Maybe we can tone down the FUD a bit. Wikipedia is flat wrong sometimes. Google is flat wrong sometimes. LLM’s can be flat wrong sometimes. No different to trusting an LLM’s output to Google’s output. Good as a starting point but not something I’m going to base my medical and legal decisions on. I don’t see the zeitgeist of LLM’s being any different. I don’t see some trend of legal or medical professionals blindly trusting LLM output either, even if some very rare and cherry picked examples would want us to believe otherwise.
I must be living in a different world or using a different version of the model, but I seem to be getting garbage from OpenAI products, specifically ChatGPT 4o. The most recent example; I tried to recreate a similar scenario that Sal Khan did where he used photos/screenshare with ChatGPT to help his son learn geometry. I did the same thing with Chess.
I started with having it walk me through a few chess puzzles. It straight up couldn’t figure out the solutions and frequently referenced coordinates that were well outside the bounds of a chess board (Knight to Q8, for example).
At this point, it feels like I’m being gaslit by the community of AI obsessives who see LLMs as the second coming of Jesus. I get nothing but garbage from them. Sure, maybe it can occasionally write me a line of syntactically correct code but that’s it. It feels like this is being shoved down my throat and I’m criticized for ever expressing skepticism. I really don’t like how this discourse is progressing.
Thank god I am not the only one. I reverted back to GPT-4. 4o is complete garbage. It seems that it's geared toward always giving you an answer regardless whether it can determine an answer or not. It's hallucinating 80% of the time and with high confidence.
I’ve experienced excessive garbage production, but also excessive verbosity, even when it’s giving helpful responses. “Here’s the code I just wrote in the previous message!”
