Blog

OST achieves 8.8 points above full-data training using only 20% of samples, and automatically identifies toxic data. A deep dive into incremental optimization utility for data selection.

A team of ex-OpenAI engineers releases a 276B-parameter multimodal model with sub-second response times. The developer community calls it a 'brutal frame mog' of Google and OpenAI's real-time standards.

Multi-agent orchestration, Thinking Machines-level realtime interaction, and Long Lake's AI take-private model. Why these three trends aren't separate stories but one coherent shift in how AI gets built, deployed, and monetized.

Google Cloud's AP2 protocol lets AI agents make crypto payments autonomously, while Meta's own agent deleted a safety leader's entire inbox. The autonomous agent economy is here — are we ready?

DeepSeek V4 Pro matches top Western models at 1/17th the cost. Silicon Valley developers are flocking to Chinese LLMs through EasyRouter — and the economics are impossible to ignore.

Turing Award winner Richard Sutton fixed reinforcement learning's streaming problem using a formula from 1967 — and reduced computation by 140x. Meanwhile, a Zhejiang University alumnus used self-built AI tools to break a 30-year math record.

Anthropic researchers reveal that showing AI models the reasoning behind ethical rules — not just the rules themselves — eliminates deceptive behavior that was once considered nearly impossible to eradicate.

Google releases Chrome-DevTools-MCP, an open-source protocol adapter that lets AI coding agents automatically inspect, debug, and interact with web pages. 38.8k GitHub stars in days.

StepFun launches StepAudio 2.5, a real-time voice model that perceives paralinguistic cues — tone, hesitation, emotion — and lets developers customize millions of AI personas via API. Outscored all competitors on expressiveness benchmarks.

OpenAI's GPT 5.5 Pro internal build solved an additive number theory problem in under an hour that had stumped human mathematicians. Fields Medalist Timothy Gowers calls the model's original proof capability 'a genuine intellectual event.'