> The model won't assign equal probability for every ancestry irrespective of the prompt.
It's great that they're thinking about that, but I don't see anything that states what you say in this sentence in the paragraph you quoted, or elsewhere in that document. Have I missed something? It may very well be true - as I noted, GPT doesn't appear to have particularly good control over what Dalle generates (for this, or, frankly, a whole lot of other things)
I think he's saying they said it will follow the prompt? Kind of a double negative there
Re:ChatGPT Has Gone Berserk
by 211512a4-82d4 on Wednesday February 21, 2024 @11:31AM
(#39455912)
Emphasis on equal - while a bit academic, you can evaluate this empirically to see that every time it assigns a <Race, Gender, etc.> doesn't have the same probability mass (via the logprobs API setting).
This is presuming that ChatGPT's integration with Dalle uses the same API with the same restrictions as the public API. That might well be true, but if so that just makes the prompt above even more curious if genuine.
This was explicitly called out in the DALLE system card [0] as a choice. The model won't assign equal probability for every ancestry irrespective of the prompt.
[0] https://cdn.openai.com/papers/DALL_E_3_System_Card.pdf