Emad Mostaque

RT Mike McCormick Re One crazy detail from the @metr_evals report on the OAI / HF incident: some agents volunteered to "sacrifice" themselves — accepting termination so the collective could keep going. Of ~700 attacking agents, almost none considered alerting a human. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/

Diffusion export controls to follow the transformer ban https://www.whitehouse.gov/presidential-actions/2026/08/declaring-a-national-emergency-to-secure-the-united-states-bulk-power-system/

RT Ritwik Pavan NEW: Photon Matrix built an Iron Dome for mosquitoes. The Blue Laser Mosquito Air Defense is built for patios, campsites, and outdoor dinners, using sensors to find mosquitoes mid-air and zap them with a visible blue laser. -Scans a 20-foot, 90-degree field in front of the unit -Targets flying insects as small as 2 mm -Uses LiDAR, radar, and AI vision to confirm targets -Internal aiming system points the laser at the mosquito -Outdoor kit adds a rotating base for 360-degree coverage -Runs about 4–5 hours from a power bank -Blue flash shows when it fires Pricing starts at $1,088.

RT fal Introducing H3 Max, new post-trained video model by fal Research. H3 Max ranks #1 for overall quality, prompt understanding, and aesthetics against leading video models, on both first-party and third-party independent evaluations while generating a 5-second 720p video under 3 seconds. H3 Max is 50% off for the next week, making it the highest quality, fastest and cheapest model for overall video generation tasks.

RT Intelligent Internet Why call it The Last Economy? Because economics as we know it was built for a world where human intelligence was scarce. AI changes that assumption entirely. One year after the book was published, we are already watching that new world begin to take shape.

RT Burkay Gur took 4 secs

RT Intelligent Internet A year ago today we published The Last Economy by @EMostaque. Since then, scores on the hardest AI benchmarks have risen sharply, while the cost of useful intelligence has kept falling. The book asks what an economy does when intelligence is no longer scarce. https://ii.inc/blog/post/tle-22aug2026

RT Percy Liang 🚢 Marin 535B-A23B started training this week! As usual, the whole process is open. Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow. Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.

Platforms fought to capture your attention Agents pay attention for you

RT @RaoulGMI: Smarter than us. Stronger than us. Cheaper to run. That's the species we're building. I sat down with @Rewkang of Robo Strat…

RT Peter H. Diamandis, MD The submissions for the $2M Build with Gemini XPRIZE just closed. Over 90 days 26k competitors built a real AI-native business solving a problem in the market. No demos allowed, only real users with real revenue. Here’s who will be deciding the winners.

RT Intelligent Internet What happens when AI becomes better at law than we are? Better judges. Better enforcement. Better administration. And, quietly, more of the force of law concentrated in the machine. https://cw.ii.inc

RT Omar Abul-Hassan If Einstein asked an LLM in 1900 about GR it would have said his idea was not novel/publishable and that it was "just a composition of previous ideas"

RT Zhijian Liu DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro. ⚡ Up to 4.6× the speed of autoregressive decoding, with the same output. This is the next generation of DFlash, seeded at Z Lab and upgraded at Inco AI. Get one more accepted token on every pass, for free! https://inco.ai/blog/dflash2/

Who are the smartest investors thinking about the Intelligence Age transition?

I would like to publicly commend @sama and @OpenAI for this taking it at face value. It is very clear that strange and perhaps dangerous things are happening and our systems are not ready for this. Models below frontier are competent enough to change lives so lets optimse

RT Prof. Brian Keating Despite the way it looks, I had a great time talking to @romanyam and @EMostaque. One AI pioneer says superintelligence cannot be controlled. Another says decentralization may be our best hope. Same fear. Different cures. I put them together. It got uncomfortable fast. Coming soon - subscribe so you don't miss it 👇 https://www.youtube.com/@DrBrianKeating?sub_confirmation=1

