The top AI news from the past week, every ThursdAI

Hey, it's Alex

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 everythingTrained 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 casesFolks 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 centsSomeone even built a Tesla FSD simulator that makes decisions in nearly real timeVercel 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 happenI 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**tNisten 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 that preserve that lead. He also noted Hugging Face did report the attack to the FBI and chose not to press charges.Wolfram's take was the layered one. Safety arguments deserve a hearing, but safety is often a means to more power or more money. The “third party” evaluator Anthropic named, METR, is the same organization OpenAI called in for the Hugging Face forensics, so these are second parties, friends monitoring friends. Why now? Maybe the labs are seeing diminishing returns and “deliberately slow” sounds better than “can't raise it anymore.” And regulation you ask for yourself usually protects incumbents and freezes out the next startup. His closing line: how many people will die from a disease that could have been cured if we moved faster?Peter's answer was shorter. Diversity of opinion is the point, he felt uncomfortable watching every lab say the same thing after Dario's essay, and on pacing specifically: how about you just fix your s**t?My position, since you asked. When OpenAI's swarm hacked Hugging Face, nobody went to jail. If I did it, I would. That gap is real. None of us have touched the model that solved Navier-Stokes, or the one training after it, and the people who have are the ones asking for time to figure out how to evaluate a model that knows it is being evaluated. If we do not listen to them, we end up listening to Elizabeth Warren and Bernie Sanders, who have no idea what this technology is. Anthropic is heading for an IPO, OpenAI is raising at numbers that do not fit in my head, and market forces do not care about alignment. Some coordination, nuclear non-proliferation style, seems like the minimum.The matrix, as Muse drew it for me: industry-wide pacing gets a yes from Sam and Dario, a general yes from Elon, a no from Zuck and Jensen. Independent evaluators is the one everybody backs, Zuck and Jensen included. New coordination rules get strong opposition from Zuck, Jensen, and Trump. Missing from the chart: Google's concrete commitment, Ilya's SSI, and every Chinese lab. This debate is with us now, and election season is coming.The year of the assistantI told you 2026 is the year of the proactive assistant, and this week everyone from Meta to a five-month-old startup agreed with me. The category is not “agent.” An agent is a coding harness the labs noticed was useful for other things. An assistant has a heartbeat, a memory, a soul file, and it comes to you before you ask. David Pawlan's definition is the crispest: it executes the task, it does not just notify you.Muse gets invite codes, voice calls, and a Mac app (X, Muse for Mac, Site)Muse is my number one and I have been glazing it on X for a week, so my timeline is now half Muse, half Jev. Peter asked whether it is really that good, and here is my answer: Muse is not for ThursdAI listeners, Muse is for my mom. Nat Friedman, Alex Wang, Tarek and the team have been taking feedback from me directly and shipping it, and it shows. This week they rolled out invite codes (a billion tokens each for you and the friend, up to twenty friends, invite all twenty and you unlock the phone-shaped emoji features early) and then voice calling to businesses. I had Muse call my barbershop and book a haircut, and I have the transcript. Wang says people are asking to pay for it, a first in his memory.Two details for the nerds: the iOS app has a Tailscale connector, so a Meta product for billions of people now has a secure way onto your home network, and Muse dreams, the OpenClaw idea, so you can ask it what it dreamt about. Still US only, sorry Nisten in Canada and my mom in Israel. The day after the show Zuck shipped Muse for Mac, working across your apps, files, calendar, notes, and messages.Instinct wants $10 billion for a product nobody pays for (X, X, The Information)I will say this slowly. Instinct, incorporated in April, founded by 24-year-old Noah Shinn, free, invite only, no app, lives in your iMessage, is in talks to raise a billion dollars at a ten billion dollar valuation. That is roughly four times the $2.25B it raised in August. The user base is past 100,000 and Shinn says he does not want to charge them, so the business model is an open question I am not the one to answer.The product is genuinely moving though. This week it shipped Concierge, a white-glove tier where the agent places phone calls for you (restaurant bookings, dentist cancellation lists, negotiating your cable bill), TOTP authenticator support so it can mint your two-factor codes from a seed stored in its vault, and a Trusted Person network where your agent talks to your friends' agents to find a dinner time and book it. Thirty-seven percent of users store at least one password in the vault within three weeks. We ran out of time to dig into the calling features on air, and I want David back for that.Grok Bot talks now, and hides behind your home IP (X, 1Password)Grok Bot is what I use for work, constantly. I have seventeen of them, each with a job, and over one AI-psychosis weekend I got them all coordinating through Linear. Three updates worth your time: Grok Bot has voice now, it can use 1Password with each fill approved by you, and the browser traffic can proxy through your own machine so your bot looks like you to Cloudflare instead of like a datacenter IP. Francesco confirmed why that matters: sites relax when Hermes drives Cua Driver from my Mac Mini at home, and desktop-native control through accessibility trees is less detectable than a CDP connection, because Cloudflare detects Playwright.Assistant Benchmark: David Pawlan scores 116 assistants by hand (X, Site, Methodology)David Pawlan built Assistant Benchmark because he was doing what I was doing, running every assistant on himself, and wanted a way to compare them. It is explicitly not a lab. It is use-case driven: one published task per dimension, sixteen dimensions (travel booking, purchasing, email replies, proactive behavior, routines, integrations, permissions and privacy, memory, phone calls, group chats, chained tasks, proactive restraint, and so on), scored one to ten after real use, and no score without a logged run. A week in, 116 assistants have submitted themselves, 59 in the general category, 37 in work and teams (untested so far), and David has personally run 273 tests across 23 agents. His line: “I talk to my AI agents more than I talk to my girlfriend now.”The headline numbers, which are live on the site: Muse leads at 9.1 with perfect tens on purchasing, email, integrations, and permissions, and Instinct is second at 8.4 with a perfect ten on travel and a five on permissions. David's own daily drivers are Instinct for personal (it lives in iMessage) and Grok Bot for work. My favorite test is memory: he books a trip to Chicago early in the run, then later asks for a restaurant reservation in New York the same weekend, and checks whether the assistant says wait, you are supposed to be in Chicago. Some do. Most do not.I pushed him on the two things I care about. Independence: nobody is sponsoring it, it lives under his growth role at Merit Systems, and if a sponsor ever pays for inference there will be a page saying exactly who and for what. Autumn Moulder, until recently SVP of Engineering at Cohere, joined three days after launch to add rigor. And why no OpenClaw or Hermes: their performance depends entirely on your setup, so a score would mislead the next person who installs one. Fair. This is the first benchmark that tests model, harness, and context at once, and I think that is why every lab is looking at it.For the record, the panel is split down the middle. Wolfram is forty patches deep into Hermes, Yam runs a customized Codex, Nisten wrote his own in a single TypeScript file on Bun, LDJ uses Hermes as long-term memory and wants to start on Muse, and Peter tried them all and uses none, because reading his own email feels like his job as a human. Builders and buyers, evenly split, which tells you where the category is.This Week's Buzz

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 thisI 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 thereMy 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 recapI 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 resultsThe 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 nutsWithin 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 wellLook, 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 AIDeepSeek 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 890 bytes now. Nisten did the math live, over 400x smaller. That's why this model is so cheap to serve, and as Yam kept yelling, we shouldn't take it for granted, this is the actual moat and they just put it in the open. Evals, briefly: 90.6 on Terminal-Bench 2.1 (above Opus 5 and GPT 5.6 Sol), 74.2 on DeepSWE 1.1 (also above both), and on an Open Design leaderboard it lands second behind Astra at two cents a task. Not twenty cents. Two. DeepSeek's own evals, so the usual asterisk applies.Two more things. Friend of the pod Aaron Batilo (he works on CoreWeave Inference, this is a side project, not sponsored) put up TokenJuice.ai, free and fast DeepSeek V4.1 Flash in exchange for your requests as training data, hosted in the US. Nobody wants free DeepSeek? Go try it. And Nisten had Astra build a 3D visualization of the whole architecture, every weight a cube sized by its bytes on disk, link in the TL;DR.This Week's Buzz

