Example

What the 2023 Weak-to-Strong Experiments Did and Did Not Show

OpenAI’s 2023-12-14 report used weaker-model targets for further training of models already trained on broad data. The question was whether training could draw out the stronger model’s existing capabilities rather than merely copy the supervisor’s mistakes. Some methods yielded models that outperformed their weak supervisors on tested tasks. The authors call this a proof of concept—an initial demonstration with limits: the reported method did not work on ChatGPT preference data, judgments about preferred answers. Model supervision also differs from human supervision. This experimental analogy does not establish reliable human review of unfamiliar scientific discoveries.

0

1

Updated 2026-11-10

Contributors are:

Who are from:

References


Tags

Scientific Oversight of AI-Assisted Research

1Cademy