nwyin

i think this take gets misinterpreted and misapplied. here are some other sentences to think about.

to become a good chef, eat a lot. to become a great producer, listen to a lot of music. to direct great films, watch a lot of movies.

these sorts of advices superficially soothe the reality that you're spending too much time consuming and not actually creating. or even practicing creation.

very cool! I had a similar idea a few weeks back and built something much smaller scoped and shaped to my personal idiosyncrasies. my approach was giving claude a CLI interface to manage worktrees and spin up other claude instances to work in them, but have all the orchestration handled via a daemon rather than trying to prompt claude to do everything.

chatml looks very polished and promising. good luck with the project and hope it succeeds.

cool!

anthropic mentions MCPs eating up context and solutions here: https://www.anthropic.com/engineering/code-execution-with-mc...

I built one specifically for Cognition's DeepWiki (https://crates.io/crates/dw2md) -- but it's rather narrow. Something more general like this clearly has more utility.

too many things in flight!

inspired by the karpathy/twitter posts on running (semi) autonomous research loops, I build https://github.com/tnguyen21/labrat to be able to try and replicate some paper results over night. still early stages but I'm getting some use out of it already.

also spending a lot of time thinking about how you "close the loop" on software projects. right now figuring out how you can combine static analysis + review heuristics to let LLMs course correct the codebase when they over-engineer or produced unwieldy abstractions.

it's a bit dated, but Terence Tao has a video of formalizing a proof with LLMs from 9 months ago which should be illuminating

https://youtu.be/zZr54G7ec7A?si=-l3jIZZzfghoqJtq