OpenAI chief scientist Jakub Pachocki says chain-of-thought monitoring — watching models write out their reasoning to catch misalignment — is becoming unreliable, and no lab is ready to scale safely.
Yahoo is rolling out AI answer engine Scout this year, using Anthropic’s Claude and Bing to win back Gen Z users.
Nvidia is shelling out $6 billion to license Poolside's Model Factory and scoop up 109 employees who worked on the Laguna model — while Poolside keeps its independence.
OpenAI previews Private Safety Processing, a zero-retention system that scans for abuse without letting OpenAI employees see prompts or responses. It's a direct challenge to Anthropic's 30-day data policy.
Herdr, a tmux-like tool built for managing coding agents, has joined Y Combinator's F26 batch. The solo-developed open-source project racked up 25K GitHub stars and 340K downloads in four months.
China's AI models took eight of the top ten token-usage slots on OpenRouter this week, with DeepSeek leading the pack.
Hold the Fn key and speak: Gemini for macOS now transcribes and edits text in any app, cleaning up filler words and even rewriting content on command.
Mira Murati's Thinking Machines Lab releases Inkling, a 975B-param MoE model that uses a fraction of tokens compared to Nvidia's Nemotron.
Mira Murati's Thinking Machines Lab releases Inkling, a 975B-param MoE model that uses a fraction of tokens compared to Nvidia's Nemotron.
Google’s TabFM can predict customer churn or fraud from structured data without any model training or feature engineering — just feed it a table.
