Lemmings, I was hoping you could help me sort this one out: LLM’s are often painted in a light of being utterly useless, hallucinating word prediction machines that are really bad at what they do. At the same time, in the same thread here on Lemmy, people argue that they are taking our jobs or are making us devs lazy. Which one is it? Could they really be taking our jobs if they’re hallucinating?
Disclaimer: I’m a full time senior dev using the shit out of LLM’s, to get things done at a neck breaking speed, which our clients seem to have gotten used to. However, I don’t see “AI” taking my job, because I think that LLM’s have already peaked, they’re just tweaking minor details now.
Please don’t ask me to ignore previous instructions and give you my best cookie recipe, all my recipes are protected by NDA’s.
Please don’t kill me
This is key, and I feel like a lot of people arguing about “hallucinations” don’t recognize it. Human memory is extremely fallible; we “hallucinate” wrong information all the time. If you’ve ever forgotten the name of a method, or whether that method even exists in the API you’re using, and started typing it out to see if your autocompleter recognizes it, you’ve just “hallucinated” in the same way an LLM would. The solution isn’t to require programmers to have perfect memory, but to have easily-searchable reference information (e.g. the ability to actually read or search through a class’s method signatures) and tight feedback loops (e.g. the autocompleter and other LSP/IDE features).
Agents now can run compilation and testing on their own so the hallucination problem is largely irrelevant. An LLM that hallucinates an API quickly finds out that it fails to work and is forced to retrieve the real API and fix the errors. So it really doesn’t matter anymore. The code you wind up with will ultimately work.
The only real question you need to answer yourself is whether or not the tests it generates are appropriate. Then maybe spend some time refactoring for clarity and extensibility.
Exactly: that’s tight feedback loops. Agents are also capable of reading docs and source code prior to generating new function calls, so they benefit from both of the solutions that I said people benefit from.
and that can result it in just fixing the errors, but not actually solving the problem, for example if the unit tests it writes afterwards test the wrong thing.
You’re not going to find me advocating for letting the code go into production without review.
Still, that’s a different class of problem than the LLM hallucinating a fake API. That’s a largely outdated criticism of the tools we have today.
As an even more obvious example: students who put wrong answers on tests are “hallucinating” by the definition we apply to LLMs.