I'm not sure I agree that LLMs solve for the statistical mean

But more importantly: isn't what you call "intent" just a series of optimization goals? You want your code to satisfy the constraint of being correct™, while also maximizing various other goals like being maintainable, easy to understand, having few lines of code, as little tight coupling as possible, etc. Goals that often conflict, but when given two implementations you could likely tell which hits the better tradeoff (in your engineering experience)

Those are all things that theoretically - with a tight enough specification and enough compute - a constraint solver could solve. No human intent necessary.

The issue is more that we can't fully specify all those side goals, and even if we could the LLM would struggle following them. A classic paperclip maximizer problem (where nobody told the paperclip maximizer to keep the planet inhabitable and all the other side conditions we implicitly assume)

> I'm not sure I agree that LLMs solve for the statistical mean

The good thing is that you don't have to agree with that, as the fundamental technical reality does it for you. LLMs do work like that. They are just statistics and probabilities.

Doesn't the universal approximation theorem apply to typical LLM architectures? LLMs solve for the shape of the input data, as well as they can approximate it in the constraints of their architecture, the optimizer and the amount of compute spent in the training process

And where those constraints prevent us from matching the shape there is no guarantee that we match anything like the mean. Though maybe we can agree that that's where the optimizer would tend to steer towards when it can't do any better

If you built a task management system, you'd have very different code bases depending on whether it's for internal use at a mid-size development org or as a SaaS.

So I wonder whether, in your experience, the results you've seen, could have improved by providing sufficient context? - or what context was given.

I.e. if you have the agent that same context, as one of your colleagues would have/require to solve a problem.

"You're holding it wrong"

You've missed my point. I didn't dismiss agents. I did dismiss the industry.

I don't need to add more context to a statement that operates on a layer above where context injection would influence it. It is a conceptual impossibility. Not a technical roadblock.

This is a weird framing, but the user is the "intent". If you can produce "great code" by yourself (whatever that is), then you can get your agent to produce it for you. If you are talking about vibe-coding then yeah, it is going to be trash

> because great code requires intent

If can put properly engineered intent in the prompt that is verifiable, it works wonders. Anything that can defaults to the llm doing its way, you're right, it just can't converge to good, not with proper constraints.