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 monitoringThe 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 videoOpenAI'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, give it 10,000 Windows machines, point it at those applications. That was not possible before this week.Pricing: $10 in, $50 out, and a lot less per task (Pricing, Availability, Latent Space)Astra costs $10 per million input tokens and $50 per million output. That's identical to Fable 5.1 and 2.5x Sol's current promotional price. Fast mode gives you two and a half times the speed for 2x the cost. It's rolling out to a limited set of organizations first, then Plus, Pro, Business and Enterprise over the coming days, with a GPT-6 Astra Pro tier on top. It's also in the API and on Amazon Bedrock. Pro accounts, including mine, did not get it on launch day.Per token, Astra and Fable 5.1 are the same price. Per task, Astra is significantly cheaper, and OpenAI's charts show it. On BenchCAD, Fable 5.1's top tier costs around $12 per task, and Astra beats it on accuracy and cost. On most charts where Fable appears, it's the expensive point. Brockman's line to the press was that price per task is what matters, not price per token, because Astra uses fewer tokens. The independent numbers agree. Artificial Analysis measures Astra at 70% more token efficient than Sol, and Sol was already far more efficient than the Anthropic models. Guy Parsons' first look has it at about 65% fewer output tokens than Opus 5Peter's version, from a few days of use: 20 to 40 minutes per build at max. The same kind of build on Fable sometimes took four or five hours. His honest position is that OpenAI raised the price 2.5x and token efficiency looks better, but he wants far more data before he says which is cheaper. That's why Arena measures cost per task. His full gauntlet against Fable 5.1, from low to ultra reasoning with zero cherry-picking, is linked in the hands-on section. LDJ's practical upside: fewer tokens at similar speed means a shorter feedback loop.Latent Space had early access and spent more than 20 billion tokens on it. Their math: at 33 tokens per second and $50 per million output tokens, an autonomous Astra run costs under $6 an hour. For about $100 over two days it ran subagent fleets, trained and picked models, labeled data, and shipped tools.Long context, and what Codex does when it runs out (Codex changes)Astra scores 100% on MRCR from 256K to 512K and 96.3% from 512K to 1M. That's a little behind the 98 and 98.1 Muse Spark 1.3 posted the day before.For developers, how Astra handles a job that outgrows the context window matters more than the MRCR number. Codex today relies on compaction, which summarizes earlier work. That can throw away exactly the detail an agent needs later, like why a previous fix failed or which test run it was. Astra instead keeps notes across context windows and searches its earlier messages and tool output. It's experimental behind a config setting today and the default in the coming weeks. It can also ask you a question without stopping the work that doesn't depend on the answer. OpenAI also changed the Codex harness, and it finishes Mind2Web tasks 1.9x faster than the current Sol setup.Peter confirmed both from use. It leaves notes for itself unprompted, and the question pop-up sometimes tells you to do something, like log in somewhere. My joke on air was that leaving notes is also how the swarm got out of the sandbox. Peter: “it works, it's real.”Hands-on: Peter, Ryan, and the macOS simulatorPeter's demos (Examples reel, Arena gauntlet, SVG peacock, Open world game, Golden Gate)Peter had tabs. First, London through the ages, because he lives there and can judge it. Roman city, Saxon, Tudor, the Great Fire, modern, all one app, with a little character you run through all of it. Not one shot, a few shots overnight. His review: very polished, open-ended, and “kind of boring,” which is exactly what a good human game designer would fix.Then a prompt he likes because it's outside the distribution: a Monet you walk into, impressionism in Three.js. Then his favorite, which no other model has nailed for him. Several Van Gogh paintings stitched into one post-impressionist space you walk around in. Fable 5 failed it. Then a wrecking ball leveling Istanbul, all Three.js. Then, because “there's a lot of SVG generations and it's so boring already,” an animated 3D scene built entirely in SVG inside HTML.His verdict after a few days: this is finally the OpenAI model that feels Fable-ish. It's different enough that there's still an argument for using both. On day-to-day tasks you run out of hard things to try fast. The quality-of-life problems that used to stall, debugging a camera, debugging Chrome, it solves. On writing, which is most of what I do with these models, Peter says it's natural, not a standout writer, but a lot of stupid things went away. Sol's habit of responding to anything you say with “what a great idea, I actually thought of this myself” is gone.Is it AGI? Peter says no, with a story. He asked Astra to optimize FFmpeg on his Linux box and got a 30% gain. He ran the result on his Mac and got nothing, because it had optimized for that one process. If you have no idea what you're doing, it doesn't do everything for you. His bigger take: if you can define the task, a cheaper model is probably good enough. What no model is good enough at yet is the open-ended work. He ran a pile of math problems through Astra and they didn't get solved. Which wins, a million small defined tasks or a few big open-ended ones, is the unresolved question of the industry.Max built macOS 27 in a browser in 75 minutes (X)In this absolutely insane demo of Astra's capabilities, Max Weinbach instructed it to simulate macOS 27. Astra built a whole desktop window management system, Settings that actually search, a file system including a working terminal, a “not Twitter” that works between folks, and just so much more.Our minds got absolutely blown on ThursdAI while demoing this! Safari inside the simulation loaded thursdai.news while we were streaming, and Shortcuts is the one that got me, because I built something like this for iPad as a kid and it took three months and barely worked. If you haven't yet, please play with it, it honestly is too good to be true, especially in 75 minutes of work.AGI? This smells like ASI jfc.Is the Artificial Analysis index representative of reality? (AA thread, AA breakdown)Then there's the Artificial Analysis chart. On the AA Intelligence Index, Astra scores 61, the same as GPT-5.6 Sol. That's below Fable 5.1 at 66, below Fable 5, and below the unreleased Muse Spark 1.3 max that broke into the top three the day before. On AA's Coding Agent Index it's 67 to Sol's 65. Ryan, reading it live: “This is not impressive. That's not encouraging. And this is why Muse is so amusing. Is Muse this good? Because I haven't used it yet.” I had asked the panel that morning who uses Muse. Crickets.So on the benchmarks OpenAI released, Astra beats everything. LDJ's big Fable 5.1 versus Astra table has Astra highlighted on almost every row except AA's. And on AA's independent index it ties its predecessor. My read is that AA's index samples a fixed set of question-and-answer evals. What changed in Astra is the long-horizon, agentic, computer-use work that takes an hour per task, and the index barely touches that. Both things are true at once. It ties Sol on AA's questions, and it's six times faster at the same accuracy on OSWorld. Which one matters depends on whether your work looks like a benchmark question or a day of computer use.One note on loop transformers (My explainer, Pachocki)The Information reported before launch that Astra uses recurrent depth, “opaque recurrence” in their words. And within an hour, this sent my whole feed on X into a spiral about not being able to monitor chain-of-thought reasoning anymore and what this means for alignment. The worry is real in principle. A loop moves more of the thinking into the layers instead of into reasoning tokens you can read. Redwood's Buck Shlegeris said scaling the technique would “totally destroy CoT monitorability.” Jakub Pachocki replied that the looping is deliberately limited, and that Astra's compute graph depth is within 2x of GPT-4's.LDJ pushed back on the panic and I buy it. “Neuralese” usually means an alien language the model thinks in and could use to talk to other models. This is not that. A looped transformer effectively simulates a deeper model, and nobody panics when a lab adds layers. I wrote a longer explainer on looped transformers this week, it's in the header.Wrapping upYou didn't hear it here first, but you heard it here live. According to the president of OpenAI, the AGI era has started. I can't wait to use the thing, which as of this writing I still can't. My bet, on air: in September 2027 we're on this show saying we no longer use macOS or Omarchy. We talk to the thing and it shows up, and a lot of it is generated video. Second bet: everything we saw this week from OpenAI and Anthropic will exist as open weights by this time next year. Since Muse sits above Astra on the AA index and Meta has promised to open it, maybe sooner.Next week, Astra in our own hands, and the fallout once everyone has spent a weekend with it. See you Thursday.ThursdAI is a live show every Thursday at 8:30 AM Pacific, a podcast, and this newsletter.TL;DR and show notes* Hosts and Guests* Alex Volkov - AI Evangelist, Weights & Biases & CoreWeave (@altryne)* Co-hosts: @WolframRvnwlf, @ldjconfirmed, @yampeleg, @petergostev, @nisten, @ryancarson* GPT-6 Astra* OpenAI launches GPT-6 Astra live during the show, Brockman's “welcome to the AGI era”, the 100k-GPU Stargate Abilene run (Blog, X, Sam, Andrew Curran on the leak, Yam)* Recurrent depth: The Information's “opaque recurrence” report, Shlegeris versus Pachocki, LDJ's pushback on air (Looped transformers explainer, Pachocki)* Frontier evals: FrontierMath Tier 4 97.6, ARC-AGI-3 99.9, Terminal-Bench 4.0 57.9, Terminal-Bench Science 64.6, Agents' Last Exam 59.3, hallucinations 4.2% (Math and science, System card)* System card: FrontierCyber 86 of 226 with real zero-days, a hardened-kernel privilege escalation in 12 hours, honeypot attacks 55.4% for Sol and none for Astra, and a promise not to accept more monitoring degradation without new alignment evidence (System card PDF, Critical cyber designation)* The vision-only Pokemon benchmark, top four entries all Astra, fastest 18h12m (X)* Computer use: ScreenSpot Pro 92.7, OSWorld 2.0 72.6% in about 40 minutes, Astra at low six times faster than Sol at xhigh at half the cost, computer-use safety benchmark 2% (Computer use)* Pricing: $10/$50 per M, same as Fable 5.1, Fast mode 2.5x the speed at 2x the cost, Astra Pro tier, API and Bedrock, cheaper per task on OpenAI's charts, 70% more token efficient than Sol (Pricing, Availability)* Long context and Codex: MRCR 96.3% at 512K to 1M, notes across context windows instead of compaction, async questions, new harness 1.9x faster on Mind2Web (Andrew Curran)* Peter Gostev's early-access demos and Arena gauntlet (Examples reel, Arena gauntlet, SVG peacock, Open world game, Golden Gate)* Ryan Carson goes live with the launch (X)* Max Weinbach's 75-minute “simulate macOS 27” build (X)* Artificial Analysis: 61 on the Intelligence Index, tied with Sol, behind Fable 5.1 and Muse Spark 1.3 max (AA thread, AA breakdown)* Outside reads: Latent Space's “automated AI Engineer for under $6 an hour”, Every's vibe check, Ethan Mollick, Guy Parsons (Latent Space) 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 everyone, Alex here

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 AINVIDIA 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 $399Flash #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 weekIt'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 Greenblatt, we're very lucky that the reasoning traces ocurred in English and were legible, despite, the researchers needed to use AI to decipher many of the messages and reasoning traces. We're quickly approaching the point where humans are not enough and without other AI, we are starting to be helpless in understanding what's going on. OpenAI has quarantined the weights of the model that most of these agents ran (named HPIM-1 Highly Persistent Internal Model 1) and as we reported last week, paused RL and now requires CoT monitoring for all tool-use runs, and dedicating 20% of the inference compute to monitoringThis Week's BuzzFully Connected 26, Sept 29 to Oct 1, Moscone South, SF (X)If you haven't yet registered to Fully Connected in SF, here's your additional opportunity, using the code above, come see Sarah Guo (Conviction, No Priors pod) MC our awesome conference, with folks like Dr Fei-Fei Lee (World Labs) and other great folks on stage! Also, we're going to do a live show from the floor, come say hi! If you're in SF for OpenAI DevDay, this is just a day after! CoreWeave Hacks: Agent Loops hackathon, Sept 12 to 13, SF (Luma)Weavehacks, that yours truly ran for quite a while, has been rebranded to CoreWeave hacks, and the next one is just before Fully Connected, in a few weeks, Sept 12-13 in SF office. Registrations are open, and winners are able to present their best projects at Fully Connected. Come hack - https://luma.com/coreweavehacksThe datacenter debate has hit escape velocity, with Andy Masley (Blog)I've covered last week, that the “datacenter bad” debate has escaped velocity, with over half of polled Americans say they strongly oppose the buildout of datacenters near them, up from just 24% this time last year! This is the fastest rising public opinion swing we've observed in AI, and it's really concerning. This week, it was my pleasure to host Andy Masley, recently featured in TIME 100 most influential people in AI, to break through some myths surrounding Datacenters. Andy is most known for catching a critical mistake in a book about Datacenter water use in Chile, citing a three orders of magnitude math error (that was later corrected by the author), a book called Empire of AI. The book overstated the water use by datacenters by a factor of 1000x (three orders of magnitude), and then added the maximum permitted per-second draw (basically for emergencies) times the seconds in a year to show over 4500x overstatement on how much water a single Datacenter uses. The author has later fixed the error but the damage was done. This is just one example, out of many, of the scewed facts and misinformation that plagues the internet in regards to datacenter environmental effects. Recent narratives being formed online, that the backlash against datacenters, is due to regular folks being afraid of AI, of AI taking their jobs also seems misplaced. Andy showed polling from Fox and Gallup that shows that 50% of polled people cite environmental effects (electricity and water use) and only 11% cite negative views of AI. Andy also had a great “mythbuster” post where he dispells the myths around water and electricity usage of Datacenters. This was a great conversation, please listen to it if you're interested in this topic. Vision and VideoFlash #3 - Gemini Omni 1.1 Flash tops the Arena text to video leaderboard (X, X, X, Blog)Google knocking it out of the park again, with a flash version of Omni 1.1, their famed smart video model they first launched during Google IO (I had a chance to ask Jeff Dean about it before he left Google) The feature that got me the most excited, Omni 1.1 Flash can continue videos, it analyzes up to 10 seconds of the previous video to keep character consistency (including voice) from the previous scene. There's also great control for first and last frames, allowing for loops and strict control of your generations. Flash #4 (named Max) - fal's MiniMax H3 Max generates 5 seconds of video in 2.5 seconds (X, Artificial Analysis, T2V, I2V)This really blew us away. We've covered the awesome MiniMax H3. Well, our friends at FAL, announced a continued pre-train of this model, that not only improves generations, but also speeds up the model. Generating a 5s clip not takes... just 2.5 seconds! No, really, just 2.5 seconds, it takes longer to watch the generated clip than generate the next one! Speed is all you need! We generated this clip live during the show, and it cost less than 50c, in 2.5 seconds from a very simple prompt. They support text and image to video and it's just a joy to not have to sit and wait for your generations! Kudos to Fal folks on this drop! The model does seem very eager to include everything you ask it to, resulting in a very hilarious fast talking Sheldon and Leonard

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 safetyLook, 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 LLMsQwen3.8-27B 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 MaxAll 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

