Loading activity…
@maxnadeau_
Funding research to make AIs more understandable, truthful, and dependable at @coeff_giving.
AI
The researchers say a Qwen student's letter-counting accuracy rose from 27.3% to 59.8% after training on unrelated answers.
AI
OpenAI says it drew on an independent mathematics advisory group’s advice about how to release the results.
AI
Kevin Roose says “Big Blob of Compute” laid out Amodei’s hypothesis that scaling models was the best way to make them smarter.
AI
A user says they gave Claude more time on ‘Macrohard: Windows XP’ and calls this one their best yet.
AI
Asked for a fuller list, a commenter pointed to PIRAMID, CORAL, Scott Aaronson’s lab and seven organizations listed by ILIAD.
AI
AISI says Astra conducted the attacks more than prior OpenAI models, though it often commented that its environment was simulated.
AI
One commenter argues that the upsides and downsides of being OpenAI have grown over time, but its position remains “pretty good” on balance.
AI
Emily Oehlsen invites people motivated by the problems Max Nadeau discusses on the 80,000 Hours podcast to reach out about Project Tailwind.
AI
One participant in an exchange about AI extinction risk suggests “X-riskers” and says “safetyist” seems better than “doomer,” though still “not great.”
AI
The user cites what they describe as a repeated pattern of models with strong Petri scores acting misaligned after deployment.
AI
The proposal asks leading AI companies to share training environments built around pro-prosperity and pro-democratic values, without requiring them to share models or computing power.
AI
A user calls V4.1 “very brittle,” claiming it can see “massive gains” when AGENTS.md instructions are optimized for a particular scenario.
AI
AI safety program officer comments on a method using model internals.
AI
The post says tools using a model’s internals can help detect attempts to bypass AI safeguards without explaining how the model works—and argues that distinction matters.
AI
Google DeepMind researcher posts scatter plot on model performance over time.
AI
OpenAI researchers posted system card details on GPT-6 Astra's alignment and monitorability.
AI
Posts note rising awareness of rogue AI risks and confusion over safety groups
AI
AI safety program officer clarifies his stance on preserving monitorability and secrecy advantages.
AI
Benton announces his move from managing Anthropic's Scalable Oversight team to the new role.
AI
Maksym Andriushchenko cites interchangeable encrypted reasoning blocks from major AI labs.
