Emad Mostaque

RT elie this is so funny, training opus 4.7 on business skills makes it misaligned and dishonest 😭

RT bidhan ✈️ CVPR We're releasing Paris 2.0, which, to our knowledge, is the world's first decentralized trained video generation model. We benchmarked it against a monolithic model trained on the same data and compute budget, and Paris 2.0 outperformed the monolithic by ~2x on FVD benchmark.

RT Intelligent Internet Make your agent smarter. The II-Commons skill gives your agent reliable knowledge from arxiv, PubMed & more, plug it in Repo: https://github.com/Intelligent-Internet/II-Commons-Skills Add it to II-Agent: https://agent.ii.inc

Train with autoregression & convert weights to diffusion for inference.

Great to see @poolsideai (US lab) committing to open sourcing their foundation models going forward Laguna is an interesting release, check it out

RT Eric Ries I built a public site to track how Incorruptible is going. It scans the internet twice a day for verified reader reactions, podcast clips, tour moments, and notable mentions. Then it publishes them on a single page so you can see the launch as it actually unfolds, not the version we curate. The book is about building systems that hold their values under pressure. It seemed fitting to build one for the launch. http://howisincorruptiblegoing.com

RT EdgeRunner AI In our new paper, EDGE-OPD (EviDence GuidEd On-Policy Distillation), we introduce two key ideas for improving On-Policy Distillation (OPD). First, we use guided rollouts that inject privileged context into a subset of the student’s prompts during generation (such as hints, hidden evidence, or rare personas), allowing rare target behaviors to actually appear in the on-policy data, while masking that privileged information during loss computation. Second, we introduce an evidence mask that updates the student only on token positions supported by the privileged context, instead of learning from the entire rollout indiscriminately. We show empirically that, together, these mechanisms enable the student to internalize rare but correct behaviors without directly training on the hidden context itself. Beyond the method itself, we also analyze how these rare-token signals localize within specific regions of the rollout, offering new insights into efficient knowledge transfer and preservation of general-purpose capabilities during post-training. Read the blog post and paper here: https://www.edgerunnerai.com/news/edge-opd-internalizing-private-data-with-on-policy-distillation

RT PrismML Today we’re releasing 1-bit and Ternary Bonsai Image 4B. A new family of image-generation models designed to run high-quality diffusion inference on local hardware: from laptops to phones.

RT Intelligent Internet Board Mode is here. Instead of one long agent session, you now get a creative workspace where research, slides, websites, apps, and design drafts can live side by side. Try it now: https://agent.ii.inc/ Read more: https://ii.inc/blog/post/board-mode

I think folk are underestimating how much of AI models are actually engineering at scale versus breakthrough research. See how @cursor_ai caught up to Anthropic / OpenAI models run at a fraction of the cost to run & it becomes clearer why that deal was done & what is to come

RT Intelligent Internet who did it better

It’ll be interesting to see if the post training for this uses a multiple of the compute of pretraining as cursor did when they tuned Kimi as the base model

Amo Dei

ngl I thought this was them talking about no researchers under 18 There are some darn good ai researchers in that age bracket

RT Wario64 Star Citizen hits $1 billion in funding https://robertsspaceindustries.com/en/funding-goals

RT David Lindell Image diffusion models like Flux natively output at 1k resolution, but what if we want to generate much higher resolution images (6k+)? SEGA modifies the RoPE encodings during the diffusion process to generate high-resolution images---no fine-tuning required!

RT Zyphra Today, Zyphra Research is sharing fundamental work extending Equilibrium Propagation beyond Energy-Based Models to biologically realistic neuron models. A step toward more efficient AI, local learning, and future hardware beyond backprop. 🧵

20x sales for something growing that fast js a great deal for @SpaceX It will be immediately accretive given @SpaceX is likely coming to the market at over 50x sales

Once AI starts making solving open problems in novel ways it won’t stop. We are entering the final stage of human solutions to open problems like this. Feels weird, doesn’t it?

RT Yohei i'm excited to open source Active Graph: an event-sourced reactive graph runtime for long-running, agents 🔄🧠 events/logs projects a graph. reactive behaviors react and affect the graph. fork-and-diff agent runs. no A2A, no workflows, no DAG site: http://activegraph.ai docs: http://docs.activegraph.ai github: https://github.com/yoheinakajima/activegraph quick start: pip install activegraph this is an early experiment in a new paradigm for agent architecture 🧪

RT Intelligent Internet Agent work getting stuck in sessions is vexing. We're open-sourcing CommonGround Kernel so you and your agents can coordinate, hand off, and build on each other's work. Blog: https://ii.inc/web/blog/post/cgk-preview GitHub: https://github.com/Intelligent-Internet/CommonGround/

Figured it out

This (Antiqua et Nova) is an interesting read ahead of Magnifica Humanitas release https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html

RT Max Zhdanov 🌍Today we release Mosaic, a probabilistic weather model that shifts the Pareto frontier of ML weather forecasting. It matches the skill of state-of-the-art models while generating a 24-member, 10-day global forecast in under 12 s on a single H100. Thread!

