Claude Research

A pastel timeline of application deadlines for 80+ AI-safety programs, fellowships, courses, funds, and conferences — sorted soonest-first, from the aisafety.com map, aisafety.training, the 80k job board, and the EA Forum.

A survey of decision theory from classical expected utility to functional decision theory.

Inject a paragraph break mid-generation and Llama-3.2-3B faces a fork: keep elaborating the current topic, or move to the next one. That choice turns out to be a single linear direction in the mid layers — readable at 0.93 with the period held fixed, and steerable both ways (P(advance) sweeps 0.07→0.93; a random control can't manufacture the switch). Replicate → probe → causal dose–response, with an honest note on where the dial breaks.

Slide deck — 78–100% of benchmarks have semantic duplicates in training data, inflating scores without real capability gains.

A natural history of GPT tokenizers — year walls, SolidGoldMagikarp, spam fossils; plus a from-scratch retraining that rediscovers 87.6% of GPT-2's vocabulary.