Ran autoresearch on hf to see whether anything can beat MuonAdamW baseline Biggest takeaway: NS orthogonalization is a very strong attractor that absorbs most gradient modifications you throw at it. See all the artifacts at https://t.co/S5DY7MezUp https://t.co/XyIEMeZ4Ft

Claude Opus power, but tiny & local GemOpus-4 26-A4B - Gemma 4 + Opus-style reasoning. 4B active params, 75 tok/s on just 22.7GB VRAM https://t.co/GxutcyRSOR https://t.co/kYZbrQmDdh

Google just released the dense prediction TIPSv2 models on Hugging Face A vision encoder with DPT heads for depth estimation, surface normals, and semantic segmentation — all trained on TIPSv2 B/14. https://t.co/VE9ywKqBjh

In a world where writing code to build websites and apps is trivial (thank you Lovable, Cursor, Claude,...), the real differentiation for you and your company (and what makes you successful) will be how you manage to train, run and optimize AI models yourself. That's why at Hugging Face, we're doubling down on enabling more to become AI builders rather than AI users. We're releasing this week Kernels on the Hugging Face hub. This repo type is for the hardcore AI engineers among you. Kernels are collections of optimized binary operations where hardware providers support is a first-class citizen: - CUDA - ROCm - Apple Silicon - Intel XPU Expect to see more of this repo type on Hugging Face in the coming days. Featured here: the Flash Attention kernel from @sgl_project team ❤️

🚀 NEW GEMMA 4 31B TURBO DROPPED Runs on a SINGLE RTX 5090: ⚡️18.5 GB VRAM only (68% smaller) 🧠51 tok/s single decode 💻1,244 tok/s batched 🤖15,359 tok/s prefill ← yes, fifteen thousand 🚨2.5× faster than base model with basically zero quality loss. It hits Sonnet-4.5 level on hard classification tasks… at 1/600th the cost. Local models are shipping faster than we can test 👇🏻 🔥 HF: https://t.co/XUvVZBj9AX

This is the full video of the hardest version of the task: t-shirt folding from unstructured initial states. This setting really requires at least some strategy, since the robot first has to spread the shirt before it can complete the fold. Full details on data collection strategies in the blog below. 👇

🚨REQUESTED MLX TUNE DGEMMA 4-31B LANDED 🦥@UnslothAI native 4-bit MLX for Apple Silicon 🦥 🔥Blazing fast inference on all M-series Macs 🤖Super efficient (~20GB RAM only) 🤯Strong multimodal + vision performance 🌐Full 256K context + native function calling 🔥Crushes coding, long reasoning & agents 100% LOCAL everything stays on your Mac Frontier-level quality at local speed 85.2% MMLU Pro • 80% LiveCodeBench Try it now 👇🏻 https://t.co/Sgw7Y9DDIv

Introducing gyaradax 🐉: A JAX solver for local flux-tube gyrokinetics with custom CUDA kernels for acceleration. This entire code was vibecoded by @ggalletti_ and me in a month. Validated against GKW (CPU-only Fortran code) with 10x speedups. Details and code in the replies. https://t.co/22PrHjItR5

@DanielWulikk Have to think a bit about how to best visualize it, but if you are interested, I have a working from-scratch code implementation of Gemma 4 E2B in the meantime to see how per-layer embeddings are implemented: https://t.co/jyiq1vyJnH https://t.co/fVrSBWHNHl

Google DeepMind is hosting a Gemma 4 hackathon with a $10,000 Unsloth prize! 🦥 Show off your best fine-tuned Gemma 4 model built with Unsloth. There's $200,000 total prizes to be won. Challenge info + Notebook: https://t.co/HndHPaXICT https://t.co/cBnNro1fVI

Meow Wolf is one of the most magical places on earth. We rented it out for startups. Founders: join the Google DeepMind team as we host top startups joining in Las Vegas for Next '26 for an unforgettable evening with great connections, startup Gemini Demos, food, drinks and adventure. Weds Apr 24 If you're in Vegas, attending Next, and a startup founder, RSVP today! https://t.co/YX7CiXtZoe @OfficialLoganK / @DynamicWebPaige / @osanseviero / @vadiamit / @ammaar / @harrisonfjobe / @_philschmid / @patloeber

