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

A Hacker News user asked how realistic a modern-day natural language understanding (NLU) system would be for tool-calling tasks, proposing a targeted 1B to 3B parameter model that handles tool selection without code generation. The user described a workflow in which an assistant responds to a Rust compiler error (E0382: borrow of moved value) by invoking tools such as rustc_explain, read_file, ripgrep, and run_tests, and theorized that removing code-writing capability would cut model size and latency. The post, submitted to Ask HN, had 1 point and 0 comments at the time of the article.

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 AVAILABLE TOOLS: - read file path - ripgrep pattern - rustc explain code - run tests ... ASSISTANT: rustc explain {"code":"E0382"} TOOL: Use of moved value... ASSISTANT: grep {"pattern":"foo"} 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. Comments URL: https://news.ycombinator.com/item?id=49655550 https://news.ycombinator.com/item?id=49655550 Points: 1 Comments: 0