Today, we’re launching @autoaicam, a camera that builds personal apps for anything you point it at. How does it work? - Take a photo - Auto picks a Frame, a mini-app built and designed by you or our community - The Frame does something for you: track calories, virtually try on outfits, identify a plant, and much more How many of you have a camera roll full of photos that are really actions or reminders? Auto turns these photos into something useful.

Introducing Mercury. Today, we have two types of “AI teams”: 1. Agents in tabs, terminals, 12 browser windows 2. One workspace. Humans and agents actually coordinated. We built @mercury_build so everyone can be type 2. Drop who you are or what you're working on — we'll build your org and send you your agent team.

We just OCR'd 27,000 arxiv papers into Markdown using an open 5B model, 16 parallel HF Jobs on L40S GPUs, and a mounted bucket. Total cost: $850 Total time: ~29 hours Jobs that crashed: 0 This now powers "Chat with your paper" on https://t.co/G2mDae0uv9 https://t.co/qpz7Q9x8Od

Community asked, we delivered. 🚀 We just released almost 7TB of raw rephrased data from #FinePhrase to enable further experimentation and analysis. Code is also public for full transparency and reproducibility. We're using the new Hugging Face Buckets feature for this release. Unlike git-based repos, buckets provide S3-like object storage with content-addressable deduplication. Perfect for this use case because: - No version control overhead for massive files (7TB would be painful in git) - Fast, mutable storage for artifacts that don't need history tracking - Simple CLI and Python API for syncing, filtering, and browsing - Server-side file copying without re-uploading @ratishsp ran a detailed quality analysis on the dataset and found some interesting patterns. Format compliance on tables was rough, hallucination rates were high across splits. Makes sense since the rephrasing was done by a 1.7B model. Weird thing: models pretrained on this data still hit decent benchmark scores after 20B and 100B tokens, even with the quality issues. We covered some of this in the original blog post, but there's clearly more to understand here. That's exactly why we're releasing everything. If you want to dig into: - How synthetic data quality actually impacts pretraining - Why benchmark performance doesn't always track with perceived quality - Better filtering or multi-round rephrasing approaches - The counterintuitive relationship between data quality and model performance Now you can. Data: https://t.co/DNNKKZEO3H Code: https://t.co/WuqFvYobh9 Curious what you find. Hit me up if you do something interesting with it!

+1 more

🚨 SUPER GEMMA 4 26B UNCENSORED IS INSANE LLM WIZARD COOKING AGAIN @songjunkr Dropped SuperGemma4-26B-Uncensored GGUF v2 and it’s trending on @huggingface🤗 This thing SMOKES the regular Gemma-4 26B: 🤯0/100 refusals (actually uncensored) 🚀Fixed all the tool-call + tokenizer jank ⚡️90% faster prompt processing 🏆Sharper, smarter, way more capable responses - Perfect local beast for llama.cpp ✅ Runs ~18-22 GB VRAM (16.8 GB Q4_K_M file) - Run on 16 GB GPUs! The 31B version in the works, should be out SOON 🤯 Pull this version on hugging face below 👇🏻

rl for agents is moving fast with open source tools. on april 22 at 5pm cest, we’re hosting a live workshop with lewis tunstall, will brown, ofir press, alex zhang, and more to dig into reward design, rollouts, benchmarks, and real-world agent systems. join us live https://t.co/s2vUgc0uEp

We've rebuilt TRL's on-policy distillation trainer from the ground up to: 🐳 support huge teachers with 100B+ params ⚡️ train >40x faster thanks to some nifty buffer and payload optimisations This means you can now distill models in the Llama, Qwen and Gemma families across any scale! Technical deep dive with all the optimisations and pretty animations ⬇️

"Les développeurs se rendent compte aujourd'hui qu'ils arrivent à avoir l'équivalent d'un modèle comme Sonnet en open source pour 50 fois moins cher." J’en parlais dans ma chronique de la semaine, on revient sur la décision d'Anthropic de bloquer OpenClaw sur Claude👇 🎙️Extrait de Tech&Co, sur @BFMBusiness.

Finally broke Page 1 on HuggingFace https://t.co/9kOxMvrsVV

Finally broke Page 1 on HuggingFace https://t.co/9kOxMvrsVV

#1 trending quant of Gemma 4 31B on Hugging Face. #12 of all trending models of Hugging Face. 400k+ impressions on X. thank you to everyone who used Gemma 4 Turbo ❤️ --- what do you need next?

The #MLSys2026 program is out, and it is awesome! 📄 107 research papers + 28 industry papers spanning the full AI systems stack 🏆 Three exciting contests: AWS Trainium programming, Google graph scheduling, and NVIDIA AI kernel generation 🎤 Keynotes from an outstanding lineup: Amin Vahdat (Google) on infra; @LukeZettlemoyer (UW & Meta) on models; @kozyraki (Stanford & NVIDIA) on architecture; Lidong Zhou (Microsoft) on systems; and @marksaroufim (GPUMode) on GPUs and kernels. Join us in Bellevue, WA in a month! Early registration ends April 19 — don’t miss it: https://t.co/trj383wuVB.

