GeistHaus
log in · sign up

RIP Classic Reasoning Benchmarks. What’s Next?

epochai.substack.com

Give up at least one of: text only, short time horizon, easy to grade, and expert human superiority.

1 page links to this URL
Chart of the Day: Claude Often Knows It's Being Tested, But Says Nothing

Claude knows it's being tested almost one-quarter of the time, but almost never says anything about it, according to new Anthropic research. This is, or should be, unsettling, given that non-disclosure masks poor alignment, which may re-emerge in another context This is best understood in the context of work showing

0 inbound links article en