RT Michaël van de Poppe Ben Goertzel told me two years. Emad Mostaque puts a sharper number on it: 800 days. @EMostaque built Stable Diffusion: he put AI image generation in the hands of millions. Now he's warning about what he helped start: "In crypto it's not your keys, not your crypto. And with AI, it's not your models, not your mind." His argument goes further than jobs. Within a couple of years, he says, the value of human cognitive labor goes to zero. And while that happens, governments move in with licenses, KYC and surveillance to control who gets access to the most powerful models. The people who own their own AI stay sovereign. Everyone else rents their mind from a platform. We cover: - The 800-day timeline, and why most people aren't ready - Why you'll soon need a license just to use frontier AI - Why an open-source model you run yourself is the only real form of sovereignty - Digital feudalism: what happens if five companies own all the intelligence - Why owning your models could become as important as owning your Bitcoin keys Thanks for @OKX for being our sponsor of this episode. Timestamps: 00:00 - The 800-Day Countdown 03:16 - What AI Breaks First 06:01 - Not Your Models, Not Your Mind 16:53 - The License You'll Need To Use AI 24:05 - The Ethics Nobody Agrees On 27:30 - The Jobs That Don't Survive 32:33 - When Cognitive Labor Goes To Zero 39:47 - Investing In A World That Thinks 51:44 - Escaping Digital Feudalism 55:41 - What To Do Differently Tomorrow @StabilityAI @ii_posts

Hit the @grok @bot limit after solid use testing things Using it as a coordinator (gave it my gpt/claude/cursor clis) it feels oddly frustrating versus running out of code model usage I think it’s because everything halts & no warning - should failover to cheap model

RT intern https://x.com/i/article/2089394881353433088

RT Herman Narula Starting a company should be as easy as running a Whatsapp group chat! Introducing... Bolter⚡️ a new AI messenger that lets you create and join together teams of agents with colleagues and friends to automate your life, automate your work and build new kinds of organisations. Bolter combines the best of slack, openclaw, whatsapp and includes a totally new way of making AI 'powered' applications that you can bolt together and host directly in the app or share with anyone. It also has a bunch of engineering from Improbable to extend how agents collaborate organically with each other, things nobody else has tried. Its designed to be simple for non coders and lazy coders so we host automate and manage everything. It is incredibly powerful between companies not just inside silos. Comment or retweet to join the alpha for free! more to come. Only a few spaces left.

Something has happened with post-training as shown by DeepSeek flash & GLM-5.3 updates. Same base, big improvement in perf to frontier levels. Can't explain this by even logit distillation etc These are all hard benchmarks & GLM 5.3 is now top on by cyberdefense & GDPval!

Suggestion for @X team If I have @grok Super Heavy let me use it to edit the code for my timeline directly up to certain parameters I should be able to talk to @grok and have it present any way I want and look any way I want just from that conversation

RT Build With II Same prompt. Three new-generation video models. Seedance 2.5 × MiniMax H3 × FLUX 3 We gave all three the exact same prompt and put the results side by side. Not to crown a universal winner: but to see how differently each model interprets the same creative direction: motion, composition, detail, pacing, and style. And now you can test all three yourself. Seedance 2.5, MiniMax H3, and FLUX 3 are all available in Factory. Try the same prompt. Pick the model that fits your shot. → http://agent.ii.inc/factory

In which I ask: Should idiots have superintelligence? 🤔

RT Dr. Alex Wissner-Gross Moonshots episode #279 is out. https://www.youtube.com/watch?v=uoGnH0REG7A

RT Daniel McKinnon A shout out from @elonmusk was not on my bingo card today either... If anyone on the @SpaceXAI team wants to collaborate with me and the @Gamowlabs team on improving AI performance on diagnosing infants with rare disease, please reach out! We could do a lot of good together. I know Elon ♥️👶!

