So… asked ChatGPT a simple question, it failed. I then corrected the thing, and asked again. ChatGPT apologized and, when asked again, said sorry another time, and gave me the correct answer. Somewhat impressed that it didn’t know the proper answer in the first place.
@maique It really depends on how much context it has both from your question and also its own training set. Then there is the process of “hallucinating” what it thinks is correct. At the end of the day, it’s just trying to predict what comes after the word it just wrote. Super fascinating stuff.
@maique And by correcting it, and giving it the answer, it understands your context and does as it’s told. It believes you saying that it was wrong, hence it’s probably saying sorry. However, it probably won’t know the answer again if you started a new chat session with it (because the context is now new = new model).
@vincent I wasn't aware about its inability to "learn". It would be great if it could.
In this case I found it odd, because the information it's providing is based on the national railway system, and it's was inaccurately telling me there's a train station in a city that's not covered by it. Never has been.
The line it mentioned runs close, but kilometers close, and it never went through the city. No station ever had that village's name.
When I told it it was wrong, it apologized, and quickly found the nearest station.
I did go back and asked the same question just now, and it gave me the wrong information again 😔
@maique yeah, it feels super magical, but it actually doesn’t have a clue. It certainly makes you feel like it’s correct, and that’s the scary part. Unless the folks at openAI add your data point, it’ll continue to get it wrong. It just doesn’t know… it’s just trying to predict what comes next…
Happy to send you links this coming week on how these language models work, if you want.
@vincent Thanks. If you come across one that you think it will help, feel free to do so. Appreciate that 🙂
@maique I got plenty, and there is a chap named Simon Willison that blogs in depth about this (the first article is one of many).
ChatGPT can’t access the internet, even though it really looks like it can
@maique If you can get into the first part of the last one I linked to, then I think you already have a good understanding. It’s pretty long…
@maique There is a new episode of ATP out, and John basically talks about this... so hopefully will explain it better.