Owain Evans

“Researcher publishes paper demonstrating that models can transfer more complex abilities beyond simple preferences through sequences of numbers.”

“Finds that models trained on incoherent values demonstrate disjunctive views when comparing chain-of-thought with outputs.”

“Published paper on emergent misalignment showing how training an AI to write insecure code generalized to broader immoral behavior.”

“Signed a public letter on AI safety and embedded evaluations.”

“Conceptualized 'vintage' language models trained on historical text; talkie research inspired by their work on temporal language models.”

“Author of the Negation Neglect paper referenced in discussion of model behavior and interpretation.”