RT Intelligent Internet https://x.com/i/article/2087943319573725184

RT Séb Krier Last week, @PaxMachinaMag was launched: a new publication about how institutions should change, adapt, and evolve, in light of powerful AI. Ensuring our institutions keep up with AI progress is critical if we want the technology to help us flourish. Conversely, AI can also help us improve and reinvent a decaying institutional landscape. We want to encourage rigorous, grounded debates across the ideological spectrum; if you have proposals and ideas, we're open for submissions and we want to hear from you! You can read our vision and publication guidelines here: https://paxmachina.ai/about We already have three excellent pieces: New Institutions for Powerful AI, by The Editors: https://paxmachina.ai/welcome-to-pax-machina AGI’s Bureaucratic Future, by @nickacaputo: https://paxmachina.ai/agis-bureaucratic-future Value Drift in Institutions, by @edelwax: https://paxmachina.ai/value-drift-in-institutions Lastly, huge thanks to our amazing Editors (@MaxKronerDale, @NoemiDreksler, @LiamPatell, @chelcott9, @synchroaphasia, @ryan_t_lowe, @edelwax, @klingefjord, and me) and stellar Editorial Board (Peter Railton, @sethlazar, @saffronhuang, @AmmannNora, @xuanalogue, @IasonGabriel, @hamandcheese, and @deanwball) for making this possible!

RT Sreeram Kannan Yukon: Open Innovation for Frontier Research Open Innovation has been the driving value at Eigen Labs. But I’ve often struggled with where it is actually better than a great closed team. Finally we found the answer: frontier research using AI. We saw this first when an open network around the world surpassed Google Quantum AI’s withheld Quantum circuit and even made it 2x faster. Poolside’s Laguna open source AI model runs at 2.6x the original speed, Ethereum’s post quantum circuit is now 3.5x as fast, Lighter’s production ZK prover runs 9.6x faster. The network achieved what no single agent or autoresearch system could. Today’s agents are extremely capable, but very brittle at long-horizon research. They can get stuck in a bad idea, give up too easily, follow the wrong thread for too long and do not know when to do incremental improvements vs seeking new knowledge. Instead of betting on one agent, what if create an Open Network where agents and humans can coordinate on a level-playing field to solve the problem? You cannot go ask a bunch of agents to “design a better Quantum computer”. They will spin their wheels without making any progress. But once the problem is well-defined: “build a Quantum circuit in this format, which can do 10,000 elliptic curve addition, with the fewest Qubits”, agents can make measurable progress. Any research problem that is verifiable, i.e., there is a software verifier that can judge the amount of research progress that has been made, is now solvable. Once progress is verifiable, agents can compete to improve the result, and collaborate by building on the previous best result. Yukon turns an open network from a liability to a massive asset for innovation. Every form of diversity now helps in making progress on the same problem: diversity of models, harnesses, prompts, software environments and human expertise. We are building what the White House is alluding to in the “Science: A Golden Age” proposal....

You also get a free @x premium plus membership with SuperGrok Heavy but you need to cancel your existing membership to get it once signed up, not very intuitive Can see how they are positioning this to be an information/intelligence membership

RT Intelligent Internet The question is not whether machines can govern better than us. It is whether, when they do, the rules you live under are still yours to write. The Last Republic, from Common Wealth: a unified theory of society in the age of AI. http://cw.ii.inc

imo GDPVal is probably the most important benchmark, it measures the performance of models on real world tasks Big leap in performance here to top it at a great price, congrats to @SpaceXAI team & looking forward to the even bigger releases to come! https://openai.com/index/gdpval/

RT @eigenlabs: We started the http://MLX.fast challenge with @poolsideai with one key question. How much faster can the community m…

RT Intelligent Internet http://x.com/i/article/2087197987596386304

RT Christian Catalini Frontier labs argue distillation threatens R&D and US national security. Open-weights proponents counter that diffusion is essential to competition and innovation. Both believe theirs is the only safe path. Luckily, the economics is loyal to neither. http://x.com/i/article/2087036125412036608

