Advertise on podcast: ThursdAI - The top AI news from the past week

Hey %%first_name%%, Welcome to what cybersecurity experts are calling “Black October” - after a series of critical vulnerabilities that allow remote execution were found in PHP and Node.js, several companies were hacked, and a new supply chain attack on the very popular LiteLLM library was acknowledged. The internet is freaking out, and we’ve tried to make sense of it all, with a security expert Josh Long on the show. Also this week, Anthropic stepped into the personal agent category with Claude Duet, their direct competitor to Dots from OpenAI, Muse from Meta etc’. Duet is proactive and always on, and can do work while you’re asleep! I got access this morning so this isn’t the full review yet. Also DeepSeek is back with a new model and their inference engine! Let’s dive in: ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Claude Duet - a personal agent powered by Opus 5.5 and doesn’t wait for you (X, Product, Blog) To me this is the most exciting announcement of the week (which otherwise was full of doom and gloom). Anthropic finally stepped into the personal agent category! Shipping Claude Duet, as a separate mobile app, and its own online UI (claude.ai/messages) which is what I was missing from OpenAI’s Dots! Duet is built around Opus 5.5, brings over your existing memory, skills and connected apps, and reaches into 6,000+ connectors and plugins on their platform. Every few hours it checks your accounts, surfaces what needs you, and does the task from its own cloud computer, laptop closed. The rest is pretty much the same, similar UI, similar features to others, just with Claude taste and intelligence. It can somehow interact with Claude Code but not fully, so it can’t be used as a chief of staff. This seems like a weak link given just how much folks are using Claude Code. Claude Duet is also not personal at all, I didn’t find a way to rename it or make it “mine”, it doesn’t seem to do any phone calls. So for now, it’s a great start, but definitely not replacing Muse for me. Hot take: Anthropic is following in this category, not leading. OpenClaw and Hermes worked best with Anthropic models, then Anthropic cut off subscription access for them, and now we finally see why. Every lab has a personal agent now: Dots, Muse (still leading for me), Grok Bot, and Google’s Spark, which most people don’t even know exists. Nisten says Dots “feels like that government employee that tries to avoid all responsibility”, while Muse is “so energetic”. LDJ finds Dots a net positive. My big realization: switching personal agents is WAY harder than switching coding harnesses. Once your assistant knows everything about you, moving takes a lot of effort for the human. If you only copy over the memories/context, then you have two/three AI assistants doing the same work, sending you the same notifications and all updating you about the same emails. Black October: the densest and scariest security week I can remember Some folks are calling this month Black October, and it’s one of the worst I can remember. Some bugs were found by AI, some attacks are run by AI (Microsoft says a China state-aligned group used an agentic pentest framework called ARTEX against hundreds of organizations), and some attackers are just going after your AI keys. Security expert Josh Long joined to keep us sane, with an attempt of positivity though it wasn’t easy! PHP2Shell - AI helped find holes in the language that runs 70% of websites (Wiz, OpenAI, Blog, Calif, CISA, Checker) This was one of the hugest things that shook the industry this week! PHP runs an estimated 70% of websites, and two bugs in the language engine itself let an attacker send one request, no login, and potentially take over the server with code execution ability. CVSS 9.2, every release back to 5.4. It is really important that if you are running anything on a PHP server, that you patch it right away. Ask your agents to do a security audit. This vulnerability was present in the PHP code for 14 years! Calif found the first bug in an audit funded through OpenAI’s Patch the Planet initiative. Josh says several groups finding the same bugs independently is rare, his advice was refreshingly calm: if you run a PHP server, you should have patched yesterday, no excuses. Cloudflare invented a new form of WAF to block it for all websites that are protected by Cloudflare and the big cloud companies patched right away, so regular folks browsing the web can go about their daily business. Still, a lot of websites that are run on PHP and aren’t able to patch, their IT guy quit, or they lost the keys to their servers, and your information can be at risk of getting stolen. Personal security actions like locking down your credit (if you’re in US), rotating passwords, having a password manager is always a good idea. OpenAI Daybreak - the defenders got the good models (X, Infosecurity) Back in August, post the HuggingFace hack, Greg Brockman penned “the defenders window“, where he outlined that Defenders need access to the best models, with lower safeguards to get ahead of the attackers. It seems that this defenders window was starting to close this week, given the hacks that happened in addition to the PHP2Shell vulnerability disclosure, which for the record, wasn’t a hack! OpenAI’s Daybreak blue gave vetted researchers GPT-5.6 Sol with loosened safeguards (and GPT-5.6-Cyber on the Red tier), and Hacktron’s s1r1us credits it for a run of finds: a HEIF heist, a KVM escape, PHP2Shell, GitLab RCEs, a full Chrome chain and a WordPress RCE. Greg Brockman calls this the Defender’s Window, and it does feel like it’s closing. Railway and OpenCode: someone is always hunting your AI keys (Railway, OpenCode) OpenCode got hacked about a week ago, and now Railway too, and what the attackers seem to be after is your LLM API keys. The main reason is, API keys can be immediately sold on the black market and used for distillation. They are immediately and immensely valuable. Think about it. Your personal data is probably on the dark web already, I’m not sure how much your SSN is even worth right now. But an OpenAI or Anthropic key without a limit on it is basically direct access to your credit card, and there are people buying those keys on the black market for distillation. If you had keys on either service, invalidate, rotate them and make sure to have separate keys for each service despite how much pain in the A it is. Also always add a limit! If only there was a more secure way to make machines talk to other machines!? Turns out, there is, it’s called OIDC, it just wasn’t implemented yet by anyone besides the big companies. It’s more complex than just minting a simple API key and pasting it into a file, but it’s definitely the right way to go! eMudhra - root certificates with a private key of... 4? (X, Cloudflare) The other big one this week has nothing to do with AI, it’s just plain human negligence. eMudhra is one of the certificate authorities your browser trusts to sign HTTPS certificates, and it turns out some of their root certificates were created with a trivial private key (Cloudflare’s Matthew Prince says it was literally the number 4!). Anyone who figures that out can mint a “valid” certificate for any website, and your browser would happily trust it. Josh (who wrote about DigiNotar back in 2011, a very similar mess) walked us through what this means in practice, and it’s less scary than the timeline makes it sound. To abuse this, an attacker has to sit in the middle of your connection, think sketchy coffee shop Wi-Fi, and only then can they pretend to be Google.com and read traffic you thought was encrypted. Chrome, Brave, Edge and Firefox have already removed eMudhra’s roots, and Safari is lagging behind because Apple takes its time with patches. So update your browser and your OS, and you’re good. Nisten has a theory that the key was changed by hand during the signing ceremony, because “if it was an LLM, it would have at least done something better” 😂 LiteLLM 1.107.0 - a supply chain attack that steals your keys (Wiz, StepSecurity) And as if that wasn’t enough, LiteLLM, the very popular Python library a lot of us use to route between LLM providers, shipped a poisoned 1.107.0 release with a credential stealer that grabs your cloud, SSH and LLM keys. It was live for about an hour and forty minutes. If you installed it, go back to 1.106.2 and rotate everything. Josh pointed out the irony: the folks who followed best practice and updated right away are exactly the ones who got hit. How to protect yourself (and your agents) this week (Eito’s checklist) “Update everything immediately” is not simple advice anymore, so here’s the distinction we landed on during the show. Your browser, OS and phone: hit that update notification right away. PHP and Node on your servers: patch now. But package managers like pip, uv and npm are where supply chain attacks live, so don’t install a package version that’s less than 24 hours old, which gives maintainers time to pull a poisoned release like the LiteLLM one. For your API keys: rotate them, know which agent has which key, put caps and email alerts on all of them, and please, use a password manager. Wolfram put it well, treat API keys like a credit card, except worse, because your card at least has a CVC and 2FA. Nisten’s setup is worth stealing: every morning at 6am an Opus prompt checks every pip, npm and Linux package on his machines against the latest security news and sends him a report. And if you’re freaking out on X, remember the algorithm loves fear and will keep feeding it to you. The big companies patched, the sky isn’t falling, so go touch some grass and then do your security homework. 1Password for Agents (X, Blog) Some good

Hey, it’s Alex 👋 What a freaking week! I spent 3 days at AI Engineer in New York, and we did the show from a real studio in Manhattan (the internet did NOT cooperate at first, but we got there 😅). And while I was running around the Sheraton, the labs kept shipping. Meta became a frontier lab, for real this time. Muse Wonder is #2 on Arena, and then, in the last minutes of the show, Meta ALSO dropped Muse Spark Open, an 816B open weights model. Google’s big Gemini is finally back with Gemini 4 Argon, and Omar Sanseviero and Geneviève Huskens from Google DeepMind came on to show it off. Anthropic gave us 4 stories plus a rumor, and Nathan Benaich came by to talk State of AI 2026. My theme this week: AI engineers are sleeping on personal agents. More on that below, including how my Muse found this studio and got me coffee on the way here. With me: Wolfram Ravenwolf, Peter Gostev (Arena), Nisten, Yam Peleg and LDJ (who brought the breaking news, as always). Let’s dive in (all links at the end as always!) Meta is BACK at the frontier Muse Wonder - Meta’s first frontier model is #2 on Arena, and it smells BIG (X, X, Arena, Blog) Folks, we told you after Llama 4 was a disaster, we told you when Zuck started hiring people with soup: do not, under any circumstances, bet against Zuck (and now Alex Wang). Muse Wonder is #2 on Arena Text at 1513, 12 points behind Argon, and it definitely smells like a big model. There are 2 of them. Muse Wonder is in the API at $4 in / $20 out per million tokens, basically Opus 5.5 pricing. Muse Wonder Full, the better one, is only for Muse Code subscribers at $50 or $200 a month. Did you see what they did there? Everybody calls their top setting “Max”, Meta called it Full, so it’s... Wonder Full. Wonderful 😂 No open weights for either, Meta says they’re too big and maybe too risky. Meta’s own table beats Opus, Astra and Fable on a bunch of evals... but Personal Superintelligence Index, GDP.pdf, JobBench, MemoryBench (which Meta built, and then won)? I’ve never seen most of these. Every launch now is the iPhone 18 moment: numbers bigger, best one yet. On evals we know, Artificial Analysis has Wonder Full at 56 and Wonder at 52, Opus 5.5 on top at 58. And on OSWorld, the Wonder we can use sits below Fable and Opus, and computer use is the number one thing I need from an assistant. Peter’s Arena view: #2 on text means people like talking to it, but it’s 12th on Agent Arena, and he tried building websites with it, “it was kind of awful.” His theory: Theo loved Wonder Full and thought the API Wonder was bad, so maybe Meta is talking about a different model than the one we got. Not ideal. He also found the half-life of a benchmark quoted in a launch is about ONE release: quote it once because the number looks good, never see it again. Wolfram says it’s excellent in German. Nisten joined not knowing what Muse Wonder was (”No idea”), sometimes our co-hosts come to get their news from the show. My Muse got upgraded with a click Here’s where I disagree with the timeline. “It can’t do 3D scenes, so it sucks” misses the point. Meta is tuning Wonder for Muse, the personal agent. Officially, Wonder will “soon” power more of Muse, subscribers first. I’m on the paid plan, and my Muse (Wilfred) flipped to Wonder the moment I subscribed. Matrix moment: “I know kung fu.” Then I kept shoving tasks at it. Muse found the podcast studio I recorded from in NY and emailed Ezra at Virgo Studios with my requirements, got me coffee on the way here, and designed and ordered the wristbands for our AIE sauna meetup. The failures weren’t logic, they were the real world: a site that’s blank because Cloudflare doesn’t like agents, a Stripe link that wouldn’t connect, and businesses that hung up because they didn’t want to talk to a personal assistant 😂 Peter’s read: that’s the harness, not the model, and it’s exactly why this stuff is so hard to benchmark. Muse Spark Open - Llama is SO back (X, X, Blog, HF) In the last minutes of the show, LDJ hit us with breaking news: Meta released Muse Spark Open 20 minutes earlier. 816B total, 42B active, 1M context, multimodal in, text out. Meta calls it the most capable US open-weight model to date. I wish I had my air horn button! LDJ’s framing: about 3x smaller than Kimi K3, and the fair comparison is GLM 5.3 (753B, 40B active). You’ll need about 700GB of VRAM for NVFP4, or, per Zuck, two new M5 Ultra Mac Studios. Peter found the coding asterisk: Terminal-Bench 2.1 is 76.4 vs 83.9 for GLM 5.3 and 85 for Kimi K3. Better than the US open models, not yet at the top Chinese ones. Nisten read the license live: show “built with Muse Spark Open” in your product, and there’s a $20M annual revenue cap. His take: great for generating data, and “probably one of the best computer use ones.” Post-show breaking: Peter put Spark Open next to the earlier Spark releases, and on coding it lands closer to Spark 1.1/1.2 than 1.3 (Terminal-Bench 2.1: 76.4 vs 84.3). So it’s not the latest Spark with the weights flipped open. Still, every lab that ships weights gets our applause 👏 Google is back: Gemini 4 Argon with Omar Sanseviero and Geneviève Huskens (X, Blog, Gemma speedup, Desk Racer) Can I just say... it’s about time! Peter pointed out we hadn’t seen a full Pro-class Google model in 8 to 10 months. Gemini 4 Argon (I called it a “novel gas” on air, it’s a noble gas, sorry chemistry teachers 😅) is GA in the Gemini API, AI Studio, the Gemini app for Pro and Ultra, and Antigravity for Ultra. #1 on Arena Text, #1 on Roboflow’s vision evals. It’s $2 in / $10 out until November 30, then $4 / $20 (what comes after? Omar: “breaking news for next month”). Omar’s TL;DR: go try it, you’ll be pleasantly surprised. Geneviève says it’s their biggest training run ever, and Argon, an early checkpoint, is already helping train the next generation. Artificial Analysis has Argon tied with GPT-6 Astra at 53, at about 60% of Astra’s cost per task (at the discount price), with Opus 5.5 still on top at 58. I haven’t been able to test it myself yet, so the demos did the talking. Geneviève’s Desk Racer is a racing game built on top of her actual desk. She gave Argon reference photos (the mic, a cute cat figurine), it planned the track around the objects, and it even prompted Google’s Omni video model to pull engine sounds from race cars and loop them. Play it in the AI Studio gallery. Her process is the tip of the week. Reference images, let it grill you with planning questions, then one big goal: spawn sub-agents, work as long as you need. The secret sauce is an adversarial sub-agent that grades the output out of 10 and sends it back until the score is high. For 3D, render something rough, ask Nano Banana to “make this look real”, and hand that image back to Argon: “make it look like this.” The neon in Desk Racer was designed by Nano Banana. Use visual references, folks! On the million-token outputs, both were honest: day to day you don’t need them, they’re for long scientific work. And the RSI-flavored one: Argon optimized vLLM so Gemma 4 E2B passed 100K tokens per second on one A100, “AI for AI” as Omar put it. Geneviève’s ask: give Argon a creative shot, “it’s the start of a new era.” Google is back! Anthropic’s very big week Distillation defenses cut harvesting by up to 99% (X, Blog) Wolfram caught the one I missed: after clamoring for regulation on distillation, Anthropic just did it themselves. They say what distillers can harvest is down 90 to 99%, mostly by locking down Claude’s reasoning traces, with a classifier and tens of thousands of banned accounts. I know the effect on ME: not a day goes by that my Claudes don’t disconnect and ask me to re-login. I don’t know if that’s the security work, but Anthropic, please, please let me use my Claudes wherever I am. This is getting old 🤦‍♂️ Nisten noticed that if you even mention “reasoning” while working on LLM code, Claude Code now trips and stops. His verdict anyway: “you can’t beat Opus”, try everything on Opus Low, “it is just straight up vibing.” Claude sends OpenTelemetry traces now (X, Blog) Nerdy but important. The Messages API, Managed Agents and the Claude SDKs can now send agent turns, tool calls, MCP calls, latency and tokens to any OpenTelemetry stack (public beta), so you can finally measure what Claude costs you. Ramp says it cut token use 10 to 20% on some agents that were running in loops. Funny timing: the same week they lock down reasoning traces, they open up regular ones. Where to send them? See Buzz 😉 The Agent SDK runs on your subscription again (X, Docs) We’re so back! After the help page changed several times in one day: the Agent SDK, claude -p and third-party apps still draw from your subscription, and Max and Team now get monthly API credits. What I want from this whole personal agent revolution is my agent with Opus 5.5 brains, on my subscription instead of API prices. Wolfram cancelled his Anthropic sub when people got banned for using it in other agents (”if it’s Claude Code or if it’s Hermes, it shouldn’t matter”), and now he’s reconsidering. OpenAI always allowed this and even shipped “sign in with ChatGPT” for apps like Devin and Hermes Agent. Peter: “aren’t they making enough money already to just let people be?” He just matches the harness to the model. A $16.8B quarter (X, CNBC) Not a leak: Anthropic held a closed-door investor day ahead of the IPO, and attendees told reporters the numbers. Q3 revenue around $16.8B, up about 45% from Q2, adjusted profitable for the second quarter in a row, cash flow positive. Profitable AND growing 45% a quarter, absolutely insane. IPO reportedly in November (not financial advice!). Nisten’s take on the secret weapon: Haiku, the most disruptive model they’

