Pulse (last 3d) · Agents 90 · Research 74 · Models 58 · Products 51 · Enterprise 43 · Open Source 29 · Legal 24 · Industry 23 · Inference 17 · Media 17

Trending · OpenAI 23 · Anthropic 14 · Hugging Face 13 · Claude 12 · Claude Code 12 · Google 10 · Codex 9 · ChatGPT 6 · Cursor 5 · Gemini 5 · Gemini 3.5 Flash Cyber 4 · Gemini 3.6 Flash 4


AMD partners with Claude creators Anthropic, investing up to $5 billion to deploy 2 gigawatts of data center GPUs. Anthropic partners with AMD, with up to $5B investment to deploy 2 GW of AMD Instinct MI450 GPUs for Claude. Source

Google has released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Google DeepMind releases Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, halving time-per-task and improving token efficiency, with 3.5 Flash-Lite gaining 11 Intelligence Index points. Source

ChatGPT Voice is now in the desktop app. OpenAI brings ChatGPT Voice to the desktop app, letting users control their computer and coordinate multiple agents hands-free via GPT-Live. Source

Claude's voice mode is now available for Opus and Sonnet. Anthropic rolls out voice mode to Claude Opus and Sonnet subscribers. Source

Launching Health in ChatGPT. OpenAI launches a dedicated Health experience in ChatGPT. Source

GPT-5.5 Scores 10.6% on ActiveVision, Humans Hit 96.1%. ActiveVision benchmark shows GPT-5.5 scoring 10.6% and Claude Fable 5 at 3.5% versus 96.1% for humans, with models unable to recover via code. Source

LLMs Get Lost in Evolving User Intent. Paper showing LLMs struggle to track evolving user intent across multi-turn dialogues. Source

ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders. Benchmark for evaluating coding agents as interactive project builders across full project lifecycles. Source

SLPO closes the latent-reasoning RL gap. First method to apply outcome-reward RL to autoregressive latent reasoners, fixing bottlenecks that kept Coconut/CODI/CoLaR from composing with RL. Source

AMD takes on Nvidia with its Helios AI rack-scale system. AMD unveils Helios, a rack-scale AI platform aimed at competing with Nvidia's NVL offerings. Source

Deepseek V4 Flash ~105 t/s on two Nvidia 4090d 48G (ada) in vLLM. Triton re-implementation of Blackwell-only DeepSeek kernels runs V4 Flash at ~105 t/s on two RTX 4090D 48GB (Ada) via vLLM. Source

Fei-Fei Li's World Labs acquires real-to-sim robotics startup SceniX. The deal expands World Labs into physical robot manipulation. Source

Startup founders urge U.S. government not to shut off Chinese open weight AI. Startup founders publicly oppose a US ban on Chinese open-weight AI models, citing distillation concerns and innovation policy. Source

Prompt Injection in NeurIPS 2026?. NeurIPS 2026 OpenReview downloads appear to contain prompt-injection text and formulaic phrasing in reviews, possibly added by the platform. Source

AI #178: A Fire Alarm For General Intelligence. Zvi Mowshowitz's weekly AI newsletter covering the latest major developments and analysis. Source

The first known runaway AI agent - or a very bad marketing stunt?. Simon Willison investigates claims of a runaway AI agent, weighing a real incident against a PR stunt. Source