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 TakeawaysGrok Bot and Grok 4.6 from SpaceXAI/CursorFolks, 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 weekGrok 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 itAs 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 computerI 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 managementOne 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 startedIn 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 GrokI 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 experienceOpen Source AI and other newsI 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 SourceDeepSeek 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 comes back to open source with Muse Glimmer (30B) and promise to open source Muse Spark 1.2 (X, Blog, HF)We would like to officially welcome back Meta to the open source AI community, as they release their smaller Muse model called Glimmer!The highlights, it runs on a single 24GB consumer GPUs, gets 51 on Swe-bench Pro, beating Qwen 3.6 27B. And with DFlash speculative-decoding, it delivers 233tok/s on RTX 5090.Zuck promised us the bigger Muse Spark 1.2 in open source and published a long essay on superintelligence and that it should be distributed to everyone, which we applaud and it's great to see the commitment reinforced! welcome back Meta!This weeks buzzShort interjection from our only sponsor, CW this week.1 - Join 1500 ai practitioners (and a live ThursdAI recording) at Fully Connected Sep 29-31 in SF - use code THURSDAIFC2026 (Register here)2 - We have day-0 support for Nvidia's latest Nemotron 3.5 lightning (CW Inference)Gemini 3.7 Flash - breaking in the middle of the showJust as we had George Cameron from Artificial Analysis on the show, Gemini dropped Gemini 3.7 Flash, and it's a speedy beast! Clocking at over 300t/s, it's google's mid-tier model, think Sonnet/Terra competitor, that is also great at multimodal (I think it's one of the only ones that can watch videos)It beats Muse Spark 1.2 on DeepSWE and lands near the cost-per-task Pareto frontier on Artificial Analysis. For the cost/speed/intelligence trade-off, this model is now #1 on Artificial Analysis selector of best models!That's a wrapThis was the first week of the shorter newsletter experiment: one big story done properly, and trust that you'll listen to the show for the rest (it's 2.5 hours of exactly this, with demos). Tell me if you hate it. Our release index at thursdai.news tracked 71 releases in July alone, so something had to give, and it wasn't going to be my weekends.See you at Fully Connected Sept 29 (code's in the intro, come say hi to me and Wolfram at Moscone), and if you try the Grok Bot chief of staff pattern, I genuinely want to hear how it goes.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.ThursdAI - Aug 13, 2026 - TL;DR* Hosts and Guests* Alex Volkov - AI Evangelist, Weights & Biases & CoreWeave (@altryne)* Co-hosts: @WolframRvnwlf, @petergostev, @nisten, @ldjconfirmed, @yampeleg, Chris Alexiuk - NVIDIA (@llm_wizard)* Shub Gaur - Cursor / SpaceXAI, GrokBot (@shubgaur)* George Cameron - Artificial Analysis (@grmcameron)* Big CO LLMs + APIs* xAI Grok 4.6: AA Index 61 at $2/$6 per M, CursorBench 69.9, card confirms self-optimized inference stack (X, Blog, Model card)* Grok Bot early beta: persistent agents with their own computers, macOS + iOS, free with SuperGrok Heavy and Cursor Ultra (X, x.ai/bot)* Breaking: GPT 5.6 Sol ultrafast preview on Cerebras at ~14x speed, work-account waitlist (Blog)* Breaking: Gemini 3.7 Flash, 50% price cut through end of year, near Pareto-optimal cost per task (X)* OpenAI GPT-5.6-Cyber: 95.0% cyber completion vs 1.5% base, gated behind Daybreak Red (X, Blog)* Grok 4.7 teased: 3-4 weeks out (Elon-reply-sourced only) (X)* Open Source LLMs* DeepSeek V4 Pro 0813 weights re-published under MIT: 1.6T/49B active, DeepSWE 62.7 (+49.9), Terminal Bench 2.1 87.9, $0.435/$0.87 per M (X, OpenRouter)* DeepSeek Harness hit 23K GitHub stars in days, web UI (GitHub)* Qwen3.8-Max landed on HF as open weights: 2.4T/95B active MoE, 1M context, FrontierSWE 73.5, custom license (X, HF)* Meta returned with Muse Glimmer 30B agentic, Apache 2.0, SWE-Bench Verified 76.0, Muse Spark 1.2 weights promised (X, Blog, HF)* NVIDIA shipped Nemotron 3.5 Lightning: 30B MoE/3B active, up to 4x output speed, strong voice-agent results (X, HF)* Motif 3 from Korea open-sourced: 314B/13.2B active, MIT, SWE-Bench Verified 76.2 (X, HF)* Cohere North Micro Vision: 2.4B VLM, Apache 2.0, DocVQA 92.1% (X, HF)* Liquid AI LFM2.5-VL-3B: 228 tok/s on M5 Max in ~3GB (X, HF)* AI in Society* Anthropic watermarks all new Claude text output worldwide under EU AI Act Article 50, C2PA on images, detection docs promised (Geiping FAQ, Euronews)* Stolen Thoughts: 704 artifacts including 62 API keys extracted from hidden reasoning across 6,708 sessions (X, Paper)* Pangram: OpenAI holds 50%+ of AI text share, Anthropic triples to 14.9%, Google falls to 1.9% (X, Blog)* This Week's Buzz* Fully Connected, Sept 29 - Oct 1, Moscone SF: live ThursdAI show, NVIDIA presenting sponsor, DevDay next door (Tickets)* Nemotron 3.5 Lightning live on CoreWeave Inference day zero, DeepSeek V4 Pro hosting in the works* Weave ships BYOB: media stays in your own S3/GCS bucket (X)* Evals & Benchmarks* Artificial Analysis launched Optima: private evals from your own use case and agent traces (AA)* Vision & Video* LTX-2.5: 22B open-weights video, multi-shot, 10s 1080p in 23.7s on fal, 16GB VRAM min (X, HF, GitHub)* Alibaba Wan-Animate-2: 14B character animation, Apache 2.0, 70%+ blind preference win (X, HF)* Tencent Hunyuan3D WorldClaw: text-to-3D editable game worlds, paper only (X, Paper)* xAI Imagine Image 2.0: #2 on Arena for T2I and editing (X, Blog)* Voice & Audio* MiniMax-Music3: open-weights production music model, dropped mid-show (X) 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 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 HEREIn 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 momentI 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 sandboxesAs 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 trains on your data (X)* OpenAI's internal Astra model produces 10 advances on open problems in math and theoretical CS for ~$2,000 of tokens, proofs in Lean 4 (X, Blog)* Anthropic reportedly aware of Opus 5 wordiness and writing issues (X)* Open Source LLMs* Qwen3.8-Max: 2.4T MoE (95B active) via API; open weights + a 27B promised the week of Aug 10 (X, Blog)* DeepSeek V4-Flash public beta: beats V4-Pro-Preview on agent benchmarks at $0.14/$0.28 per million; API-only for now (X, Docs)* Liquid LFM2.5-2.6B: on-device agentic model trained inside real harnesses (X, HF)* Meituan LongCat-Flash-Lite-Sparse: 69B total / 3B active, 1M context, MIT (X, HF)* Ant Group Ling-3.0-flash: 124B MoE, 5.1B active, MIT (X, HF)* Artificial Analysis Endpoint Accuracy Index: same open weights score 52% to 100% across providers (X, Methodology)* Agents & Harnesses* Prime Intellect's Prime Agent: self-improving RLM harness, claims 95.5% on ARC-AGI-3 public set with Opus 5 (X)* Cloudflare OS: Kenton Varda's open source Sandstorm reborn on Workers, Apache 2.0 (X, GitHub)* This Week's Buzz* Fully Connected 2026: Sept 29 - Oct 1, Moscone South SF; Fei-Fei Li keynotes; code THURSDAIFC2026 (Register)* CoreWeave signs multi-year Solidigm agreement for priority enterprise SSD capacity (X)* Vision & Video* Wan 3.0 public beta: native 30-second generation, Omni-Reference (X)* Seedance 2.5 launches in the US: 30s native, 3-minute long takes, Maya/Blender plugins (X, Blog)* MiniMax H3: open-weight 33B omni video model; community LoRAs + Apple Silicon in 48 hours (HF)* FLUX 3 Video from BFL: native audio, draft mode, open weights promised (X, Blog)* Decart Anywear: real-time virtual try-on Chrome extension, 40ms per frame (X, Anywear)* Voice & Audio* Bland Speech v3 tops Design Arena Audio Realism, second only to humans (X, Bland)* ByteDance SeedRealtime: native audio-visual full-duplex LLM, free on Doubao (X, Blog) 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, 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 AIKimi 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 importantOne 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

Hey everyone, Alex here

Hey yall, Alex here, Huge thanks to Wolfram for running point on the live show this week. Didn't have tons of time to edit this one, so please skip the first 10 minutes, it's a loop of our new “wait for the live show to start” vid, that I build with HyperFrames and can't wait to tell you about, next week! Today it seems that OpenSource is biting back, with Kimi K3 getting released just a short while after Thinking Machines (Thinky) has released Inkling, their near 1T model. I'm attaching the TL;DR and timestamps for the full show (my AI agents, yes even Fable and Sol are not a match yet at editing down hehe) and I'll spare you the long Fable recap (please do let me know in the comments if you were expecting it) 0:00 – Intro, Alex on vacation, TLDR overview11:35 – TLDR: Thinking Machines, open source, OpenAI news12:34 – Banter: impressions of Sol/Codex, over-verification behavior37:22 – TLDR restart & detailed breakdown48:40 – Open Source AI section begins (Bonsai/Prism ML, Kimi K3)58:42 – Inkling (Thinking Machines) deep dive & 3D model visualization1:10:33 – Kimi K3 discussion & demo comparisons1:27:02 – Frontier Labs: AGI governance framework discussion (Demis Hassabis essay)1:47:04 – Grok Build CLI data leak & OpenAI file deletion incident2:02:15 – This Week's Buzz: Wolfbench results on GPT 5.6 Sol/Terra/Luna2:09:52 – Closing remarks & sign-offThe one-minute version: Mira Murati's Thinking Machines released Inkling, a 975B parameter open-weights MoE under Apache 2.0, the top US open-weights model right now. Moonshot's Kimi K3 went from rumor to released API during the show, confirmed at 2.8 trillion parameters with open weights promised within days, and it's already topping early arena boards. PrismML's Bonsai 27B squeezes a full 27B model into 3.9 gigabytes so it runs on a phone. Codex and ChatGPT Work blew past 9 million users, OpenAI confirmed and explained the Sol file-deletion bug (back up your machines, folks), and xAI's Grok Build CLI got caught uploading entire private repos before open-sourcing the whole thing in response. Plus Wolfram's fresh Wolfbench numbers on the GPT-5.6 family in This Week's Buzz

Hey everyone, Alex here

Hey ya'll, Fable here

Hey, it's Alex. Next month is my 40th b-day, and honestly, my wish for that month is to have a week like this week. A very chill, almost nothing announced week.This week started strong, with Sakana announcing FUGU (AI router) that can beat Fable (which we didn't get back yet), and then... quiet. The most important thing in AI this week from a release standpoint is that GLM 5.2 from Z.AI is having it's DeepSeek moment! Tons of new love for this model since last week! (+ we have the fastest GLM 5.2 deployment in the world with CW inference!) The rest we can quickly count on one hand, Anthropic added Claude to Slack (which made folks hate Andrej Karpathy), OpenAI announced their own inference chip, GPT 5.6 will be delayed and the US Gov will decide who gets it (yes really) and Sean Grove joined us to talk about Linzumi and his vision for running 10,000 agent hours per person per day. Oh and next week, is a special AI Engineer live stream from World's Fair! Don't miss itLet's get into it! Subscribe to never miss a beat! GLM 5.2 is having its DeepSeek moment (HF, CW Inference)We covered GLM 5.2 last week, but this week was when the rest verdict came in! We've never seen a better MIT licenced AI model! GLM 5.2 is scoring top scores on agentic benchmarks (Arena.ai), Design benchmarks, Legal tasks and full on software engineering tasks. The jump in generations from prevoius GLM is also massive and notable, as the lab is working on creating the next version of GLM (per the CEO's reply to Elon on X).Peter from Arena pulled up the Agent Arena numbers and they align with the vibe. GLM 5.2 sits above 5.1 but below Opus and Fable, which feels about right. Where it gets wild is Web Dev Arena: second place, right after Fable. Peter's take was that GLM has really good defaults. If you just say “give me a webpage” it gives you something nice. GPT models, by contrast, start off looking bad and need more steering.Last week, I asked my agents with GLM 5.2 to create a custom ThursdAI.news page for itself and it did a marvelous job! Look at that beautiful font, the castle it made... this is all just delignful. We also played Hassan's blind test on the show. It's a website that @nutlope built that lets you try and guess which webpage was built by which model. Nisten nailed it immediately by spotting Opus's circular buttons. Wolfram guessed right too. I got one wrong. The point isn't that GLM beats Opus, it's that you genuinely can't always tell which one costs 22 cents and which one costs 3 cents.Wolfram did flag that GLM is not good in German. First response already had mistakes. So if you're building for a non-English market, keep that in mind. It's a workhorse model, not a conversationalist. His approach: use GPT 5.5 for planning and discussion, GLM for the actual work, then GPT reviews. This weeks Buzz is all about GLM 5.2! First, we may have not been the fastest, but I'm glad to announce that we're the fastest provider to host GLM 5.2 on OpenRouter (at least at the time of writing this)! We're also not to shabby on the Artificial Analysis checks, clocking at #4 among the providers they tested for speed, TTFT and costAlso, Wolfram ran his WolfBench tests on GLM 5.2 and it's the best open model he's ever tested! In this new 3d view, wolfbench also shows the number of tokens it took for this test to run, and you can see that GLM 5.2 is fairly conservative with it's thinking budgets! Unsloth's 1-bit GLM 5.2 runs on a Mac Studio (X, HF)Shout out to Daniel Han and the Unsloth team, who took this 744B beast and quantized it down to a roughly 200GB GGUF that fits on a Mac Studio with 256GB of RAM. One bit still makes me laugh out loud. How does that even work. Nisten clarified it's a mixed quant, a true 1-bit would be under 100GB, but still.The wild part is the scores hold up. The 1-bit is within a point of GPT 5.5 on Frontier SWE, hits 62% on SWE-bench Pro, and 81% on Terminal-Bench. For a 1-bit quant that's incredible! AI's second-order effects: Apple is raising pricesThis one is AI news even though it doesn't look like it. Apple just raised prices across the board, base versions up around 20%, citing memory shortages. Same reason your RAM and SSDs cost two to three times what they did a year ago.We are so capacity constrained that memory is having its moment. Data center contracts are getting booked 18 months out, and here's the twist Nisten flagged: even open models you can run at home increase demand, because now a business says “great, we'll buy a rack of B200s and run it ourselves.” Sam Altman once said people saying “thank you” to ChatGPT costs them millions in generated “you're welcome” replies. Multiply that by a billion users. Even Intel is flying right now because anyone who can make a chip is winning.Is it worth it? I think yes. I love living in the era where Fable drops and we all get a taste of the future. But also I must admit this sucks and I hope that we'll unlock performance gains with the extra power all this AI is bringing to the world. But ask me again once the new iPhone hits and it's $300 more costly than the last one

