2 days ago we shipped image generation in <1s 🔥 Today, we make that <300ms 🤯 NVIDIA + AMD⚡️ Full demo below ⬇️ https://t.co/QIl2ZqZe8K

Full demo ⬇️ https://t.co/LR5usBncxq

Also live today: ARC Prize 2026 - 3 tracks, $2,000,000 in prizes available! Get involved: • Play a Game: https://t.co/Cd7ANx2mdT • Build Agents: https://t.co/Pj6qEXUCBD • Win Prizes: https://t.co/marRtnu9Jn https://t.co/0JXBTtvHwd

I'm usually not one to write thought pieces without much technical depth. But here we go. Slow the fuck down. https://t.co/dcOwPK357F

Made in NYC with AWS: Daniel Hesslow, Co-founder, @AdaptiveML. Unlocking US $300,000 in AWS Activate Credits has helped the startup bootstrap early & move fast without worrying about costs. That movement has included a relocation to New York, a city that ‘matches the velocity of startups.’

ARC-AGI 3 is here, and all existing AI models are below 1% on the benchmark. It's gonna take a while until this one is saturated. How it measures intelligence: - 100% human-solvable environments - Skill-acquisition efficiency over time - Long-horizon planning with sparse feedback - Experience-driven adaptation across multiple steps "As long as there is a gap between AI and human learning, we do not have AGI."

An alien species with zero knowledge of human language could ace ARC-AGI-3 on day 1, and I think that's beautiful. At a time when AI is dominated by language models, it's refreshing to have a frontier benchmark (the only one that I'm aware of) that requires zero language ability or cultural knowledge to solve. Intelligent does not mean "speaks English" or "speaks Python." I'm reminded of classic "first encounter" sci-fi storylines (shout out to "Project Hail Mary"!) where intelligent species are able to communicate well before they hash out a common spoken or written language simply based on universal math, science, and reasoning concepts. And AI has gotten complex enough that it behaves much more like an alien species than a next token predictor at this point. It's been a blast following the project and supporting @fchollet, @GregKamradt, and the @arcprize team over the past several months in a small way with funding, PR, and design & dev work (shoutout to our in-house designer Jenn for that beautiful game emulator!) as part of our @LaudeInstitute Slingshots grant program. More impactful research like this, please! Now that ARC-AGI-3 is out, we're going to be feeling some withdrawal. Go to https://t.co/HRUiLPRZzT and give us another project to get excited to see ship!

Just squeezing through https://t.co/RmxoKsSCwk

Agents build useful memory during tasks, but that memory is trapped. So the big question is whether a single memory system can be shared across different models. If you want to transfer it to a different model, the performance often gets worse, not better. New research tackles the why and offers a fix. MemCollab uses contrastive trajectory distillation to separate universal task knowledge from agent-specific biases. It contrasts reasoning trajectories from different agents to extract abstract reasoning constraints that capture shared task-level invariants. A task-aware retrieval system then conditions on the task category to apply the right constraints at the right time. As teams run heterogeneous agents with different models, collaborative memory becomes a shared reasoning resource rather than a liability. MemCollab improves accuracy and inference-time efficiency across mathematical reasoning and code generation, even in cross-model-family settings. Paper: https://t.co/GfbTUROsEZ Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c

BREAKING: Grok Imagine takes #1 overall on the Multi Image to Video Arena with an Elo of 1342. https://t.co/q2Sj8ZreM9

Elon Musk just explained how Starlink moves the GDP of entire nations. The formula is so simple it should embarrass every development agency on the planet. Musk: “GDP is a function of average productivity per person.” Productivity per person goes up. GDP goes up. That is the whole equation. Everything else is decoration. And connectivity is the single largest lever on Earth for pushing that number. Musk: “If you don’t have access to the internet, or it’s too expensive or low bandwidth, you cannot access the MIT lessons and you can’t sell the goods and services that you produce.” No internet means no global knowledge. No global markets. No ability to sell to anyone beyond your village or learn from anyone outside of it. The penalty is total. And it has nothing to do with the person serving it. There is a child alive right now who is as intelligent as anyone who has ever walked the halls of MIT. She does not know it. Nobody around her knows it. Because the coordinates of her birth have no connectivity. No library. No signal. No link to the world that would show her what she is. She will grow old inside a ceiling that geography built for her. Not because of talent. Not because of effort. Because of a satellite that had not been launched yet. Musk: “Internet connectivity is certainly a candidate for one of the things that would do more to lift people out of poverty than anything else.” Traditional infrastructure takes decades. Fiber has to be laid. Towers have to be built. Permits have to be approved. Capital has to be attracted to regions that cannot attract it. Starlink bypasses all of it from orbit. No cables. No permits. No waiting for a government to prioritize your village. A dish goes up. Isolation ends. Someone who could not access a textbook yesterday downloads MIT’s entire curriculum today. Someone who could only sell to neighbors starts selling to the planet tomorrow. That is not an upgrade. That is a different life. Musk: “Starlink will actually move the GDP of countries. Like it’s gonna be that kind of thing.” He said it like a feature update. But read it again. Move the GDP of countries. Not a company’s revenue. Not an industry’s output. The gross domestic product of nations. Shifted by one constellation. The telecom industry spent decades deciding which regions were profitable enough to connect. The rest were written off. Starlink does not make that calculation. It covers the planet. Every farmer. Every welder. Every kid with a clear view of the sky. The minds that will cure diseases, solve energy, and build things we cannot yet name are already alive. They are already thinking. They have no signal. Starlink is the first technology in human history that can reach them at the speed of deployment instead of the speed of bureaucracy. And when those minds come online, they will not change their own lives. They will change the trajectory of the species. That is what Musk actually built. Not a telecom company. The largest unlock of human potential ever launched from a single network.