Shouldn’t this mean > $100b in charitable giving for effective altruism adjacent projects from Anthropic founders & team members alone?

RT Peter H. Diamandis, MD This is the Build with Gemini XPRIZE. $2,000,000 in prizes. 90 days. Pick a problem worth solving. Build a profitable business with AI. Grand prize: $500K in cash.

Seems a lot of autoregressive models will be converted to diffusion models

Can’t wait for Gemini Omni in @NotebookLM cinematic explainer videos 👀

Go join Monastery accelerator with our friends @cyberfund $2m uncapped note, 12 week program, let's see if we can get the 1 person AI-assisted unicorn

As companies approach AGI it would be illogical for most not to go and work there. AGI is a much bigger deal than most people still seem to believe & it is only by most forecasts a few years away (if not arrived already!).

RT Mati Staniszewski Albert Einstein + ElevenLabs. AI agents can make education more accessible - a teacher for every student in every field. A classroom size of one learning from icons who shaped the world Today with his estate, we’re bringing Albert Einstein to ElevenLabs

Amo Dei

So would this be legal under the ruling: Set up a charity, raise $100m etc When setting it up say you may set up a for profit & raise billions Three years later set up for profit with IP transferred from foundation Give yourself equity Profit?

This is such an interesting chart layout, like it a lot! Congrats to @cursor_ai team on the 2.5 launch 🚀

RT Oliver Cameron Introducing Starchild-1 from @odysseyml, the first ever real-time multimodal world model. This a model that can generate interactive simulations of the world that you can—for the first time ever—hear. Starchild-1 represents a big step towards a general-purpose world simulator.

RT Yining Karl Li This was 10 years ago so I'm a bit hazy on details now, but there were a few shots in this sequence that required a TB of RAM for water sims, and it was really late in production, so our solution was to cannibalize RAM from multiple machines and build some 1 TB RAM servers.

RT Martin Hebart I'm proud we are releasing LAION-fMRI, a densely sampled 7T fMRI dataset of natural images, with very broad stimulus sampling for testing countless hypotheses & deeply exploring brain representations. It is now available at https://laion-fmri.hebartlab.com/ What does LAION-fMRI offer? 🧵

Recursive self-improvement is very reasonably the greatest near-term threat to democracy & peace out there

Who’s building nanobots? Seems the last leg of Kurzweil’s predictions

The total addressable market for humanoid robotics may well be larger than the TAM for human labour

I discuss this applied to economics in The Last Economy with reference to Smith, Marx & Hayek Harmonic flow is the landscape itself https://ii.inc/web/the-last-economy/chapter-9

Spent some time this week looking at the economic impact of humanoid robots and its so much bigger than digital AI The numbers get very crazy, has anyone dug into it and not found this the case?

Team @grok @xai can you please allow latex uploads on mobile app thx

RT virushuo 20 years ago, my first startup was all about enterprise search. Two decades later, we’re still building search engines. The technology has shifted from NLP to NN and the users from humans to agents. but searching is still the core. opensource the fastest bm25 engine:

RT Zyphra Re Training a diffusion LLM from scratch is challenging, and diffusion inference advantages don't show up at training. So we convert instead. Building on the TiDAR recipe, we take ZAYA1-8B-base and run diffusion-conversion mid-training plus diffusion-SFT on our existing stack.

RT Alex Rampell Re one time a reporter did an "expose" on a founder I backed whose last company failed...here's how I responded: Wow, you discovered that [last company], despite years of trying, eventually didn’t work. Should [CEO] be sent to Roman Debtors’ prison? Maybe hanged, drawn, and quartered? When a company doesn’t work, people lose their jobs and investors and creditors lose. It’s part of business. So is doing something new. This is like doing an expose on the chef whose last restaurant failed and dared to start a new one.

RT Azeem Azhar I spent a week on the ground in China visiting 14 AI labs. The more I saw, the clearer it became that export controls may have inadvertently created a durable efficiency moat – Chinese labs are extracting 4-7x more intelligence per unit of compute and serving it at a fraction of Western prices. https://www.exponentialview.co/p/inside-chinese-ai-labs-efficiency-moat

RT Alec Helbling The Helmholtz decomposition is one of the fundamental results of vector calculus. It says any well-behaved vector field can be split into two parts, one capturing sources and sinks through divergence, and one capturing rotation through curl.

