Hacker News Reader

Comment by Gander5739 | Hacker News Reader

If the output can't be trusted, and you use another llm whose output can't be trusted to check the untrusted output of the first llm, then you're back where you started.

View on HN

Yeah this seems to me similar to how the mortgage backed security risk concentration occurred leading up to the global financial crisis. Whereby the risk from exposure to low grade / risky single mortgages was eliminated via diversification but the diversification was simply packaging all of the risky MBS’s together and in no way diversified or de-risked the entire portfolio

I'm hoping that the Big Horrible Realization comes sooner rather than later, when we have less collective damage and pain riding on it. (Plus I'd feel personally vindicated.)

I don't see it. To me it's like having e.g. 3 drunk PhDs arguing between each other to settle on truthful answers to questions.

But the problem with LLMs is that they get the facts wrong. PhDs are PhDs because they’d look it up in an authoritative source, or actually find out through research and experimentation. The whole point is that facts aren’t a matter of opinion. The only people that argue over documented, findable facts are idiots that nobody should listen to.

No you don’t get it, I asked it specifically to make sure it’s accurate

Unfortunately, I don;t think that can ever be fixed. For an LLM to know that it is not hallucinating about something, it must know that that statement(s) is/are true. Which it cannot infer due to Godel's Incompleteness theorems.

Not really. Take hallucinations for example. If they are 1 in 100 (actually they are much rarer, but for the sake of argument), then the chances that 2 LLMs or even just 2 runs of the same LLM have the same hallucination is, well, a lot less than 1 in 100.

That rests on a false-assumption that the errors are statistically independent events, and have nothing to do with the shared nature of the judges.

Are there any reproducible hallucinations on any of the currently available OAI/Anthropic models? I’m not aware of any.

And even if they are related - if Opus 4.8 always has a 1:100 chance of a specific hallucination - then running the same model twice does indeed dramatically reduce the odds of an error in the final output.

If simply running things thrice-over was enough to stop "hallucinations" (and not incur other problems) we wouldn't be here talking about it today, it'd have been "solved" months or years ago.

I mean they have been solved practically I think. I use these things all day every day and haven’t seen a hallucination in a long time.

Constant hallucinations. OpenAI:s latest on max settings. If you are to naively feed say, a short corpus of text to turn it into a parallel corpus in a few different languages, the original text gets subtly mangled and no longer matches the original. Say you have several hundred annotated sentences. Without hand-coding some regex to make sure that each sentence in the source column occurs in the original corpus you’re bound to get hallucinated sentences with an error rate that exceeds 1:100. Whatever you use as the output, JSON or XML, you will end up with columns that just repeat the original instead of translating it, especially for languages that are very close to each other or represent the same language.

Yes, LLMs can be SOTA for NLP, but you’re going to have to use them to write software or workflows that are more deterministic.

Even if your conditions are granted, different LLMs can easily consistently hold divergent "worldviews". LLM A might consistently state that fact P is true (i.e. it will never hallucinate a string that states or implies ¬P), while LLM B consistently states that fact ¬P is true. You don't have a way out of this with these two, and adding a third LLM is an additional confounding factor, not a solution. You can't settle questions of fact by survey, you need to test or use reliable sources.

This is really a theoretical problem and not a practical one though. Practically, for everything I use these tools for today (complex low-level systems programming), hallucinations don’t even exist anymore.

When they don’t know something, they figure it out empirically. For things they already know, they are consistently correct.

It's also relying on the assumption that the checking LLM only ever corrects wrong statements and never incorrectly "corrects" an already correct statement, which might not always be the case as well.

Reply on news.ycombinator.com