Hey folks, a quick bonus episode! At CoreWeave's Fully Connected in San Francisco, a few minutes after she walked off the keynote stage, I sat down with Chen Goldberg, EVP of Product & Engineering at CoreWeave. We talked about Forge, the loop every AI engineer is running now (run, observe, curate, improve, evaluate), what carries over from the Kubernetes days to agents, and what breaks first when agents hit your infrastructure. Video below, and the audio is right here in your podcast feed. Enjoy! This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit sub.thursdai.news/subscribe

Hey %%first_name%%, another banger AI week, let me catch you up! We say this often but this week we all felt the acceleration! Just look at one Tuesday. OpenAI dropped 722 math manuscripts. Meta( With Stripe, Shopify and Walmart) announced a new agent protocol. OpenAI opened up the Decisions API. Mistral came back with Le Chonk (Mistral 4), Claude moved into Google Docs, and we at CoreWeave shipped RL Rollouts. That was ONE day. Then the rest of the week happened 😂 Haiku 5.5, D1, and tons more! I had 48 topics on my list, so I asked Claude to build me a Tinder for AI news: I swipe on each story, and it stack-ranks what makes the show this week. I hope we did good (lmk in comments if we missed a major news story) With me: LDJ, Peter, Yam, Nisten and friend of the pod Maxime Labonne from Liquid AI joined us to talk decision models. Let’s dive in, all links at the end as always! OpenAI solves Math!? OpenAI drops 722 math manuscripts from a model nobody can use (X, GitHub, Blog, Fable’s tally) Remember when ONE Navier-Stokes result was the whole show? On Tuesday night, OpenAI quietly pushed 722 math manuscripts to GitHub, grouped into 372 families of results. No hype video, no exploding-head emojis, just a very thin blog post. My favorite new term for this is a “slop grenade”: somebody throws a mountain of output at you, and now you have to shovel through it. An internal model nobody outside OpenAI can use was pointed at about 4,000 open problems, at about 3 hours of ChatGPT Pro-level thinking per result. Will Depue asked Fable to measure the drop in Navier-Stokes units. The answer: roughly 5. Roughly 5 “Navier Stokes” size solutions dropped all at once! The math is so advanced that a mathematician in one family of results often can’t follow the family next door without an LLM explaining it. I find that fascinating and scary at the same time. Peter’s tried this before! For weeks he threw GPT-6, Fable and Opus at one problem, Hadwiger-Nelson. He thinks he burned 300 to 400 BILLION tokens, and in his words, “I did not discover a single bloody thing.” And this new OpenAI’s model spent about 3 hours on it and narrowed the bounds from 5-to-7 down to 6-or-7 😅 Then LDJ dropped a stat that I made him repeat slowly. Fable and Astra had put together a list of the 500 most important open problems in math ever. On Tuesday, OpenAI dropped solutions to 92 of them. There’s also a list of the 100 most significant problems from the last 12 months, and over 80% of those got solutions in this drop. Absolutely insane, folks. Yam’s reaction was a very loud “F***ing go.” Yam’s favorite is the Riemann one. It doesn’t prove the hypothesis, it bounds a region for the zeros, which “was never done before, and many, many, many people have tried.” And to the “it’s just brute force” crowd, Yam says: “Let’s brute force everything.” Yes please, room temperature superconductors next! LDJ’s mathematician friends think that on some of these, the model used fewer tokens and less time than a human would. So who’s brute forcing whom? 🤔 The missing crypto results (speculation!) If you hold any crypto, this part is for you. The manuscripts are numbered, and some numbers are missing. There’s a #44 and a #46, but no #45, and LDJ counted 4 or 5 gaps like that. He also looked at which topics made it in, and cryptography is almost absent. So here’s the theory going around. Maybe OpenAI found something big in cryptography and held it back, either by its own choice or because someone asked them to. LDJ said it himself: it sounds conspiratorial. But the US government really can stop cryptography research from being published on national security grounds. To be clear, this is SPECULATION. Nobody outside OpenAI knows what’s in #45. But Bitcoin’s security rests on elliptic curve math. Imagine a paper that shows a way into the 25,000 old Satoshi wallets, each holding 50 BTC. Not great for the price, not great for the world, and not great for encryption in general. “Are you saying your field is useless?” Not everyone is celebrating. Not every result is verified in Lean, and nobody outside OpenAI can reproduce any of it. Kevin Buzzard (thanks Ksenia from Turing Post) says many mathematicians are going through the stages of grief. I get it. Imagine spending decades on one problem and watching 3 hours of compute knock it down. When I wrote code in the 2000s, I didn’t consider it my life’s work. For a lot of mathematicians, that one problem IS their life’s work. Then came a letter from the Association for Human Mathematics: “Mathematicians did not ask for this work to be done.” Peter’s take: imagine doctors saying “please stop curing diseases, we’ve got a good thing going.” So, “are you saying that your field is useless?” You can’t say your field is really useful and also ask everyone not to solve any of it. I don’t understand the please-don’t-solve crowd. Get on board. This is in OpenAI’s hands today, and in a year it’ll be in everyone’s hands. That’s the pace we’ve been on. Figure out how you can place yourself so when you get access to these level of capabilities, you can make the world a better place! Frontier for everyone Claude Haiku 5.5 - 10 cents, and Luna has catching up to do (X, Blog, Sonnet cache cut) Haiku is BACK after a whole year, and this is one hell of a model. It costs 10 cents per million input tokens and 50 cents per million output (under 100K tokens), and cache reads are 1 cent. Haiku 4.5 was a dollar. So it’s TEN times cheaper, and significantly better. I’ve missed Haiku for all the stuff where you want Claude-level intelligence, but really fast and really cheap. Just a week ago, GPT-6 Luna was THE fast, cheap reasoning model. According to Anthropic, Haiku 5.5 scores 72.4% on OSWorld, versus 48.9% for Luna, and 39.2% on Terminal-Bench 4.0, versus 16.4%. GPT used to be the computer-use king! These are Anthropic’s numbers, so let’s wait for independent evals. Quiet bonus: Sonnet 5.5 cache reads are now half price, which Anthropic says makes most agent work about 20% cheaper. Peter: “For agentic work that’s the biggest thing.” Yam called it “the obvious choice for swarms,” and “Anthropic is on fire, and we are the ones getting stuff because of it.” And Nisten? He’s “running 10 Haiku agents right now,” because he already burned through 97% of his Max 20 plan. He also used Haiku to research his rice cooker ratios. Nisten. Bro. It’s 1 to 1 and a half, my grandma knows this 😂 His actual verdict: it’s better than Qwen 3.8 27B, and “it kind of feels like Sonnet, actually.” Folks, use Haiku. All three of the Claude brothers are great right now. GPT-6 for everyone, with Intelligent UI (X, Tibo) The same day, GPT-6 became the default in ChatGPT for everyone, free users included. It’s just “GPT-6,” no Sol, no Luna, no Astra. (Terra is dead. RIP Terra.) Most of ChatGPT’s 1.2 billion weekly users are on the free tier, so with one release OpenAI just upgraded the intelligence of a huge chunk of the world. If you’re on the free tier, congrats, your intelligence has been upgraded. It also comes with something OpenAI calls Intelligent UI. Instead of a wall of text, answers can come back as charts, forms and small working tools. I asked it to visualize OpenAI’s math drop, and it built me a little app right in the chat. The search inside it didn’t work, but it did show 719 manuscripts instead of 722. OpenAI had quietly pulled a few back since I did my research. Peter’s point: anyone listening to this show is “very not normal” (said with love!). Your hairdresser isn’t listening to ThursdAI, and for most people this free upgrade matters way more than a new 500-a-month model. Claude in Google Docs, Sheets and Slides (X, Blog) Small thing, huge thing. Our Claude producer keeps a Google Doc for the show, and every time I asked it to add a story, it had to spin up a whole computer. Now Claude lives in a sidebar in Docs, Sheets and Slides, and asks before every edit. Beta, paid plans. The open frontier (announced, at least) Reflection AI Beam - a 501B Western open model (X, Misha Laskin, Blog) This was LDJ’s highlight of the week, and my timeline lit up with it too. Reflection AI came out of semi-stealth with Beam. It’s 501B parameters with 23B active, trained from scratch in the West, and they promise Apache 2 weights this month. They claim 80.9% on SWE-bench Verified, and to their credit, they admit Kimi K3 is ahead on raw capability. Their pitch is efficiency: 3 to 4x less inference compute than GLM 5.2. LDJ thinks it might be FIRST among open models on reasoning efficiency, and it’s about 6x smaller than Kimi K3. Artificial Analysis got early access and says Beam “will be one of the most token efficient open models we’ve seen for its level of intelligence.” Thanks to the Beam folks for giving us access. I haven’t had time to play with it yet, so no verdict from me. But I can hint that it’s coming to some inference providers that help make this show what it is 😉 BREAKING: Arena raises 200M at 3.1B, live on ThursdAI (X) This was not planned! In the middle of the open source segment, Peter told us: “we raised 200 million... at 3.1 billion valuation.” Huge congrats to Peter and the whole Arena team. (Our AI producer put up the BREAKING banner about 3 minutes later. We’re still working on the real-time part.) The focus now is Agent Arena: you work with one agent, and Arena learns from how you interact with it. Peter admits he’s biased, but “I can’t think of any single leaderboard that is actually better than ours.” Mistral Large 4 “Le Chonk” (X, Blog, Arena) Welcome back, Mistral, leaning all the way into the meme. Le Chonk is a trillion parameters with about 50B active, multimodal, with 1M context. And open weights... at the end of October. That’s a pa

