@neuro_stack@thdxr I’m not specing the whole PRD myself, if that’s what you mean. But I like having them to direct, tweak and understand whats happening. I also like having them in the codebase to record/maintain intent.
Maybe it’s not SOTA (curious what is), but it seems to work well.
@thdxr I'm late to the game/sure other harnesses are good too, and not trying to single out claude code necessarily, btw, is just the one I've been using most and finding success with.
@thdxr Agreed. I was reluctant to dive in for a long time because the things heavy users were saying seemed dumb, the tooling/harnesses were bad, and the results seemed bad
Claude code plus a detailed, doc driven research/PRD/review/refine flow seems capable of producing good results
@denbvk@ThePrimeagen Was a bit harsh with the gif, and I get the whole "shift to a more review centric workflow", fwiw/that's reasonable.
Am just tired of people thinking you don't need to open the guts or ever tweak things yourself anymore. You need that for deep understanding.
@denbvk@ThePrimeagen I still find code that looks good in an isolated portion of a diff often looks a lot worse if you see more of the surrounding context.
I convinced myself of the same thing for a bit, btw (diffs are enough). But I always find more issues on closer inspection in an editor.
@denbvk@ThePrimeagen > Yes. LLMs will only be cheaper over time
The amount of subsidization right now is insane.
I agree local will eventually get close enough/can't wait.
But it's silly to use an LLM for EVERYTHING. You don't need a gpu to change a line of css.
@denbvk@ThePrimeagen I find myself discovering more issues when reviewing in my editor than on a gh diff. I'll hop of cc if I want to see surrounding context, and editors are just better for human reading/writing code.
I also do a ton in cc, I just think it's silly to do EVERYTHING there.
@denbvk@ThePrimeagen So there's no longer any circumstance under which you ever see yourself writing code, even in the context of a review... and you want to train yourself to be entirely dependent on a subscription service to do even the most trivial possible task...
@denbvk@ThePrimeagen What if you want to trace/tweak code changes an agent made? Or hand code critical logic? Or make an easy change *for free*/without paying a model provider?
A good harness is an awesome new tool. It's not an omni hammer that turns everything in the world into nails.
@vasuman > This is the slowest, dumbest, and most expensive that AI will ever be.
AI generated data is polluting next gen training sets and AI inference is highly subsidized.
It’s not a given that we haven’t reached a peak.
But yeah, last two statements are true regardless.
@meaning_enjoyer Comparing it to “theatre” isn’t quite right, as the temptation was real/fully experienced.
It’s like when a father eats some food before asking their child to copy them: the food and the eating is real, but it’s a given they can do it.
He’s demonstrating how for the child.
@BretWeinstein@PageauJonathan I think the difference is that Jonathan views religion as the foundation of reason and perception rather than a “thing” you can reason about directly.
People need a top level “meta” perceptual framework. For you, I think it’s rationalism/evolution. For him, it’s Christianity.
@davidabecket@untimelysalts This is interesting/where I partially disagree, but yeah, I think we agree more than not.
I view religion as a glasses prescription for seeing the whole, and agree Nietzsche was corrective. But I view Christ as the focal point of a good prescription rather than a spoke.