Comment by y1n0
14 hours ago
The problem I’m finding, at least with today’s models, is that it produces disposable software. It’s not good at making well-architected, durable software. Stuff that could be maintained and bug-fixed.
I mean you can get good stuff out of it if you know what you are doing and guide it, but if you just say “here’s a regression suite. Write an implementation that passes” you will get something that works for a while but ages quickly and will need to be thrown away.
With today’s technology I’d still want a GPU driver developer guiding the LLM rather than some rando who is out of their element. But cutting down the exploration cycle time and giving the developer massive parallelism (have 10x agents exploring different hypotheses or features) is the real win. We don’t need to skip all the way to slop just to squeak out a little more effort savings.