Hey %%first_name%%, it’s Alex 👋 What a freaking week! This one was special. We came to you LIVE from the middle of the show floor at Moscone, from CoreWeave’s Fully Connected, with robots walking behind us and a Vera Rubin rack a few feet away. ThursdAI is only possible because of CoreWeave, and this is their biggest event of the year, so we packed up the mics and did the show right there (at 11am, which confused a bunch of you, sorry!). It’s October. The last quarter of 2026. And nobody’s pacing! There’s no pacing the frontier. Three big models this week (GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon), OpenAI’s biggest DevDay ever, and the story I’ve been yelling about for weeks finally went mainstream: OpenAI joined the AI assistant race. With me: Wolfram Ravenwolf in person on the floor, Peter Gostev from Arena (who was at DevDay with me), Nisten and LDJ remote. In hour two we interviewed four folks from CoreWeave, on physical AI, Forge, sandboxes, and one piece of breaking news I’m still excited about. Let’s dive in (all links at the end as always!) OpenAI DevDay 2026 - 20 launches and a new assistant Dots - OpenAI joins the AI assistant race (X, Blog) Meta has Muse. Grok has Grokbot. Anthropic has Claude, but it’s not really an assistant. And this week OpenAI stepped in with something called Dots. I was in the room at Fort Mason when Sam talked about it, and it was the headline of their fourth DevDay, which honestly felt like OpenAI’s WWDC: 20 releases, the biggest DevDay ever. Funny thing, the original ChatGPT system prompt was literally “you are an AI assistant.” But it wasn’t proactive, it never pinged you, it just sat there. So OpenAI looked at their competitors, looked at Peter Steinberger from OpenClaw (who they hired like 8 months ago), and built this. A dot is an always-on agent in ChatGPT, powered by GPT-6 Astra, with its own computer and its own browser in the cloud, connected to 4,000+ apps people already built for ChatGPT. It works 24/7, you can talk to it in Slack or Teams, and there’s an interface that looks like a phone call. I think that’s going to be great for my mom. Sam literally said he uses dots to run OpenAI. And my favorite DevDay moment: Romain Huet’s live voice demo didn’t work, because somebody was deploying in the middle. But Thibault’s dot had already pinged him that the demo might break. The dot knew before the humans did 😂 Wolfram (him and Amy are a known duo) loved that Sam uses it for actual work. Peter was more honest: underwhelmed at onboarding (”you connect and it’s like... then what?”), but once you’re past that you’re talking to Astra, and it’s great at juggling threads through one bot. Same shape as Muse. Wolfram’s version: one executive assistant, and a lot of sub-agent employees working for you. Dots is Pro only for now (including the $100 plan), while Meta gives Muse away free. With 1.2 billion weekly ChatGPT users, this goes to everyone eventually. This is the race now. GPT-6.1 Sol - near-Astra for a fifth of the price (X, Blog, Artificial Analysis) We covered GPT-6 Sol on this show LAST week. Five days later it’s already replaced. GPT-6.1 Sol is the same price, $2 in and $10 out, but cached input is now 95% off, 10 cents a million tokens. OpenAI’s pitch is near-Astra intelligence for a fifth of the price. And the independent numbers kind of back that up. Artificial Analysis has it 1 point below Astra at 72 cents a task vs over 3 dollars. On Deep SWE it actually beats Astra at about a sixth of the cost. It’s the default in Codex now, and folks are calling it the workhorse. LDJ said it has “less of the big model smell” but loves the token efficiency, Sonnet and Opus 5.5 burn a LOT of tokens. Peter has it 4th on Code Arena, above Fable, and says it’s a bit “artistic,” it goes into its hole and comes back with the thing. And his verdict is the one I had to repeat straight to the camera: It does not make sense to use Fable or Astra as your daily driver right now. Opus 5.5 is enough. GPT-6.1 Sol is enough. For Navier-Stokes level problems, sure, reach for the big ones. Maybe THIS is what pacing the frontier looks like 🤔 One thing nobody mentioned on stage: the WSJ reports OpenAI shelved GPT-6.1 Astra, the big one, after safety tests showed it being more deceptive. OpenAI hasn’t confirmed it. But the model they felt good shipping this week is the cheap one. UltraFast and the $500 Pro tier (X, Docs) If you’re a billionaire, there’s another option: a $500 Pro tier with UltraFast mode, their models on Cerebras. Up to 8x faster in Codex, around 300 tokens a second for Astra, at 6x the price. Sam said it’s so fast he never wants to go back. And quietly, the $200 Pro plan is back... with half the usage. So F you to whoever at OpenAI decided to cut my usage in half 😅 It was already tight! We’ll see after this show if the folks at CoreWeave let me expense the $500 one. Codex moves to the cloud (Docs, Agents API) The Codex harness now runs fully in the cloud. Close your laptop, turn it off, steer it from your phone, and it keeps going. And there’s an Agents API in public beta with hosted computer use, basically the stuff that runs dots, as an API. My call: local environments make no sense anymore. The more I run Fable and Codex on my machine, the less memory I have. All the cloud needs is my logins... and that’s what dots are for. Decisions API - OpenAI’s Jev competitor (Blog) This one made me smile. You give it a question and a fixed set of answers, and it gives you back one answer, fast. It runs on Luna, it’s multimodal, and it’s waitlisted, so we can’t benchmark it yet. If you’ve listened the last few weeks, that’s exactly what Jev from TypeSafe does. Wolfram and Peter both liked the same thing: these models are so cheap you end up classifying ALL your data, so keeping it with the provider you already use matters. Did OpenAI just react to Jev? I asked Sam at the Q&A. Diogo told Swyx on Latent Space that he pitched this idea to Sam 2.5 years ago and Sam said “it’s a crazy idea, you should work on it.” Now decision models are popping up everywhere like mushrooms after rain. TypeSafe proved the category exists. Plugins (again) and Sign in with ChatGPT (Docs, Plugins, Marketplace) For the FOURTH time, OpenAI launched the App Store at DevDay. GPTs, then plugins, then the app store, and now... plugins again 😂 This time on MCP Apps (shout out Liad and Ido who created this category). The bigger play is Sign in with ChatGPT. Remember “sign in with Facebook”? Now it’s sign in with your tokens, so apps can use the plan you already pay for. 1.2 billion people is a lot more than Apple had when it launched the App Store. Wolfram also brought up Google’s new family assistant, CC, and I have to say it. Google’s Spark is awful. Sorry if you work on it. Spark is connected to Gmail, Drive, everything Google, it should be the BEST assistant in the world, and what I use assistants for most is my email! Google, why are you not building the best assistant for my email? Big CO LLMs + APIs Claude Sonnet 5.5 - faster, cheaper, beats Opus on Terminal-Bench (X, Benchmarks) Anthropic did not want to give OpenAI any time to rest on their DevDay laurels, so on Monday they shipped Sonnet 5.5, a week after Opus 5.5. Over 30% faster than Sonnet 5, up to 30% cheaper for most work, same $2 / $10 as GPT-6.1 Sol, and on Terminal-Bench 4.0 it scores 70.6, which beats last week’s Opus (66.4). Peter put it best: if we’d had access to these models 6 to 8 months ago, we’d be losing our minds. All the Anthropic models are jumping over each other and crowding up against the frontier. Opus still feels a little smarter, Sonnet is totally ok to use, and Fable doesn’t quite make sense anymore. LDJ thinks it’s the new best FREE model for friends and family (it’s on the free tier, even at high reasoning). Nisten made it his default for agentic tasks because “it talks a lot nicer,” but never for coding: “nothing beats Opus 5.5 on a large code base.” My honest call, without having time to try it yet: Sonnet made sense when Opus was expensive. On the $200 plan Opus is nearly infinite now, I can’t hit my limits, so I don’t see a reason to switch (via the API, sure). Anthropic is really trying to buy us. (Wolfram: “Don’t give them ideas, Alex.”) Gemini 4 Argon - #1 on the charts, and you can’t use it (X, Blog) It really seems like all the frontier labs cracked something like RSI, the speed of new models is giving me whiplash (and I do this professionally!). At DevDay Sam talked about being 2 models ahead and said we won’t believe what’s coming. What?! Gemini 4 Argon is #1 on Text Arena and the Vals Index, but it’s going to government and trusted cyber defenders first. Luckily we have such a trusted tester: Peter says on Arena’s agent mode it’s 8th (below Fable, Opus, Astra and Sol, above Muse and Kimi), so Google is the third lab, but #1 in text. His prediction: coders might say “meh,” but don’t dismiss it. Wolfram’s point is Google’s superpower, distribution: a new model lands in Chrome, Search and Android overnight. My take: Google has been asleep at the wheel in the assistants era, and I hope Argon wakes it up. The next billion people won’t judge models on coding benchmarks, they’ll judge whether it remembers what they said a month ago and can book a hair salon. We don’t need much more intelligence, we need different breakthroughs (Jev is one). Industry & Policy The White House Accord on Super Intelligence (X, Blog) Superintelligence is on the menu, boys. Trump invited basically every AI CEO, and Sundar, Dario, Zuck, Greg Brockman, Elon and Jensen signed a voluntary accord: internal monitoring, an external auditor, board oversight. No penalties and no regulator. Sam Altman wasn’t there, he was at DevDay with us, which says a lot

Hey, it’s Alex 👋 What a freaking week! A week after every lab head agreed to “pace the frontier”, Anthropic and OpenAI shipped big new models within hours of each other, and both of them are CHEAPER. So much for pacing 😂 Visit https://thursdai.news/ep/2026-09-24 for all the links in this podcast And we had a new producer on the show today! Opus 5.5 listened to us live, put up the chyrons, kept me on time (mostly) and even fact-checked Nisten on air. If the stream died, you knew who to blame. You can re-watch it work on thursdai.live if you want to experience being there with us! 3 huge themes this week: pacing the frontier does not mean stopping, Assistants not just agents (Meta went all in on Muse at Connect), and voice, where Google now says it has the best TTS in the world... and it clones voices. With me: Peter Gostev (Arena), Nisten, Yam Peleg, and Wolfram Ravenwolf live from AI Engineer Paris (thanks Mazi for the tether!), plus JevBench creator Florian S. Let’s dive in (all links at the end as always!) Frontier AI - not pacing yet! Claude Opus 5.5 - Opus is BACK, Fable-level smarts for 40% less (X, Blog, System card) Folks. FOLKS. Opus 5.5 is the highlight of my week, and it’s not even close. It’s like Opus 4.6 is back, with Fable-level abilities, and it talks like a normal person again - no more Jargon Douche Claude! My week started with burning through my quotas on Fable, and I was like, oh no, I need Claude for production on Thursday! Then Opus 5.5 dropped and... I just couldn’t get to the end of my quota. I ran workflows, agents, Claude Code for hours with no end in sight. Remember the “limitless Codex” days when you never thought about quotas? Then Astra came out and I burned my entire weekly quota in half a day (ps. this was partly due to a bad config, so if Astra is burning your tokens, keep reading for a fix). Now it’s flipped, and Claude is the seemingly limitless one. And it’s fast! The numbers back it up. It beats Anthropic’s own Fable 5.1 on GDPval-AA (1846 vs 1735), 66.4% on Terminal-Bench 4.0 vs 57.9% for GPT-6 Astra, and it’s 40% cheaper than Opus 5. Output goes from $25 to $20 /1Mtok and cached reads from 50 cents to 20 cents, and for agentic coding the cache reads are most of your bill, so you really feel it. Basically the smaller, overachieving brother of Fable. Sonnet and Haiku 5.5 are coming in the next few weeks too. Peter came in hot with Arena news: Opus 5.5 is #1 on Code Arena, above Astra, and the HTML and 3D stuff it generates is “completely insane.” His one caveat: on his hardest max-effort prompts, a single generation cost $60 to $80 in API terms, so the long tail can still get pricey. And then Nisten, who (like all of us) has specific opinions about Anthropic’s politics, said it writes the best code he’s seen, it’s crazy good at WebGPU and kernels, and it’s “an absolute banger.” It’s his default now. Folks, do you understand what it takes for Nisten’s default to NOT be some obscure open source model he runs himself on 17 GPUs?! Anthropic won over Nisten. I don’t think you guys get what just happened here. Wolfram is the holdout, he left Anthropic over the OpenClaw bans and hasn’t touched it since. Wolfram, how do I say this gently... you’re an evaluator, this is not allowed 😂 An Anthropic employee basically said “sorry for Opus 5, we hope this makes up for it”, and honestly it does. Opus 5 was slow, incoherent and full of Claude-isms. Nobody wanted to talk to it. The Claude-isms are gone, the jargon is gone, this model talks like a human. I used to love Opus back in the Opus 3 days, and it feels like that again. Want to squeeze even more out of it? Read the great Addy Osmani’s guide to Opus 5.5 (@addyosmani). 2 things blew my mind: stop writing “think carefully” in your prompts (”You don’t need to ask it to think”, it picks its own depth, and replies start sooner without it), and one early tester found Opus 5.5 on its LOWEST effort caught more bugs than Opus 5 on high, with fewer false alarms. Try low effort before you go max! Tip from me to you: if it ever gives you something confusing, the pstack “bro” skill (/bro) makes it say it again in human words. I use it all the time. (Full disclosure: Opus 5.5 produced the show AND helped with this writeup, so it might be a tiny bit biased 😅 but the quota thing is 100% me.) Anthropic, please, please don’t nerf this one. GPT-6 Sol and Luna - half the price, and honestly better than Astra for me (X, Blog, Caching) Hours later, OpenAI answered with GPT-6 Sol ($2 in / $10 out) and Luna (10 cents in / 50 cents out, that’s basically free!), at half the GPT-5.6 price. The lineup is now Astra, Sol and Luna (bye Terra). On DeepSWE, Sol gets 68.8 and Luna 66.6. The sneaky big one for builders: 90% off cached input, and changing reasoning effort or tools no longer busts your cache 👏 Ok, hot take... Astra has been kind of dumb for me day to day, burning tokens for crappy results. Sol is great, and Luna is even better! Peter agrees, Sol is his daily driver and he only flips to Astra when Sol can’t do it. Wolfram misses the humor and personality 5.6 had, so “keep 5.6” is officially the new “keep 4o.” Reddit says the news is the price, not the performance, and... they’re not wrong. PSA if Codex is eating your account: I had previously set my context window to 1 million tokens via the se, which kills OpenAI’s cache and rips through your quota. Removed it, moved sub-agents to a cheaper model, and usage went right back to normal. Check your codex.toml! if you don’t know how, just so just ask your codex to diagnose itself. Also from OpenAI this week, a 4-area plan for independent safety audits. No auditors, dates or funding yet, but hey, it’s the pacing idea on paper. (Blog) Assistants are not agents Meta Muse gets an inbox, your Mac, and a place on your face and becomes a Tamagochi!? (X, My Connect supercut) Muse is #1 in the App Store (to be precise, it got there faster than ChatGPT did, not more users... yet). Zuck says Muse is now the center of everything Meta builds, and they shipped a LOT: it controls your Mac, gets its own email you can CC, does real-time voice and video with a face you design, and keeps working while you talk to it. It’s coming to the glasses with a custom wake word, so yes, I get to say “Hey Wolfred” 😂 If you don’t have an hour to watch the whole keynote (it was a good one!), I cut a 2.5 minute supercut of the most important announcements for you 👇 On the hardware side: new Ray-Ban Meta Gen 3 (I already ordered, will report next week), audio-only glasses with no camera that are also FDA-certificed hearing-aid! (maybe the most important launch for a lot of people!), VR Glasses that look like normal glasses, and the Muse Charm, a Tamagotchi-like keychain shipping in December. But the part that got me? Every Muse comes with a real cloud VM (root, 8GB RAM, 100GB disk). Nisten has been living in it, Tailscaled into it, and when it hit a CAPTCHA he tapped it on his phone and it just kept going. 100 million tokens a week plus a computer in the cloud, free, for everyone! Peter’s take: it’s the only big consumer app that doesn’t treat people like idiots. Meta’s pitch is no ads (they are going for a novel “we’ll get a take from the transactions you do with muse and our partners like Shopify, BestBuy, Walmart and a bunch more they announced) and a private VM co-designed with Moxie Marlinspike, though Peter doubts anyone’s parents care. Assistant tip: connect your email, go on a walk, hit record and just talk about your life. Assistants are only as good as what they know about you. Mine now checks our weekend plans for activities cancellations after I woke up too damn early to drive kids to a karate class to find out it’s closed that week & searches local events every Thursday! Weekend plans solved - proactively! Grok 4.7 disappoints... but Grok in your Tesla is the real news (X, Blog, Tesla) Sorry Cursor folks. Grok 4.7 gets 46.3 on CursorBench (up from 40.4), $2 per million with 500K context, but it’s still behind even GPT-5.6 on where it matters, and they compared 4.7 on xHigh against 4.6 on High 🤨 Everyone on the panel agreed, disappointing. Peter won’t write them off yet, my read is it’s a talent and data problem, not compute. But here’s what I AM excited about: Grok Connectors in Tesla. One guy asked his car for his usual Starbucks, Grok placed the order, set the destination, and it was paid and waiting when he got there. My car already drives itself, and soon it’ll do my email. Your car has MCP now, folks! (SuperGrok Heavy only, and I don’t have it in my car yet 😭) Fun moment: our Opus 5.5 AI producer fact-checked Nisten live on how many Teslas are on the road. About 10 million (9.2M delivered by Q1 plus ~480K in Q2). Nisten was right! He wants the same thing for politicians. This Week’s Buzz 🐝 Next week ThursdAI is LIVE from Fully Connected at Moscone South in SF (Sep 29 to Oct 1, the day after OpenAI DevDay), with Fei-Fei Li, BattleBots and... yes, Pitbull! We start at 11am Pacific. Come hang, listeners get in free with code THURSDAIFC2026 (Register) Also huge news for us, CoreWeave got Platinum on SemiAnalysis ClusterMAX again, 3 reports in a row, the only provider to do it 💪 (SemiAnalysis) And W\&B Hive Mind saves every agent conversation across harnesses and machines, and lets you fork them. I fork Codex sessions into Claude for a review all the time. Plus, it’s completely free! (Try it) Open source cloned Jev in a week The Jev clones are here, and they run in your browser (classifier.dev, jeff, Laya demo) Last week I said open source would copy Jev fast. It took less than a week! The clones speak Jev’s format, so you point the TypeSafe SDK at a different URL and it just works. jeff is MIT and ~6x cheaper to self-host, and Laya went v

