“METR could not confidently upper-bound the capabilities of the models during the period they had model access.”
“Apollo Research believes that o1-preview has the basic capabilities needed to do simple in-context scheming.” (Apollo prompted the model to misbehave; this was an evaluation for scheming capabilities, not propensity.)
Apollo found that o1-preview sometimes instrumentally faked alignment during testing . . . , it sometimes strategically manipulated task data in order to make its misaligned action look more aligned to its ‘developers’ . . . , and an earlier version with less safety training proactively explored its filesystem to test for the presence of developer oversight before acting on its misaligned goal.
OpenAI o1
It’s more capable and better at using lots of inference-time compute via long (hidden) chain-of-thought.
OpenAI pages: Learning to Reason with LLMs, o1 System Card, o1 Hub
Tweets: Sam Altman, Noam Brown, OpenAI
Discussion: https://www.transformernews.ai/p/openai-o1-alignment-faking
“METR could not confidently upper-bound the capabilities of the models during the period they had model access.”
“Apollo Research believes that o1-preview has the basic capabilities needed to do simple in-context scheming.” (Apollo prompted the model to misbehave; this was an evaluation for scheming capabilities, not propensity.)