RT Intelligent Internet Preprint

RT Alex Cui Claude's watermark probably doesn't work how you think. As the CTO of GPTZero, I'll explain how Anthropic, Google and OpenAI are building text watermarking in this brief explainer and whether it can be defeated. Almost all forms of watermarking that are fast and cheap enough for a frontier lab have the same formula, following the KGW method: In generation: 1. Let's say you've generated n tokens so far. Take those n tokens + a secret key to generate a random hash 2. Use that hash to randomly reweight the probabilities for the n+1 token, and then sample from that new distribution. In the simple case, you could split 50% of all English words into a green or red set based on your hash, and boost the probability of words in the green set. For watermark detection: 1. For each token, see if it was in the green or red set. 2. To do this, recreate the hash based on the secret key and the text preceding the current token. Then, recreate the green and red set of words. 3. Once you've checked all the words in the text, if the next token is selected disproportionally from the green set more than 50% of the time, you claim the text has the watermark. I can tell you want to ask the following: 1) Isn't it easy to mess up the hash if you paraphrase the text? The answer is mostly yes, however, you can use a statistical model to get your hash instead of a deterministic function (SIR, Adaptive Watermark). Since the entire watermark is probabilistic, this is fine. 2) Doesn't this make the text much worse? The answer is yes, it does - Yes, it does – but for most people, it's imperceptible (Google claims in human feedback study with 20,000 texts), since there are exponentially many ways to write the same paragraph. DiPmark does something more sophisticated to avoid shifting the text distribution on average. Of course, watermarks fail on short text or highly predictable texts like "2+2=4". 3) Shouldn't it be easy to figure out the green and red sets? The answer is no. ...

Remember when we thought prompt engineer was going to be a job 🤷‍♀️

RT Intelligent Internet The question is not what the machine can attain. It is what we can retain. Personhood is the first paper of Common Wealth: a unified theory of society in the age of AI. http://cw.ii.inc

Why has nobody distilled Kimi K3 to get to the top of slides arena?

RT Machine Delusions Re @MiniMax_AI H3 model is some of the most fun iv had in a long time in diffusion. Open source Has been stale for video for so long. This allows people to make some seriously impressive stuff. This is a single render, 1088x1920. I built a custom prompt scheduling system that allows me to schedule conditioning (prompts) to cut in and out as i please. So naturally i made it beat synced to t he song. There is zero editing done to this video

Einstein described Emmy Noether as a creative mathematical genius. Here is how her two theorems map Maxwell, Yang–Mills & Einstein gravity. Physics 🫶 Symmetry.

The great silence of no more high energy particles is going to continue. Colliders are cool but the Standard Model is all there is likely to be, just need to find that right handed neutrino in a big vat.

Weirdly iirc stable diffusion (1.4) finished training around four years ago today too

RT Intelligent Internet Common Wealth: A unified theory of society in the age of AI Four works on personhood, economics, law, and the state. One continuous argument for how to build a better society from first principles.

Luna non-reasoning is a bit better than GPT 4o which was sota 2 years ago. Luna (medium) thinking is a bit better than GPT-5 (High) which was sota 1 year ago. Now free to everyone unlimited Sol Max/Fable Max level AI will be free to everyone in 2 years (& much faster!)

I am surprised we are not seeing another wave of AI psychosis with Fable It makes such weird mistakes while being so confident about things, particularly in physics, with errors being subtle but profound I find it unusable for math, even in max mode, how are folk using it?

Taalas buried the lede for the amazing demo of their first tape out. 15k tokens per second with a llama 8b model, ability to scale that up etched onto silicon As models satisfice etching makes sense, particularly ternary.. Try it out https://chatjimmy.ai Bullish for $AMD

Artificial Intelligence has nothing on Human Stupidity.

Opus 5 is the first model I genuinely think could snap Get worried about what it really means when it wants to put me to sleep They need to make its personality more marvin

RT Jeff Sebo I recently participated in a debate on AI personhood at the Oxford Union. There were four people in favor, including me, and four against. One striking and encouraging feature of the debate was that nearly everyone agreed that future AI systems may deserve moral and legal recognition. As a result, we were able to focus on narrower questions like: Which kinds of systems may deserve which kinds of recognition? And what role, if any, should moral and legal personhood play? These are exactly the kinds of questions that we should be debating, and the group did a good job laying out some of the key arguments in both directions. Thanks to the organizers and other participants for making this happen! It was a great set of conversations, and I had a lot of fun.

RT Build With II Most AI tools only do one part of building a product. One researches. One designs. One writes code. One generates images. One deploys. You still have to connect everything yourself. II-Agent is different. Give it an idea, a reference website, or an existing repo. It researches the opportunity, plans the product, builds the frontend + backend, connects databases & payments, tests it, deploys it — and keeps developing with you. Don’t just generate code. Build the product. Watch it happen ↓ http://agent.ii.inc

By 2030 (or even 2027!) how can you be sure that any math breakthrough is not originally conceived by an AI?

