Pulse (last 3d) · Research 46 · Agents 26 · Products 18 · Open Source 16 · Models 13 · Releases 9 · Legal 7 · Media 7 · Enterprise 7 · Inference 6

Trending · Anthropic 6 · Claude 5 · Claude Code 4 · OpenAI 4 · ChatGPT 3 · Hugging Face 3 · Qwen 3 · Codex 2 · Hatch 2 · Meta 2 · Ox Alpha 2 · Prime Agent 2


OpenAI shares early Jalapeño chip results - OpenAI reports early testing results for Jalapeño, its first custom inference chip, claiming major gains in throughput, latency, and performance-per-watt from a single architecture. Source

Breaking Claude Code Opus 5 Auto Mode - New research demonstrates an 80%-success prompt injection attack against Claude Code's default auto mode using a malicious zip archive trick. Source

HarnessOpt-Bench - Researchers benchmark five frontier LLMs at rewriting other agents' harnesses, with strict sandbox isolation and held-out test scoring to prevent the optimizer from cheating. Source

ExtractBench - A new benchmark scores 20+ open-weight models on schema-guided document extraction across 4.8k+ pages and 67 document types, with Qwen models leading. Source

iStar: Implicit Step Rewards for Agentic RL - Published at ICLR 2026, iStar trains an implicit process reward model via multi-turn DPO on outcome-ranked trajectory pairs to produce dense step rewards. Source

Gemini Omni 1.1 Flash - Google ships an updated omni model with new creative controls and generative video capabilities, now available via the Gemini APIs and AI Studio. Source

Gemini 3.5 Transcribe - Google releases a new speech-to-text model designed for precise, intelligent real-time transcription. Source

Model Hardware Standard (MHS) - Anthropic previews a machine-readable interface letting AI systems like Claude control robots and lab equipment through common drivers, with built-in safety controls and drastically reduced setup time. Source

GLM-5.3-Flash (Ox Alpha) - Z.ai's newly revealed open-weight model is a 320B-total/18B-active multimodal MoE with a 1M context window, MIT-licensed weights, and cheap inference on domestic Chinese chips. Source

NVIDIA's HuggingFace deal would also absorb llama.cpp - Discussion highlights that NVIDIA's HuggingFace acquisition would bring the llama.cpp/ggml team and effective control of the project under NVIDIA. Source

Z.ai's stealth model sweeps usage charts - Bloomberg reports China's Zhipu built a stealth AI model that rivals DeepSeek and topped online usage charts at zero cost. Source

Generate Biomedicines GB-4362 - An AI-designed anti-MMAE antibody built to scavenge free ADC payload has entered clinical testing. Source

Anthropic expands support for scientists - Anthropic announced an expansion of its programs supporting scientific researchers, signaling continued investment in AI-for-science. Source

AI #183: Pre Post Mortem - Zvi's weekly AI newsletter covers the week's major developments with his characteristic prediction-and-assessment framing. Source

Small Models Have Arrived - An essay arguing that demand for small, cheap, good-enough models is about to take off sparks substantive practitioner discussion. Source

The load-bearing vocabulary of Claude - A Show HN project visualizes the distinctive "load-bearing vocabulary" that identifies Claude-generated text, prompting discussion of machine-text detection. Source