@adam_karvonen
Research at Truthful AI. Prev: Anthropic Fellows, MATS (Neel Nanda and Owain Evans streams)
AI
Anthropic says comparable chip work took dozens of engineers months, versus about three days with Claude.
AI
FlashProve aims to verify replies against a provider’s committed model without revealing weights, but proof generation takes substantial GPU time.
AI
The reported MATH-500 score rose from 7.9% to 48.1% with the same sampling settings.
AI
Astra's claimed robotics edge led one observer to speculate about transfer at scale; a reply on X questioned the surprise.
AI
The new team member says AI safety needs more basic research and calls Truthful AI a great place to do it.
AI
Adam Karvonen posted results from training models on self-generated explanations of in-the-wild behaviors.