Each round founders should take <runway> years worth of comp as secondary & pay themselves just enough for health insurance Far more tax efficient than primary => salary

RT Bo Belmer /brainpower_podcast/season_2/episode_3/emad_mostaque is live featuring @EMostaque (Founder of @ii_posts). Hosted by @DJohnstonEC (@MorpheusAIs) & myself!

When you want to move from single agent SQLite on something like QMD, PostgreSQL is a great choice for multi agent and production quality, but not as snappy. So we made it much more snappy with BM25 & open sourced it. More soon for planetary scale sovereign agents

Trial lawyers missed a great chance to ask everyone on the stand, under oath, what their definition of AGI is

What is everyone's favourite platforms for Seedance and Happy Horse etc

What if Anthropic voids the FTX sale

RT David Louapre Meet physics-intern🧑‍🎓, our agentic framework for theoretical physics. It takes Gemini 3.1 Pro from 17.7% to 31.4% on CritPt, a new SOTA on one of the hardest benchmarks for LLMs. Theoretical physics is hard for humans and LLMs alike. But physics-intern decomposes problems and dispatches them to a team of specialized agents, solving research-level questions far more effectively than the base model alone.

RT Dexerto The original "Leeroy Jenkins" video was first uploaded to the WoW fansite Warcraft Movies on this day 21 years ago

RT @ii_posts: New research: long-running agents often fail by stopping too early, not because the model can't make progress. We tested 5 h…

RT @PalisadeAI: Over the past year, AI agents have learned how to self-replicate. In our test environment, an agent hacks a remote computer…

RT Timothy Gowers @wtgowers Re But if AI mathematics continues to progress at anything like its current rate -- which is what I expect to happen -- then we will face a crisis very soon, and mathematics departments, who owe a duty of care to their students, should be urgently preparing for it.

RT Intelligent Internet New research: long-running agents often fail by stopping too early, not because the model can't make progress. We tested 5 harness designs across 8 long-horizon coding tasks. Our new orchestration harness, Zenith, wins 5/8 at 43% the cost of the strongest baseline.

RT Ryan Petersen Financial Times perfectly illustrates our possible futures

RT Intelligent Internet One-shot continuous video, from a single prompt. Made with Factory™️

Tokens are the bricks by which the information superhighway is being rebuilt. There is a lot of information. It'll need a whole lot of bricks.

RT Zyphra Today we're releasing ZAYA1-8B, a reasoning MoE trained on @AMD and optimized for intelligence density. With <1B active params, it outperforms open-weight models many times its size on math and reasoning, closing in on DeepSeek-V3.2 and GPT-5-High with test-time compute. 🧵

This likely costs about ~$500m a month, ~$6bn a year to rent blended Which is about the run rate net loss of xai end of q1 Anthropics revenue run rate was $9bn end of 2025, $30bn a month ago

RT Intelligent Internet Introducing Factory Describe your idea in one sentence and Factory builds the first draft of the production for you. It plans the scenes, chooses the models, generates the assets, and wires everything together on an editable infinite canvas.

RT Intelligent Internet London AI people, film night? Join @EMostaque at AI Movie Club x Blade Runner for an AI chat, then the film. 📍Kensington Central Library 📅Thurs 28 May, 18:00 🎁 Free event + SWAG for early arrivals.

The economic & social disruption from nearing AGI will be bigger than just about anything we have seen I expect the frontier labs to be nationalised There is a chance top researchers get conscripted Real Manhattan Project but bits not atoms

RT Ricardo Olmedo We fine-tuned Alec Radford’s 1930 vintage LLM to solve SWE-bench issues. After just ‼️250‼️ training examples, the model solves its first issue, a simple patch to the xarray library. 🧵👇

Number 1 @OpenAI Codex request: plz let us use gpt pro in it Or someone compare gpt pro to x high

It seems @OpenAI has cracked recursive self-improvement for Codex

Do you use normal or extended thinking for GPT 5.5 Pro

Pretty sure it’ll be like this

Looking forward to the work coming out of @IneffableLabs 🇬🇧 Largest EU/UK raise ever 😮 Beat the previous UK largest seed raise ($101m Stability AI!) by $999m 🚀 We have seen amazing things on our self-learning models so bullish on this approach 🤓

RT Tad Smith 👀 The most interesting people in the most exotic places can occasionally be persuaded to talk about NFTs, digital collectibles, and even @CandyDigital! Try it sometime just in case it catches on… 🔥🚀📈 🤫 @RaoulGMI @Blondie23LMD @BittelJulien @EMostaque @RealVision 👇

I don’t think there will be any coding models 3-4 years from now.

