mufeez

I post-trained Qwen3-Coder to fix bugs using an actual debugger. The result: Solve rate: 70% → 89% Median turns to fix: 46 → 19 (-59%) Instead of just reading code or print-debugging, it: - reasons from execution - inspects live variables and call stacks - sets breakpoints, steps, and evaluates expressions

ok london kinda mogs nyc ngl

@_lychrel @bilaltwovec @_arohan_ walked in for dinner around 9

@_lychrel @bilaltwovec @_arohan_ ambassador clubhouse is their sister restaurant a few blocks away, might offend someone but it’s similar food and vibes

How do I buy spv on dishoom and hoppers?

@swyxio yikes

@ily_8750 lol @bilaltwovec

Canary Wharf will always be my favourite part of London

@tenderizzation inshallah the scaling laws hold

tuning into the gdb testimony like it's the super bowl

@willccbb @linear you should check out linear.app/diffs

let's just let @linear do it

@arnie_hacker @chesterzelaya @OpenAI nice! excited to see the results

@chesterzelaya @OpenAI guessing you mean my cs2 project? compute+cloud service quotas💀 i'm rendering out 1K-hours cleaned dataset atm and moving to training baselines this week though :)

Okay so the @OpenAI image-gen team have definitely rendered out a bajillion cs:go games

@__morse @mingjie hmm

GitHub if Linear designed it

Don't tell anyone (to enable, reconnect GitHub integration with new permissions). Agents which output draft PRs also get code diffs inside Linear now

@bilaltwovec caffeinate -di -w ftw

sudo pmset disablesleep 1 ftw

@tenderizzation ootl, why did they go for an old school dense model?

highlighting the lowest score is crazy

@gilescope imo DAP would be a good fit, since most people are using debuggers within editors: microsoft.github.io/debug-adapter-… the harness could use that instead of a CLI tool; it’s designed to be pretty extensible

@moofeez Just wondering what the ideal agentic debugger interface would be… is there an MCP for a language server? Or is there a more efficient mechanism…

I post-trained Qwen3-Coder to fix bugs using an actual debugger. The result: Solve rate: 70% → 89% Median turns to fix: 46 → 19 (-59%) Instead of just reading code or print-debugging, it: - reasons from execution - inspects live variables and call stacks - sets breakpoints, steps, and evaluates expressions

@bartolomeo_diaz @MaximeRivest nvim!

@moofeez @MaximeRivest What is the terminal editor/ide u using?

@JedidiahMain i’ll cover this in my blog post – a lot of the project was taking proven approaches and adapting them to my specific training objective/shape

@moofeez coming from a web SE, what do i need to know to be able to do this?

@GrapotteM @PrimeIntellect thanks! x.com/moofeez/status…

@moofeez @PrimeIntellect Fantastic work! How much did it cost you ?

@EliasLumer @willccbb I explored the standard harness + tool call approach, though there’s definitely room for experimentation here

@moofeez @willccbb Interesting. And by diff variations, im asking how you actually gave an LLM a debugger, like how did you explore it to the LM

@EliasLumer @willccbb will publish when i’m done the blog post!

@moofeez @willccbb This is great, do you have the implementation (tools for debugger, etc) open source? Would love to see/implement this

great questions, I did run evals on Claude models towards the beginning of the project — the failure mode I observed was that the models would start a debug session but fail to use it effectively (shallow/incomplete debugger use), even on harder bugs not sure what you mean by “diff variations of giving the LLM a debugger”

@moofeez @willccbb Did you give opus 4.6/ gpt5.5 a debugger? And see results there? Also did u try diff variations of giving the LLM a debugger?