Skip to content

Latest News

OpenAI Releases MedChemBench, Reports GPT-6.2 Astra at 84.5%

OpenAI Releases MedChemBench, Reports GPT-6.2 Astra at 84.5%

Nov 20, 2026
OpenAI

OpenAI releases MedChemBench, a 99-problem benchmark built from 117 papers and patents and 761 molecular structures that tests whether AI models can reconstruct molecules and predict compound properties. OpenAI reports GPT-6.2 Astra earns 84.5% of available credit at the xhigh setting, versus 58.2% for GPT-5.6 Sol. OpenAI says the results do not establish whether models can discover new drugs.

Anthropic's Levent Alpöge Claims Claude Proved Riemann Hypothesis, Lean Proof Compiles

Anthropic's Levent Alpöge Claims Claude Proved Riemann Hypothesis, Lean Proof Compiles

Nov 20, 2026
Scientific American

Anthropic mathematician Levent Alpöge claims Claude produced a proof of the Riemann hypothesis, and independent researchers including MIT's Andrew Sutherland have compiled the Lean version without errors. Sutherland says the chance anything is fishy is near zero, but a comparator run checking the proof's statement and kernel is still pending.

Google Says Gemini for Mac Can Click and Type Across Apps

Google Says Gemini for Mac Can Click and Type Across Apps

Nov 20, 2026
Google

Google says computer use in its Gemini app for Mac now clicks, types, and fills forms across local files, desktop apps, and the web in an isolated window users approve step by step. It is limited to English in the U.S., Apple Silicon Macs on macOS Sequoia 15.0, and Google AI Plus, Pro, and Ultra subscribers.

Chrome Engineers Say Mirror Dimension Yields 1,000 Attack Tests Daily per Engineer

Chrome Engineers Say Mirror Dimension Yields 1,000 Attack Tests Daily per Engineer

Nov 20, 2026
Google

Chrome engineers Aaron Selya and Charles Qian say the team built a synthetic web, called the mirror dimension, to test its auto browse agent without exposing the live internet. Minted TLS certificates and a local interception proxy keep real URLs intact. They claim the setup generates over 1,000 attack candidates daily per security engineer, about 30 times manual testing.

METR Says AI Research Speedups Doubled AI Progress in 2026

METR Says AI Research Speedups Doubled AI Progress in 2026

Nov 20, 2026
substack.com

Independent research group METR, using redacted Anthropic evidence, reports a 2x acceleration in AI capabilities in 2026, mostly because AI speeds up AI research. METR researcher Ajeya Cotra estimates a 10x research speedup, and Anthropic staff report 12x. The Linchpin warns this could produce superhuman AI by 2028 or 2029.

DeepMind Researcher Says Four Paths Could Carry AGI to Superintelligence

DeepMind Researcher Says Four Paths Could Carry AGI to Superintelligence

Nov 20, 2026
DeepMind Institute

Google DeepMind researcher Tim Genewein argues in a DeepMind Institute essay that four routes could carry AI from human-level AGI to superintelligence: scaling models and data, paradigm shifts, agent coordination and recursive self-improvement. He says global effective compute stock has grown at least 10x per year, and that cognitive industrialization could commoditize much knowledge work.

Page 1 of 626
Next
Showing 1 - 10 of 6251 articles