alt Hacker Newsfair point, this is the thing I struggled most to extract out while writing it - if you can propose an RL environment that penalizes a model for bad design, then I'm all ears - right now there's no fast oracle/verifier for this (as stated in the post)
My current evolving take on "how would you build such a thing" is you need to tee up a roadmap of 20 features and feed them to a model one at a time, so it can't design up front for what's coming.
That way if it builds the first 10 features and the codebase goes to slop, it get's penalized when it can't build features 11-20, or when those features take wayyy more tokens/time/cycles than a model that maintains a clean codebase can do.
This is how most real software is built by most teams - incrementally, getting feedback from users along the way, and steering goals in response.
Replies
Isn't that how AI written software gets better too? By steering the model towards the goals of the user?
I don't know if it's a question of how many features to feed to the model, either. Of course, overwhelm the context with too many features and it will get confused. But that's where the memory management idea that G. Huntley talks about is helpful. You're trying to steer the model within its memory limits towards a certain goal.
The problem is getting it to produce "good" code. Formalizing what that means is the task of the programmer. How do you steer a model towards always, or more often, producing good code so that you don't have to do rework? That's the same problem as with a junior engineer, but the way you do it is different. Right now we're trying to do it with mountains of prompts — which sort of works but has diminishing returns — and with onerous code reviews. We've seen this get better over time, but I think some more mechanical methods will help as we figure out how best to steer the models.
I’ve experimented a bit with this approach for quasi-benchmarking models. Most models nowadays will keep trying until you cut them off, they’re happy to go into millions of tokens for an easy task, digging deeper and deeper into crap design. Compared to a clean reference implementation you get easily 20x token difference at task 11+ for the same model. Somewhat ironically I never published the results because the both the harness and the task set were vibe coded and I didn’t have the energy to clean it all up.