Ask HN: How realistic is a modern-day era NLU

I'm looking for people who are a bit more familiar with neural networks than me (I follow kaparthy's tutorial once, that's about it).

Lets say I like using LLMs for quality/complex tool-calling to e.g. solve bugs, or just be a more sophisticated researcher.

But I don't need it to write code for me. What would this look like? Would it just be a fancy NLU? Could you get a 3B tool that's targeted towards tool-calling? or maybe a 1B?

It wouldn't be more than just

```rust USER: cargo build fails with: error[E0382]: borrow of moved value: `foo`

CONTEXT: foo: Vec<String>

AVAILABLE TOOLS: - read_file(path) - ripgrep(pattern) - rustc_explain(code) - run_tests(...)

ASSISTANT: <tool>rustc_explain {"code":"E0382"}</tool>

TOOL: Use of moved value...

ASSISTANT: <tool>grep {"pattern":"foo"}</tool> ```

I have a theory that scraping off the "code-writing" part would eat up a lot of time and size. I really don't care that it writes code for me, it's the best way to stay in the loop anyway, so off-loading this to a model is pretty useless for me. You could even test this against multiple tools. Maybe you could even use a frontier model to dig through earlier sessions and autoresearch better ways to improve this model.

Please let me know if something like this has been done before, and what the possibilities are today. I would love to dig deeper into this if there's a slight chance this is possible.