Post #3166118

2025-08-12 06:20 UTC

There is a hope that incorrect output can be fixed with better 'alignment'. Alignment is a euphemism for hacking and filtering input/output of an LLM to stop it from producing results that are considered unacceptable. In other words, to stop the LLM producing what the LLM would normally produce. It's basically a bug filter that by its nature needs to be constantly updated in response to feedback and complaints. It is by its nature volatile and never complete.

Replies (3)

@kevlin@mastodon.social 2025-08-12 06:26

Code is a repository of determinism. That is its value. While it is not necessarily perfect, it has a consistency, comprehensibility and level of correctness that is inherent to its design. By design, LLMs do not have this and are, therefore, unsuitable as a direct replacement. That does not mean they cannot be useful. It does not mean they cannot be used to generate code. But it does mean they cannot be used to replace code.

Open ##3166119
@svelmoe@hachyderm.io 2025-08-12 06:49

@kevlin@mastodon.social And funnily enough, with the 'bug filter' - one way of doing that is knowing what the actual output should be, and thereby, in essence, making the LLM process redundant :D

Open ##3166132
@OmegaPolice@hachyderm.io 2025-08-12 08:25

@kevlin@mastodon.social And the bug filters are most likely, just like 'RAG', good old deterministic code, aren't they.

Open ##3166134