wkey22

• wkey22 27 Nov 2020 9:10 UTC

  LW: 10 AF: 5

  AF

  in reply to: Steven Byrnes’s comment on: Why GPT wants to mesa-op­ti­mize & how we might change this

  I used your idea of “a” vs. “an” as the basis of a GPT-3 experiment to show that GPT-3 indeed probably does do lookahead. Details are at https://​​www.reddit.com/​​r/​​GPT3/​​comments/​​k0mvf3/​​experiment_that_shows_that_gpt3_can_probably_plan/​​