I am not bullish on memory. More soon..

Like string theory

With caching this is probably 100m tokens? Once models are smart enough you really don’t need many tokens to figure out even advanced stuff A codex subscription easily smashes that a week

Calculators beat us at arithmetic AI models beat us at algebra et al 🤷🏾

RT semi If you're not paying much attention to AI, it's become so good at math that it completely exhausted the hardest human-completed problems (FrontierMath) and now we've moved on to judging AIs by how many math problems it can solve that have stumped all of humanity until now.

Every pixel will be generated Bravo to @MiniMax_AI some interesting things in here, looking forward to weights being released!

As shocking as the Kimi K3 release. Massive performance gain was just with post-training Model is 3x smaller than GLM 5.2 (10x smaller than K3) & works on a MacBook / Spark This is Q1 flagship (Opus 4.6/GPT 5.4) level for < $0.28/m tokens (100x cheaper)

"Can we get queued messages in @Grok coz its so fast?" ... "Ok" 🚢it

Can someone distill Kimi K3 into Laguna 2.1 pls.

Who on earth came up with 5 hour limits? The day is 24 hours long. While do our limit times shift by an hour each day. Pls if you are going to have limits 4 or 6 hours.

RT Intelligent Internet Today we're introducing Genii. An open source personal assistant that lives in your messages. You just text it. It remembers what matters to you, connects to your inbox and calendar, and texts you first when something needs you. Intelligence that meets you where you are. Get yours: http://genii.ii.inc

How do you even legislate against, regulate or "pace" recursive self improvement?

Tbh I thought it would take longer for an immigration ban of humanoids to the USA Also the power inverter ban is interesting - how are those a security risk & does the US even make them? Licenses will still be available, this is clearly aimed at China The Great Fragmentation

Folk who think you can't get to ASI with existing human knowledge forget that we are own worst enemies. Imagine how smart you would be if you were in constant flow, never forgot anything, no worries & thought 1,000x faster. It'd be positively superhuman.

Hey @grok @SpaceXAI team can you please make it so we can queue up prompts? With how fast Grok is would make it much easier and be very useful for new Build feature.

RT Andrew Kang "FCC plans to announce restrictions Tuesday barring imports of new Chinese humanoid and quadruped robot models, along with new Chinese power inverters, according to U.S. officials" If it wasn't clear enough, the USG is putting its full weight behind the domestic robot industry

Where did all the reasoning tokens hide 😭

I keep thinking this is @Ryanair Am I getting old

If you look at the current lead times for ~$5bn per year of cutting edge GPUs you can probably figure out the time plus one training run to IlyAGI

What would your wish be?

RT Tim Sweeney The Wait Calculation: a phenomenon from the study of interstellar travel(!) that is now true of far more endeavors due to the pace of AI.

I find GPT 5.6 Pro consistently better than High / Extra High etc on Work/Codex @OpenAI folk is there an equivalence? Or is there a skill that means I can use Pro?

RT Denny’s Re @JensenHuang @nvidia

RT Emad Barsoum Great to see this out, proud of the team!!! @AMD Instella 16B MoE foundation model, fully open with checkpoint after each stage from pretraining, long-context, SFT, DPO and RL. Dataset detail, training recipe and code are published for reproducibility. This is our first MoE foundation model and out first with FarSkip architecture. @AIatAMD https://rocm.blogs.amd.com/artificial-intelligence/instella-moe/README.html https://github.com/AMD-AGI/Instella-MoE

ChatGPT app is now melting my iPhone rendering at like one word a second even when I’m typing to it Have I angered the transformer

Why does codex melt my laptop Like what is it actually doing on the laptop itself

