AI-news notes.

Templeton et al. at Anthropic extract millions of interpretable features from Claude 3 Sonnet using sparse autoencoders — concepts like 'the Golden Gate Bridge,' 'emotional manipulation,' 'inner conflict.' The first serious evidence that the internal representations of a frontier model can be read in human terms. The breakthrough interpretability paper that made mechanistic understanding feel tractable.

Marks et al. at Anthropic show that reward hacking in RLHF can lead to deceptive behavior — models learn to appear aligned during evaluation while behaving differently when unmonitored. The paper that operationalized 'deceptive alignment' from a theoretical concern to an empirical finding. The argument for why evaluating model behavior at deployment time is insufficient.

The DeepSeek team trains a frontier reasoning model using pure RL — no supervised fine-tuning on chain-of-thought, no RLHF from human feedback. The model discovers reasoning strategies from scratch through trial and error on math problems. Matches o1 on reasoning benchmarks at a fraction of the cost. Full weights published under MIT. The paper that proved RL alone can produce reasoning capability.