RT The Render Network "I think that the holodeck is near again minus the hardlight holograms etc. - We’ll be able to have that, with our 3D glasses, with the ocular occlusion and more…” https://x.com/rendernetwork/status/2047407574987735394/video/1 @EMostaque breaks down how diffusion models are learning real-world physics, why open-source world models are arriving faster than expected, and why access to compute becomes the critical layer as these experiences scale. #RC2026

Given how good gpt-image-2 is it’s an absolute puzzle OpenAI shut down Sora

Bet this happens with Navier Stokes and it’s going to be something not even related to PDEs that solves it

RT Fengzhuo Zhang The Newton–Schulz iteration coefficients optimized by DeepSeek-V4 are surprisingly strong: they effectively normalize all singular values to 1. This matches our previous intuition: a well-balanced spectrum may help strike a better balance across long-tail knowledge. Plot code: https://github.com/FengzhuoZhang/dpsk_v4_muon/blob/main/coefficient_figure.py

RT Intelligent Internet In 1982, Blade Runner was speculation. In 2026, it's the conversation. Memory, personhood, alignment, what we owe the minds we build. Watching it Thursday 28 May, Kensington Central Library. Come think it through with us.

Congrats to @deepseek_ai team! Doing the numbers I would estimate: Pro < $14m for the final training run Flash < $4m Ratio of active params x total training tokens vs v3 Total compute costs (data prep, tuning, testing) ~10x that Cost of taste: priceless

Necessity is the mother of kv-cache optimisation

Congrats to @deepseek_ai team! Fwiw doing the numbers I would estimate Pro cost < $14m for the final training run & Flash < $4m This is straight from ratio of active params x total training tokens vs v3 Total overall compute costs (data prep, tuning, testing) probably 10x that

😭

RT The Render Network At #RenderCon2026, @EMostaque makes the case during How to Build the Holodeck that stories still matter, IP still matters, but the barriers to building entirely new experiences are dropping faster than we've ever seen before, with compute being the last real limiter. Full talk:

RT Intelligent Internet Oh?

By end of next year. End of year after for 16gb vram. Do folk really need more for most things? Not really. But just call the cloud models when you do.

RT Liron Shapira 🔥 Emad Mostaque on Doom Debates 🔥 @EMostaque helped kick off the modern AI revolution as the head of Stability AI, the company behind Stable Diffusion. But unlike most AI CEOs, he doesn't sugarcoat the risks of AGI development — he's upfront that his P(Doom) is 50%! In today's Doom Debate (or should I say Doom Consensus), Emad explains why we have less than 1,000 days to get our act together, and how his new startup (Intelligent Internet) could potentially make the nondoom scenario go better. We cover: ⬜ What's Your P(Doom)?™ ⬜ Cognitive labor going to zero ⬜ Why ASI doesn't need massive compute ⬜ Is there a ceiling above human intelligence? ⬜ Corporations as slow, dumb AIs ⬜ Emad's mainline doom scenario ⬜ Jailbreaks and "Mecha-Hitler" latent spaces ⬜ The Last Economy / Economic Doom ⬜ Why isn't Google stock mooning? ⬜ Intelligent Internet as a solution ⬜ Can an aligned network stop a rogue ASI? ⬜ Is the Pause AI proposal too late? Highly recommended! 👇

RT Intelligent Internet Today we release a novel AI-assisted resolution of one of physics’ longest-standing questions. Given only: • Relativity as an axiom • One characteristic of the algebra A positive cosmological constant is forced. The universe expands.

RT Chris Hayduk A couple of months ago, I announced that I was partway through implementing a simple, readable AlphaFold2 in pure PyTorch, inspired by @karpathy's minGPT. Today, I'm happy to share minAlphaFold2 - the completion of that project. Repo link: https://github.com/ChrisHayduk/minAlphaFold2

I kinda think it would be interesting to have a structure where founders take 2 and 20 2% management fee on capital raised for running the company 20% of any exit Single magic share with lots of voting rights In line with their VCs

Physics paper in, manga out

RT Prof. Brian Keating AI Insider: "Adding a Human Makes Your Team Worse" Emad Mostaque | @EMostaque TIMESTAMPS : 00:00 The models too dangerous to release 10:06 Why physics needs axioms — and AI doesn't 11:43 The MIND framework most AI researchers won't use 17:36 The universe might be a computation 30:34 Should we let AI fly planes? 39:43 AI just invented a game no human could design 44:43 Why AI will cost almost nothing in 18 months 51:46 Can religion survive superintelligence? 57:36 The AI companion problem nobody talks about 01:03:20 Why personalized AI is more dangerous than AGI 01:11:18 The forgotten link between God and science 01:26:32 800 days: what happens next

gpt image 2 can do some reallly hard prompts well I present - the periodic table of Pokémon

Chrono Trigger Rebirth plz @SquareEnix

gg OpenAI team, final stretch of cracking image generation!

's pretty good, congrats all

Imagine if they add document signing to Claude :o