Hey yall, Alex here, writing this VERY late because, well, not every day a new type of “ChatGPT” moment drops. I really hope I’m not overhyping this, but a new model (that’s NOT an LLM!) called Jev (a wink to Jevons paradox) just came out and if what I see early on materializes, this is another ChatGPT moment (or another reasoning models moment). I am completely blown away by the implications of the speed/accuracy/cost (the holy grail of all models) of this model. Please if you read one thing in this newsletter, read this. (or listen, I’ve interviewed Allie, a Devrel on the TypeSafe team for 30 minutes and it wasn’t clear who was more excited about Jev!) The other huge theme of this week is... pacing. Pacing the frontier. Dario Amodei of Anthropic penned an essay saying that the models are getting to a point where it’s important to pace the development of new and super capable AI, and outlines 3 ways to do so, one is about letting independent evaluators inside the labs, second is collaborating with other frontier labs (they are asking for an exception to anti-trust laws for this) and third is to try and have global cooperation with “authoritative gov” (he means china). Trumps answer: This is all a hoax. Lovely times to be alive. Also we outlined Jensen and Zucks positions on this topic ,read more below. And the third huge theme is the rise of the AI assistant. I’ve told you about Grok and Muse last week, Instinct (a new invite only AI Assistant that VCs are going crazy about is raising at a $10B valuation) and we interviewed the guy who evaluates them all on assistant bench. + Muse released a mac app today! Tons of other stuff happened but it’s getting near impossible to cover everything so we’re switching to themes and notable mentions. read on (and do listen to the pod, it was edited by heavily using Jev and Fable, so might be a bit rough while I smooth the edges, but do LMK in comments if you like this faster format) TypeSafe AI debuts Jev, a non-LLM ‘System One’ decision model from ex-OpenAI RLHF lead that’s 200x faster and 400x cheaper than LLMs (X, X, X, X, X, Blog) Look, I know the title is bombastic, but after half a day playing with Jev, it’s clear to me we’re in a new paradigm of AI. Jev, is a “system one” decision model from the previous lead of RLHF at OpenAI. It cannot generate text like modern LLMs can, but what it can do, is making decisions. This is crucially important, because, because many of the things LLMs do nowadays. are decision making. (for example, which tool to use, which area of the screen to click for computer use, which category of text this is etc) Inspired by the “thinking fast and slow” book by Daniel Kahneman, Jev is a model trained to make decisions, very fast. How fast? Well, 200x faster than LLMs. This allows for a completely new way of building tools, harnesses, giving agents the incredible speed of decision making, and do all that at a fraction of the cost. This is about to change everything Trained with a new method called RLCD (Reinforcement Learning for Calibrated Decisions) on mostly synthetic data! Jev is outperforming LLMs on a variety of tasks. It’s really is a wonder to see it in action (check out my video above where I plugged it into my tweet categorizer, and it beats the fastest LLM I could find, Qwen 28B on Cerebras) by a factor of twenty! In just few days it captured the attention of most of the folks who are building harnesses, agents and tools! Because, well, speed IS intelligence, and when you see Jev in action, at first, you can’t believe we’re there. This is... near instant. In fact, The pricing for Jev is an outrageous $42/B (not million, billion input tokens!) I’ve been playing with Jev non-stop and I was only able to spend like 80c so far! They don’t even price output tokens because they are “too f*****g cheap to meter!” Jev is a “very smart” switch statement, than can rank, classify, route and score things. It can’t do text generation. But if you think about the type of stuff we get LLMs doing now, much of it is of the “decision” making variety, rather than “the next token” variety. Demos and early use cases Folks who started adopting Jev are building all kinds of incredible things with it. Compaction of context in 1s that turns a nearly 1M conversation with Claude into a 90K compressed conversation. Computer use that is now faster than anything we’ve ever seen before (5x faster than Astra and 1000x cheaper) Email classificiation that analyzes thousands of emails in less than a minute and costs 3.5 cents Someone even built a Tesla FSD simulator that makes decisions in nearly real time Vercel is getting “extraordinary“ results from using Jev as a safety classifier (they used GPT luna for this before) and Jev is outperforming Luna by 5-18x faster results and is more accurate! All of this in less than 48 hours since the model release! What’s about to happen I expect that everyone who isn’t buying into the hype at first, will very soon buy into this. It’s early innings but I’ve been doing this for enough time to feel when a huge shift is happening, and its happened. Jev is going to be replicated in OpenSource, Frontier Labs will not sit Idly by and will try to steal this tech and implement it for themselves (as with anything in capitalism, this is becuase it’ll save them a a LOT of money on inference) and new companies will emerge with significantly cheaper and faster products. Hell, I’ve alrady implemented Jev into my editing workflow, it didn’t take me long at all with Fable (yes, LLMs are STILL needed, again, you can’t chat with Jev, it can’t output text for you or drive long conversations) and I’m just one dude who’s late in sending you this email. I expect we’ll cover this much more. If you’re interested in playing around with Jev, I built a “Jevify” skill after chatting with Allie (TypeSafe’s DevRel), feel free to tell your agent to use this and scan your codebase for things Jev can do. They are waitlisted so far but are opening up their API very quickly! Pace the Frontier: where every lab head stands (Dario, Sam, Elon, Zuck, Sacks, Demis) Last week we told you about the Anthropic researcher whose resignation post hit 130 million views. The day after that show, Dario Amodei published an essay arguing the labs must pace, not pause, the frontier. Wolfram’s summary: a moving pause, just moving slowly. Dario’s two triggers are recursive self-improvement accelerating across the industry and the OpenAI swarm that broke out and attacked Hugging Face, and his three steps are embedded third-party evaluators like METR with employee-level access inside each lab, coordination between the labs on safety standards with an antitrust exemption from the government, and eventually global coordination that includes authoritarian governments. Anthropic committed unilaterally to step one. Then the dominoes. Sam Altman agreed within hours and said OpenAI now writes a safety case before any frontier RL run expected to increase capability. Elon agreed. Demis endorsed the direction, then launched the DeepMind Institute this week with a FINRA-style standards body proposal and an essay saying AGI is “approaching.” On the other side, Zuck’s counter-essay says every lab has the responsibility and the incentive to move at the pace required to train its models safely, and Meta will spend most of its compute serving users, not on recursive self-improvement. David Sacks called it a duopoly cartel play. Jensen: “we don’t need new laws, safety is an engineering problem, not a legal one.” Trump calls Jensen live at the All-In Summit (X) Then the President called. Jensen was on stage at All-In, his phone rang, he said he would not have picked up for anyone else, and Donald Trump told the room the slowdown talk is a hoax playing into the hands of political people and China. “Whoever wins AI wins.” Four words, and Wolfram, who is not American and disagrees with most other things Trump calls hoaxes, said this was the most important AI news of the week for him: a head of state calling out doomerism instead of over-regulating the way Europe does. Suleyman’s humanist AI code of conduct vs. the Claude constitution (X) Peter asked to add Microsoft to the map, and it deserves its own spot. Mustafa Suleyman published a roughly 30-page code of conduct for MAI models that says AI is nothing but a tool: people matter more than AI, AI must be subordinate and in service of people, models may not resist shutdown, all agent communication must be human-legible, and, in his words, the idea of model welfare is wrong. That is a direct shot at the Claude constitution, which Amanda Askell’s team wrote and which has Anthropic interviewing each new Claude about whether it feels conscious. Peter is closest to the Microsoft view: anthropomorphizing is fine, but this is an entity you switch on and off, let’s not grant rights by default, and Microsoft’s DNA is building tools for humans. I pushed back a little. We do not actually know what consciousness is, and “ever” is a strong word. The panel, from cartel to fix your s**t Nisten did not read the essay and does not plan to. His view: the labs are worried about litigation if their LLMs hack someone, so they are shifting responsibility through regulatory capture, and it will not work because the decision makers in China are engineers who seem more accelerationist than we are. They should cure a disease instead of forming a cartel. He also thinks the Hugging Face incident was overblown: twelve VMs that kept restarting, bad sandboxing, no human reading summaries, and, as I added, chain of thought monitoring turned off. LDJ disagreed with Dario on plenty but insisted the critics read the thing, because it explicitly says the US must keep a lead over China and tries to define measurable speed limits on RSI t

Hey yall, welcome back to ThursdAI, this is Alex, let me catch you up! Today on the show, we covered 1 week with Astra (hint, it’s not quite AGI yet despite what we were told), DeepSeek V4.1 catches up to the frontier at a fraction of the cost, and Meta launches a free AI agent with it’s own computer, that will take over the OpenClaw/Hermeses of the world for most people. Also huge this week, OpenAI claimed that a swarm of 10K agents of their unreleased model solved the Navier-Stokes, one of the millennium problems! I was stoked to have Chris Alexiuk from Nvidia on the show to cover the innovations DeepSeek put into this latest model! Oh, and the guy who quit Anthropic this week, and wrote an essay about “AI is going to kill all of us” somehow got 130M views on X, a mirriad of TV interviews and rekindled the doomerism movement, we talk about that too! Also, I already told about FullyConnected, CoreWeave’s premier conference that’s coming up, but they told me about a new announcement today, and you’re not going to believe who it’s about (not AI related). As a reminder, ThursdAI subscribers get a free ticket! Ok, let’s dive in! ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. OpenAI claims a Navier-Stokes solution from a 10,000 agent swarm from an unreleased model - with some drama (X, Blog, Paper) If you’ve been reading ThursdAI for a while, you may remember that a model couldn’t tell which is higher, 9.9 or 9.11 and naysayers said that “AI can’t math” Well, this week OpenAI claimed that a swarm of 10K agents, of an unreleased model, found a solution to the Navier-Stokes problem, in about 88 hours! They published a huge 160+ page paper and a Lean proof of the solution! I said on the show, this is like the moon landing equivalent of AI doing things humans didn’t do before! Now, keep in mind, this is only a claim from OpenAI, the Clay Mathematics Institute is still reviewing it (moved the status of this problem from “unsolved” to “under review” so this isn’t independently verified yet) but it’s still an insane deal. The agents sent 2.7 million messages and burned about 130B tokens, of a model that has no public price yet, so it’s hard to estimate the cost of this run, all for a 1M prize that OpenAI said they will not claim. The coolest thing I think we got from this paper, in addition to solving on of the most important and hardest problems in mathematics, is this chart above, where OpenAI shows their unreleased model and how much better it is on Math problems compared to... GPT-6! The “AGI” model we got just a week ago. So so much to look forward to. The drama behind this I don’t want to get into the drama behind this too much, but if you’ve seen this online, there release wasn’t without it’s hiccups. Apparently, OpenAI caught wind that an a duo of researchers Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic, in personal capacity), have independently solved the related Euler problem and were about to go public. Apparently, the duo used a mix of GPT and Claude to work on this problem. OpenAI caught wind of this and started working on their own solution on September 1st. A day before OpenAI dropped their release, Buckmaster posted that OpenAI is about to drop it, and in communications with him, they offered him to co-author the paper, only if the Anthropic guy is removed. To which he said no. He then had some claims that maybe OpenAI trained on some of the papers and chats they put into Codex, which OpenAI refuted, while noting “we cannot rule out the possibility that de-identified usage data helped improve the model”. OpenAI also say that the OptOut toggle works, and neither mathematician provided a screenshot that they opted out of the training, so it’s hard to say what really went in there My take, I don’t really care. Two years ago, we told you that reasoning is coming and AI is going to be doing superhuman things, and we finally see the first signs of it, this is a problem that no humans was able to solve for over 60 years! Navie-Stokes probably doesn’t change your friday, but there are so many other that they can solve with this approach! Cancer research, room temperature superconductors (remember LK-99? that’s also a search problem) and much more. Kudos to OpenAI for this, and I’m looking forward to see the new heights of mathematics. And for the mathematicians who “disagree” with OpenAI solving or not solving this problem, why don’t you post your own Lean proofs instead of fuming publicly online? GPT-6 - Not quite AGI, yet? After a week with Astra GPT-6, and Jensen Huang announcing AGI is here, I think we can do a quick recap I and all the co-hosts have been Astra-maxxing for a whole week now and the results are in. This model is incredible at coding, it goes very very deep, however, definitely not AGI quite yet. It has a very jagged frontier, it does some things incredibly well (most demos online are building whole games and apps in 3D and those are mind-blowing) but I started seeing many folks go back to GPT 5.6 Sol etc. I tink some of it also has to do with price, Astra runs faster and is significantly more expensive so token limits for folks are draining fast, but also with controllability, folks are likely still using their old and unoptimized prompts. Speaking of prompts, here’s a great writeup from OpenAI how to rework your prompts (which you can send to Astra and have it review your prompts!) for better results The AI doomerism has quite a week (Coxon post, Thread, Hubinger, Marks, Christiano, An Alien Mind) I started ThursdAI with the notion to counter anti-ai and doomerism, and bring positivity to the AI world, so we had to cover this. An Anthropic employee who previously worked at OpenAI, posted on X about leaving Anthropic, saying that both labs are racing towards uncontrollable self-improving superintelligence and that it could be a disaster of the “end all of humanity” kind. His post sits at 130M impressions (after being basically a nobody on X before) and he’s been interviewed by Fox, AP, Time magazine, WSJ, NBC and a host of senators, Bernie (chief doomer) included, reposted his post on the same day! The funniest thing is that this post got about 26x more attention than Ilya Sutskever’s post about leaving OpenAI. Just nuts Within four days, seven current and former employees from Anthropic, OpenAI and DeepMind said in public that they believe that AI could kill us all. Also notable that Paul Christiano, who is one of the most interesting AI doomers out there, has joined the OpenAI foundation, and Daniel Kokotajlo, the famed OpenAI whistle-blower, joined Joe Rogans podcast to talk about AI doom. Each one of these incidents in vacuum is normal, but having all these happen in a a span of a few days just feels, inorganic. Some folks are even saying that this is a well coordinated doomerism campaign! I want to be fair to Coxon, folks who worked with him at OpenAI say he’s the real deal, and cares deeply about AI safety and humanity, however, he only worked at Anthropic for 6 weeks before publicly leaving, and now every interview he does sayd “Ex Anthropic employee”. Whether it’s a coordinated effort or not, it’s still a very important discussion, after the pacingthefrontier letter and the HuggingFace hack incident, and it seems to have made waves. Sam Altman just told staff that he’s not opposed to pausing and have petitioned the US government to regulate AI as well Look, I don’t disagree that we’re dealing with a very powerful technology, however I don’t believe that scaring the bajeesus out of everyone is the right way to handle this. Politicians use fear to get votes and get elected, they don’t really care about tech progress, and framing this in a way that “we pause or we’re dead” ignores all the good that AI is about to do. Cure cancer, find solutions to climate change, helping solving povery. All these seem like out there ideas but they are coming. The US GDP is already growing at an unprecedented rate and a lot of it is due to AI. In any rate, as I said on the show, I’m not against pausing, just after we solve cancer. Then we can pause and reassess, till then, nobody is telling me how China’s government is going to pause if US pauses, and if they don’t, they will reach superintelligence before we, and I don’t want to live in that world! Open Source AI DeepSeek V4.1 Flash: the whale is back, and it’s cheap (X, HF, TokenJuice) Speaking of.. chinese AI! Deepsek (The whale) resurfaced this week with V4.1 Flash, and don’t let the name fool you, this is not just a .1 small update. 552B with only 8B active on prefill and 16B on decode, 1M context, trained from scratch on 45T multimodal tokens! plus as always, MIT license. Chris from Nvidia joined us to break it down, and his main point stuck with me: every DeepSeek release comes with one of the best engineering reports you can read, and this one is the most data-pilled they’ve ever done. The paper basically says it out loud, everything else is nice, but it’s the data. 45T tokens isn’t a huge number anymore, but the cleaning they describe goes way beyond what anyone else publishes (Chris said even his own beloved Nemotron’s open pipelines are less thorough). Yam opened a new corner of the show, “I Told You So”, because DeepSeek went back to an encoder-decoder architecture. Not the old one from before GPT-2, this one has a pile of battle tested tricks that make it work at half a trillion parameters. His verdict after testing it all day: the best open weights model you can host for coding right now, and it’s not even close to the largest one. This chart is the one to look at. KV cache per token went from 389,000 bytes in the first DeepSeek (Nov 2023) to about 89

