this has pushed me a little bit toward the "markdown files are executables" camp

but using this still requires a pretty clear understanding of what and where everything is, in addition to the scope of what "everything" is itself

i think the swarm stuff is probably more of a problem of management/delegation

e.g., i've been having a good time with spec-ing + implementing in parallel with github.com/jtan2231/viz...

but this is primarily for delegation w/optional oversight, and heavily dogfooded on a 20k line repo

things like moltbot seem neat but i think are marred by the obvious security issues + whatnot

if there was something like true verification (0 human interference) that's when i think things would be more interesting. otherwise in mixed/ambiguous spaces online i just thing everything's fake

+ it has all the automated testing + architectural docs + whatnot

stet.ink is a better example of something entirely llm-generated, with similar criteria

but frankly this just seems like a provocative misunderstanding of how actually useful LLMs are

Ive long since stopped writing code on github.com/jtan2231/viz...

It’s been dogfooded for a while now. My prompts at this point are either “how do we currently do X?” or “what would Y look like in this repo?”

Followed by drafting + implementation by the project itself

i'll say there's real dopamine for me in seeing that codex worked for 60+ minutes on a problem

i think it's some visceral heuristic of something like effort expenditure + tension + success = dopamine + trust or whatever

obviously inefficient but my dopamine centers don't seem to care too much