That is the thing, they are not “only going to get better” because the training has hit a wall and the compute used will have to be reduced since they are losing money with every request currently.
Technology these days works in that they always lose money at the start. Its a really stupid feature of modern startups IMO. Get people dependent and they make money later. I don’t agree with it. I don’t really think oir entire economic system is viable though and that’s another conversation.
But LLMs have been improving exponentially. I was on board with everything you’re saying just a year ago about how they suck and they’re going to hit a wall even. But the don’t need more training data or the processing power. They have those and now they’re refining the LLMs. I have a local LLM on my computer that performs better than chat GPT did a year ago and it’s only a few GB. I run it on a shitty laptop.
I experimented with quite a few local LLMs too and granted, some perform a lot better than others, but they all have the same major issues. They don’t get smarter, they just produce the same nonsense faster (or rather often it feels like they are just more verbose about the same nonsense).
I don’t know what to tell you. I have them successfully compiling tables of search outputs to compare different things for method development and generating code, saving me hours of work each week. It all needs to be checked, but the comparison comes with links and the code is proofread and benchmarked. For most of what I do it’s really just a jacked up search engine, but it’s able to scan webpages faster than me and that saves a lot of time.
As a hobby, I also have it reading old documents that are almost illegible and transcribing them pretty well.
I really don’t know what you’re doing that you’re just getting nonsense. I’m not.
One other comment pointed me at one issue that might be a major difference. Is the code you generate in one of those ultra-verbose languages like Java where we had basically IDEs generating code from much shorter descriptions already 20 years ago? I could see LLMs doing well with those.
I tend to try to generate code mostly in Rust or sometimes shell or config files or DSL for various programs and 99% of the time the code does not even come close to what I wanted it to do, mainly because it just hallucinates itself some library interfaces that do not exist.
That is the thing, they are not “only going to get better” because the training has hit a wall and the compute used will have to be reduced since they are losing money with every request currently.
Technology these days works in that they always lose money at the start. Its a really stupid feature of modern startups IMO. Get people dependent and they make money later. I don’t agree with it. I don’t really think oir entire economic system is viable though and that’s another conversation.
But LLMs have been improving exponentially. I was on board with everything you’re saying just a year ago about how they suck and they’re going to hit a wall even. But the don’t need more training data or the processing power. They have those and now they’re refining the LLMs. I have a local LLM on my computer that performs better than chat GPT did a year ago and it’s only a few GB. I run it on a shitty laptop.
I experimented with quite a few local LLMs too and granted, some perform a lot better than others, but they all have the same major issues. They don’t get smarter, they just produce the same nonsense faster (or rather often it feels like they are just more verbose about the same nonsense).
I don’t know what to tell you. I have them successfully compiling tables of search outputs to compare different things for method development and generating code, saving me hours of work each week. It all needs to be checked, but the comparison comes with links and the code is proofread and benchmarked. For most of what I do it’s really just a jacked up search engine, but it’s able to scan webpages faster than me and that saves a lot of time.
As a hobby, I also have it reading old documents that are almost illegible and transcribing them pretty well.
I really don’t know what you’re doing that you’re just getting nonsense. I’m not.
One other comment pointed me at one issue that might be a major difference. Is the code you generate in one of those ultra-verbose languages like Java where we had basically IDEs generating code from much shorter descriptions already 20 years ago? I could see LLMs doing well with those.
I tend to try to generate code mostly in Rust or sometimes shell or config files or DSL for various programs and 99% of the time the code does not even come close to what I wanted it to do, mainly because it just hallucinates itself some library interfaces that do not exist.