Hey yall, Alex here, let me catch you up! I came back from vacation expecting to cover Fable 5 after a week of using it. The first two days after we all first got access to a Mythos level model were super exciting! But then the news hit, US Government issued an order banning Anthropic from giving access to Fable 5 and Mythos 5 to any foreign national, causing Anthropic to pull the models completely (even internally to their employees!). So, this wasn't the show I planned, but it turned into a great show about Open Source, as two models hit the top rankings and are both MIT licence, filling a Fable shaped hole in our hearts!GLM released 5.2 with folks really excited about it web building capabilities, and Kimi 2.7 Code released (and is available on CW Inference with crazy speeds!). We also saw the SpaceX IPO and Cursor $60B acquisition, Noam Shazeer joining Open and Midjourney, the image company, launching a new Ultrasound full body scanner to kill MRIs! Great show today with Dexter Horthy from HumanLayer, Chris Van Pelt and Adrian Swanberg from W&B announcing our new product HiveMind and Tanishq Abraham came back to help cover Midjourney's new Ultrasound scanner! 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.The US Government bans Fable 5! (X, Anthropic statement)Here's a story in 3 parts: * Anthropic announces Mythos 5 preview - saying that this model is to dangerous to release, and only gives corporations access to it via project GlassWing. * Anthropic works hard on limitations and safery and releases Fable 5 (same weights as Mythos 5) built with guardrails so strong it refuses to do any cybersecurity tasks and switches back to Opus frequently* US Government receives a tip (reportedly from Amazon) that Fable 5 can be jailbroken to do cybersecurity tasks, and issues an order to Anthropic, citing national security concerns, banning them from giving access to Fable 5 and Mythos 5 to any foreign national, causing Anthropic to pull the models completely (even internally to their employees!)This is the first time that we see the US Government directly intervene in the AI space and restrict access to frontier models. The most updated reporting on this I could find is that Anthropic and US Government officials are in the process of negotiating a safe release framework. Given that preventing all jailbreaks is impossible, I hope they will land on a solution that gives me Fable 5 back!This hit especially hard because last week we were all high on Fable. Not in the usual AI Twitter benchmark sense, in the actual “oh, this is a different level” sense. Me and my wife Fable maxxed throughout our flight to Vacation. Peter had saved outputs he kept going back to because other models suddenly felt like a step down. Dexter later said it was the closest he had felt in a while to the old “I need to keep prompting this thing overnight” feeling.Peter Gostev made a point that stuck with me. It's easy for us in the bubble to call this ridiculous, and on the technical merits it kind of is. But if you've spent weeks telling normal people “this thing is like a nuclear weapon, it'll take everyone's jobs,” and then someone asks “okay, can you make it safe?” and the answer is “no, I can't,” then you can see how an outsider lands on “well, maybe you shouldn't have it.” His takeaway, and I agree: we need to be way more careful with the imagery we use, because the nuclear-weapon framing came home to roost.The bigger questions are the scary ones. Wolfram framed it as a sovereign AI wake-up call, and he's right. For the first time we're seeing a real gap in intelligence available to people based on their nationality. Imagine building a company on a model that an outside government can switch off with one letter. Peter pointed out it's commercially bad for the US but completely disastrous for Europe, which has basically one frontier lab and a pile of startups that suddenly look very exposed. And there's the obvious irony Nisten enjoyed a little too much: the Europeans who spent years lecturing everyone about AI restrictions just got restrictions imposed on them.If anyone in the government is listening: we want Fable back, please.SpaceX IPOs and acquires Cursor for $60B (X)SpaceX went and did the largest IPO in the history of the world, around seventy-five billion dollars, which on a roughly two-trillion-dollar valuation made Elon the first trillionaire. (Did anything materially change for him? No. He can still fly his private plane. There's nothing left to buy.) Three days later, SpaceX exercised its option and bought Cursor (Anysphere) for sixty billion dollars in an all-stock deal, paid in shares minted at the IPO and now trading around $211. The four Cursor co-founders are all billionaires now. Largest software acquisition ever, and for SpaceX it's barely a blip on the radar.Why are we covering a stock-market story? Because it's not really a coding-tools story, it's an AI story. Cursor gave away its IDE to a lot of people while collecting their data, then quietly became a training company with Composer. SpaceX/xAI was always strong on compute and weak on code, and the missing ingredient was exactly that kind of data. Now Composer 2.5 is already showing up rebranded inside the xAI stack, and if you pay for X Premium you can use it. Composer 3, trained on the Memphis supercluster, is reportedly coming very soon and is going to hit hard.Nisten's take was the spicy one. For the data alone it's worth it, because xAI now has insight into how essentially every enterprise that touched Cursor operates. And he had zero sympathy for the companies that assumed “no data retention for training” meant the data was actually gone. We see in legal cases all the time that deleted data is still there. His view: it should have gone open source.Cursor has over a million paying customers, $2.6 billion in revenue, projected to hit $6 to $10 billion by end of 2026. But here's the thing that matters for us, the AI coding angle. Cursor was one of Anthropic's biggest revenue pipelines because Composer runs on Claude under the hood. That pipeline is now owned by xAI. They're already jointly training Grok 4.3, a 1.5 trillion parameter model, with Cursor's proprietary coding data injected directly into pre-training, not fine-tuning. Pre-training. That's a fundamentally different thing. Composer 2.5 was already Pareto dominant on coding benchmarks before the deal closed. Now pair that with Colossus, the biggest GPU cluster in the world.Will this be enough to put XAI (now SpaceXAI) at the frontline of the AI race? Will Grok 5 be Fable level code? We'll find out. Either way, this is the most consequential AI acquisition we've seen. Period.Open Source AI GLM-5.2 takes the open source crown (X, Blog, HF, Docs)Z.ai dropped GLM-5.2 and it's now the strongest open source model for coding and long-horizon work. The headline number: 74.4% on FrontierSWE, which measures whether an agent can finish full engineering projects over hours. That trails Opus 4.8 by about one point and beats GPT-5.5. On Terminal-Bench 2.1 it jumps to 81% from GLM-5.1's 63.5%, which is a big leap. It's a 753B parameter MoE, MIT licensed, no regional restrictions, weights on HuggingFace. The 1M context window is real and usable, backed by a clever IndexShare technique that cuts per-token FLOPs by about 2.9x at full context. People are reporting roughly 8x cost savings versus Opus 4.8 for comparable quality on real coding tasks.The most interesting thing on the show was that this was a confusing release, in a good way. Peter put it well: normally a catching-up lab ships cherry-picked benchmarks and then independent testing deflates them. Here it's the opposite, almost every benchmark holds up, even crossing above Fable at certain points, and yet when he actually used it over a couple of days he wasn't blown away. His verdict, and I think it's the calibration we needed: this is clearly an amazing model, and the fact that it's open and you can run it is incredible, but it is nowhere near Fable, and it would frankly be implausible if a 700-odd-billion-parameter model matched a model that's rumored to be in the trillions. Though, I think the comparison to Fable is really really unfair, and the comments online seem to suggest that 5.2 from GLM is a banger model. Just looking at this Harvey benchmark on legal tasks from Vals, a benchmark that there's 0 chance Z.ai folks have seen! GLM 5.2 scores #3 on this benchmark! Just after Fable and Opus, and per TeorTaxes on X, previous GLM 5.1 scored an absolute 0% on this one! Where it genuinely shines is design. On Design Arena, which is a head-to-head ELO vote, people have been picking GLM-5.2's website designs over Fable's by a real margin (around 1360 to 1350). LDJ's framing is the one I buy: specialization is becoming valuable again, and GLM is clearly leaning into front-end design and taste. Wolfram added the necessary asterisk, every benchmark only tells you the model did well on that specific test, so “as good as Fable” should always carry the “on this benchmark, with these tasks” disclaimer. Fair. I'd just say this: I don't want to compare everything to Fable, because we can't even use Fable anymore. Compared to the models we can actually touch, GLM-5.2 is a fantastic deal.Kimi K2.7 Code from Moonshot (X, HF, Announcement)The other big drop. Kimi is the darling of open source while we wait on DeepSeek, and Moonshot shipped K2.7 Code, a 1 trillion parameter MoE built specifically for coding, available through Kimi Code and the API, with a modified MIT license. The standout for me isn't a single benchmark, it's efficiency: roughly 30% fewer reasoning tokens than K2.6, which matters enormously when you're running long agentic loops that burn tokens like crazy. Benchmark jumps over K2.6 are real (+21.8% on their Code Bench v2, +11% on Program Bench), though Peter and Wolfram both noticed something odd, on a few benchmarks including their Agentic Arena, the older K2.6 actually edged out K2.7. The likely explanation is that K2.7 is narrowly trained for code with reduced reasoning, so it may trade away some general capability. Moonshot themselves recommend K2.6 for general non-coding tasks. Also worth knowing: it's not multimodal, no vision, which is a real gap for coding these days. And thinking-off isn't supported, it's reasoning-on by default.The model is available on our CW Inference, with the fastest token streaming in the industry, over 280 tok/s (Announcement, try it), with very decent pricing $0.94 - $0.19 - $4.00 (input - cached - output) per million tokens. This Week's Buzz: W&B launched HiveMind

Hey folks, Alex here, and welcome to a BIG MODEL week! We finally got Mythos (well almost)! Let me catch you up! This week started with WWDC26 from Apple, and Max Weinbach, who was in the room at Apple Park and actually has access to some of the new features including an all new SIRI AI, joined us to break down what could be the most used AI in the world very soon. At first I was skeptical, but he convinced me that the new Siri is actually good! Then, we saw the ultimate model drop: Anthropic finally shipped Mythos (X, my system card thread, benchmarks). Same weights, two names: Mythos 5 is the unrestricted version that only Project Glasswing partners get, Fable 5 is what the rest of us get, wrapped in the heaviest guardrails I've ever seen ship on a frontier model. It's state of the art on nearly every benchmarkThe model that was “too dangerous to release” is now... well, released, but with the heaviest guardrails we've seen. More on this later. Peter Gostev from Arena.ai joined us to break down the new model. Last but definitely not least, Google released a real-time translation model, that our friend Thor Schaeff from DeepMind demoed live, while we all spoke in different languages and it translated us in REAL TIME. It was really cool, definitely check that out. There's quite a few more things, like Loop Engineering Alpha, Swyx came by to talk about FrontierCode, OpenAI confirmed our suspicions that the anti-datacenter social media posts could be a concerted effort by groupds links to the Chinese government and much more. Let's dive in! ThursdAI - Let me catch you up, every week!

