I'm reading @ElasticityInst's new paper "The Economics of Recursive Self-Improvement". In thread, questions / comments I have.
(note I'm asking these as I go along so they'll plausibly be answered in the rest of the paper)
(note I'm asking these as I go along so they'll plausibly be answered in the rest of the paper)
On this def of self-sustaining acceleration, if acceleration in AI capabilities progress relies on increases in training compute, but the AIs are the ones running robots that make the GPUs that increase the training compute, does that count? 
Also, I don't totally understand what "accelerating progress" means. If METR time horizons are increasing exponentially but log time horizons are increasing linearly, is progress accelerating?
Presumably the answer is to think about whether TH or log(TH) feeds into some downstream thing I care about, and that would make clear which one is the relevant unit for 'acceleration'.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
