Thought experiment: Algorithmic progress outperforms brute-force compute scaling

tszzl · x · 2026-08-16

CoreAutoAI proposes a thought experiment comparing algorithmic progress since 2020 against raw compute scaling. Scenario A applies today's algorithms to GPT-3 levels of compute, while Scenario B scales GPT-3's architecture with GPT-5.6 levels of compute. The conclusion is that the model from Scenario A beats Scenario B on every axis by a significant margin, challenging the dominant narrative focused on datacenter spend and highlighting widespread credit assignment errors.

Original post →

More from AGI Musings

AGI Musings channel →