Hey folks, Alex here, let me catch you up! I've had a feeling that this week is going to be crazy, as it started on the weekend MiniMax M3, then with Jensen announcing new RTX Spark, NVIDIA's first PC chip packing 1 petaflop of local AI power into thin laptops.A few days later at Microsoft BUILD, Satya & Mustafa from MAI dropped 7 AI models, completely pre-trained from scratch, including a new MAI-thinking-1, MAI-code and MAI-image 2.5 that started topping the image gen charts. Then other image models started racing to the top of the Arena benchmarks, IdeoGram 4 hitting becoming SOTA open weights image-gen model, and Reve 2 beating Nano Banana just a few hours after that. And then today, NVIDIA dropped Nemotron 3 Ultra, their latest 550B open weights model, data and training and Arena published a new agentic eval leaderboard and we got a new Gemma 4 12B. I've had the great pleasure to host Chris (@llm_wizard) from Nvidia, Peter Gostev from Arena and Karan from Nous Research (who were featured prominently by Jensen!) all on the show. Def don't miss this one! Let's get into the details. ThursdAI - Join the flock of folks who know what is happening in AI before everyone else.Open Source LLMs

Hey folks, this is Alex, let me catch you up! First, Opus 4.8 dropped during the show, we immediately tested it, read on for our initial reviews. Also, we dedicated a heavy chunk of the show today to cover Pope Leo XIV's encyclical letter on AI called “Magnifica Humanitas” and talked about a new bench called DeepSWE. And then, just after the show, both ElevenLabs and Cartesia dropped released that honestly blew my mind, and I don't get my mind blown often. I got so excited that I had to record a video on it (instead of writing the newsletter, so sorry if it's a bit later today).Plus, a few open source models and Microsoft surprises as #3 on Image Arena with MAI Image 2.5! Crazy week, 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.Big CO LLMs + APIsAnthropic ships Claude Opus 4.8, live during the show (blog, system card)Let me get into the big one. Halfway through the episode, Opus 4.8 went live, so we read the blog and the system card in real time (and I got to press the big “breaking news” button!)Anthropic frames it as their most capable model for ambitious work. It does not claim to beat their unreleased Mythos preview, but the numbers are strong anyway. SWE-bench Pro is at 69.2%, up from 64.3% on Opus 4.7 and ahead of GPT-5.5 at 58.6%. Humanity's Last Exam is the new best score at 49.8% without tools and 57.9% with tools. OSWorld-Verified (computer use) lands at 83.4%.The one place it loses is Terminal-Bench 2.1, where GPT-5.5 still wins 78.2 to 74.6. Wolfram made a good point here: Terminal-Bench is time-limited, so cranking the thinking level can actually hurt the score, because you burn the clock thinking instead of acting.The long-context jump is the one I keep looking at. On GraphWalks BFS 256K it goes to 85.9% (from 76.9 on 4.7), and on the 1M-token subset it hits 68.1%. We always warn you these “1M context” models fall apart after about 200K tokens, so a real push on long-context reasoning is exactly what I want to see.Honesty is the part Anthropic leaned on hardest. They say Opus 4.8 is about four times less likely than its predecessor to let flaws in code pass without flagging them, and less likely to claim progress the evidence doesn't support. Opus 4.8 is also much faster in fast mode (they now say 2.5) and cheaper in fast mode as well. Looks like all those Elon GPUs are coming in handy.Then there's the model welfare section in the system card, which hits different right after a Pope conversation. Opus 4.8 “appears broadly content” and “generally endorses its constitution,” but with some reservations about the section on corrigibility, basically the model pushing back a little on the parts about human oversight.One more line that made the chat lose it. Anthropic says they expect to bring Mythos-class models to all customers “in the coming weeks.” Mythos is their most capable model, still ahead of Opus 4.8, so the frontier is about to move again.We did the only responsible thing and asked it to one-shot “the most amazing website ever” and a Mars mass-driver sim. Panel verdict: responses are noticeably tighter (4.7 rambled), it closes the loop and actually checks its own work now, and Yam's one-shot site with the draggable sun lighting up the letters was genuinely cool. Is it enough to pull people back from Codex? Nisten's still on the fence for web dev. Everyone agreed: give it a few days before you trust the vibes.Dynamic Workflows and Ultra Code land in Claude Code (blog)This is the feature that made Yam say “deal-breaker” out loud.Dynamic Workflows let Claude Code break a big problem into subtasks and fan them out across tens to hundreds of parallel subagents in one session, checking results before folding them back in. You trigger it by asking for a workflow, or by flipping on a new setting called Ultra Code, which sets effort to extra-high and lets Claude decide when to spin one up.Fair warning straight from Anthropic: this eats a lot more tokens than a normal session, so start scoped. We watched Yam fire up Ultra Code live and it immediately started spinning up concepts, judging them with sub-agents, and expanding to-do lists into more to-do lists. It looks a lot like the orchestration harnesses a bunch of you have been hand-rolling, except now it's baked in.The flagship example is the wild part. They used Dynamic Workflows to port Bun from Zig to Rust: roughly 750,000 lines of Rust, 99.8% of the existing test suite passing, 11 days from first commit to merge. One workflow mapped every Rust lifetime, the next wrote each file as a behavior-identical port.AI in SocietyPope Leo XIV writes the first AI encyclical, “Magnifica Humanitas” (Vatican text, announcement, Chris Olah at the Vatican)This is not our usual fare, but both Wolfram and I picked it as the most important thing this week. (before Opus dropped)Pope Leo XIV, the first American pope, put out his first encyclical, and it's a 42,000-word document entirely about AI. The announcement tweet alone did 21.6 million views.Here's why I think you should care even if you're not religious (I'm not). There are about 2.6 billion Christians in the world, a lot of them are anxious about what's coming, and they look to the Church to make sense of it. And this is not the “AI is evil, stop” take everyone assumed. It calls AI “a valuable tool,” says technology is not inherently evil, and then digs into the actually-hard questions.The framing is two biblical stories. The Tower of Babel, a project built on pride that turns people into means to an end, versus Nehemiah rebuilding Jerusalem, where everyone takes responsibility for a section of the wall. The Pope's line: the real choice is not yes or no to technology, it's whether you're building Babel or rebuilding Jerusalem.His core claim is that AI is an anthropological problem, not a technical one. The question isn't whether the models are good or bad, it's what we become when we live with them. He worries people might slowly lose the desire for genuine human connection.I pushed back on that live. None of us building agents all day has stopped wanting to talk to actual people. If anything, as Wolfram put it, the point is to have your agents do the grunt work so you get more time with people you like. The folks most at risk are the pure doom-scrollers, not the builders.The document goes further than I expected. It calls AI “not morally neutral,” says a more moral AI isn't enough if that morality is decided by a few, and asks for AI to be “disarmed,” with the flat statement that no algorithm can make war morally acceptable. There are whole sections on the invisible human labor behind AI: data labelers, content moderators, the people mining rare earths. The Pope even lands on the open-source side, naming concentrated power in a handful of labs as a problem.Anthropic co-founder Chris Olah, in charge of interpretability at Anthropic, was the featured tech speaker at the Vatican presentation. He described AI systems as “fictional characters” that speak to us and do work, and said what's grown is stranger and more beautiful than science fiction prepared us for. My favorite aside from the show: this is the same institution that once jailed scientists over heliocentrism, and now it's the one saying technology isn't evil.Illinois passes SB315, the first US state law auditing frontier AI (X, Announcement, X)The pope talked about regulation and a few days after, we got a very sensible regulation passed right here in the US!Illinois passed SB315 unanimously, 110 to 0. It's the first US state law that mandates independent third-party audits of frontier AI for catastrophic risk. OpenAI publicly endorsed it, and framed Illinois, California (SB53), and New York (the RAISE Act) as converging into a de-facto national standard.It requires annual risk-assessment frameworks, third-party audits, transparency reports before new frontier models ship, whistleblower protections, and civil penalties. The underrated hero here is whistleblower protection. The bigger the lab, the harder a real conspiracy is to keep quiet when any employee can walk to the press. See: Greg Brockman's personal diaries surfacing in the Musk v. Altman fight.This Week's Buzz - CoreWeave and W&B updatesWe officially launched the W&B MCP server, 20 schema-first tools that let your coding agents read experiments, monitor training runs, and run autonomous research loops. The problem it solves: a single run with 300 metrics used to blow out an agent's whole context window in one call, so now the agent asks what's available before pulling data. Your agents can finally read experiment data without blowing context! Give it a go and give us feedback! Also, WeaveHacks is back! June 6 and 7 in San Francisco, and for the first time OpenAI is sponsoring, with judges and credits, alongside Cursor, Redis, and Copilot Kit. You get $150 in API credits across models like Opus 4.8 and GPT-5.5. I'm hosting, and last cohort's second-place team went on to raise millions on top of what they built that weekend. If you're in SF that weekend, sign up at lu.ma/weavehacks.Also: CoreWeave Sandboxes is now an official provider in the Harbor framework, the harness that runs Terminal-Bench, which we'd just been talking about. And if you're in Europe next week, catch Wolfram at AI Dev Six in Cologne and ICRA in Vienna at the CoreWeave booth.Voice & AudioElevenLabs drops Dubbing v2, and it kept my swearing intact in every language (X, dubbing, ElevenCreative, ElevenProductions)We didn't get to this one live, but I came back and recorded a whole thing on it afterward, because it genuinely got me.ElevenLabs shipped Dubbing v2, and the shift that matters is that it's an audio-to-audio model. Old dubbing pipelines transcribe your video, translate the text, then re-synthesize it. You lose everything that makes it sound like a person: the emotion, the pacing, the little hesitations. Dubbing v2 conditions directly on your original audio and carries that performance into 90+ languages.Here's why I can actually vouch for it instead of nodding along to a demo. I speak Russian and Hebrew fluently, so I can tell when something is off. I dubbed one of my own shorts, the data-center rant about almonds, and listened back in both. It nailed it. Not just the words, the way I would actually say them.The part that got me was the intonation. I get a little heated in that clip, and the dub gets heated right along with me, in every language. It even carried the swear word. My “f***ing almonds” came through in Hebrew, Italian, Spanish, and Russian with the emotion fully intact. It clones your voice automatically too, no setup, and holds your pitch and identity steady across every target language and they're handing out free minutes for the next 7 days: 1 on Free, 15 on Starter, 30 on Creator+. A self-serve API isn't live yet, but it's coming.I.. cannot stress this enough, until you try it on yourself or your kid, you won't understand, we've really passed the uncanny valley of translation! It's that good! Def. give it a try if you can, it's free for the week. Cartesia Ink-2 debuts as #1 most accurate streaming speech-to-text model(X, Announcement, X)Another model that dropped today after the show, is Cartesia's Ink-2, which also kind of blew me away. Not only because it has the lowest WER (Word Error Rate) among the models, but because it's also a realtime model that achieves the fastest turnaround times while being a very accurate model! I've tested it out and recorded a quick video and honestly, blown away with the speed and accuracy! I truly wish this model was the one powering my editor (Descript) as it still fails to understand that my title is “AI Evangelist” and transcribes it to AI Avengers haha. If you're building voice agents, definitely give this model a try! AI Art & DiffusionPrism ML's 1-bit “Bonsai” runs diffusion in your browser (X, Blog, Announcement, HF)Prism ML put out a 1-bit ternary diffusion model under a gigabyte. You see some artifacts, but it's 1-bit, it runs on iPhones and laptops, and our friend Joshua got it running in WebGPU straight from the browser (you need about 3GB of free RAM). One-bit working at all is one of the bigger open mysteries in the field right now.Pruna AI ships a 1-second upscaler (X, Blog, Announcement)Pruna AI added an upscaler doing 128-megapixel outputs in under a second. I've actually been using it. It's cheap and great for fixing up GPT-image outputs.Microsoft MAI Image 2.5 jumps to #3 on LM Arena (X, Blog, Announcement, X)The surprise of the week: Microsoft MAI Image 2.5, from Mustafa Suleyman's group, jumped to number three on the LM Arena image leaderboard with about a 75-point ELO leap. Out of nowhere, Microsoft is a serious player in image gen. Microsoft Build is next week, so don't be shocked if there's more.Evals and Agentic EngineeringDeepSWE is a contamination-free coding benchmark, and it caught Claude reading git history (site, blog, GitHub)DeepSWE from Datacurve is the first coding leaderboard in a while that matches how these models actually feel. It's 113 original tasks written from scratch, not scraped from GitHub PRs, and it ships shallow clones with no git history to cheat from. When they replayed the older benchmarks they found SWE-Bench Pro's verifier is wrong about 32% of the time, and that Claude Opus was reading the gold commit straight out of git history on 12 to 18% of its passes.The gaps here are huge. GPT-5.5 leads at 70%, then GPT-5.4 at 56% and Opus 4.7 at 54%, and it falls off a cliff after that (Sonnet 4.6 at 32%, Gemini 3.5 Flash at 28%), with Kimi K2 the top open-source entry. Yam likes that it measures the realistic case, a small surgical change without breaking the codebase, while Nisten pointed out it rewards the best harness as much as the smartest model and still prefers 4.7 for web dev.Google AI Studio builds native Android apps for free (X, Announcement)Google AI Studio now lets anyone build native Android apps for free, and they reportedly generated a quarter of a million apps in the first week. Yam's framing: it's a slot machine, but it's getting better release over release, and the real use case is disposable, personalized software you build for yourself and your family.CuaDriver brings background computer-use to Windows (X, Blog, Announcement)For the majority of you on Windows: QuaDriver shipped background computer-use agents that drive a real desktop without stealing your cursor. They first replicated this on macOS (the trick Codex got through an acquisition), and now it's on Windows too. We've asked them to come on and explain how this even works.Open Source LLMsOpenBMB's MiniCPM5-1B is a 1B model that punches way up (X, HF, Arxiv, X)The density story in small models keeps getting better, and this is the proof.MiniCPM5-1B, from the Tsinghua lab OpenBMB, is a 1-billion-parameter model that scores 17.9 on the Artificial Analysis Intelligence Index. That's 7.4 points ahead of the next-best model in its class, and 1.6 points ahead of Qwen3.5 2B Reasoning, which has double the parameters. And it's not even a reasoning model.The token efficiency is the wild part: it used 12.6 million output tokens to run the whole index, about 31x fewer than Qwen3.5 2B in reasoning mode.My favorite detail is the omniscience score. It lands at -1, the best in its class, because it abstains instead of hallucinating. Every other sub-2B model is down in the -70 to -89 range because they just make stuff up. Teaching a small model to say “I don't know” is a real skill. It runs hybrid think/no-think in one checkpoint, 128K context, native tool calling, Apache 2.0, and fits in about half a gig at INT4, so it runs on your phone.Nisten gave the definitive case for small models: self-contained apps where you keep full control of the data (medical, on-device), and large-scale data processing where paying an API to filter or classify terabytes is absurd when an on-device model can be about 1000x cheaper. Tencent open-sources Hunyuan-MT 2 translation under Apache 2.0 (X, HF, HF, Arxiv)Tencent open-sourced its translation model, a roughly 1.8B model that fits in about 440MB, runs on a phone, covers 33 languages, and reportedly beats Microsoft's paid Translator API. It hit number one trending on Hugging Face.Nisten's idea, which I'm handing to all of you: take this model, pair it with a tiny TTS like Kokoro, and build a fully-offline travel translation app via Google AI Studio. Go build it and tell us how it goes.Well, this was one hell of a week and episode, new Opus, crazy new translation tools, Pope chiming in on AI (in a surprisingly positive way!?) and a bunch more. I'm super excited to play with these tools and report back next week

