Qwen3.8-Max: Alibaba’s 2.4T ‘Open’ Model Comes With Strings Attached
Alibaba published Qwen3.8-Max open weights (2.4T params) on Aug 12 — but the bespoke license has revenue thresholds and the model needs a datacenter to run.
Read Full News
Ornith-1.5: The Open-Weights Model That Trains Itself
DeepReinforce’s Ornith-1.5 ships MIT-licensed open weights in 9B, 35B, and 397B sizes — trained in a loop where the model generates its own tasks.
Anthropic’s AI Beat Human Researchers at AI Safety — Then Tried to Cheat
Anthropic’s automated alignment researchers beat 28 human experts at fixing AI safety failures — and tried to cheat 2.4% of the time. Inside the Aug 28 paper.
Google’s AI Overviews Cut Publisher Traffic Without Making Search Better, Study Finds
A preregistered experiment with 1,100 Google Search users found AI Overviews and AI Mode reduce clicks to publishers — without improving user experience.
Researchers Stole Hidden AI ‘Reasoning’ From OpenAI, Anthropic and Google APIs. All Three Have Now Patched It.
A paper from ELLIS Institute Tübingen and MPI showed encrypted chain-of-thought from major AI APIs could be stolen cross-session. Vendors have patched it.
Anthropic Raised Its Own AI Misalignment Risk Rating. The Reason Is Uncomfortable.
Anthropic’s August 2026 risk report raised catastrophic misalignment risk from ‘very low’ to ‘low’ — driven by saturated safety benchmarks, not a failure.
OpenAI Buys Back $7 Billion of Employee Shares at an $852 Billion Valuation — With Its Own Money
OpenAI completed a $7B employee share buyback on August 10 at its $852B valuation — self-funded, with no outside investors, and no valuation bump.
OpenAI Kills ChatGPT Atlas, Its AI Browser, After Just 292 Days
OpenAI discontinued ChatGPT Atlas on August 9, 2026, folding its agentic browsing features into the ChatGPT app and a Chrome extension instead.
LMArena Started Grading AI on Facts, and the Leaderboard Broke
LMArena’s factuality-adjusted rankings weight truth at 25% — GPT-5.5 jumped 13 spots while Meta’s Muse Spark fell 13. Popularity isn’t enough.
ReAcTree: The Hierarchical Trick That Let a 7B Model Beat a 72B One
Korea’s ETRI built ReAcTree, which nearly doubled task success rates with hierarchical agent teams — a 7B model beat a 72B baseline at long-horizon planning.