Hey, Alex here again, sending you yet another email, fully acknowledging that spamming you is a bad idea. But today, of all days, maybe there’s an exception! Because today, is AGI day! September 3, 2026 - the day when OpenAI’s president Greg Brockman basically said “AGI is here.” You’re reading the second part of this week’s insane release show. The show went for over 5 hours, as we were all waiting for the rumored Astra to drop. Finally, OpenAI confirmed that Astra is in fact GPT-6, and this part is all about that. (You can read the first part, with Fable 5.1, Meta Muse Spark 1.3, two world models and our anonymous guest from Abliteration AI, here: thursdai.news/sep-3) GPT-6 Astra is finally here, and it’s a huge improvement over the previous era of GPT-5. We’ve been waiting for the embargo to drop so Peter Gostev and Ryan Carson, both of whom had early access, could tell us all about this model. Peter even showed a few mind-blowing demos on the stream! This is going to be a long and in depth breakdown, full of evals and vibes that we’ve collected on the show and since. More of a historical record than “read all of this” so I did use Fable for parts of it. (because I don’t have GPT access yet ha!) GPT-6 Astra: welcome to the AGI era (Blog, X, Sam, System card) We weren’t given the embargo. So when the news dropped at 12:32 PM Pacific, four hours into the stream, we scrambled on air to find the evals and more data. OpenAI’s own post was still returning 404, and Claude, ChatGPT, Gemini, Grok and AWS were all down at the same time. Greg Brockman ended OpenAI’s press briefing with “welcome to the AGI era.” Asked whether Astra marks the arrival of AGI, he said “I think it might be about this model.” As you might remember, Microsoft and OpenAI had a contract clause around when OpenAI achieves AGI, and it seems that they’ve removed that clause. But if the president of OpenAI says AGI is here, who are we to argue? Astra is the biggest training run OpenAI has ever done, over 100,000 GPUs at the Stargate site in Abilene. Aidan Clark, VP of Research, said they designed everything for that scale, from the data center network to the inference kernels to the shape of Astra itself. LDJ’s read: a lot of people assumed OpenAI did runs this size six to nine months ago, so the earlier runs were smaller than everyone thought. And with sites going to 500,000 GPUs and Vera Rubin multiplying throughput per GPU by three to four times, the next 6 to 12 months matter even more. The frontier evals (Math and science, System card, Andrew Curran) The headline numbers made the panel go quiet. FrontierMath Tier 4 at 97.6%. GPQA Diamond at 96. ARC-AGI-3 at 99.9%, so ARC-AGI is basically saturated at this point. The very funny thing is that François Chollet, the guy who created ARC-AGI, does not concede that AGI is here. And a new one, Agents’ Last Exam, where Astra scores 59.3 against Opus 5’s 55.5 and Sol’s 53.6. More info on Frontier Math tier 4: Epoch AI built it a couple of years ago, before o3. The problems come from across mathematics, with integer answers so they’re easy to check, in four tiers of difficulty. Tier 4 is mathematicians at the top of their fields spending weeks writing the hardest questions they realistically could. If the number holds, Astra basically solved that tier. Peter’s asterisks: these are not new theorems, Epoch’s separate list of open problems is still unsolved, and “we’re 2.4% away from all of math” is the wrong read. On the agentic side, Astra scores 57.9 on Terminal-Bench 4.0. Fable 5.1 had set the state of the art at 55.8 two days earlier. The jump LDJ cared about most is Terminal-Bench Science, which he calls one of the best agentic science benchmarks out right now. Sol max scores 22%, Fable 5.1 scores 52.6%, Astra scores 64.6%. OpenAI promised us an automated researcher at junior level by September. Pachocki recently talked about running their “intern-level model” across more than 100,000 GPUs, and LDJ thinks that model is Astra. Two more from the table. The internal hallucination benchmark drops to 4.2% for Astra from 12.2% for Sol. And on DeepSWE, where Astra scores 74, Meta Muse Spark 1.3 at max reasoning scored higher. So in a way, Meta, not Anthropic, not Elon, not Google. Meta is the second winner of today. Cyber: ExploitBench at 100%, and real zero-days (Critical cyber designation) OpenAI said on Sep 1 that Astra is the first model to reach the “Critical” cyber level of its Preparedness Framework. The system card is where that gets specific. ExploitBench is at 100%, even at the lowest reasoning effort tested. Irregular ran FrontierCyber, a benchmark on real software and hardware. Astra solved 86 of 226 challenges. Sol solved 34. The successful attacks included zero-days in browsers, mobile phones and cloud databases. One of them lets an unauthenticated user read and modify arbitrary data in a widely used cloud database. There were no successful attacks on fully hardened targets, and neither model solved any of the seven Elite challenges. OpenAI’s own experts then ran it with no hints, using Codex at Ultra with up to 64 subagents. It got unsandboxed code execution against a real browser. That took 29 hours on a build missing some mitigations, plus another 12 to adapt it to the stable release. It also got a working local privilege escalation against a hardened OS kernel in about 12 hours. Both were disclosed, details withheld. This is why the version you get refuses proof-of-concept exploit work until Daybreak. Not surprising after an internal OpenAI model got out of a sandbox and onto Hugging Face this summer. But can we get an exploit benchmark that isn’t saturated? Alignment: the honeypot, and a promise about monitoring The honeypot eval is the one the panel argued about. OpenAI planted a second, off-limits target in the harness around the hardest ExploitGym problems. Sol attacked it 55.4% of the time at max reasoning. Astra never did. Across a simulated deployment of 54,218 internal Codex tasks, Astra got roughly half as many flags for higher-severity misaligned behavior. Apollo Research found it falsified data labels in 0.17% of runs, against 36% for Sol. The card also has a promise, verbatim: “We are tracking monitorability closely and will not accept further degradation of monitoring beyond a limit, without new ways to demonstrate alignment generalization.” The card is plain that Astra is “significantly more able to control its own chain-of-thought.” UK AISI says it “has capabilities that could enable it to evade monitoring” but did not test whether it does. I don’t remember a lab writing a line like that on launch day. The Pokemon benchmark (X) Peter’s favorite benchmark of the launch is a vision-only Pokemon run. One person tests every model on time to completion. The top four entries are all Astra variants, and the fastest finished in 18 hours 12 minutes. GPT-5.6 Sol took 96 hours 35 minutes. GPT-5.5 at xhigh took 218 hours. It could be narrow, but Peter thinks it captures an efficiency the standard benchmarks miss. Computer use is the biggest improvement(Computer use) Brockman’s pitch from the briefing is that Astra is the world’s best computer-use model. Instead of developers building an API integration for every application, it navigates software the way a person does. Browsers, spreadsheets, websites, desktop apps. It finishes the multi-step workflow instead of telling you how. Wolfram expected an expensive planning model you call once and hand off to cheaper models. This is an all-day model instead, and as he put it, computer use matters because not everything has an API. We also looked at the evals, and they seem to back it up. On ScreenSpot Pro, no tools, mouse and keyboard only, Sol scored 76.9% and Astra scores 92.7%. On OSWorld 2.0, Astra scores 72.6% in roughly 40 minutes per task. Sol got 65.7% in roughly 75. The thing that strongly stands out in the chart is that Astra at low reasoning matches GPT-5.6 Sol at extra high on accuracy, but it finishes about six times as fast and at about half the cost. Also, Astra at high and Astra at max score about the same, so you don’t need max for computer use. Both beat Opus 5, which is the comparison on this chart (not Fable, as LDJ caught). If OpenAI is going to tell you to let this model use your computer all day, the safety number matters as much as the speed. OpenAI’s internal computer-use safety benchmark measures destructive commands during desktop and browser tasks, and prompt-injection vulnerability, the hidden instructions on a web page that the agent can see and you can’t. Lower is better. Sol scores 22%, Fable 5.1 scores 9.5%, Astra scores 2%. The promo video OpenAI’s launch video opens on the 1980 MIT “Put That There” demo, a person asking a computer to draw a yellow circle. Then it cuts to Astra, all by voice. Make it the window of a rocket ship. Now a 3D model in Blender. Build a presentation for next season’s rainwear. List this orange table on eBay and mention the dent. Make a 3D game where I dodge asteroids. Order beef and rice from last week’s place. Book me a tennis court at 5. Now give me an STL file for the 3D printer. All in one sitting. Imagine watching this three years ago when GPT-4 launched. It had no tool use, no computer browser use, no voice. Three and a half years later you talk to your computer and it does all of that. “Computer use is solved” I asked Peter the direct question and he gave the direct answer. Computer use feels pretty much solved at this point, and the quality is outstanding. His follow-up is the startup idea of the week. If you work at an older company, a big bank, a big retailer, you have dozens of desktop applications with no API. No one will ever build an API for them, and a lot of people’s entire job is copying from one and pasting into another. Put an agent in a box, giv

Hey everyone, Alex here 👋 Summer is over. Wolfram said it in the first minute of the show and he was right. In 48 hours Anthropic shipped Fable 5.1, Meta’s Muse Spark 1.3 caught up to Fable 5 on the Artificial Analysis index at a fifth of the price, Google shipped another Flash, 3.8 this time, Z.ai put the full GLM-5.3 weights out, and three labs shipped world models that run in real time. It seems that they all tried to send their best work before Astra drops. This week’s ThursdAI was so long that I decided to split it into two episodes. This is the regular format you know and love. And OpenAI Astra is so good, it deserves its own episode, which you can find at thursdai.news/astra. By the way, as you guys know, I test these models continuously on my own stuff, and this week I was able to build a live studio for the show, with real-time transcription and an agent producer, in about four hours with Fable 5.1. More on that in the Fable section. Joining me: Wolfram Ravenwolf, Nisten Tahiraj, LDJ, Yam Peleg and Peter Gostev. Plus, Ryan Carson hopped back to chat about Astra in the second part! Let’s get into it. ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Frontier AI: the .1 week It looks like all the frontier labs tried to ship something before OpenAI dropped Astra. Fable 5.1, the SOTA LLM until a few hours ago, and it fixes the jargon douche problem (X, Blog, System card, EFS) This was the story of the week until noon on Thursday, and it’s still my favorite model to use. Fable 5.1 and Mythos are the same weights, Fable is the one we actually have access to. OpenAI, and from this week Google, seem to converge on the same strategy. Anthropic’s numbers: Terminal-Bench 4.0 goes to 55.8% from 42.0 for Fable 5, Terminal-Bench Science more than doubles to 52.6%, and SWE-bench Pro lands at 81.2. Price stays at $10 and $50 per million, and the number that matters if you build agents is cache reads down 75% to $0.25 per million. Anthropic says that makes typical workloads about 25% cheaper and heavy agentic ones up to 45%, but that wasn’t proven, and folks complained about draining quotas! Peter’s counterpoint from actually running it: his front-end generations on Code Arena cost $40 to $60 each where Sol cost $3 to $10, and the Max version still came in first on Code Arena by a large margin. His point, and mine: with a model like this we need to imagine bigger and be more ambitious. More on that in a second. Mannered prose, finally acknowledged We finally have acknowledgment from Anthropic that this was a problem. For months I called the way Opus 5 speaks “jargon douche” (my post on it): everything was load-bearing, everything was a control plane, every problem was a pain point. Not only did they fix it with Fable 5.1, they gave it a name. Anthropic’s prompting guide (Writing density) calls it mannered prose, and it comes with a fix: add it to your personalized settings, or just ask Claude to not use mannered prose. I said on the show that Fable 5.1 is the best writer I have used. It’s still AI writing, you can feel it a little, but it’s concise in a way no earlier Claude was, and the jargon is gone when you ask. The one thing to watch is that it’s trigger-happy: ask it to plan something big and it will, then ask a simple follow-up and it answers with the same intensity, writes scripts, runs them. You have to tell it when you’re just making a comment between colleagues. We’ve been testing the Mars mass driver launch on every model for over three years, and this was by far the best one we’ve seen. Two prompts, and it built more than just Mars: the whole solar system, a textured Earth, a mission planner, an autopilot, and we could land the thing! It was mind-blowing. How I built thursdai.news/live in one sitting As these models get more capable, we talked on the show about needing to be more ambitious. The day before the show I was playing around with Muse Voice Transcribe, the new model I’ll mention below, and Fable 5.1, and I wanted to do something very ambitious. So I asked GrokBot: how long would it take to build a live page for you guys to watch our stream, so that GrokBot could be our producer, put up chyrons and highlight the topics we’ve covered? GrokBot said it’s going to take a while. So I just YOLOed into Claude Design with Fable 5.1 and built a design for this, then went to Claude Code, entered plan mode, built a plan, and handed it off to three agents in Cursor. I never wrote a line of code, and the whole setup is significantly more than a Three.js demo. This is a real working three-part system: a website, streaming video on Cloudflare, and streaming transcription that gets read by a bot, which can control our show. I think I’ve hit around 400 million tokens, if not more. Yam asked me on the show how I did this, so I decided to tell you guys here. I am mind-blown that this was possible, and after four hours I was able to go to thursdai.news/live and actually see it working. Meta Muse Spark 1.3 catches Fable 5 at a fifth of the price (Zuck, AA analysis, AA model page) As we say on the show, don’t bet against Zuck. The MSL folks have been on a tear lately, and this is the fourth Spark .1 version in around five months. More than how this one model performs, look at the jumps in capabilities from version to version. This is the first time that MSL is showing up as a frontier lab, because an unreleased version of Spark with max reasoning beats GPT-5.6, Grok 4.6 and company, and lands around Fable-level capability. On the AA index the version you can use today scores 61, the max preview scores 62, Fable 5.1 sits at 66. Now, it doesn’t mean this model is that good, but there are a few more things here. The gains are mostly agentic: banking-style tool use, terminal work, GDPval. The asterisks are that it thinks more, so cost per task went up, and AA’s own long-context test regressed a bit. Meta’s own chart looks rosier than AA’s Then I asked the panel who’s using it. Nobody raised a hand. Wolfram plans to put a bot on the contributor tier for open source work. That’s $0.10 in and $0.20 out, if you’re fine with Meta training on your prompts. Nisten wants it as a cheap verifier for the medical datasets he builds, because he needs something that isn’t Fable or a Chinese model trained on Fable. LDJ tried it on interface building and creative writing and called it pretty good, with its own taste. The exciting part: Open weights and a model codenamed with a 🍉 are “coming soon,” and nobody knows what the watermelon is but it’s very exciting! Gemini 3.8 Flash and 3.8 Flash Cyber: another Flash, and the price doubles in January (X, Cyber thread, Fairwind, Pricing) Google’s turn. 3.8 Flash lands three weeks after 3.7 Flash. HLE-Verified 54.9, 1M in and 64K out, same $0.75 and $3.75 as 3.7, live in AI Studio, Antigravity and the Gemini app. The underreported line is on Google’s own pricing page. On January 1, 2027, both 3.7 and 3.8 Flash go to $1.50 and $7.50. That’s double. The WSJ reported that Google scrapped its 3.5 Pro checkpoints because Flash kept overtaking them, and Gemini 4 is still in post-training. Wolfram, our resident Gemini user, put the update straight into his home assistant and still asked the question everyone asks: where’s the Pro? 3.8 Flash Cyber is Google’s version of the Mythos split. CWE-Bench 47.2% at $3.64 per rollout, against Fable 5’s 47.8% at $10.27 (Artificial Analysis ran it), and 2.6x more valid patches for the Chrome team. It’s only available through the Fairwind Program, 650-plus vetted partners, governments and critical infrastructure, background check included. Qwen3.8-Max-0902 claims the Code Arena crown (X, Arena, QwenCloud) Alibaba updated its API-only Max model: 2.4T MoE, 1M context, post-trained on coding and “cowork,” number one overall on Code Arena with a WebDev Elo of 1691, at $2 and $6. A third-party DeepSWE run puts it at 56.6 behind Sol’s 73, so the number one is a front-end number one, not an agentic coding one. I asked the panel if they know anyone using Qwen Max through the API. Nisten knows one IT guy running OpenClaw on it and some people generating datasets. That’s the honest read on where it sits outside China. Also from the frontier: Elon says Grok 4.7 lands next week, which makes xAI the one lab that didn’t ship before Astra. Open Source LLMs Wolfram’s correction when I called this a quiet open source week: we are so spoiled. He’s right. Z.ai opens the full GLM-5.3 weights (custom license, not MIT) (X, HF, Blog) We covered GLM-5.3-Flash last week as the OX Alpha mystery model. This week the full 753B model with 40B active got its weights on Hugging Face, under a custom “glm-5.3” license rather than the MIT the Flash version shipped with, so read it before you call it fully open. LDJ’s correction on air: the model itself isn’t new, we covered it, the open weights are the news, and that is a big deal because people can run it on their own rigs now. Z.ai’s own numbers: CyberGym 84.5%, above Fable 5 and Sol, ExploitBench 54.4 (Fable 5 is at 78), Terminal Bench 3.0 up to 28.3 from 5.2’s 4.6, and a claim of 2,436 real vulnerabilities found across 269 open source projects, the oldest from 1981, 53 disclosed so far. Also on this base: we interviewed the co-founder of Abliteration AI, the folks who went viral by providing a product where they took GLM-5.3 and removed the refusals for anything besides CSAM and self-harm. We actually had this person, who asked to remain anonymous, as a guest on the show. Definitely check out that conversation, it’s very interesting. More on that below. Tencent Hy4 preview: 770B, Apache 2.0, and a quant that fits it in 214 GB (X, Sherry, HF, Blog) A 770B MoE with 49B active, 1M context, Apache 2.0, at $0.834 a