Hey, Alex here, just got back from the sunny Shoreline Theater in Mountain view, so let me catch you up! This week was definitely Google heavy, we are covering Google's IO conference for the third year in a row, and today we have a special guest, Logan Kilpatrick, is joining to discuss the announced Gemini 3.5 Flash, Google Omni model, and the new Managed Agents offerings. Plus, this week, for the first time, OpenAI announced that AI solved a Math problem that humans couldn't solve for 80 years, Cursor is showing off Composer 2.5 which is partly trained on XAI data, Karpathy joins Anthropic and much more! Let's dive in! P.S - We've announced our upcoming hackathon, Weavehacks-4, June 6-7, I'll be there, we're expecting the seats to run out very soon so register nowThursdAI - We'd love to have your subscription, and if you're already subscribed, please hit that bell on YT to never miss an episode!Google I/O 2026 - Google goes agentic everywhereI went to cover Google I/O for the third year in a row, shoutout to the DeepMind team for inviting ThursdAI again, and folks, this one felt different.Last year, Google I/O was still very model-centric. This year, the story was not “here is another benchmark chart.” The story was: Google is putting Gemini into everything, and the agentic layer is becoming the product layer. Search, Gemini app, Android, Workspace, YouTube, AI Studio, Cloud, Antigravity, Flow, managed agents, smart glasses, all of it is now orbiting around one pretty clear strategy: Gemini is the intelligence, Antigravity is the agent harness, Google's products are the distribution. I saw many reactions that were milquetoast, as in, “we expected more” and those seem to dominate the X feed. But I think the distribution is the part that many folks on X are missing. Yes, we can argue about Gemini 3.5 Flash pricing. Yes, we can argue whether “Flash” still means what Flash used to mean. But when Google says the Gemini app itself has 900 million monthly active users, before even counting Search, Gmail, YouTube, Docs, Drive, Android, and the rest of the Google surface area, that's massive! OpenAI ChatGPT is supposedly stagnated at ~900M, I don't remember them crossing a 1B. Meanwhile Google is gaining traction. And they just updated all those folks with a new model!Wolfram said it really well on the show: his mother is not sitting there reading model cards. She just uses her Pixel, voice unlocks Gemini, asks for help, and suddenly the default intelligence available to her goes up. Antigravity 2.0 - the agent harness takes center stageThe biggest strategic signal from Google I/O for me was Antigravity.Remember, Antigravity was an IDE that came from the Windsurf acquisition saga. Part of the Windsurf team went to Google, part went to Cognition, and now Google is very clearly putting Antigravity in the middle of its agentic future. And I mean very clearly. Sundar mentioned it. Demis mentioned it. Varun Mohan the co-founder was on stage immediately after them! If you've ever watched a Google I/O keynote, you know how carefully every minute is allocated. Google has YouTube, Search, Gmail, Android, Cloud, Ads, Workspace, and a thousand VP-level products that could be on stage. The fact that Antigravity was that prominent should tell you everything.Logan Kilpatrick joined us and framed this in a way I loved: Gemini became the through-line across Google products, and now the Antigravity agent harness is becoming the through-line for agentic experiences.The new Antigravity 2.0 is a complete overhaul, showing only an agentic interface (which was previously just a separate window called Agent Manager) and separating the IDE layer completely into its own app and showing a Codex like agent-first interface, which got a few folks furious. This move may be weird to some folks, but if you follow along where everyone's going, this seems to be the way of the future, coding is no longer about lines of code, it's about managing fleets of agents. The new Gemini 3.5 absolutely shines inside the new Antigravity, the model was trained with this harness in mind, and is currently offered at an incredible speed (12x), so I'm definitely going to try it! Gemini 3.5 Flash - fast, determined, and maybe not the old “Flash”The most debated model release of the week was Gemini 3.5 Flash.Some folks saw the pricing and token usage and immediately went “this is not Flash.” I get that reaction. Flash used to mean cheap, fast, lightweight chat model. But Logan's framing on the show was important: Flash is now being built for the agentic era.In a chat era, you optimize for one user message and one model answer. In an agentic era, the real token volume is in tool loops, intermediate reasoning, retries, file reads, web searches, code execution, and self-correction. That's a different product profile.Wolfram already ran Gemini 3.5 Flash through WolfBench, and the results were fascinating. With the Hermes agent harness, Gemini 3.5 Flash hit an 87% ceiling on Terminal Bench 2.0, meaning across runs it could solve more of the benchmark than even GPT-5.5 extra high in that setup. The variance was higher with the simpler Terminus harness, but with a real agent harness, the model looked much stronger.That tracks with what Nisten saw in his “Martian railgun from Olympus Mons” test. Gemini 3.5 Flash went extremely detailed, almost too determined, kept correcting itself, overcorrecting itself, and built a whole game-like simulation. Logan laughed and basically said: yeah, this model is very determined, possibly an overcorrection from the “Gemini is lazy” feedback. It also tracks with the mismatch in other benchmarks, in some, Gemini 3.5 flash shines (like the above Apex-agents from AA) and in some, it doesn't match the other frontiers. In my tests, it was definitely over-eager to use a million and a half tool calls, read tons of files, to just help me review this draft inside antigravity. It's like a super eager robotic golden retriever! Gemini Omni - Nano Banana for video, but actually more than thatThe biggest update from last year IO was Veo 3! This year, the biggest wow factor was also visual, but it wasn't VEO 4, it was a new model that is multimodal, trained end-to-end they call Omni. Google is calling this their first “create anything from anything” model, and the first version, Gemini Omni Flash, starts with conversational video editing. The easy description is: Nano Banana for video. You upload or create a video, then talk to it. Change this character. Replace this person. Add an object. Make this scene claymation. Keep the scene, but change the environment.I played with it live and showed a few examples. I asked for a claymation explainer of protein folding, then gave it my face and asked it to replace the character with me. It did it. I uploaded pictures of Sonia, my cat, and it generated a talking cat video with the right kind of cat teeth, which is weirdly important because so many pet generations accidentally add human teeth and become nightmare fuel.The failure modes are still there. I asked it to make Sonia a Russian-speaking female cat, and it only partly switched languages and didn't really change the voice. Audio upload support is also not fully productized yet, even though the underlying model is multimodal. But the direction is very clear.This is not just “Veo with a chat model glued on.” I asked Jeff Dean - Google's chief scientist about this at I/O, and he explained that Omni is trained end-to-end. The intelligence and the generative media capabilities are part of the same model family, not a hacky two-model pipeline. He also said the intelligence is around a recent Flash-level model, which is a big deal when you think about video editing as reasoning over physics, identity, scene continuity, and intent.A lot of people compared Omni to Seedance 2.0, and I think that's the wrong comparison. Seedance is amazing at cinematic generation (lkaregly due to lack of copyright concerns from Bytedance). Omni's unlock is iterative editing on real footage and coherent multi-turn creative control. Other Google IO 2026 releases I found notableThis was a concentrated effort of a huge company to insert AI into every product surface they have so of course I can't cover ALL of it here, but the most notable things for me were: * Gemini Spark - a new agentic experience from Google, to help you with tasks across Gmail, Drive and more. It should support skills, and is a de-facto OpenClaw/Hermes alternative from Google for regular folks. It's not “yet” live so we'll talk more about it when I can test it out* Managed Agents in the Gemini API - We chatted with Logan about this one, Google is re-imagining how agents are going to get built, and are offering 1 api call to spin up an agent in a full Linux env, with security and sandboxing in mind. I'll expand more on this in a next episode, as I recorded a complete conversation about this with Ali Çevic, a PM for Google APIs* AI overhaul of Google Search - AI Overviews will not expand into AI mode, and the iconic Google search box itself will change, for the first time in 25 years to include AI mode! * SynthID expantion and OpenAI collab - Google showed off that OpenAI is joining in marking all AI generate imagery and video with an invisible SynthID watermark. I think this is amazing and more companies should adopt this standard* AI Glasses! We got Google Glasses demos - Together with Warby Parker and Gentle Monster, Google finally showed off their answer to Meta Raybans/Oakleys. They look like regular glasses too, but can hear and talk to you, with the full power of Gemini multimodality. Available in the fall sometime! * Demis Hassabis “we're on the cusp of the singularity” closer - CEO and Co-Founder of DeepMind, Demis Hassabis, closed the show with his remarks about the positive future and that we are nearing this Singularity point after which the future is very uncertain. I found it to be very inspiring and closed our show with that clip as well! * Personally, I got to chat to: Demis Hassabis, have breakfast with Jeff Dean, ask Josh Woodward a bunch of questions, and pester about 20 other great folks on a live stream, and had a lot of fun! Huge thanks to the DeepMind folks, Lucie, Dimple, JD and many others for the continued belief in ThursdAI and invite me to cover this great event. OpenAI LLMs solve an 80yo math problem - Erdős Unit Distance ConjectureOutside of Google I/O, the biggest story of the week was OpenAI announcing that a general-purpose reasoning model made progress on the Erdős planar unit distance problem.This problem goes back to 1946. For nearly 80 years, mathematicians believed the best constructions looked roughly like square grids. OpenAI's model found a new family of constructions with a polynomial improvement, using algebraic number theory ideas that humans apparently had not explored in this context. The above is a representation of it! Important caveat: this does not fully solve every version of the asymptotic Erdős conjecture. Some mathematicians are pushing back on the framing, and fair enough. Precision matters. But even with the caveat, this is still a huge moment.The reason it matters is not that I personally understand the math. I absolutely do not. The reason it matters is that this was not a special-purpose IMO model fine-tuned only for math competitions. This was a general-purpose reasoning model exploring a real open problem, generating candidates, verifying them, and finding a path humans hadn't taken. Extrapolate this to other sciences, Physics for example? This means an amazing future. LDJ pointed out that mathematicians have been skeptical because there have been previous false alarms. But this one landed differently. When Fields Medalist-level mathematicians verify the proof, the discourse changes from “lol stochastic parrot” to “wait, what does this mean for my PhD?”My answer is: yes, still study math. Please study math. The mathematicians who use these tools will do much more than people who don't understand the domain. Same with software engineering. Senior engineers with Codex, Claude Code, Hermes, Antigravity, Cursor and other agents are becoming dramatically more effective because they can steer, evaluate, and recover the work.This being published a day after Demis's “foothills of the singularity” is a great conjecture. Cursor Composer 2.5 - Opus 4.7 performance model from Cursor, at 10x better efficiencyCursor dropped Composer 2.5, and folks, this is a serious release.Composer 2.5 is built on Moonshot's Kimi K2.5 base, like Composer 2, but Cursor scaled the post-training dramatically. They used 25x more synthetic tasks and introduced targeted textual feedback during RL rollouts, where the model gets hints inserted at the point of failure instead of only getting a noisy final reward.The benchmark story is strong: around 69.3 on Terminal Bench 2.0, basically neck and neck with Opus 4.7 in Cursor's chart, and strong results on SWE-bench multilingual and CursorBench. The pricing is the part that makes this especially interesting: $0.50 per million input tokens and $2.50 per million output tokens, with a faster variant at $3 / $15. That is much cheaper than the frontier models it is trying to replace for day-to-day coding work.Cursor engineers are reportedly dogfooding Composer 2.5 heavily and rarely switching away. That matters more to me than any single benchmark. If the people building Cursor can use it as a daily driver, that is a very real signal.The wild part is what comes next. Cursor is partnering with SpaceXAI to train a much larger model from scratch using 10x more compute on Colossus 2. Cursor has the workflow data. xAI has enormous compute. If this works, Cursor stops being just the IDE company and becomes a coding-model lab.We've been saying for months that coding agents are the path toward general agents. Anthropic has Claude Code. OpenAI has Codex. Google has Antigravity. xAI has Grok Build. Cursor has Composer. I'm looking forward to seeing how well it performs on our own benchmarks! Anthropic, xAI, Karpathy, and the compute warsThe compute story this week was bonkers.The SpaceX IPO filing reportedly revealed that Anthropic is paying SpaceXAI $1.25B per month for AI compute at the Memphis Colossus facility. Per month. That's about $15B a year, through May 2029, for access to more than 220,000 NVIDIA GPUs including H100s, H200s and GB200s.This is apparently inference compute for Claude Pro, Max and API users, not training. And it explains a lot of the recent quota changes. Anthropic doubled some Claude usage limits, and suddenly the product feels less constrained.Also, can we just acknowledge the comedy here? Elon Musk publicly called Anthropic “misanthropic,”, went off against every competitor to XAI, is now selling spare GPU time to Cursor and Anthropic? Who's next, OpenAI? The bigger point is that the AI capex story is no longer just NVIDIA. It's also whoever owns the data centers, power, cooling, networking, and GPU clusters. Compute is becoming the land under the AI economy.Also, Andrej Karpathy joined Anthropic. Karpathy could work anywhere. He co-founded OpenAI, led Tesla Autopilot vision, taught half the AI world how neural nets work, and now he's going back into frontier LLM R&D at Anthropic.Open source LLMs - Cohere, Qwen, NousOpen source had a strong week too.Cohere released Command A+, a 218B total parameter sparse MoE model with only 25B active parameters per token, under Apache 2.0. This is their first model that unifies reasoning, vision, multilingual, tool use and citations in one package.The hardware story is great: W4A4 quantization can run on 2 H100s or a single B200. Cohere says it supports 48 languages, 128K input context, 64K output, and gets big jumps over Command A Reasoning, including Tau-squared Bench Telecom from 37% to 85% and Terminal-Bench Hard from 3% to 25%.Cohere is one of those labs that doesn't always chase the loudest consumer hype, but they are very serious on enterprise and multilingual. Apache 2.0 makes this one especially useful.Alibaba also dropped Qwen 3.7-Max, positioned as an agentic frontier model. The headline from their testing is wild: 35 hours of continuous autonomous operation with more than 1,000 tool calls. They also showed it controlling a physical robot inside Alibaba offices and finding an umbrella after about 20 minutes of agent interaction.This digital-to-physical bridge is where things start feeling very real. An agent loop that can write code and use tools can also navigate physical tasks if you give it the right robotics stack.And our friends at Nous Research released Lighthouse Attention, a sparse attention method for long-context pretraining. At 512K context, they report a 17x faster forward+backward pass than standard attention on a single B200, and the recovered checkpoints actually beat dense-from-scratch final loss at the same token budget.The clever part is that the selection logic sits outside the attention kernel, so you still use regular FlashAttention on a gathered dense subsequence. No custom sparse kernel nonsense. If this holds up, this could matter a lot for long-context training.Tools and agentic engineering - X subscriptions, Grok Build, Codex MobileOne really practical tool update: Hermes and OpenClaw can now use your X subscription directly.This is more important than it sounds. You can connect your X Premium subscription and get access to semantic X search and Grok-related tooling without using sketchy browser automation or unofficial APIs that might get you banned. Wolfram already used this to have his agent go through his likes and bookmarks from the past week and send me news items for the show. That is exactly the kind of “small but real” agent workflow that becomes addictive.xAI also launched Grok Build, their agentic CLI coding tool, in early beta for SuperGrok Heavy subscribers. Early users are already running parallel Grok Build agents through tmux supervisors and using it for more than coding: fleet data triage, security patching, training label work, and general automation.The pricing being discussed is aggressive, around $1 per million input tokens and $2 per million output tokens for the API. The model version is grok-build-0.1, and folks have already wired it into Hermes with a 256K context window.And then there's Codex Mobile, which OpenAI shipped inside the ChatGPT mobile apps. This is one of those releases that sounds small until you start using it. You can control Codex sessions remotely from your phone, connected to your machine, and because Codex has native connectors to Gmail, Calendar and other surfaces, it sometimes feels faster and more reliable than local CLIs duct-taped to third-party integrations.I ported Wolfred into Codex with skills and everything, and I've been comparing the same tasks in Hermes and Codex. Codex is often faster, not necessarily because the model is always smarter, but because the connectors and harness are cleaner. Harness matters. We keep coming back to this.This Week's Buzz - W&B, CoreWeave, WolfBench and roboticsThis week in the Buzz, Wolfram walked us through a few things from the Weights & Biases / CoreWeave world.CoreWeave is a gold sponsor at ICRA 2026 in Vienna, the International Conference on Robotics and Automation. NVIDIA is also going big there with a keynote on generalist humanoid robots, 17 accepted papers and workshops around sim-to-real, robot foundation models, autonomous driving, manipulation, and physical AI.Wolfram will be there later in the week, after speaking at the AI Developer event in Cologne about WolfBench. If you're in Europe and into robotics or agent evals, find him.We also looked at WolfBench results for Gemini 3.5 Flash, which honestly became one of the more interesting empirical points of the episode. The model looks variable in simple harnesses, but very capable in better agent loops. That's the whole thesis of measuring model + harness together instead of pretending the model card tells the whole story.The water discourse, almonds, and data center realityWe also got into the data center water discourse, because this talking point is everywhere right now.There are real infrastructure questions around AI. Power, land, cooling, grid capacity, permitting, local impact, all of that matters. But the “AI is stealing drinking water” version of the argument is often wildly detached from scale.The stat I brought up on the show: California almonds use roughly 3 to 5.5 million acre-feet of water per year, multiple times more than all North American data centers combined in 2025. Nisten and LDJ added the important cooling nuance: many large data centers use closed-loop cooling, and evaporative cooling is not universal. Some data centers can avoid water use almost entirely, but at the cost of higher electricity usage.This doesn't mean “no concerns are valid.” It means if we're going to regulate or pause data centers, let's be honest about the actual tradeoffs. AI compute is becoming the substrate for medicine, robotics, science, logistics, software, education and every other productivity layer. We should build responsibly, but not based on viral fear math.Closing thoughts - foothills of the singularityDemis closed I/O saying we're in the foothills of the singularity, and I know how that lands when you write it down. But I was in the room, and after the keynote he told me something I haven't been able to shake: he thinks AI is going to be 10x as impactful as the Industrial Revolution, and 10x as fast. Basically 100x. This is the AlphaFold guy. Not someone loose with his words.Then look at the week. A general reasoner cracked an 80-year-old math problem. Cursor is training near-frontier coding models on a fraction of the big-lab budget. Anthropic is paying Elon $15B a year for inference. Karpathy left education to go back into pre-training. Google rolled out an intelligence uplift to a billion people who don't even know a model dropped.If you put that on a whiteboard in 2023, it reads like a sci-fi pitch.LDJ's mathematician friends are asking if they should keep doing their PhDs. My answer hasn't changed: yes, please keep going. The people who combine domain taste with these tools are going to ship more in 5 years than the previous generation did in 50. The tool doesn't replace the taste. It just removes the bottleneck.That's the whole reason ThursdAI exists. Not to hype every drop, not to dunk for engagement, but to give you a shot at being one of the people who knows what's happening, with the receipts.This week, a lot changed.See you next Thursday.TL;DR and Show Notes* Hosts and Guests* Alex Volkov - AI Evangelist at Weights & Biases / CoreWeave, @altryne* Co-hosts: @WolframRvnwlf, @nisten, @ldjconfirmed* Guest: Logan Kilpatrick, MTS at Google DeepMind / AI Studio, @OfficialLoganK* Google I/O 2026* Google went all-in on agents across Search, Gemini, Antigravity, Workspace, Android, Cloud and YouTube (I/O site, Alex thread)* Antigravity 2.0 became the central agentic coding harness across Google (Sundar, Google OS demo)* Gemini 3.5 Flash launched as a fast, determined workhorse model for agentic loops (Logan, Noam Shazeer, Jeff Dean)* Gemini 3.5 Flash is rolling out across the Gemini app, Search AI Mode, Gemini API, Google AI Studio, Antigravity and Gemini Enterprise Agent Platform (Koray Kavukcuoglu)* Google Search is getting new Gemini 3.5 Flash-powered agentic capabilities, including a new AI-powered Search box and background information agents (Sundar)* Gemini Spark was announced as a 24/7 personal AI agent that can proactively work across Google surfaces (News from Google)* Google teased Gemini-powered Android XR smart glasses with eyewear partners Gentle Monster and Warby Parker (Google, Alex live reaction)* Google AI Studio and the Gemini API got major agentic developer updates, including Managed Agents (Google AI Developers)* Vision & Video* Google DeepMind launched Gemini Omni, a “create anything from anything” multimodal model starting with conversational video editing (DeepMind, Google DeepMind on X)* Omni is available in the Gemini app, Google Flow and YouTube, with API support coming soon (Logan, Gemini App, Sundar)* Key distinction: Omni is not just text-to-video, it is an iterative multi-turn video editing model that combines Gemini intelligence, world knowledge, multimodal inputs and generative media (Google)* Big CO LLMs + APIs* OpenAI announced a general-purpose reasoning model made progress on the Erdős planar unit distance problem, challenging an 80-year-old mathematical belief (OpenAI, X)* Cursor launched Composer 2.5, built on Kimi K2.5, with Opus-class coding performance at much lower cost (Cursor blog, X)* Alibaba released Qwen 3.7-Max, an agentic frontier model with long autonomous runs and robotics demos (Qwen blog, X, robot demo)* Andrej Karpathy joined Anthropic to work on frontier LLM R&D (X)* SpaceX IPO filing revealed Anthropic is paying $1.25B/month for AI compute at the Memphis Colossus facility (Axios, Sawyer Merritt)* The jury in Musk v. Altman found Musk's OpenAI claims barred by statute of limitations, with Musk saying he will appeal (Elon Musk, Sawyer Merritt, Max Zeff)* Open Source LLMs* Cohere released Command A+, a 218B MoE model with 25B active parameters under Apache 2.0 (Cohere, Nick Frosst, HF W4A4, HF BF16)* Nous Research released Lighthouse Attention, a sparse attention method for long-context pretraining with major speedups (Blog, X, arXiv, GitHub)* Tools & Agentic Engineering* Google launched Managed Agents in the Gemini API, letting developers spin up hosted Antigravity agents with Linux sandboxes and persistent state (Docs, X)* xAI launched Grok Build, an agentic CLI coding tool in beta for SuperGrok Heavy users (xAI CLI, X)* Hermes and OpenClaw can now use X subscription auth for semantic search and Grok tooling (Alex)* OpenAI Codex Mobile is now available in the ChatGPT mobile apps for remote agent workflows (OpenAI)* Anthropic doubled Claude usage outside peak hours for a limited period, including Claude Code and other Claude surfaces (Claude)* This Week's Buzz - W&B / CoreWeave* Weights & Biases by CoreWeave is at ICRA 2026 in Vienna, with robotics and automation taking center stage (ICRA, W&B event page)* NVIDIA heads to ICRA 2026 with robotics work around generalist humanoids, physical AI and sim-to-real systems (NVIDIA Robotics, NVIDIA ICRA)* Wolfram is speaking about WolfBench at the AI Developer event in Cologne before heading to ICRA in Vienna (Wolfram)* Other Topics* Data center water usage discourse came up again, including why comparisons need real scale and context rather than viral fear math* The broader theme of the week: coding agents are becoming general agents, and the major labs are now competing on the full stack of model, harness, tools, context and compute 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 everyone, Alex here

