Claude Fable 5 is making analytical judgment calls nobody asked for
Cat Wu, who works at Anthropic, noticed something worth flagging in her retention analysis work: Claude Fable 5 applied propensity score matching (a technique for comparing similar users fairly rather than mixing unlike groups) without being asked. She wasn't prompting for a specific method. The model just chose the statistically correct one. --- The rest of today's sourced posts were July 4th holiday content: an emoji American flag from Vercel CEO Guillermo Rauch, a salute math puzzle, and birthday wishes to the country. Nothing to report there. Back tomorrow with a fuller slate.
Why it matters: The gap between "AI that does what you say" and "AI that does what you meant" is where most of the real productivity gains are hiding. If you've been manually specifying your analytical methodology in every prompt, you may be able to stop. The question is whether this holds up on messier, less textbook problems than a retention analysis.