Interpreting the compute multiplier chart in Anthropic's IPO roadshow video: the author speculates each grid step is 10x, with an endpoint of about 30000x
Trying to understand that compute multiplier graph
AI summary
A LessWrong author interprets the compute multiplier curve chart in Anthropic's IPO roadshow video, speculating that the endpoint represents roughly a 30000x compute multiplier. He believes the four points are Claude 3.5, Claude 4, Mythos and an unreleased model; if each grid step on the Y axis represents 10x, then the compute needed for pretraining at the same quality is about 1/30000 of that in 2024. The author says the actual value could be much higher or much lower, and the chart concerns only pretraining.