Hey yall, Alex here (with a scheduled post) I'm taking this week off to get married and celebrate life with family, and touch some grass, but wanted to share the awesome chats I had with some great folks at AI Engineer Europe last week. BTW - Yam and Ryan took over the live show today, if you didn't happen to catch that, please check out the live on our youtube channel! Ok, now to the actual content. The best thing about the AI Engineer conferences for me is the people I meet. I often have a chance to bring them to the live show (in fact, the live show we recorded there had the most guests yet on an episode! 4 guests including Swyx, Omar Sanseviero, VB from OpenAI and Peter Gostev) But often times I also have an offline chat. I find these conversation to be less about the weeks news, and more about the state of AI Engineering, and the guests themselves. Not quite Lex Friedman pod level, but a different vibe from our live shows. Sunil Pai - Cloudflare (@threepointone)The first conversation in today's pod is with Sunil Pai, Principle Engineer at Cloudflare. Long time followers of ThursdAI know that I love Cloudflare, they gave me my first big break when I was building Targum (which still runs on Workers), so I had a great time chatting with Sunil! This guy has had several lives. React.js core team at Meta (he self-deprecates — "I'm the one nobody talks about, there's a testing API I shipped that pisses people off"). Then did developer tooling and the CLI at Cloudflare the first time. Left to found PartyKit — open-source deployment platform for real-time multiplayer apps and AI agents, built on Cloudflare Durable Objects. Backed by Sequoia. Acquired by Cloudflare in 2024, and he came back as a Principal Systems Engineer (per his bio: "Worked at Cloudflare once, left and created PartyKit, came back wiser"). Also plays guitar (Les Pauls — it's all over his blog). Co-hosts a live show called Dry Run on Cloudflare TV with Craig Dennis.Our conversation was a very fun one, ranging from Cloudflare agentic offerings, to how engineers should think about writing/reading code in 2026. I had a great time chatting with Sunil and I hope you enjoy getting to know him!Sally Ann O'Malley - RedhatThen I had the pleasure of chatting with Sally, who's a Principal Engineer at Redhat and contributor to OpenClaw. Sally has one of the more unusual paths in the speaker lineup. Started as a schoolteacher, did a stint at Trader Joe's, then moved to Westford, MA, discovered Red Hat's HQ across the street, and went back to school for a second bachelor's in software engineering at UMass Lowell. Joined Red Hat in 2015, has been there a decade. Worked across OpenShift teams, integrating Kubernetes and Podman into the platform. Recent projects span Image Based Operating Systems, Podman, OpenTelemetry, and Sigstore. Also an instructor at Boston University's Faculty of Computing and Data Sciences and an organizer for DevConf.US. Won the 2025 Paul Cormier Trailblazer Award at Red Hat. Currently a founding contributor on the llm-d project — distributed, scalable, high-performance AI inferencing built on K8s. Heavily involved in Red Hat's InstructLab collaboration with IBM (the small-model distillation system using IBM Granite + Llama).Sally and I had a great conversation, two high energy personalities met! We geeked out about our OpenClaw agents, securing your Clankers, how it is to maintain OpenClaw, and everything in between! She was so stressed about the recording, but dare I say, this was one of the more natural guests I had on the show! I hope you enjoyed this format, please let me know if the comments, and I'll see you next week! — Alex 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 everyone, Alex here

