July 14, 2026
AI taught AI and HN chose chaos
Show HN: I RL-trained an agent that trains models with RL (for –$1.3k)
Coder says his AI taught another AI for minus $1.3k — and the comments instantly split
TLDR: A developer says he built and open-sourced a system where one AI learns to train other AIs, using real rented computers and rewards for better results. Commenters were split between impressed and deeply skeptical, with debate raging over whether this is a breakthrough or just flashy “vibe-coded” automation.
A Hacker News maker rolled in with a very internet-core flex: an AI agent that learns to create training jobs for other AIs, then sends them off to rented GPUs to see what actually works. Even better, he says he open-sourced the whole thing — code, weights, scripts, failures, receipts, the lot — while claiming the system improved over time and even handled a fresh task it hadn’t seen before. In plain English: he built a robot coach for robot students, then trained the coach by rewarding it when the students got better.
But the real show was the crowd reaction. One camp basically went, “Cool idea, why is this getting buried?” with one commenter openly baffled that the post got “downvoted into oblivion.” Another group immediately hit it with the classic internet side-eye: is this genuinely new, or just another AI agent with access to more compute and a lot of sparkle? The emoji-heavy readme even caught strays, with one commenter dryly asking if it was “mainly codex,” which is about as Hacker News as a raised eyebrow gets.
Then came the skeptics swinging hardest. The sharpest critique was that in a messy field like reinforcement learning — basically trial-and-error training with rewards — it’s dangerously easy to “vibe code” something huge without really understanding what it’s measuring, where it breaks, or whether it’s actually good. So while fans saw a wild open-source experiment, critics saw a potential black box wearing party confetti. Either way, the comments made one thing clear: people are fascinated, suspicious, and absolutely unable to look away.
Key Points
- •The article describes a two-loop RL system in which an outer agent writes RL training jobs for smaller inner-loop models.
- •The full project is presented as open source, including weights, harness, task families, reward code, GPU orchestration, and pilot write-ups.
- •The outer loop uses Tinker and tinker-cookbook with importance-sampling GRPO, while the inner loop uses prime-rl with GRPO on Runpod GPUs.
- •Each episode includes task specification, sandboxed job authoring, validation with capped retries, dispatch to GPUs, and reward based on trained-model performance.
- •Reported results show reward increasing from about 0.0 to a peak near 0.63 over 54 training steps, including transfer to a held-out task family.