Color me skeptical, but I’m always willing to update my priors. He’s got p(AI doom) at 70% according to his bio (he was safety Philosophy PhD at OpenAI, I think?)
His predictions for AGI (expert level, it sounds like) is either this year or as far out as 2029.
The bit about Llama 3 70B is particularly intriguing given the Q* stuff that I believe got some folks fired. Q* can do interesting things around long range planning, at least the standard Q* algorithms that are well established in the industry.
But my issue is - why doesn’t he think something like that could be put around llama? I really don’t see any of the GPT4 models being able to do that much better. Any ‘doom spiral’ is going to come from your Graph of Thoughts agent architecture at this point.