In a controversial 2021 Nature paper, Google’s reinforcement learning (RL) approach to designing silicon chips created a stir. The paper’s claims, which lacked proper documentation, raised eyebrows and attracted critical media attention. To make matters worse, the Nature paper failed to provide key inputs and omitted crucial steps in the methodology. Two separate evaluations were conducted to fill in the gaps, and both demonstrated that Google RL falls behind human designers, a widely recognized algorithm called Simulated Annealing, and even readily available commercial software. Crosschecked data reveals significant errors in the conduct, analysis, and reporting of the study, substantially undermining its integrity.
https://arxiv.org/abs/2306.09633