Comments on Ajeya Cotra’s draft report on AI timelines

Abstract

The training of large language models and performance in complex tasks such as video games suggest that current artificial intelligence systems require vast amounts of data and compute to generalize to new tasks and environments. It is still unclear how future improvements in algorithmic efficiency will ultimately affect the computational cost of achieving transformative AI (TAI). This paper proposes a framework for thinking about training compute requirements, derived from the observation that brain training is likely constrained by the lifetime of individual organisms. It combines elements from several proposed anchoring hypotheses that relate biological and artificial intelligence, including estimates of the number of synapses (parameters) in the brain and the amount of biological computation per second performed by those synapses. The paper also introduces a new variable, the effective horizon length for training, which represents the amount of data it takes to effectively update learned parameters. The authors find that their framework can explain several orders of magnitude difference in estimates of the compute cost of TAI, arguing that the choice of anchoring hypothesis and other parameter values is crucial. They also propose a two-stage model of AI training, with one stage corresponding to hyperparameter training across many agent lifetimes. – AI-generated abstract.