3 ms·

I don't think this is quite correct. If the LLM has seen lots of instances of usage of an API, it can write code to target the API. It can generalize to some d

I don't think this is quite correct. If the LLM has seen lots of instances of usage of an API, it can write code to target the API. It can generalize to some degree, but things go off track the further your requirements are away from the training data. If your code is a lot of duct tape between well-documented, or at least well-named, APIs, that code can be automated. Which is great. That kind of code was always boring to write. I'm less convinced that LLMs will be great at inventing new abstractions to map to a problem domain, and wiring up these new abstractions in a large codebase. They'll need augmentation, fine-tuning, guidance, and it's not clear how well it'll all fit together, and where the limitations of the tech will show up as capability cliffs.

precompute 4y ago
Yes. Outsourcing engineering to LLMs is like building bridges based on structural integrity in minecraft. The real product here is just a "language calculator", that also does "code generation" because it makes financial and PR sense. That people even believe these models can be novel makes one look to the way this thing is marketed. It's also a good time to really take our heads out of the sand and re-evaluate how we expect people to learn civil engineering if their only teacher is a minecraft world. You might get some people that are perfect in minecraft. The rest will be hopelessly stunted. Pretty soon it'll pivot to materials engineering to figure out how exactly a minecraft block adheres to a surface because we lost the original irl way to build a bridge.