2 ms·
This is, really, yet another example of the trouble with the "LLMs are correct 90% of the time, and only go totally off the rails 5% of the time" marketing line
by rsynnott 2y ago
This is, really, yet another example of the trouble with the "LLMs are correct 90% of the time, and only go totally off the rails 5% of the time" marketing line. There are remarkably few use cases, it turns out, where that is okay; you really need it to _not matter at all_ if the output is arbitrarily wrong.
(I suspect character.ai was originally conceived precisely because it appeared to be a usecase where LLM unreliability would be okay, the creators not having thought sufficiently carefully about it.)
- JumpCrisscross 2y ago> remarkably few use cases, it turns out, where that is okay; you really need it to _not matter at all_ if the output is arbitrarily wrong It's probably true in most cases. Those cases don't cover kids.
- rsynnott 2y agoI think there are a lot of cases where it _seems_ to be true, until you think through the details. Most cases where it actually _is_ true are, in practice, very low impact; the big proven one seems to be, essentially, generation of high-volume spam content.