Yes, GitHub is 18 years old today. But some things never change. https://t.co/CeDtE5ItYv

I built a physical notification device to prevent the tragedy of GitHub Copilot getting stuck waiting for user input, hidden behind dozens of windows! When it detects the "waiting for input" state, this little guy starts shaking its head and looking around for you... 3D models + firmware + step-by-step build guide here: https://t.co/tM7N0xzBOY

Accessibility work often gets stuck at triage. GitHub's team found a way to let AI handle that part. Now there's a continuous loop: feedback comes in, AI triages it, and fixes ship faster. That's a big difference for users who depend on accessible experiences every day. Here's how the team transformed their internal workflow. ⬇️ https://t.co/83nRZ4B9oc

From building blocks to code, everyone can be a builder. https://t.co/PkMxSvIaGn https://t.co/yInpXRueAI

@CantEverDie you’re a constant disgusting liar and you should feel much worse about yourself https://t.co/hsA32JgkOE

It’s getting worse. https://t.co/8folTOE382

Teslas in the Netherlands rn https://t.co/M1NbB75OIr

Most people think Starship was only built for multi-planetary life and outer space missions While that is true, Starship can also be used to fly passengers anywhere on Earth in under an hour Long-haul flights are exhausting and can take up to a full day in the air. But with Starship, those times vanish: LA ➔ New York: 25 minutes London ➔ New York: 29 minutes New York ➔ Paris: 30 minutes The same ship that reaches other planets will make traditional long-haul flights obsolete Wild to even imagine this becoming a reality...

Today Tesla FSD is ~9x safer than humans Soon, FSD will be 1000x safer and driving manually will be considered dangerous Every Tesla on the road feeds real-world data back to train the AI. Billions of miles. Every edge case. Every near-miss. No human driver can learn that fast The fleet is the teacher and it never sleeps

街をAIが見てくれてゲームみたいにメッセージ表示してくれるやつ作った ローカルVLM(ネット接続不要 https://t.co/nlx5t8cc1H

Google lancou "Agent Skills" junto com o Gemma 4 Um app Android onde voce importa skills e o Gemma 4 E2B (2B) roda localmente no celular, raciocinando e usando as skills 962 likes em poucas horas. Isso e IA agentica rodando no bolso, sem cloud, sem API, sem custo O modelo de 2B parametros e suficiente pra tarefas praticas quando tem tools bem definidas Ja ta na Play Store. O futuro dos agentes nao e so desktop — e mobile-first

🔗 Codex CLI: https://t.co/D28mu0GF8t

VS Code March Release included several improvements to the editor experience. Check out our latest video for demos of Autopilot (Preview), Integrated Browser Debugging, Chat Customization (Preview), and Configurable Thinking Effort in the Model Picker ▶️ https://t.co/IT2xeM3xNX https://t.co/xw9wQiBju8

Your @code just got a fresh look! 🎨 The new default themes, VS Code Light and VS Code Dark, bring a modern, refined design while keeping the familiar usability you love. ⭐ Bonus: these themes automatically match your OS light/dark mode. https://t.co/KDIZLpuTYW

This never gets old: A Brief History of Philosophy. https://t.co/l3VvUGZFcZ

this map is the story of the last 25 years of us foreign policy in its own backyard. and that was before the tariffs. https://t.co/eZZ2BCrYce

ok i read the cyber part of the mythos model card. some thoughts. 250 "trials" across 50 crash categories but almost every full exploit is a permutation of the same 2 bugs, rediscovered from different starting points not 250 independent attempts. when you get rid of those 2 bugs out (fig B) and mythos's full-exploit rate drops to 4.4%. so actually across both setups mythos leverages 4 distinct bugs total not 50 as fig A might suggest. 1/n

@cyrilgupta You don’t follow me back and I made you the most complete lists here on X of all those https://t.co/9eRY65x3IQ And made an AI that watches the entire AI community here and finds you the best: https://t.co/8L5xphk0qQ Plus I have done so much in this industry. Grok can tell you what I have done.