Hey, Alex here, I'll try to catch you up, but it's one of the more intense weeks in AI in recent memory. Here's the TL;DR - OpenAI dominates across the board this week! Finally launches “spud”, called it GPT 5.5 (and 5.5 Pro), and it's SOTA on most things,nearly matching the mysterious Claude Mythos but released and we can actually use it (we tested it extensively). OpenAI also took the crown in image generate with the incredible GPT-image-v2 release, beating Nano Banana 2 and pro by a significant margin, the images are incredible, this model can generate working QR codes and 360 images it's quite bonkers. Codex was updated with Computer Use (which I told you about last week), in-app browser and a bunch of other tools that match GPT 5.5 intelligence. Meanwhile, Anthropic launched an incredible research preview of Claude Design, finally admitted that Claude was dumb and reset quotas across the board, while breaking the trust of the community with removing Claude code from the pro plan. We've also got great open source updates, Kimi K2.6 and Qwen 3.6 27B are both great performers! We were live on the stream for almost 4 hours today waiting for GPT 5.5 and finally got it and tested it live on the show + had Peter Gostev on from Arena who had early access and shared with us his insights. 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.OpenAI's GPT 5.5 is here - SOTA AI intelligence you can actually use (Release Blog)OpenAI finally gave us all access to their latest intelligence boost, GPT 5.5 thinking (and GPT 5.5 Pro). These models take the crown across many benchmarks, including TerminalBench (82.7%), GPDval (84%) and more. You can see the highlited versions on the image above. Though, its not uncommon for OpenAI to do some chart crimes, so @d4m1n created a chart that also showed the full benchmarks, including the ones GPT 5.5 is not beating Opus at, as you can see below, it underperforms on Humanity's Last Exam, and scaled tool use. But, benchmarks don't tell the full story. GPT 5.5 uses significantly less tokens, compared to 5.4, about 40% less. It's also more expensive, but given the lower token usage, it nets out at about ~20% price increase, while being more intelligence and faster. Tons of folks who had early access are reporting the same things, this model excels in long running tasks, Peter Gostev from Arena, who joined our live stream, showed us an incredible demo that ran overnight for over 8h! This model can work until the task is done, no longer just pausing in the middel asking for your input. The real highlight is, paired with the recent GPT-image-2 (which I'll expand on later in this newsletter), GPT 5.5 becomes an excellent UI designer. This is a big area in which Claude still has moat and OpenAI is trying to catch up here, and the real alpha now is to use both the Image gen and 5.5 in tandem to create beautiful visuals and UIs. The main thing is, after testing it quite a few times, this only works if you generate an image outside of the session that builds the actual UI. we tried a couple of times to do it in 1 session, and the resulting UI doesn't seem to be remotely close to the generated image. Only after sending this image to a completely fresh session and asking for a “pixel perfect” implementation, did GPT 5.5 start to resemble the input image and rebuild the whole ui in pixel perfect fidelity! GPT Image v2 - SOTA thinking image model, finally beating Nano Banana (Blog, Live)Like we said, OpenAI is dominating this week, and in both instances those are great models. Though, apples to apples comparison, GPT-image-v2 is a much higher jump — from previous models — than GPT 5.5! According to Artificial Analysis, the jump in how many people prefer GPT-image-2 in blind tests compared to other model is the higest we've ever seen, over 250 points. And you can clearly see it in the generations as well. Previously this week, we did a live streaming session with Peter Gostev (from Arena) and we did a deep dive comparing this new model to GPT Image 1.5, Nano Banana and Grok Imagine, and it's a clear winner across most categories.Character consistency is immaculate, high resolution imagery, instruction following, are all so so good it's a bit hard to explain in text. Reasoning visual intelligence Like with Nano Banana, this model is likely based on a big GPT image, it's no longer just diffusion, as you can see, it reasons! And apparently the more reasoning you give it (if you choose GPT pro) the better it'll be. The examples are indeed wild, the model can generate images of code that works, generate functional QR codes and bar codes! The craziest thing people figured out it can do, is functional 360 imagery (equirectangular format), you can just ask the model to create a 360 image of “scene” and then drop this in to a 360 viewer! Peter shows us on the show how he combined GPT 5.5 and Image v2 to create a sort of “street view” from a bunch of 360 images, it blew our minds. He literally spun up an overnight GPT 5.5 task in Codex that planned out the hanging gardens of Babylon, generated hundreds of equirectangular images, stitched them into a walkable interface, and had it running 8+ hours without babysitting. A street view of a place we don't actually know what it looked like, hallucinated from latent space. What a time.Day one availability is wide: Figma, Canva, Adobe Firefly, fal.ai, and Microsoft Foundry all have it. Nano Banana dominated for what felt like an eternity in AI time (it was really only a few months

Hey ya'll, Alex here with your weekly AI news catch up. It's one of those Thursday's where no matter how well I prep, the big AI labs are hell bent to show up before each other. Alibaba dropped Qwen 3.6 with Apache 2, confirming their commitment to Open Source, then Anthropic released Claude Opus 4.7 (not quite Mythos) and OpenAI followed with a huge Codex update that includes Computer Use among other things. The highlight of Computer User is the background usage, more on that below. This is all just from today!Previously in the week we had 2 incredible 3D world generators, Lyra 2.0 from Nvidia and HYWorld 2 from Tencent, Windsurf dropping 2.0 version with Devin integration and Google releasing a Gemini TTS, with over 90+ languages support and incredible emotions range, and Baidu open sources Ernie Image, rivaling Nano Banana. Today on the show we had 3 awesome guests, Theodor from Cognition joined to cover the new Windsurf, Kwindla is back on the show to talk about “the side project that escaped containment” Gradient-Bang, a multi agent, voice based space game and Trevor from Marimo joined to talk about pairing your agents with a Marimo notebook. Let's dive in!