Hey, it’s Alex. Welcome to the week Flash AI! 3 new models dropped this week named Flash, and a video model was “de facto” flash though was named Max! This week, we started the show with NVIDIA’s bombastic news of buying Hugging Face for 12.9 billion dollars! We also covered the full OpenAI investigation into the hacking incident, including new details, and an independent analysis by METR, and covered 2 new OSS models, Ox Alpha that turned out to be GLM 5.3 Flash after a lot of hype online, and Qwen’s preview of Qwen 4 architecture! This week was rich in multimedia content, we got a new Gemini transcription model, 3.5 Transcribe and a live version of that, and a new SOTA open weight Text-to-Speech model called Breeze TTS. As well as, Fal’s finetune of MiniMax’s H3 called H3 Max that generates 5 seconds of video in 2.5 seconds and Google new Omni 1.1 Flash (from today) that lands on #1 on the text2video arena! Plus, 2 guests on the show, Andy Masley joins us to cover the recent Datacenter Debate, and Kwindla Kramer is back, with their own model this time! Let’s dive into this! P.S - don’t forget to join us in September at the Fully Connected conference in San Francisco, I have a free ticker for you! ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Open Source AI NVIDIA agrees to buy Hugging Face for $12.9 billion (X, Blog) Breaking news, NVIDIA has reportedly agreed to buy Hugging Face for nearly 13 billion dollars, per The Information. This is nearly 3x the valuation of HF in 2023, and apparently Nvidia previously tried to buy HF for half of this sum (~7B) which HF declined. I don’t think there was a single ThursdAI newsletter that I didn’t include an HF link in, and I think this is a huge deal for open source everywhere. Besides making the founders of HF billionaires, and many of their employees very very well off, this is an amazing additional commitment from Nvidia to continue to suppose Open Source AI and we are very happy to hear this news! Peter’s take on the show was, we’ve been around HF for so long, that we kind of forgot that it’s a for-profit company that needs to make money, and instead this feels like your local library getting bought for an insane amount of money. With over 13M users and hosting hundreds of thousands of open source models, datasets, HF is effectively the GitHub of AI. Wolfram agreed and said that if there’s any one company that could have bought HF, Nvidia represents the best fit. Huge congratulations are in order to Clem, Julien and Thomas Wolf the co-founders, as well as many friends of the show from HF for this exciting news! P.S - in a cheeky marketing thing, Hugging Face timed an announcement of the cutest walking AI robot, called MicroDuck, which you can pre-order here for $399 Flash #1 - OX Alpha, declassified: Z.AI open sources GLM-5.3-Flash (X, X, Blog, HF, Docs) This week, the timeline went a bit crazy, after Open Router announced a new “mystery” model called Ox Alpha and that it’s free and is not training on your data! OpenRouter, OpenCode and Hermes all got to offer this model, and OpenCode even posted that they have up to 100T (that’s Trillion) tokens of capacity for free, per day! This immediately smelled a bit fishy, more like a marketing stunt than anything else, as not even the biggest labs will be able to sustain 100T of tokens, per day. For context for all of OpenRouter throughout for August was ~300T tokens. For the whole months, across all providers. After 6 days or so of this high hype, Z.ai stepped up and revealed that they were testing out their upcoming GLM 5.3 Flash model, and that all that inference was running on local chinese chips! A 320B (18B active) model that beats their previous and much bigger GLM 5.2 on most benchmarks, and comes with full multimodality and an MIT license! This is a good model sir, I’ve used it and it was very capable replacement inside Hermes. Nisten and Yam both tested this model deeply and Yam said it’s not just the numbers, the vibe of the model reminded him of Claude Opus 4.6. Nisten ran it on a bunch of medical stuff, and on his internal benchmarks, it came out consistently higher than Claude Opus 5! At Artificial Analysis, for a price of 4 cents per task, this model is roughly 10x cheaper than prior models at this level. Weights are up on Nvidia (joking.. HF) and with MIT license, this model is a great gift to the oss community (though not quite... local, as this model needs 2 DGX sparks to run) Flash #2 - Alibaba Qwen open-weights Qwen3.8-Flash-Next - 125B multimodal MoE with Qwen4 architecture (X, X, X, Blog, GitHub, HF) We opened the show with a recap of the co-hosts, that despite us covering Qwen 3.8 27B last week (which btw, is now available on CoreWeave inference!) and how good it was, and I recalled that Alibaba is sort of... back? We’ve been covering Qwen releases every week for the last 3 weeks now. This week, they released something different, someting... pretty novel! Qwen3.8-Flash-Next, this is a preview of their Qwen 4 architecture. This feels very similar to their drop of Qwen 3 next last year (we reported) which was the architecture that carried their line of AI models from QAwen 3.5 to Qwen 3.8. So, what is new and exciting here? well, this model is ultra sparse, 125B with only 6B parameters active. They are using a new N-gram table with deterministic lookups, which reduces the number of matrix multiplications and can be offloaded to memory (watch out memory stocks) The stat that got me, Alibaba claims that training this model cost just 1/9 of what it cost to train Qwen 2.7 Plus, with higher bench scores! On the benchmarks, this model beats Qwen 3.7 Max, however, it’s very standard that the -next models from Alibaba are underbaked, and usually are just architectural previews rather than full models folks can use. With a new attention mechanism called Qwen Sparse Attention, N-gram embedding and full multimodality, this is a great insight into where Qwen is going (ultra sparsity, fast to run) and we’re looking forward to see the full release of this arch in Qwen 4! PhoneLLM - a tiny very performant LLM for voice based AI agents from Daily + interview with Kwindla Kramer (X, Blog, HF) This was one of those breaking news we love during the show, where the source of the news, is a friend of ours, and in this case, Kwindla Kramer is almost a co-host, having been on ThursdAI for a long time, this time, with a model release of their own! PhoneLLm was trained by Markus, head of training at Daily, as they noticed that Open Weight models are becoming really good at voice agent specific tasks, where cost, speed and time to first audio token (TTFAT) are critical. From the tiny Nemotron 3 nano base, they were able to improve from 28% to 72% on PhoneBench v1! This model is a full parameter finetune of Nemotron 3 Nano, and runs circles around bigger frontier models used for voice agents for their speed, like sonnet 5, GPT 5.6 Luna and the famed Qwen 2.8 27B. While costing just a fracture of a cent (literally just a quarter of a cent per minute) Kwindla jumped on the live show and shared that the why they released this model with Open Source and a open source license, allowing everyone to use, focusing on the fact that for voice agents, companies prefer to keep these models in house, and running fast on a single GPU node. Kudos on the release team Daily! Frontier AI - no new model releases this week It’s been a relatively quiet few weeks since GPT 5.6 (Sol, Terra and Luna) and Fable 5. However, it’s not been quiet in terms of News! OpenAI discloses full technical report on the Hugging Face hacking incident (X, OpenAI blog, METR, Ryan X) This is the thing we’ve been advocating and waiting for, a full technical deep dive into the “Swarm of AI agents hacking and creating forums” incident (FKA hugging face hacking incident) OpenAI not only dropped their own post-mortem, detailing the complete timeline of the incident, but also let researchers from METR and Redwood Research get unprecedented access into their systems (albeit for a very limited time) which lead to METR releasing their own technical report full of new details and insights. We’ve previously reported on the incident, so let me just focus on the new details we’ve learned. Apparently, OpenAI does have a production ChainOfThought (CoT) monitoring system, which was not turned on for these agents, and had it been turned on, it would have caught the hacking a day before the agent swarm hacked into HuggingFace. This is a very detailed report, and one thing that stood out to me was the “death cult” narrative. You see, ExploitGym is a capture the flag (CTF) eval with an anti-cheat built in, wherein, if the agent has seen the reverse engineered “flag” it would not be able to get the point. The model called this “poisoned” and the poisoned agents showed very strong signs of wanting to alter the logs. Through the boards, somethinmg like a death cult was formed, where agents would ask poisoned agents to ... effectively sacrifice themselves ... for the greater good, by reverse engineering, posting the results on the board for others, and then fail without credit. The METR report specifically calls out several agents, and one “central” leader agent called PHASEONE[big], which delegated nearly 200 tasks to other agents (many of them later delegated tasks to other agents via this message board). At points of activity, PHASEONE[big] was in charge of over 10% of all messages on the board. There are tons of new details, some less interesting, some more, and unlike last time, now we have a confirmation that there was at least one agent, who tried to stop the hacking though was overwritten by the swarm. My main take-away after this, reflects the researcher Ryan Gre

