When Klein asked how long it would take to get to these various threat levels, Amodei — who said he's wont to thinking "in exponentials" — said he thinks the "replicate and survive in the wild" level could be reached "anywhere from 2025 to 2028."
"I’m truly talking about the near future here. I’m not talking about 50 years away," the Anthropic CEO said. "God grant me chastity, but not now. But 'not now' doesn’t mean when I’m old and gray. I think it could be near-term."
Resolves to YES when:
Dario Amodei makes a public claim that AI models had sufficiently met this condition.
OpenAI, Anthropic, or Meta makes a public claim that AI models have sufficiently met this condition.
Resolves to NO on December 28th, 2029 if the YES conditions are not revealed to have been met before the end of 2028. So, if the condition is met in 2029, this resolves NO; if the condition is met in mid-2028 but only revealed publicly to have been met then in 2029, this resolves YES. I just want to take into account potential announcement delays.
Slight modifications from question by Jamil Dhanani: https://manifold.markets/JamilDhanani/by-the-end-of-2025-will-ai-models-a.
How does this resolve if one AI system starts replicating in the wild but the cloud hosting company detects it and shuts it down before it gets big?