Hacker News Reader

Comment by w-m | Hacker News Reader

I'm quite confused. I learned 20+ years ago that lead poisining from the wine was the cause, or at least the suspected cause of Beethovens deafness. It's great that they found lead in his hair (again?), but thoroughly weird that they don't mention that it was the go-to theory already.

On a large tangent, I tried asking GPT-4o to check the article whether it mentions that previously that theory already existed (could have overlooked it). GPT-4o confidently made up a complete quote from the article, then doubled down after I said I couldn't find the quote, and finally reversed its opinion 180 degrees. It's of course much too early to tell, but my feeling from a handful queries has been that GPT-4o hallucinates quite a bit more than GPT-4.

View on HN

Indeed! There was even a book on the topic I read years ago. Lead poisoning seemed quite conclusive even back then. https://www.amazon.com/Beethovens-Hair-Extraordinary-Histori...

That particular sample was shown to have been a fake last year when multiple purported hair samples were compared.

I asked it yesterday if there were any baseball games in town this weekend. It hallucinated a false schedule against the wrong team complete with times, and even linked to the (correct) schedule. I told it I googled it and they're playing a different team this weekend and it then (without searching the web again) apologized and shared the correct schedule.

It seemed like such a simple question, but it just made up a very plausible, but entirely incorrect answer. I asked it several times in different ways why and how I could tell if it is correct and all it could tell me was to look it up myself and cross-reference which enitrely defeats the purpose of using it.

Schumann believed it was a form of syphilis which causes progressive deafness, then a period before death of wonderful clarity, like the whole world is verging on a great revelation of unity and you are the carrier.

This is a real form of syphilis, but AFAIK there's no evidence other than the ode to joy.

I spent the last 20 hours using 4o writing a tauri/rust desktop app. I don't know rust, so it was a LOT of pasting errors back in copying back to my code. But I'd say after about 14 hours it got very lazy. Does AI get tired like humans?

Tell it you will give it tips for kWhrs.

Sounds productive. Don't forget to take breaks, drink water, and sleep.

I've noticed that errors seem to increase the longer the chat goes. I think having too much previous context confuses it. I seem to have much better luck if I occasionally start a new chat.

AI can get 'stuck' in local optima. At which point it will hallucinate or repeat itself over and over or both. Easy way is to start over or you have to really feed a bunch of stuff to it to get it to move.

A good way to think of AI is like a cubic spline. The formula that pops out of all of the nodes strung together make a thing that resembles a cubic spline. Everything is valid along that formula but only at the nodes where it crosses the real spline. Everything else might be close +/- some error rate. But it can also create hills and valleys that you should not reach. But are perfectly valid for the spline. The more noes you add the higher the calculation is. But the more accurate it is. But that can also create wildly weird things between the valid points.

Reply on news.ycombinator.com