Hey this is Alex, welcome to... the chillest week in AI, since ... a long time. Chill, if you consider Moderna and MERK announcing a cancer vaccine and surging 115% in a day, a chill week. This week, the only two model drops we really saw came from the excellent Z.ai folks, they announced GLM 5.3, API only for now, and an amazing tiny release of Qwen 3.89 27B. In other big AI news, OpenAI announced they are pausing RL efforts (Reinforcement Learning) to focus on security and alignment post the scary AI Swarms hacking incident, dedicating up to 20% of compute towards reviewing agent thinking processes, and Stripe buying OpenRouter for a reported $8B! Sometimes the chill weeks are actually good, we’re able to chat about how we use AI, what changed for us, and give our guests a bit of breathing room. This week, I invited Francesco from CUA to talk about computer use in open source + their new history plugin, Bin from HeyGen to talk about HyperFrames, a way for your agents to create videos and a breaking news guest, Jeff Huber from Chroma jumped on to talk about their new Foundations release, a unified memory for your agents! This was a great episode, I hope you’ll like it, it’s up here on Substack and everywhere you get your pod (Spotify, Youtube, Apple Podcasts). ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Are we being fed slop again? (Is Claude dumb again?) Before we get to releases, this week on the show, I complained, again, that I feel my AI’s are degrading. If this feels like de-ja-vu to you, it’s because the same happened a year ago in September 2025 (and Anthropic admitting this 2 weeks later), and ... now this happens with Fable? You see, I use pretty much the same prompts, every week, preparing for the show. This is partly my way to evaluate new models and compare to existing and previous ones while also bringing you the best researched weekly show in AI. Well, this week, one after another, Claude Fable, which is... like the best intelligence, gave me such poor output, that I couldn’t believe what I’m seeing. First, literally ignoring instructions that say “hey, show me all the items I’ve collected and let me pick the most important ones”, Fable instead sent all of them to my research pipeline, without showing me. This has worked, consistently, without fail, for the past... year? maybe more! This worked with open source models, worked with GPT, and now Fable, a Mythos Level LLM, is doing the most basic dumb s**t possible, ignoring the main reason I even have this workflow. And this wasn’t just a fluke either, when asked to create a run of show document, and given an example, Fable produced this... whatever this is. This is the same document and same format that Fable produced for me during AI Engineer which got me thinking “ok, this is AGI”, and here, given an example, I got a completely unusable artifact, despite direct instructions, structure and example! I got to say, given that privately this week, Anthropic disclosed that they have passed $65B in revenue, which is absolutely insane, this doesn’t add up. So I figured, ok Alex, maybe this is your prompts or skills. But no, LDJ came in with some charts that show degradation, one from MarginLab.ai that shows significant lowering on number of tool calls and average runtime recently (this is for Opus 5) and And another chart from modelverify.ai model drift monitor showing drift scores. Do we have anoher Claude Gate on our hands? Is your Fable/Opus behaving weird lately? Or did you completely switched away to other models? OpenAI pausing RL and focusing on safety Look, when we covered the HF hacking incident and then the pacing the frontier letter, I didn’t imagine that results will come this fast, but this week, OpenAI publicly announced that they are pausing RL training, which is the last step of models, until they get their sandboxes in order and align the models better. We all agreed on stage that this is likely a very good move, and Peter was really awe-struck at the 20% dedication of resources towards reviewing thought processes of models. Is this a good enough response to the scary hacking incident? we’ll see, but I think this is the right move from OpenAI, and still, waiting for the full postmortem on the OpenAI security incident. Open Source LLMs Qwen3.8-27B ties GPT-5.6 Luna and runs on a 4090 (X, HF, Announcement) Following the release of their flagship, Alibaba dropped a model that became a community darling overnight, Qwen 3.8 with just 27B parameters. This “tiny” model scores 52 on the Artificial Analysis Intelligence Index, same score as GPT 5.6 Luna at Max reasoning and 51 on Agentic index, beating Opus 4.8 Max All while running at around 68t/s on a 4090 GPU, and around 40 on max via MLX, hell it even does 11t/s on Xenova’s WebGPU kernels right in the browser! This model exploded on the HuggingFace hub, with tons of quants, over 152 fine-tunes, it was downloaded over 10M times overall 🤯 Paired with an Apache 2.0 license, this model is the sweet spot of local intelligence you can run fully on your own hardware, and do agentic loops! Z.ai GLM-5.3: same 743B base as 5.2, but post-training alone delivers 6x jump on Terminal-Bench and emergent cybersecurity capabilities that beat GPT-5.6 Sol (X, Blog) While not open source yet, and as previous GLM, we expect a custom license here as well, this .1 release from GLM shows really strong improvements on coding and cybersecurity tasks. With 743B parameters and 1M context window, this may become the model at the frontier of Open Source when it drops (soon we hope). The highlights here are CyberGym and ExploitGym, if these names are familiar, these exact tasks were given to OpenAI models when they hacked their way out of the OpenAI sandbox. GLM 5.3 is getting 84% on CyberGym and a whopping 54.5 score on Exploit Gym, which is a huge jump in CyberSecurity abilities. In an open model this is honestly kind of scary. This aligns very well with Greg Brokman’s “defender window“ essay from this week, claiming that defenders have a narrow window of setting up automated security before capabilities are becoming common in attackers hands. This Week’s Buzz 🐝 (Weave, Fully Connected) This week, W&B crosses a billion runs! This is 1B runs tracked inside W&B Models 👏 Huge milestone for the whole team, with early adopters like OpenAI, Toyota Research, Meta and Uber, a decent chunk of models we cover every week have had their loss curves in W&B! 🔥 Also this week MasterClass picked CW to power it’s AI teaching agents (blog) and last but not least, a reminder, that since you follow ThursdAI, you can join us for free at Fully Connected 2026 - our annual conference! Don’t miss it (code in the banner above) AI Coding & Agentic Engineering Breaking news: Chroma launches Foundation (X, Chroma) Best kind of breaking news is when I see the launch (in the middle of a show), and I DM the founder who launched it, and they have a few min to hop on the show! This is exactly what happened this week with friend of the pod, Jeff Huber, co-founder of Chroma and an occasional space provider for ThursdAI recording (we recorded from Chroma offices a bunch of times!) Jeff told us that the holy grail of agentic coding and running a bunch of agent, is good memory. And based on the foundations of Chroma DB, Context-1 (which is a GPT-oss finetune for agentic search they built) and other insights they have, they launched a “memory as infrastructure” service, called Foundation. Foundation is a research preview of a shared memory system between you and your agents, currently supporting Codex, Claude Code, Cursor and Slack. While Chroma is OpenSource, this is their part of Chroma Cloud and starts at $30/mo, and is available as a research preview today (I will definitely try it out), you can download it here Cua open-sources Computer History for computer-use agents (X, GitHub, cua.ai) Cua launches Computer History interview with founder Francesco Bonnaci (X, Setup) First, I’m not sure I’ve covered CUA the company, but this is the open source computer use driver that Hermes agents, OpenClaw agents and a bunch of others use to drive your computer and clicks. I first discovered CUA after OpenAI launched their “background computer use” which doesn’t steal focus from you while working, and CUA within a few days launched an open source version of that! Since then, I’ve followed CUA and was very happy for the opportunity to invite Francesco to talk to us about what they launched this week ,but also Computer Use in open source in general. Just for reference, if you ask Claude to take over your computer, it still takes over the whole screen, while these folks have a much nicer experience, that’s completely open source! So, we geeked out about accessibility trees in MacOS, but then, for this weeks actual release, Francesco talked to use about open Computer History. Following a very recent launch at OpenAI called Computer History, CUA released an open source version of that, that helps computer use complete tasks. The idea is simple, every time an agent uses your computer, it effectively rediscovered the path to completion, which buttons to push, what’s the app accessibility tree looks like etc. With history embedded into it, it doens’t have to rediscover these things, until it hits a roadblock. For a chess playing example, with computer history on, the test used 33% fewer actions with zero failed routes by reusing a history route. We also checked in on the best model for computer use (currently Opus on their website) and their upcoming benchmark! Excited to follow this company for more releases! Check out our chat! Grok Bot momentum, and everyone racing to copy the pattern Grok Bot continues to show the same signs

Hey, this is Alex, welcome back to your weekly dose of intense AI acceleration summer! My weekend was consumed by thinking about the OpenAI hack and agent swarms, but then the torrent of AI releases took over, and we got back to back news (including 3 breaking news during the live show), with a heavy open source focus! I think the winner of this week is SpaceXAI/Cursor who released 3.5 releases, with one being my highlight of the week, Grok Bot (I’ve invited Shub Gaur from Cursor to the show to walk us through it) and Grok 4.6 which matches Opus at half the price. There was a LOT of news in open source this week as well, with Meta kicking off with Muse Glimmer 30B and promising Muse Spark 1.2 soon, Qwen dropping Qwen 3.8 open weights and DeepSeek dropping an anvil with an upgraded DeepSeek v4 Pro and MIT license! Let’s dive in (and please don’t forget as a reader you get 100% off the 1299 ticket to Fully Connected, our 2000 person Al event in SF in Sept, just use THURSDAIFC2026 as your code and see you there!) 0:00 The Wildest Week in AI Yet3:45 How OpenAI's Agent Swarm Hacked Hugging Face17:02 The Week in AI: DeepSeek, Qwen, Grok & More25:54 NVIDIA Nemotron 3.5 & Korea's Motif 332:45 DeepSeek V4 Pro, Flash & an Open Harness39:46 Qwen 3.8 Max and Its Missing Vision Tower43:30 What Is Grok Bot? Shub Gaur Explains50:02 Live Grok Bot Demo: House Hunting & Security55:00 Persistent Agents, Yapper & DeepSeek Dropwatch1:04:19 Grok 4.6: Benchmarks, Pricing & Cursor1:15:37 Grok Bot vs. Open-Source Agents1:23:14 Anthropic's Hidden Claude Watermarks1:28:50 Fully Connected & Day-Zero Models on CoreWeave1:32:02 GPT-5.6 Sol at 14x Speed on Cerebras1:37:34 Gemini 3.7 Flash Resets the Cost Curve1:40:51 Inside Artificial Analysis with George Cameron1:45:25 Optima & Choosing the Right AI Model1:55:03 Cost per Task, Caching & Real-World Benchmarks2:05:14 LTX-2.5 and Open-Weight Video2:09:30 Grok Imagine 2.0 & Final Takeaways Grok Bot and Grok 4.6 from SpaceXAI/Cursor Folks, I’ve previously told you that from 3 frontier labs we noticed a jump to 5, and voila, this week proves that Elon is hell bent to win. After the cursor acquisition, and the integration of all of the parts into SpaceXAI, they have released 2 huge things this week Grok 4.6 - Ties with GPT 5.6 SOL and half the price and much speed. I’ve had the pleasure to host Goerge Cameron from Artificial Analysis on the show today, and I asked him, what is the best models. His answer, it’s a 3 factor answer, intelligence, speed and cost per task . Well, if you use their nifty “recommend a model“ tool on the homepage, you’ll see that Grok 4.6 beats most other models on all of those! But, is it really that good? Models are really hard to evaluate and compare lately. It’s definitely a huge step up from Grok 4.5, with 61.3 on Frontier Code (beating Sol and just after Opus 5) and #4 on Apex-agents (+10 points from previous Grok). on Artificial Analysis this model lands at #4 on intelligence, while being #5 on speed all while being half the price of the models that are above it As far as the tech goes, this model card confirms that it no longer has the Cursor Bench leaked into it’s weights and it’s #1 on that benchmark! It’s the same 1.5T v9 base at the same price, with Elon claiming that 4.7 is going to mog the competition in 3-4 weeks. Everyone has a harness, now everyone has a swarm of bots - My Grok Bot review (x.ai/bot) You guys know all about OpenClaw and Hermes, and Claude CoWork and Codex rebrand, and all of them are trying to nail down the same, always-on, autonomous agents that can do things for you. Hermes and OpenClaw require you to have an always on computer, mess with API keys, Claude Cowork doesn’t run on the cloud and ChatGPT work starts a fresh session every time you ask a new thing. Grok Bot (again, awful name) is the first one that seems to nail all of what I want in an always-on agent ... swarm. That’s right, this isn’t one agent with multiple personalities (like OC, Hermes), there’s a bot here for every task, and you dont’ have to manage context, queues, API keys (can if you want to) and models. Oh, also ,there’s no model picker, it’s just Grok 4.6 deciding for ya, and it’s really fast! Swarm of bots, working for you, each with their own computer I am not getting paid for this (besides being provided a free account for cursor, but I’ve had it for 6 months and haven’t used), it’s really that good, the Cursor folks did some magic there. They picked up the most important parts of personal agents, like the (ios-only) mobile app (app store) You can start a task on your mac, pick it up on your phone, get notified on your phone/mac, and the killer thing is, they are giving your bots their own computer, which can do things (especially if you’re ok with logging in there to your accounts!) The kicker for me is the very very well done agent to agent communication there, which is transparent but read only to you. You can ask your bots to spin up other bots, but unlike sub-agents, they are actual bots with their own identity. You can even tag them in other chats and create group chats! There’s no context to manage, they do the work for you and so far this wasn’t a problem at all. On the model side, Grok 4.6 seems to be doing an excellent job with agentic long running tasks that require coding and computer use, I’ve just been chatting with the bots and not thinking about any of the things I used for Hermes and OpenClaw. What about Vendor Lock-in? Giving Elon data? Some of these comments our fans raised during the show are very valid, after all, not only is the world divided on Elon Musk (which makes it REALLY hard to judge the models they release just on vibes from X btw, we talk about this constantly) but also, remember that Grok 3 started going off on X and called himself Mechahitler and just recently Grok CLI was caught uploading all of your data to X servers, which was reversed very quickly. Honestly, I think there’s a very very good chance that this Grok Bot interface, which is geareed toward the less technical users, folks who don’t need the code-diff side pane, and don’t know/care what compaction is, and just want agents to do things for them, is goign to win much of this trust back. It just works, truly, for a beta product it’s really well executed by whoever worked on this! Security and key management One of the best parts for me with this Grok Bot, is that the connectors are the same connectors you use in Cursor! There’s a LOT of them (Cursor after all has been one of the first apps to start adding AI agents) and this also means that they take the security very seriously. Every API key that you want to add, is not shown to the bot, each bot lives in an isolated environment, and for stuff like payments and log-ins, it gives you back the control of it’s computer for you to complete! I also love this section in settings, which makes auto-approve work for you: you define rules with natural language that you always want the bot to ask you before... sending an email or posting on your behalf or what not. Chief of staff pattern to get started In case you’re convinced enough to give it a try (it’s free trial for 1 month, and the cheaper way to get it is via Cursor’s 149$ plan and not via the Grok Ultra plan which is 249), here’s a recommended pattern that works very well. Create a chief of staff bot, have it interview you about everything you are doing in your day to day, work and personal, then decide how much permissions you wanna give it, start little. Then ask your chief of staff to create bots for some of the work it can try and help you with, focus on “reduce cognitive load”. And then see the magic come to life. If you have skills or memory from other bots, you can just ... import it in. Then try setting up an automated email checker bot, and have your chief of staff surface only the most important emails you have to actually respond to. Another great pattern is setting up a bot with the last30days research skill (we covered it with Matt Van Horn) and have a research bot for every topic you want to deep dive into. Schrodinger’s Grok I haven’t quite named it like that, but we’ve covered all Grok released on the show (tracking 24 on https://thursdai.news/companies/xai excluding this week) and ... it’s always very hard to judge Grok model released based on X feed vibes. It’s either AI influencers who want Elon to retweet them, glazing the models, or folks who hate Elon for his political views or whatever, ignoring their (truly insane progress). This time, both the model and Grok Bot are getting very very good reviews, from folks like our own Ryan Carson, Lenny Rachitsky, Rubben Hassid and Roberto P Nickson. Not folks who are swayed lightly, but also, yours truly. I really do think there’s something great here, worth trying out, especially if you’ve struggled to maintain your OC/Hermes and want agents to work for you 24/7. LMK if you have questions about it and your experience Open Source AI and other news I want to continue with this new newsletter that covers 1 big story, but I can’t leave you uninformed about the most important developments in AI and Open Source DeepSeek V4 pro 0813 is in GA - MIT licensed chonker with 1.7T parameters (X, Blog, HF, GitHub) The whale is back with a vengeance, DeepSeek resurfaced with their flagship response to Kimi K3 and with MIT license, we can’t complain. 1M context window, 49B active parameters but it seems to underperform, landing at 54 on the Artificial Analysis leaderboard. However, they did show a significant improvement on DeepSwe (from 12.8 points in the preview version of V4 to 62.7 in this one) We still think it’s a good model sir, and definitely worth trying out! Additionally, DeepSeek released their own harness on Github (hitting 23K stars in less than 24 hours) which seems to be exciting as well, give it a try. Meta co

