Post #2962818
2026-05-19 06:59 UTC
Replies (3)
-
@ozzelot@mstdn.social 2026-05-19 07:26
@eniko@mastodon.gamedev.place But I can! I'm different! It won't do anything to me! @jenniferplusplus@hachyderm.io
-
@david_chisnall@infosec.exchange 2026-05-19 07:52
@eniko@mastodon.gamedev.place @jenniferplusplus@hachyderm.io Well, no, but you can nudge it into a different place in the latent space which might be closer to your desired output. The problem with wanting to do this is that it requires you to have a mental model of the shape of that space, which is a high-dimensionality space with things located nearby based on a similarity metric inferred from the training set. And even if you did somehow create a mental model of how this works, that mental model won’t be portable across different ML models (and often not across different training runs of the same model). You can build a bunch of heuristics (‘if I put X in the prompt, it increases the probability that the output will look like Y’) but these are not going to be consistent and reproducible. But that’s fine, because now you’re being charged per token, wrong answers that look close to the desired output are a feature, because they encourage you to try again.
-
@mo@mastodon.ml 2026-05-19 08:40
@eniko@mastodon.gamedev.place I mean, there's a techniques to influence random, or preparing for different outcomes so in average you lose less and get more But making deterministic things random already looks like kinda dumb decision @jenniferplusplus@hachyderm.io