A LessWrong article interprets the chart showing scaling in compute in Anthropic’s IPO roadshow video. It estimates that if each Y-axis gridline represents 10x, the total scaling is approximately 30,000x. This corresponds to the compute required to pretrain Claude at equivalent quality being about 1/30,000 of the 2024 level; however, if each gridline represents 1.1x, the total scaling is approximately 1.53x. The chart covers pretraining only and does not indicate the situation for post-training.
Source:LessWrong · Read original
Leave a Reply