Meta's TLX Beats FlashAttention-4 on Blackwell GPUs by 50%
PyTorch's new Jagged Flash Attention kernel on Blackwell beats FlashAttention-4 by 13% forward and 50% backward using Triton Low-level Extensions.
20 stories
PyTorch's new Jagged Flash Attention kernel on Blackwell beats FlashAttention-4 by 13% forward and 50% backward using Triton Low-level Extensions.
Suno's new Speech beta generates spoken audio and matching background score as a single track, extending the music app into narrated storytelling.
Epoch AI released a public explorer tracking 8.3 million ChatGPT messages from 5,000 YouGov panelists, revealing how real usage intensified from 2022 to 2025.
Perplexity shipped an open-weight 27B decision model and a hosted API that returns calibrated probabilities over your options instead of free text.
A single closed-form math trick turns 2D rasterisation and gradient-based vector graphics into the same WebGPU shader, outperforming DiffVG and matching Slug.
Black Forest Labs ships FLUX 3 Image with bounding box layout control, multi-reference composition, native 4K output, and pixel-perfect multi-turn editing.
Claude Code now accepts TypeScript mods that intercept events, redraw the UI, and replace built-in features through the plugin system.
A new benchmark asks AI to find the handful of prior papers that actually inspire new research, and today's retrieval agents fall short.
VISTA gives multimodal models a lossless visual memory they can zoom into, lifting Claude Opus 5.0 to a perfect score on ARC-AGI-3.
A new training-free decoding trick for looped transformers contrasts early and late recurrent passes, boosting accuracy and cutting compute nearly in half.
Harvard researchers use MCMC sampling to reshape expert data into on-policy traces, letting supervised finetuning match or beat RL on generalization.
Odyssey's new framework pairs agents with a self-repairing world model, lifting StarCraft win rates by up to 91% and cracking the hardest robot coordination tasks.
Researchers from Tel Aviv University and MIT propose Local Support Learning, a gating-based method that lets LLMs learn new skills without erasing old ones.
Factory's Droid agent can now run recurring engineering workflows on triggers and schedules, with user-chosen models, machines, and connectors.
Goodfire's protein-embedding monitors catch more harmful biology requests with fewer false refusals, staying 3-5x more robust against adversarial attacks.
Field-test, Max effort, private repo to 512k. Reading 60/64 each. Kimi 8/8 vs DeepSeek 7/8 repairs. Bill $1.70 vs $31.75
An autonomous agent ran Hugging Face's July breach end to end — 17,000+ actions, no distinct identity checked at any step. Here's the exact point in that kill chain Ory Agent Security is built to close.
Microsoft AI's new streaming speech-to-text model tops the Artificial Analysis leaderboard with 2.5% WER delivered just 0.13 seconds after speech ends.
Most identity vendors shipped a brand-new, parallel system for agent security this year. Ory's bet is the opposite: reuse the identity stack you already run, and move the enforcement point inside the one place agents actually act — the harness event loop.
Perplexity's agentic Computer tool now renders interactive charts inline, using TradingView's Lightweight Charts for candlesticks, volume, and moving averages.