|
|
|
|
|
|
by MacsHeadroom
932 days ago
|
|
|
Non-RLHF models can do this just fine. Even non-finetuned 7B models, 3 orders of magnitude smaller than GPT-4, can unscramble text and fix typos reliably. Half, or better, of the things people discover "GPT-4 can do" can be done with non-RLHF GPT-3 from 2020 or with a model 1000x smaller. |
|
|