Your curated collection of saved posts and media
Speculative decoding for Gemma 4 31B (EAGLE-3) A 2B draft model predicts tokens ahead; the 31B verifier validates them. Same output, faster inference. Early release. vLLM main branch support is in progress (PR #39450). Reasoning support coming soon. https://t.co/PoK8zbA7li
A new PyTorch-native backend is coming to unlock the power of Google TPUs: ✨ Run existing PyTorch with minimal code changes. ✨ Get a 50-100%+ performance boost with Fused Eager mode. Read the engineering deep dive here: https://t.co/GQPRYaKz7E #TorchTPU #PyTorch #MLOps #AI https://t.co/HiIdXVw6Oh
Codex just got a lot more powerful. Computer use, in-app browser, image generation and editing, 90+ new plugins to connect to everything, multi-terminal, SSH into devboxes, thread automations, rich document editing. Learns from experience and proactively suggestions work. And a ton more.
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 ⬇️
To show off what you can do with @OpenAI Agent SDK + @modal, we built an ML research agent (inspired by @karpathy). It can: - Spin up GPU sandboxes of any shape - Run a pool of subagents - Persist memory - Snapshot state for fork/resume Here it is playing Parameter Golf: https://t.co/r7QhvNmdEq
ok i've barely used cursor the past 18 months.. but today, i pulled it up and asked it to implement one of the more ambitious projects i had in mind .. something that really pushes it, y'know? the project was a nextjs web app, with a chrome extension + tauri desktop app (mac/win/unix) spent 10 mins working on a spec for the work, and left the thing to run -- checking in at key points to make sure everything worked. all told, it worked for like 3 hours.. and in the end, i had a suite of fully working apps! and the crazy bit is ... i don't even know which model it used?! i just picked "Premium" > "Max mode" (i have $60k in credits to burn and 6 months to do so) needless to say, i'm going to be using cursor alot more from now..
We’re introducing Cursor 3. It is simpler, more powerful, and built for a world where all code is written by agents, while keeping the depth of a development environment. https://t.co/rXR9vaZDnO
New video model HappyHorse-1.0 by Alibaba-ATH debuts at #1 in Video Edit Arena. It scores 1299, leading Grok Image Video by +42 points and Kling o3 Pro by +48 points. Video editing is an emerging frontier capability for video models, and only a small number of models support it today. Huge congrats to the Alibaba-ATH team on this incredible milestone!
HappyHorse-1.0 is now live on Arena! 🚀 Early evals show exceptional performance in Video Edit. We are now in the final optimization sprint for the official launch in 2 weeks. We invite the community to get early access and test our capabilities at https://t.co/iiyfgPtib5. 🐎✨
OpenClaw 2026.4.10 🦞 🧠 Active Memory plugin 🎙️ local MLX Talk mode 🤖 Codex app-server harness plugin 🧾 Teams pins/reactions/read actions 🛡️ SSRF hardening + launchd fixes stability, but with attitude🦞 https://t.co/PW7WDumTf1
Now the masked_token_weighted is learning. We ablated the inpainting task, swapped MSE for SmoothL1Loss (more robust to outliers), and per-dim normalized the reconstruction targets, significantly reducing curvature-dim dominance. ref: https://t.co/FL5X61xpbQ https://t.co/0j03IXFXR2
Training a 10M params foundation model on 8xH100s. The regime is self-supervised pretraining on 29GB of CAD and engineering meshes with masked token modeling, contrastive consistency, and spatial inpainting. You could guess what it is for. https://t.co/y9c9v5d8fN
@reza_byt Doesn't tokenwise looped transformer have issues with pretraining since each token has a different depth and also has to learn the recursion depth?