kkkamur

I agree with everything you wrote but I would rather choose to be an optimist here because it will unlock an age of discovery where humans are going to live at 10x of their potential eventually and cure cancer and become a society bold and technologically adept enough to cut through the universe.

This is so dystopian dammmmm

Damm this is so cheap literally, I have been running 100s of subagents and it is cheap - with the contributor model ofc :)

Fine tuning is a part of it, but environment generation and evaluation so that the training data matches the production is the hardest part. After the data comes in we can try OPD, RL, SFT to understand what method is the best one. Is this something you would like to try on your production setting ?

We’re building Leibler, a harness that turns production agent traces into training data and environments, then post-trains smaller models for your actual workload.

Production traces → curated data & environments → evals → post-training → cheaper specialized model -> deployment

Our hypothesis: a post-trained Qwen3.8-27B can handle ~90% of production tasks at frontier-level performance, while cutting inference costs by up to 10×.

We have proved this with one of our pilots where we decreased their costs from $50k/month to $4.6k/month, we believe this is the closest we have gotten to self evolving agents.

We’re looking for teams running agents in production to test this with us.

https://leibler.dev