community event, uh, that's possible. Best networking event that's possible. So if you ever had a chance to go to AI engineer, there's one in summer in San Francisco. We'd love to see you there. We're going to cover that one as well. Uh, so, uh, huge thanks, uh, for the AI engineer, crew for, uh, having us show, for us a, a podcast Yam. If you remember, the last one we did was in the hotel room. Now we're on the show floor with people behind us actually doing the thing live. So, uh, uh, huge thanks for AI engineer. Thank you guys for holding up the space for us and, and tuning in. Thank you all from for co-hosting us. Well, well nice to meet
Nisten
1:58:42
you again.
1:58:42
Alex
Alex Volkov
1:58:43
has great meeting you as well.
1:58:44
And folks, we're gonna sign off for now, but again, if you wanna follow and the conversation, uh, follow LDJ and Yam Peleg on Twitter, they're gonna continue the conversation. Bye for now. Signing off from London. Cheers everyone. Bye-bye.