✳flâneur — a map of the web's best reading

An AI Which Imitates Humans Can Beat Humans | Tom Cunningham – Tom Cunningham

tecunningham.github.io · 7,712 words · saved by 1 readers

Thanks to comments from many, especially Giorgio Martini, Grady Ward, Rob Donnelly, Inés Moreno de Barreda, and Colin Fraser. If we train AIs to imitate humans, will they ever beat humans? AI has caught up to human performance on many benchmarks, largely by learning to predict what humans would do. It seems important to know whether this is a ceiling or we should expect them to shoot out ahead of us. Will LLMs be able to write superhumanly-persuasive prose? Will image models be able to see things in photos that we cannot? There is a lot of technical literature on imitation learning in AI but I haven’t found much discussion of this point (Bowman (2023) is a notable exception). In a formal model I derive five mechanisms by which imitative AI can beat humans. The evidence is unclear. There are many reasons why this could theoretically occur but I couldn’t find much evidence for superhuman performance: many benchmarks which we use to evaluate ML models have human labels as the ground truth

An AI Which Imitates Humans Can Beat Humans | Tom Cunningham – Tom Cunningham Thanks to comments from many, especially Giorgio Martini , Grady Ward, Rob Donnelly , Inés Moreno de Barreda , and Colin Fraser. If we train AIs to imitate humans, will they ever beat humans? AI has caught up to human performance on many benchmarks, largely by learning to predict what humans would do. It seems important to know whether this is a ceiling or we should expect them to shoot out ahead of us. Will LLMs be able to write superhumanly-persuasive prose? Will image models be able to see things in photos that we

Explore this link on the map →

related reading