Post #3166116

2025-08-12 06:11 UTC

To all measurable intents and purposes, LLMs are non-deterministic. The value of code is that it is deterministic. If your business is dependent on software, you want to know how it will behave and that it will behave consistently. Consider the role and necessity of consistency for software in regulated markets, the medical domain, safety-critical systems, etc. Consistency is inherently not a feature of LLMs.

Replies (2)

@kevlin@mastodon.social 2025-08-12 06:15

You also want a high level of correctness. Very high level. Consider how often you encounter compiler bugs. Compare that with the frequency with which you experience AI hallucinations. It's not even a close competition. Hallucinations are a feature of how LLMs work.

Open ##3166117
@sigmasternchen@comfy.social 2025-08-12 06:21

@kevlin@mastodon.social That’s true. However, there is value in stochastic results in some cases. To give an example from a different industry: Antigen tests have a false negative and false positive rate > 0. Still they are very useful. When you need an exact result you do a PCR test that takes longer to complete and is much more expensive. Similarly, LLMs could be useful (and I’m ignoring the moral implications here - obviously this is non-trivial) for roughly evaluating if e.g. an insurance claim is valid or not. Only on edge cases or when a customer complains, a humans looks at it. This is non-deterministic, could however save money in the long run, and also give faster responses to the customers.

Open ##3166136