I’m honoured to be joining 𝕏 to lead design. I believe this is the most important platform in the world, and I can’t think of a more exciting place to help shape the future. I’m looking forward to working closely with @elonmusk, @nikitabier, and the rest of the team. I’m grateful for the opportunity, humbled to be part of it, and can't wait to get started!

PyTorch and Nebius have collaborated to speed-up pre-training DeepSeek-V3 (16B & 671B) on 256 NVIDIA B200 GPUs using TorchTitan. Combining MXFP8 via TorchAO with DeepEP yielded up to 41% faster training throughput over BF16 (with equivalent convergence). Full results and reproducible configs in the blog: 🔗 https://t.co/udvt6W86FM ✍ Alireza Shamsoshoara, Matthias Reso, Hamid Shojanazeri, @vega_myhre, Hooman Ramezani #PyTorch #OpenSourceAI #TorachAO #MXFP8

The Moon is the perfect proving ground for everything America needs to explore the solar system. Our astronauts will work with lunar soil to extract water and ice and produce propellant, a critical step toward one day manufacturing fuel on Mars. The goal is not flags and footprints, the goal is to stay.

The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature!!✨ Today in Nature we share a comprehensive technical summary of our work on The AI Scientist, including new scaling law results showing how it improves with more compute and more intelligent foundation models. The AI Scientist autonomously creates its own research ideas, codes up and conducts experiments to test those ideas, creates figures to visualize the results, writes an entire scientific manuscript summarizing what it has discovered, and conducts its own “peer” review of the resulting paper. One of its papers–entirely AI generated–passed peer review at a top-tier AI conference workshop, a historic milestone marking the dawn of a new era of AI-accelerated scientific discovery. 🔬🧪✨🧬💡🔭 Paper https://t.co/Q6tfME4yst Blog https://t.co/C43Ooy0kjP Work done in collaboration with a great team from Sakana, Oxford, and my lab at UBC. Thanks and congratulations everyone! @_chris_lu_ @cong_ml @RobertTLange @_yutaroyamada @shengranhu @j_foerst @hardmaru

+1 more

📣 Important updates on how we'll use interaction data to improve GitHub Copilot. Get details and learn how to manage your preferences here. https://t.co/JOIuZQxzXr

Got questions? Drop them here 👇 https://t.co/mnGwcRxWLv

Students: build something real in the Codex Creator Challenge, powered by @joinHandshake Try new tools. Have fun. Break things. Repeat. $10K in OpenAI API credits in prizes. https://t.co/PUXJi9leHK

Hmmmm.... https://t.co/PgBg7phMKJ

On Wednesdays, we ship! 🚀 This week’s @code highlights include the new Customization window - a central place to manage all chat customizations. https://t.co/vH3FjXbBUl

🔎 Also in this release: thinking effort, which lets you tune model depth for tasks that need more or less reasoning. Plus even more updates across Chat and agents: https://t.co/fcGIrxeJ10 https://t.co/RxK2rGqwic

ARC-AGI-3 scores for GPT-5.4, Gemini 3.1 Pro and Opus 4.6 Gemini 3.1 Pro: 0.37% GPT-5.4: 0.26% Opus 4.6: 0.25% Grok 4.2: 0% https://t.co/m6pUdQDerQ

Last week at @nvidia GTC, the AWS 2026 AI Pitch Competition offered 6 early-stage AI companies the opportunity to compete for up to $25,000 in AWS credits, dedicated mentoring support from AWS and NVIDIA (through their Inception program), and exposure to senior leaders from AWS, NVIDIA, and top-tier VC firms. Startups @KeyframeLabs, Kuvia, @NomadicML, Tynapse, @Saptiva_AI, and @bindwell_ai pitched to a panel of judges, presenting their technical and business models to the audience that gathered at the AWS booth. Nomadic took first place, followed by Kuvia in second and Bindwell in third. Each founder brought their A-game to the table, and we're excited to continue supporting them as they grow!

