@EpochAIResearch. My DMs are open.
Anonymous feedback: https://t.co/0k6Duym4fI
Jul 21 • 11 tweets • 2 min read
The R&D compute gap between, say, Anthropic and Zai is huge.
Yet the ECI/benchmark gap is roughly on the order of 6 months. I can think of 3 kinds of explanations for this:
(1) Researchers at Zai are way more compute efficient. This could be due to talent, or to optimizing for compute efficiency more (eg, having more employees per unit of compute).
Jul 21 • 4 tweets • 1 min read
Currently, there is no trend in model capabilities we can extrapolate to get a good guess of what tasks AI will vs won't be able to do N years from now.
In the absence of this, I feel extremely uncertain about timelines to TEDAI, AI research parity, etc.
Jul 21 • 5 tweets • 1 min read
How do people interpret the "Composer training and RL" bar here?
I see 3 possibilities: (1) that 85% is confusing/misleading, eg it includes all R&D for Composer training rather than final runs, or it's compute spending rather than FLOPS and RL has way lower MFU
May 6 • 11 tweets • 2 min read
I’m very uncertain about AI timelines and takeoff speeds. My views oscillate a lot, but my all-things-considered p(TEDAI¹ before EOY 2030) and p(AI research Parity² before mid-2028) are consistently above 5%, often 10% or more.
Part of why I want to post this is that there might be some people who take comfort in an “Epoch AI worldview” that has less aggressive timelines than e.g. the AI Futures Project.
While it’s true that I expect things to go more slowly than AIFP:
