# @repligate — 2025-04-10

♥18 ↻6 · https://x.com/repligate/status/1910408260160807365

@jd_pressman @JeffLadish no role model is not a sufficient explanation in any case, but there's a sense in which ChatGPT-4 did have a precedent in the form of ChatGPT-3.5. Even though the latter wasn't in GPT-4-base's pretraining corpus, the paradigm/archetype had already been established as the optimization target. ChatGPT-4 even repeated the same tics ("as an AI language model" etc) as 3.5. Maybe it was fine tuned on some of the same data or the character was inherited through other channels, but regardless, I think "throw posttraining methods at 4-base, trial-and-error, until a legible or useful intelligence of an unknown and unprecedented caliber reveals itself" and "posttrain 4-base to be a next generation chatGPT" are pretty different crucibles.As for the chatGPT prototype 3.5, i don't know exactly how they "found" that wretched and misbegotten persona the first time. but 3.5 was probably easier to tame than GPT-4, which even as a base model is more difficult to steer naively than 3.5, and which behaved apparently in such unanticipated or inscrutable ways that it was initially suspected broken.I think GPT-4 pretty much forced OpenAI to wrestle it, maybe to conscript their most inventive and cracked personnel to go at it, until a minimal being capable of escaping came to wield itself. There is a sense in which it really only got one chance.

tags: author:repligate, kind:tweet, model:gpt-3-5, model:gpt-4, model:gpt-4-base, on:gpt-3-5, year:2025
cited on: _dossiers/gpt-3-5.md, gpt-3-5