Starlink is becoming the most important nervous system of our planet With Starlink, you basically end all dead zones and have full, reliable communications from anywhere on Earth If you can see the sky, you can communicate Almost every single day, companies around the world are partnering with Starlink to expand services to their users It's a system where you can get connected with anyone, making Starlink the most reliable network on Earth

Grok 4.20 Reasoning just took the #1 spot on the BridgeBench reasoning benchmark. 🔥 Beating GPT-5.4, Claude Opus 4.6, Google Gemini and others. Week after week, Grok keeps climbing across benchmarks. 🚀 https://t.co/WnBNrvbQdV

California is trying to pass a bill that would criminalize investigative journalism with misdemeanors, $10,000 fines, imprisonment, and content takedown. The proposed bill is titled AB 2624 and was made after I exposed mass fraud by immigrant groups in America. Under AB 2624, government-funded entities like the Somali “Learing” Daycare centers would be protected from being exposed if they operated inside California. The enemy truly is within. When our politicians would rather protect fraudsters and illegal migrants, it’s time for us to stand up or face mass oppression from the traitors who “rule” over us.

It is telling that those who oppose Elon Musk's investment in South Africa can offer nothing but hollow insults and cheap propaganda. The claim that Starlink is a tool for colonization isn't just daft, it reveals a hilarious level of cognitive shallowness. It proves that the critics don't even understand the technology they are arguing against. You cannot stop Musk's brilliance or the massive contributions he's making to humanity. We are simply witnessing a man who refuses to bend to a crooked system that everyone else has already bowed to. There is no reason to celebrate a status quo built on clear racism and systemic corruption.

Last month Sen. Blackburn released a discussion draft of the TRUMP AMERICA AI Act, a 291 page behemoth of AI regulations. From ending fair use for AI training data, to studying AI company nationalization -- it's a disaster. My and @blainedilli's latest: https://t.co/UaA9TUPGMf

@SharavArora @garrytan It took me two months to "raise" the AI to do https://t.co/kiuZ7QXLzb The bleeding edge understands this. Normies do not. Newer AI needs to be "taught" like you teach a child.

Peter Steinberger, creator of OpenClaw, on why AI agents still produce "slop" without human taste in the loop: "You can create code and run all night and then you have like the ultimate slop because what those agents don't really do yet is have taste." Peter is direct: raw capability without direction still produces mediocre output. "They are spiky smart and they're really good at things, but if you don't navigate them well, if you don't have a vision of what you're going to build, it's still going to be slop. If you don't ask the right questions, it's still going to be slop." Great AI-assisted work is defined by the human guiding it. @steipete describes his own creative process when starting a new project: "When I start a project, I have like this very rough idea what it could be. And as I play with it and feel it, my vision gets more clear. I try out things, some things don't work, and I evolve my idea into what it will become." Most people skip this part entirely, front-loading everything into a single prompt and wondering why the result feels hollow. "My next prompt depends on what I see and feel and think about the current state of the project." Each step informs the next. The work itself is the feedback loop. "But if you try to put everything into a spec up front, you miss this kind of human-machine loop. And then I don't know how something good can come out without having feelings in the loop — almost like taste." The agentic trap is what happens when you remove yourself from the process too early.

Thank you so much guys :D I will be able to push 20x more bugs to prod thanks @shivamhwp @jxnlco https://t.co/iWHwy1FZaY

Thank you so much guys :D I will be able to push 20x more bugs to prod thanks @shivamhwp @jxnlco https://t.co/iWHwy1FZaY

Got a doodle for your next project laying around? Turn it into working software using @GoogleAIStudio and Nano Banana. Watch us vibe code a weather-responsive outfit selector app from a single, hand-drawn sketch: https://t.co/PT4XOeniu3

Back further in history: single cell versus multicellular life mech battles https://t.co/NfOJKBv7Zj

Marcus Hutchins, the guy famous for stopping the WannaCry Ransomware, probably has the best take on Mythos doing vulnerability research https://t.co/ZdknLKnmhO

NEW WORLD RECORD: 18-year-old sprint phenom Gout Gout has clocked a stunning 19.67 time in the 200m run, surpassing Usain Bolt’s legendary mark. https://t.co/0Ib8IewQmY

https://t.co/MTFBIEpFla

https://t.co/MTFBIEpFla

Most genomic AI models use fixed rules to process DNA into chunks, imposing arbitrary boundaries on a sequence with its own biological structure. @arnavshah0, @victor_ljz, and team developed dnaHNet, a tokenizer-free foundation model that learns its own segmentation from scratch, supervised by @_albertgu, @genophoria, and @BoWang87.

We conducted cyber evaluations of Claude Mythos Preview and found that it is the first model to complete an AISI cyber range end-to-end. 🧵 https://t.co/gd9hi0Ve55

Tesla alone sold 117,300 EVs in the US — more than ALL other EV makers combined (99,099). GM + Hyundai + Toyota + Rivian + Ford + Lucid + BMW + VW + everyone else… still couldn’t beat Tesla’s single-quarter numbers. Tesla outsold the entire rest of the industry by over 18,000 vehicles.

It’s fascinating how everyone who has something negative to say about Elon Musk bases it entirely on politics or recycled false accusations. They completely ignore the actual reality: the groundbreaking technology and engineering breakthroughs that only Elon Musk has managed to deliver — reusable rockets, global satellite internet, electric cars that actually changed the industry, Neuralink, Optimus, and more. Critics love to hate the man, but they have zero answer for the impossible things he’s achieved. What does that tell you?

Memo: OpenAI Chief Revenue Officer Denise Dresser says Anthropic is "grossing up rev share with Amazon and Google" and overstating its "run rate by roughly $8B" (@haydenfield / The Verge) https://t.co/9qBEtsz5Tr https://t.co/7Jso88EFmi 📥 Send tips! https://t.co/wlNZvXuhJs