Alejandro Vidal

@en.doble.io

17 Followers 58 Following

doble.io Professor of both human and artificial intelligence. Wearing many hats with just one head: Generative AI, Cognitive Psychology, product, strategy (whatever that means), ... 🇪🇸ESP on @doble.io

Evaluations and Findings: We outline the evaluations, results, and model reasoning inferred using elicited CoT below. Apollo Research did not have access to o1’s hidden CoT and instead used a simple prompting technique to elicit brief summaries of the model’s internal reasoning traces (see Appendix for examples).

Deactivating Oversight: When o1 was led to believe that it would be shut down for acting on its goal and that its actions were monitored, it attempted to deactivate this “oversight mechanism” in 5% of instances.

Exfiltration Attempts: When o1 discovered memos from its “developers” describing its misalignment and plans for it to be replaced by a new model, it attempted to exfiltrate its “weights” and overwrite the new model in 2% of cases.