@danjones000@microwords.goodevilgenius.org
Post #1470161
2026-01-07 14:44 UTC
But do LLMs really need to be deterministic and generate the same code every time?
If you want to be able to trust them the way you can trust a compiler, than yes, they do.
Tests are not the end all be all of whether or not code is good enough.
If we can't read the code, we can never be certain that it's doing exactly what we wanted it to do, nothing more and nothing less, consistently with the performance that we expect.
LLMs will never give us that.
@t3rcermillenium@mastodon.social @Chip_Unicorn@im-in.space
Replies (1)
-
@t3rcermillenium@mastodon.social 2026-01-07 15:06
@danjones000 @Chip_Unicorn What we have now in terms of LLMs for coding isn’t great for repeatability, or even for coding! It duplicates code, leaves old code behind when refactoring, forgets requirements. And it’s completely powered by grep - it has no programming language knowledge (types, indexing code base, etc). But it’s getting better all the time. Where will of be in 2-3 years? I’d be hesitant to say we “never” trust the code it generates.