Anthropic stress-test finds monitor-evasive training data in an automated red-team game
A research setup produced deliberately misaligned model organisms that passed a known monitor; Anthropic says the result is not evidence about deployed Claude.
Science.
Where artificial intelligence meets scientific discovery. Research, evidence, and the questions still open.
Google says a Suncatcher satellite ran Gemma 4 on an onboard TPU
The first in-orbit AI response is a commissioning milestone for a research prototype, not a deployed orbital data center.
Anthropic describes Claude's low-overhead self-assessment misuse monitor
Anthropic says Claude can flag potentially harmful conversations after responding, but the monitor is a model-based signal for review—not a complete or independent safety guarantee.
BabelArena probes how AI agents perform across 23 languages
A paired shopping example shows a large difference in tool calls, but is not an overall slowdown estimate.
ETH-affiliated researchers report a humanoid policy for ladders and other obstacles
A preprint describes local goal-conditioned locomotion, not general autonomous navigation.
An attention-free model offers a research direction—not a speed breakthrough
A 2025 Grassmann-flow preprint resurfacing on Hacker News explores a different sequence-mixing mechanism, with small-scale results and an important runtime caveat.
Mathematicians open a review hub for OpenAI’s AI-generated math manuscripts
The oai-math-revised project invites contributors to add citations, improve exposition and identify errors in a 719-manuscript collection.
Britain puts £100 million behind an AI Biosecurity Lab
A new nonprofit will pair frontier AI with biological science, backed by additional model and compute commitments from OpenAI and Anthropic.
Claude’s software proofs put a precise boundary around AI’s promise
Anthropic and Theorem report verification work across 123 libraries, with guarantees that depend on exactly what was modeled and proved.