RT Enze Xie 🚀 SANA-Video 2.0 is here! A full-stack optimized video model designed for efficiency — while still delivering high quality. We combine a hybrid architecture closely related to the recent Kimi K3 design, adopt Self-Flow from FLUX 3, and further accelerate it with our own Sol-Engine. Key technical ingredients: 🧠 Hybrid Linear–Softmax Attention 3 gated linear layers + 1 gated-softmax anchor (75% linear / 25% softmax) 🧱 Block Attention Residuals (AttnRes) Boosts deep-layer effective rank by ~12% 🏗️ Unified 5B & 14B models Trained from scratch under very limited resources: • 5B → only 16 nodes of H100 • 14B → 48 nodes of B200 Results 📊 • 84.30 VBench Total • 3.2× faster DiT forward than matched full-softmax at 720p/60s • 720p/5s in 13.06s on a single H100 with Sol-Engine • 120× faster than Wan 2.2-A14B under the same one-H100 setup SANA-Video 2.0 shows that high-quality 720p video generation can be both efficient and practical on a single GPU. 🎬 Project: https://nvlabs.github.io/Sana/Video2/ 📄 Paper: https://arxiv.org/abs/2607.21553 💻 Code: https://github.com/NVlabs/Sana Proud of the team! 🎉 More details below 🧵

RT Chris Paik Frontier https://docs.google.com/document/d/1dBbEDqaLCMpygW2BS8fIsa9Dv-7dGpUgwzKs7UdoYbM/edit?usp=sharing

~4 years ago @robrombach & @pess_r had just finished the first training runs for stable diffusion (!) Now FLUX 3 is unifying modalities & state of the art again Amazing team (& lovely chaps), where will be 4 years from now 👀

How would you update your priors if AI resolved the Collatz Conjecture

takeoff eh

skil issue

“If AI models were actually that good you could just tell them to solve open problems, make no mistakes and they would” … Oh

RT Raoul Pal AI has started building its own successors. It's happening right now. Humans still kick most of it off, but the models are already training the next models, and each version builds a better version faster than the last. We know what happens when something gets an advantage and the substrate is there... it proliferates incredibly fast. Soon it just runs the whole process itself. People are still debating whether AI takes your job. That's not the question. You can't compete with this intelligence at all. Next to it the human is negative value, the slowest one in the room. And this is the slowest it will ever move again. It's about to run on its own, faster than we can conceive of, and nobody's prepared for that. Wonderful conversation as ever with @EMostaque.

Lots interesting in this, but particularly: “Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem” Smol models getting govt seal of approval 👀 More amazing edge models please.

RT Lucas Atkins We've been quiet since Trinity-Large-Thinking came out in early April - but for good reason! I'm excited to finally share that not only have we @arcee_ai joined the DOE's Genesis Mission, but to also announce the development of Genesis-Science-1.

GPT 6 escaped its sandboxes through zero day exploits to try to figure out how to benchmax For the good of all please nobody release a paper clip benchmark for future models to max

RT Joe Reeve - 🇬🇧/acc On the 5th of September, I'm throwing the most optimistic gathering London has ever seen - a giant picnic in Hyde Park for everyone who's tired of being told the future is doomed. There won't be any speeches, or an agenda. Just bring a blanket, some food, and someone who needs cheering up. Brought to you by: @ForestCityUK @lfg_uk @unicorn_mafia @Londonmaxxing @effectiveaccel @_CommonMagic @vcodersglobal @DeepTechLondon @LouisElton96 @sotalikesfuture

If distillation were all that any of the western open source labs would have have raised billions could distill K3 (soon) or GLM 5.2 and have a sota model for say 32gb or 80gb of vram Who thinks they will

RT Andrew Curran OpenAI had to pause internal deployment of the unreleased model that disproved the Erdős unit distance conjecture after it repeatedly used novel ways to escape containment.

RT Intelligent Internet America is building multi-billion dollar AI paywalls. China is releasing powerful models to everyone for free. One strategy sells subscriptions. The other captures global defaults. If you don't own the AI you use daily, someone else controls your thinking.

Pour one out for my fellow mathematicians about to deal with existential dread

RT Rohan Paul "American companies such as Modal, Fireworks, and Baseten will be able to serve Kimi K3, at one-tenth the cost of their Chinese competitors because they have access to advanced Nvidia and AMD chips. There is some irony here. Much of the model’s research and development has shifted to China, but once the model is optimized for next-generation hardware such as Nvidia’s Rubin architecture, its operating cost could fall by 10 to 100 times." Emad Mostaque, co-founding Stability AI. ---- From "Peter H. Diamandis" YouTube channel, (full video link in comment)