dejavucoder

fair enough

we live in exciting and scary times...

you may notice Kimi, GLM have also started telling how their model is able to optimise it's own inference pipeline

https://www.kimi.com/blog/kimi-k3

fair argument

this is true. in one of the later problems (cholesky decomposition), the organizer ran the submissions on a tiny training run to validate... and also provided code for same for our reference. most of the top solutions hit 4/8 or so. not very numerically stable.

i found out that as i learnt more domain wise, i was (obviously) able to steer better. doing a re-write can also remove lots of slop and context rot (and subsequently make it easier for both human and LLM to make solution more numerically stable, less reward hackish)

submission #2 by gau nernst is most numerically stable

1. labs have lots of inference capacity 2. they will have domain experts working on this so their efficiency is gonna be exponentially more (can direct LLM better, save money, reach same results faster)

hello author here.

yes, it gives labs edge and leads to self-recursive improvement loops.

also i was myself able to finish 7th in a later competition with 2-3 other approaches which are variants of the method discussed in this blog.

in general, having a harness as thin as possible with some problem specific instructions while controlling for context rot is the key.

point i am trying to make is there are a lot of optimisation surface areas possible.

author here!

welcome! check out my featured section

It's just a you can tell claude to make to write notes to

Can probably give access to tools like ast-grep to Claude. Will help it see all references. I still agree some dynamic references might still be left. Only way is to prompt well enough. Since I tested this on a Ruby on Rails codebase, I dealt with this.

lol yeah

thanks!

Fair point.

For context, I was using Claude Code on a Ruby + Typescript large open source codebase. 50M+ tokens. They had specs and e2e tests so yeah I did have feedback when I was done with a feature - I could run specs and Claude Code could form a loop. I would usually advise it to fix specs one by one. --fail-fast to find errors fast.

Prior to Claude Code, I have been using Cursor for an year or so.

Sonnet is particularly good at NextJS and Typescript stuff. I also ran this on a medium sized Python codebase and some ML related work too (ranging from langchain to Pytorch lol)

I don't do a lot of prompting, just enough to describe my problem clearly. I try my best to identify the relevant context or direct the model to find it fast.

I made new claude.md files.

I use Claude Code 50% of times with Cursor now due to the diff and tab. The extension is just a bit buggy sometimes otherwise I would use it much more. I hit some node related bugs today while searching stuff with it (forgot to report to Anthropic lol). Other bugs include a scroll stuttering.

Almost feels like a game as you level up!

Thanks, I will check this out