Hacker News Reader

Comment by keiferski | Hacker News Reader

I've been re-watching House lately and I can't help but constantly think that a LLM would be excellent at diagnostics. At least in the sense of, "the patient has displayed XYZ symptoms, the test results have XYZ data, and ergo the following conditions are likely. Especially in the more obscure interactions and second-order effects.

View on HN

Medical experts have done teardowns of House episodes and most of what happens in the series wouldn't happen in real life, that is, diagnoses are better than how they are depicted there.

Oh yeah I'm sure it's not super accurate, I just mean the specific task of diagnosing a patient would seem to benefit from a LLM-like piece of software.

But surely this could just be done with a constraint software, with better accuracy and less likelihood of making something up?

Possibly but in my experience LLMs are more useful for brainstorming and generating original ideas than “being accurate” all of the time. Hence my idea that they could be useful in diagnosing difficult conditions.

Yeah, it's a trade off between certainty and creativity. Constraint software would be all certainty which, if it contained all the required information, would be the perfect system. But we don't have "all the information", or at least i haven't seen anyone claim so. So some degree of out-of-box creativity is required to catch more obscure correlations and strangeness. I think AI is just about as certain as previous systems, but with a big boost to the creativity aspect. It's also free. I can imagine the omnidiagnoser would be a closely held trade secret worth billions and thus absolutely not free to the public.

Reply on news.ycombinator.com