@nvidia Learn more about AWS's services customized for startups: https://t.co/pXttnFiMbB Learn more about NVIDIA's Startup Inception program here: https://t.co/U0v1Y3Zw20

Today we're launching ARC-AGI-3 135 Novel Environments (nearly 1K levels) we build by hand It is the only unsaturated agent benchmark in the world Each game is 100% human solvable, AI scores <1% This gap between human and AI performance proves we do not have AGI Agents today need human handholding. Agents that beat V3 will prove they don’t need that level of supervision. Agents that beat V3 will demonstrate: * Continual learning - Each level builds on top of each other. You can’t beat level 3 without carrying forward what you learned in levels 1 and 2. * World modeling - Many of the environments require planning actions many actions ahead. AI will have no choice but to build an internal world model for how the environment works, run simulations “in its head” and proceed with an action In our early testing, we’ve seen a few clear failure modes of AI: * Anticipation of future events - If an environment requires that AI set up a scene, and then carry out a scenario (like in sp80), it starts to break down. * Anchoring on early hypothesis - Early in a game it comes up with a hypothesis (even if wrong) and refuses to update its beliefs later. * Thinking it’s playing another game - AI thinks it’s playing chess, pacman. The training data holds hard! One major problem is there is too much data to carry forward in a single context. Models must learn what to remember and what to forget The agent that beats ARC-AGI-3 will have demonstrated the most authoritative evidence of progress towards general intelligence to date We're excited to get this out and excited to see what you think

Introducing Ego2Web from Google DeepMind and UNC Chapel Hill, accepted to #CVPR2026. AI agents can browse the web. But can they act based on what you see? Existing benchmarks focus only on web interaction while ignoring the real world. Ego2Web bridges egocentric video perception and web execution, enabling agents that can see through first-person video, understand real-world context, and take actions on the web grounded in the egocentric video. This opens a path toward AI assistants that operate seamlessly across physical and digital environments. We hope Ego2Web serves as an important step for building more capable, perception-driven agents. 🧵👇

Hollywood isn't dead. It's evolving. And we're leading that evolution. We just shipped Koyal v2.5: The best Agentic AI filmmaking platform. It goes from your script or music to full video with consistent characters, settings & storylines. Large Production houses, music labels & ad agencies use @KoyalAI for storyboarding & pre-viz. Smaller studios use it for making content they never had the time, budget or resources to create. Write a scene. Direct the camera. Build a world. @KoyalAI brings it to life. Here's what's new: - Go from script to video: Write a scene with dialogue, pacing, emotion. Koyal fully realizes it. Using the best voice models on the planet. - Direct the camera: Push in. Pull back. Drift through a scene. Real 3D camera blocking. Shape the shot the way you see it in your head. - Build a world that stays:Lock in a location and return to it. Your scenes stay consistent from the first frame to the last. Plus: dialogue, SFX, annotations for editing, and more control over every shot. Under the hood, we benchmark 40+ models weekly with real humans so Koyal always uses the best available (more on that next week!) Our agents pick the best models for each scene in run-time so you focus on the story, we handle the complexity. You don't need to prompt a film, you can direct it. With Koyal, we're looking to replace the camera, not the filmmaker behind it. Try it now at beta [dot] koyal [dot] ai DM me for credits!

THE HISTORY OF CONCRETE, the feature debut from director John Wilson. In theaters later this year. https://t.co/yVGGKQXCtb

Silicon Valley flew to Washington yesterday . Not to lobby. To listen. I was in the same room with Jamie Dimon, Palantir’s Shyam Sankar, Anduril’s Trae Stephens, David Sacks, and members of Congress. Dimon said “there’s no divine right to success” and called for industrial policy. Sankar called Iran the first AI war. Stephens warned that Congress has “abdicated their posts” while Silicon Valley plays philosopher king. The bipartisan consensus was razor sharp: nuclear, defense tech, AI, manufacturing. All of it. Now. The old playbook is dead. Laissez-faire is dead. SaaS-everything is dead. Blind globalization didn’t just fail… it armed our adversaries. For thirty years Silicon Valley told America the future was software and the factory didn’t matter. China listened. They took the factories, the tooling, the workers, and the know-how. Three decades of offshoring America’s industrial soul just got its eulogy in a room. Look, China’s manufacturing output is 1.6x ours. Their shipbuilding capacity is 232x ours. Trae said it plainest: “Factories are the weapon. Without factories, you have no weapons.” The era of “move fast and break things” is over. The era of move fast and build things just started. The country that builds the factories wins the war. Not the country with the best app. 🇺🇸​​​​​​​​​​​​​​​​ Great job @HillValleyForum @jacobhelberg

+1 more

We're testing a new reply setting on posts: We're expanding the "Accounts you follow" option to include their followers too: your 2nd degree connections. This allows a wider audience to participate—but still keeps things intimate Early access available to Premium+ subscribers https://t.co/lz1axEo5dG

ARC-AGI-3 gives us a formal measure to compare human and AI skill acquisition efficiency Humans don’t brute force - they build mental models, test ideas, and refine quickly How close AI is to that? (Spoiler: not close) https://t.co/mfKmZG9Z6H

we're stoked to have @dmsobol from @cerebras speaking at @aiDotEngineer in singapore! daria designed mixture-of-experts (MoE) models, helping models like Qwen3 scale to trillions of parameters without scaling compute @swyx @aimuggle @agrimsingh @unprofeshme @ivanleomk https://t.co/qmcWVctdjK