← Back to context

Comment by asim

16 days ago

We need a language and a transpiler. Honestly the LLM has many uses. Agents have many uses. And we are narrowing down how to make them deterministic and predictable for programming machines and software. But that also means we need something beyond natural language for the actual implementation. Yes we've moved a level up, but engineers are not product managers, so as much as we can define the scope and outline a project like a 2 week sprint using scrum or kanban, the reality is deterministic input for deterministic output is still the way to go. Just as compilers and higher level languages opened the doors to the next phase, the LLM manages this translation and compilation, but it's missing a sort of intermediary language, a format that's going to be much better processed and compiled directly down to machine code. We're talking about LLVM. Why are asking LLMs to write Go code or Python, when we could much better translate an intermediary language to something far more efficient and performant. So I think there's still work to be done.

5 comments

asim

Reply
wtetzner  16 days ago

Am I understanding what you're saying correctly?

* We need a deterministic input language

* The LLM generates machine code

Isn't that just a compiler? Why do we need the LLM at that point?

  • CuriouslyC  16 days ago

    If the compiler only gets you 80% of the way there, but what it does is sufficient to put the LLM on rails, like programming language mad libs, I'd say that's a win.