“Research by Chris Ola at Anthropic used sparse autoencoders to successfully identify the function of individual neurons in large language models.”