No, it was not my first experiment - again with this unbelievable condescension! I have been playing around with this stuff since GPT-3. That was my first pract

No, it was not my first experiment - again with this unbelievable condescension! I have been playing around with this stuff since GPT-3. That was my first practical use case where GPT wasn't a totally useless waste of money. It works very well with AI-related papers, maybe a 1% hallucination rate. I got the idea from an AI researcher and I was intrigued that ChatGPT might actually be useful for me. But it was not. The problem, as always with ANNs, is that I went slightly off the happy path. Even a skeptic like me assumed the AI papers successes was evidence that GPT-4 was better at remembering its pretraining data; instead I think it's evidence that a data contractor RLHFed the answers and GPT had it memorized. > LLMs are terrible search engines… except for the times when they are great search engines! But note that what you said about finding papers was wrong, it works extremely well for AI research. The reason LLMs are useless to me across the board is that these unpredictable and arbitrary limitations apply to everything, not just search. "Learning when and what to use them for" is pure trial-and-error because it seems to amount to guessing what tasks the 3rd-party data contractors trained the LLM to solve. I am not a Python or JavaScript developer, nor do I write code for extremely well-known libraries. I use F# for oddball projects (often analytics), and GPT-4 was utterly useless for F# codegen. My first experiments with GPT-3.5 showed that it would plagiarize hundreds of lines of public F# projects, including from my own GitHub, without any prompt engineering or trial-and-error - it was just blind plagiarism. GPT-4 isn't quite that bad, but it's still not even close to being good enough to help me - in particular it has no understanding of high-performance F#. I would be spending far more time auditing and optimizing its crappy code. And time spent writing code has never been the limiting factor in my F# development. I also do some recreational mathematics on finite geometry and combinatorial group theory; GPT-4 was utterly useless here, even with CoT prompting, and even though it solved more complex graduate-level algebra problems without any difficulty. Of course, those problems were repeated and solved in dozens of graduate textbooks. My cute little groups, not so much. I believe CoT prompting is theoretically incapable of helping GPT here since the computational complexity is too high. What CoT prompting gives you is a bunch of insidious errors that take time and effort to unravel. Otherwise there's nothing I do that would even conceivably benefit from an LLM: it can't play guitar, it can't play with my cats, and I would never use it to communicate with friends or family. I guess I could fill my brain with shallow subject knowledge about something, a few choice sentences. But I'd much rather understand something in depth by reading a book. I'm not too busy to read a book. Otherwise... maybe I could use LLMs to write polite no-thank-yous to unsolicited recruiters. This tech truly has nothing to offer me. I think you are failing to understand that, as a Python developer who maintains one of the biggest Python web frameworks and writes a popular blog for general tech audiences, LLMs are unusually well-suited towards your use cases, due to reasons that will not extend to people working in more isolated corners of the world.