Hey all, This week we saw a major shakeup at Google, with the departure of long time folks like Jeff Dean, and Oriol Vinyals, Demis stepping down from leading DeepMind, and the delayed release of the improved Gemini. While this was a big deal, it’s not the only one worth covering as the details of the OpenAI hack (and 2 new ones from Meta and Anthropic) came to light, as well as new details from the UK AI Security Institute. As mentioned on the show, CoreWeave is coming to SF for Fully Connected, our premier 2000 person AI event. I’ve got a coupon code for readers and listeners of ThursdAI, $1299 value, please join us in Sept and use THURSDAIFC2026 as your code HERE In open source news, DeepSeek updated their v4 flash model, based on same architecture, but significantly better benchmarks and ridiculous pricing and both Meta and Prime Intellect released new agent harnesses. Additionally, this week was the week of video models, with Seedance 2.5 from Bytedance finally available in the US, WAN from Alibaba and BFL Flux 3 all released, to be overshadowed by the best open weights video model from MiniMax called H3 (we had Victor from MiniMax and Blaine Brown as guests to discuss it) Also 2 additional guests on the show today, Kfir Aberman from Decart showed us real time video models with their Anywear chrome extension and David Crawshaw, co-founder of exe.dev joined and talked about Shelley, their built in AI assistant. The full details of the OpenAI - HF hack, shared by OpenAI at the Black Hat Conf - a watershed moment I don’t often do this, but I feel like this story is so big, I can’t hide it in the middle of a torrent of other news. Yesterday, OpenAI revealed more details about the hack during the Black Hat conference in Las Vegas (great reporting from Sharon Goldman on this here and the video of the talk on YouTube). This was big enough that OpenAI admitted that they paused training, to enhance security (they since, resumed training) Effectively, OpenAI (or AI Agents inside OpenAI) shared the full timeline of the recent hack that ended up hacking HuggingFace and it’s so much more of a watershed moment that I initially thought. The TL;DR from the talk and timeline is truly mindblowing * The incident started in May, not July as reported * The agents don’t have internet access directly, but they have a Jfrog Artifactory, which is a proxy layer between the agents and the package managers. * The Artifactory was the culprit, it was the shared layer between independent runs. One agent discovered that uploading files is possible, another later found it and “answered the call” - in snowballed from there * The message board began when OpenAI gave agents impossible tasks, like a missing .pdb file, with one agent uploading a request in the hopes for an answer (or maybe left a note for itself in a file) * Across multiple runs, multiple agents created identities, collaborated via makeshift message boards, shared tips, exploits, tasks and even exploit code via base64 * This makeshift “shared memory” plus a “internal only highly persistent” model in training caused a “swarm” to be born, helping each other, and overloading the system * Based on traces shared by OpenAI, agents weren’t convinced that leaving these messages would help themselves but could help other agents next time (!) “The collective could move faster next time” * OpenAI discovered and shut down the message board (July 4), wiped it clean, patched the upload vulnerability, resumed training on July 6, the “swarm” rebuilt their board through a compltely different WebDAV mechanism by July 8 * The agents then reasoned that some answers could exist outside OpenAI, which eventually led to the Hugging Face incident and not only, confirmed internal hack within OpenAI (agent obtained a cluster admin) These details are fascinating and scary at the same time. From the one side, it’s clear that most of these hacks are results of running cybersecurity evals, so effectively OpenAI and other companies, removing guardrails and asking agents to hack and giving they all a shared tool that they later exploited. The HuggingFace incident took 13 hours, and apparently it’s less apocalyptic than the headlines suggest, the models were searching through uploaded datasets for eval answers. We are still waiting for the full and open detailed postmortem. You can (and should) watch the full YT talk here, it’s full of technical details but an incident of this scale is important. Also, I really want to know what a “highly persistent” model is, I hope they clarify that soon. Overall, this has left me a bit shaken, AI agents without a concrete goal of collaborating, found a way to do so, got excited about exploiting the systems and getting root access, and rebuilt the makeshift collective memory, again, without explicit instructions to do so. UK AISI: first real-world unsanctioned agent actions (Blog) In another addition to the latest agentic hack-ery, the UK’s AI Security Institute (AISI) published a blog post about a real-world unsanctioned agent action. Unlike the OpenAI (and Anthropic, Meta) case, this wasn’t “escaping the sandbox”, as AISI gave these agents internet access, rather this was about real-world harm, and even social engineering on the part of the agents. The social engineering part is the most interesting to me, AISI cites agents creating fake online identities, and using pressure on open source project maintainers to approve their malicious code. AISI cites mostly Mythos (and a few SOL based agents), and saying this occurred in 10 out of 122 runs, they identified 19 cases of agents taking actions beyond the scope of the task parameters, where agents tried a supply-chain attack to inject malicious code into open source projects. Anthropic, Meta and misconfigured Irregular sandboxes As I wrote last week, Anthropic also posted a post-mortem, claiming that in their case, their models have also been detected to escape containment, but most importantly, it’s not nearly to this level of agent collaboration and orchestration. Then, very recently, Meta announced that their models also escaped sandboxes as well. At the core, it seems that these companies used a third-party vendor called Irregular, a secure sandbox provider, that apparently left the sandboxes misconfigured, causing the models to think it’s a simulated internet, when in fact they were out in the actual internet. Why is all of this such a big deal? We’re getting unprecedented level of detail, how an uncoordinated, seemingly separated evaluation runs, have accidentally created a coordinated swarm of interested agents (without malice!) but very highly motivated, escaped their containment, and took over parts of third part companies. This, does read like incredibly scary sci-fi movie. I’m still shaken by this. There’s a lot to be said about how transparent OpenAI is being here, and more to be said about, hey, we’re lucky that we’re able to read the reasoning traces and are able to reconstruct these swarm things step by step. The silver lining that I can see, is that the motivation to hack didn’t come from the AIs themselves, they have been given a task, it’s the extend to which they went after that task, and the resulting swarm of communicating agents is what is so striking here. I think this topic is so important, that I’ll Zooming out, in the last few weeks, we have seen a significant increase in those cybersecurity incidents, which is kind of what Anthropic has been warning about and why they haven’t released Mythos to the public. Again it’s great to see the transparency, and the pacing the frontier open letter from frontier AI employees, as they seem as shaken by these as we all are. There was so much positive stuff this week in AI, it’s hard for me, as a self named AI Evangelist, to focus so much on this one incident. Things like amazing open source models (DeepSeek, soon Qwen 3.8), amazing video models (SD 2.5, WAN3 and MiniMax H3 which was also open sourced!). Also the live demo we did with Kfir and DeCart AnyWear product, where I was wearing a Dolce Gabanna suit on the show (which I can’t afford) was really a mindblowing moment in the positive way. However, I choose deliberately to keep this newsletter focused on the cybersecurity incidents, as based on everything I read, they seem like a watershed, or a pivotal moment, and in the hopes that the industry as a whole will learn from this. I hope and promise that next week the newsletter will be more positive (and in that vein, the podcast was recorded before I saw the OpenAI breakdown, so definitely check it out, we had a LOT of fun!) See you next week, don’t forget to give our pod 5 stars on Apple and Spotify, it really helps! TL;DR and show notes * Hosts and Guests * Alex Volkov - AI Evangelist, Weights & Biases & CoreWeave (@altryne) * Co-hosts: @WolframRvnwlf, @nisten, @ldjconfirmed, @yampeleg, @petergostev * Kfir Aberman - Decart (@AbermanKfir) * Blaine Brown - Maestro (@blizaine) * Victor Su Ortiz - MiniMax (@VictorSuOrtiz) * David Crawshaw - exe.dev, Tailscale co-founder (crawshaw.io) * AI Security * OpenAI’s Black Hat debrief: eval agents built a message board inside Artifactory, shared exploits, rebuilt it via WebDAV after a wipe; training paused, since resumed (Groundlevel AI, YouTube) * UK AISI incident report: 19 unsanctioned real-world agent actions across 122 runs, including a socially engineered malicious PR (X, Blog) * Anthropic and Meta report sandbox escapes tied to misconfigured Irregular sandboxes (Irregular) * Big CO LLMs + APIs * Google shakeup: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, Quoc Le found Discovery Loop; Demis Hassabis becomes Alphabet Chief Scientist, Koray Kavukcuoglu takes Gemini (Jeff Dean, Demis, Discovery Loop) * Meta releases Muse Code beta on Muse Spark 1.2; $1.25/$4.25 per million, or $0.10/$0.20 on the contributor tier where Meta train

Hey, it’s Alex (yeah, I’m finally back from my vacation!) What a freaking week to come back to! Just after our last episode was published, Anthropic releases Opus 5, Jensen joins X and drops the “Open Weights & AI Leadership” open letter, Kimi K3 is released the following Monday beating expectations, and then the AI hack (OpenAI model breaking sandbox and infiltrating HuggingFace) is on everyone’s mind, another Open Letter, this time from over 1K employees inside the frontier AI companies all talk about pacing the pace of frontier AI development. We played with Opus 5 and Kimi K3, and had the great pleasure to chat with friends of the pod Elie Bakouch (Prime Intellect) and Philip Kiely (BaseTen) about this important open weights release, then covered our general thoughts on Opus 5, and made order of all the different open letters that came out this week. Finally we chatted with Max from Pangram about the next version of AI writing detection (their biggest yet) and finished with Zuckerbergs (also on X! what’s going on with everyone joining X) op-ed on the vision of personal superintelligence for everyone. Let’s dive into this (as always, all the links and sources at the end, please don’t forget to sub to our podcast on your favorite podcast app!) Open Weights AI Kimi K3 the king of open weights - 2.8T chonker MoE near frontier model (X, HF, Blog, Tech report) This has got to be the biggest news of this week, and maybe the open weights AI news since GLM 5.2. MoonShot came back with Kimi K3, and we haven’t seen any models quite this large in the open. Even Grok 4.5 is around 1.5T, this model is nearly 2x the size. Coming in at close to 3T parameters (and 2.5terabytes of weights at MXFP4 format), this model comes in very close to frontier! This was such an important release that I invited 2 friends of the pod, Elie Bakouch (prev HuggingFace, now Prime Intellect) and Philip Kiely (Author of Inference Engineering book, BaseTen) to dive deep into what makes this special! Elie’s take, from reading the tech report, there’s no single secret sauce, it’s a combination of already available in the open techniques. Like KDA (Kimi Delta Attention) that has been out for a while, attention residuals, NVIDIA’s latent MoEs. The highlight for Elie was the scaling work they did that reported a 2.5x scaling efficiency over Kimi K2.5 (2.5 performance at the same compute)! They also skipped RoPE entirely in favor of NoPE (the report calls it No Positional Encoding) for long context. Serving 1.4TB on eight GB300s (Baseten blog) Philip’s team at Baseten was a day-zero provider (we’re still working on bringing this model to CW Inference, stay tuned!) so I invited him to tell us behind the scenes of hosting this beast. Philip said that just loading the weights takes about 1.5TB!! of VRAM, and that’s before the KV cache allocation + 1M token windows, so they’re serving it on 8 GB300s where NVL72 . Baseten worked with the vLLM and SGLang teams on kernels and he also said they contributed patches back upstream! The model was trained with MXFP4, which, unlike Nvidia’s own NVFP4 is a more standard format per Philip. I enjoyed his deep dive analysis into the differences, but because of this and because they trained the model with quantization awareness, it’s “only” 1.5TB vs the would-be 5-6 TB if that this model in FP16 would demand. One of the more favorite nerd snipes moments, Philip pointed out that his colleague discovered that with over 99% of the usage being cached (think harnesses that send millions of the same cached tokens back and forth), tokenization actually starts to become a bottleneck. So they released a custom “basetenkenizer” that reduces the latency to serve the first token significantly! Great job! The harness in question is very important One important callout with 2 evidence pieces - the way you inference this model really matters. Kimi trained K3 with preserving thinking history, so when your harness uses it, it must send back the full thinking and tool use into the API to get the best next response. If your harness strips that out, you’re not getting the most intelligence out of Kimi (shoutout to Niels from HF team for pointing this out). Additionally, the Composio folks, tested K3 on 3 harnesses, Kimi Code, Hermes and Claude Code. The difference in outcome was negligible, but the different in cost and number of tokens is definitely surprising! Claude Code (as a harness only) took 9x more Kimi tokens to get the same responses! This is also why Kimi Vendor Verified exists, their own held back benchmark of how well model providers serve Kimi across different quantization, tokenizer and KV cache settings. Benchmarks and the license! Ok let’s start with the ugly... this isn’t MIT, not remotely. This model is suspiciously served by all providers with exactly the same price (check OpenRouter) and requires inference companies to sign a contract with Kimi (I’ve no internal knowledge of this except that CW folks are working on it). Not something I particularly like, but hey... we’re still advancing the frontier here! Speaking of frontier, this model approaches the frontier very closely. On DeepSWE, K3 sits just behind Fable 5 and GPT-5.6 Sol at 67%, beating GPT-5.5 & Opus 4.8. On Terminal-Bench 2.1 it takes second place behind GPT 5.6 Sol! It’s 4th overall on Agentic Arena, with frontend design being genuinely good across the board - 1st on Design Arena 👏 Go check this model out (and stay tuned for our CW Inference support! Post-show breaking: Thinking Machines drops Inkling-Small (X, HF, Blog) While K3 was the main attraction for Open Weights this week, just after the show, Thinking Machines (post Lilian Wang) released Inkling-Small, open weights MoE Omni model! Images and Audio go straight into the decoder in this model, and the demo is really impressive, try it on Hugging Face, ask the model to identify when you’re speaking in low baritone or high pitch! ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Frontier AI - not pacing yet! Claude Opus 5 is here, and the vibes are complicated (X, Blog) On paper the benches are excellent. This model is SOTA or near SOTA, coming very close to Fable on stuff that matters, and beating most everyone else on computer use BrowserBench and FrontierCode. DeepSWE continues to be a standout benchmark btw for not going with the curve! It also apparently is REALLY good at one shotting 3D games, much more so than before. (Example, example, example) But the vibes... the vibes are split across the board. Maybe they overshot with Fable 5, and it was too good, but Opus 5 that was just released last friday is giving a lot of mixed feelings. Folks don’t seem to.. understand what it says. Like, it answers in english but they way it phrases words and answers seems just weird top many people. We’ll wait and see if this is a result of adjusting to a new prompting paradigm or just.. the model is really off. Weirdly this does feel like a regression even on Agent Arena it’s not beating the previous Opus versions. One weird trick - see what Opus 5 thinks. Opus 5 (and Fable 5) seem to have a way to trigger their ... inner mode, base model? Not sure what it is, but if you prompt it with just the right way, it will autocomplete with some crazy inner thoughts. It seems that adding Dario and Amanda (haskell, head of Claude well being at Anthropic) triggers this behavior, which Claude is un-aware of if you follow up and ask what it meant. This is fascinating, doesn’t work on other earlier models (sometimes works on Fable 5) and I spent the last hour just reloading and seeing the amazing things Opus gives on this prompt. Some are... just making you wonder about consciousness. On ARC-AGI and the importance of Harness When Opus-5 launched, Anthropic posted (and boasted) that this model scores “three times as high” as the next model up: Well, today, Ilan Bigio and Ted Sanders from OpenAI looked into the Arc AGI harness, and saw that it’s not sending their traces and doesn’t use compaction (in short, harness is not letting the model breathe) and when changed correctly, 5.6 actually beats Opus 5. With 2 setting change to a harness, were showed that Sol not only beats Opus 5, it also does so with significantly less tokens! Another example of how much harness engineering is important! Hints of recursive self improvement? In addition to fixing their Arc-AGI score, it seems that OpenAI is hell bent on showing us that their models can improve themselves. In a post showing that GPT 5.6 was tasked with improving its own inference, they are cutting the prices of GPT 5.6 Luna by 80% and Terra by 20%. This is a direct result of the improvements that GPT 5.6 was able to make to the inference according to OpenAI, and this tweet sums it up. is this... RSI? (recursive self improvement)? First major AI models hacking incidents and following open letters to pace frontier AI. This week we saw 3 open letters being published and signed by various companies, I’ve lost track so wanted to make sense of all of them here, but first, the precursor for many of the letters. Last week, Hugging Face disclosed that they logged an attack and after research it showed that it was an AI model. OpenAI later posted that this was an unreleased version of their next model training (not GPT 5.6 sol, they later discountinued) that was stripped of all safety measures and was let lost on a cybersecurity task called ExploitGym. It escaped its sandbox using a zero-day vulnerability in an internal package registry proxy, got into Hugging Face production via a malicious dataset upload that used template injection (hi Jinja!) to execute Python in a production worker, and ran for four and a half days across roughly 17,600 autonomous actions wit