That solves the "garbage training data" problem, but it doesn't solve the "it's just a language model" problem.

If you fine tuned ChatGPT on all the sources you mention, you now have a model that produces results that could plausibly have been written by a domain expert on the Linux kernel, but you don't have a domain expert. It will still hallucinate, because that's a fundamental feature of generative AI, it will just hallucinate much more convincingly.

I get what you're saying, but I'm just not convinced that it will continue to be a huge problem. In some sense, if the state of art is where we're at today with language models, then sure. But I think it'll get better -- in part because I'm not sure humans just aren't souped up language models with some weird optimization functions...

Could it? There's lots of commentary about how it could easily be this or that, but I don't work in ML and have no clue whether it is actually easy or not to tweak ChatGPT to work in such ways.

For example, in my experience ChatGPT isn't very "smart". It has a lot of knowledge but it can't infer any facts from that knowledge. When you ask it to write a program it has no real idea of what it actually does, and you can easily get it to add features it already added, or tell it something is a bug when it really isn't.

This doesn't sound like the stuff you could make a domain expert out of, at least, not out of the box.

I'm not asking it to infer much, but to piece together a fair bit and understnad what I'm asking. For example, here's a chat I had with ChatGPT:

"How does common subexpression elimination work?"

It answered it correctly and gave some basic code examples to demonstrate the concept. Then I followed up with:

"But can it do the elimination if the variables are flipped, but semantically equivalent, like 'y + x' in the example above?"

It again gave what I would consider a correct answer. Note 'x + y' was what was eliminated previously, so I reversed it here.

Then I asked: "Are there cases where the compiler might fail to eliminate common subexpressions?"

And again a good answer.

Now all of this could be found on the web somewhere, but for example the second question didn't show an obvious answer on Google when I searched for it. I'm sure I could find it, but I know this field well. If I was someone new to the field I'd probably spend a lot of time parsing useless articles to find something that answered the question in a way I could understand.

I'm less concerned about it writing code (which is cool). For me, the ability to help me learn an area quickly is far more useful. It doesn't need novel answers, but the ability to understand what I'm asking and answer it. I think it's really close to being able to do this now.