My first Vibe RL experience | sriraam
sriraam.me · 3,910 words · saved by 1 readers
A log of how I used Prime Intellect and Codex to Vibe RL and got Qwen3-4B to beat GPT 4.1 mini by ~2x at drawing ASCII diagrams.
March 5, 2026•Sriraam Raja A log of how I used Prime Intellect and Codex to Vibe RL and got Qwen3-4B to beat GPT 4.1 mini by ~2x at ASCII diagrams. Before we dive in, I want to clarify what I mean by Vibe RL and how it is different from Vibe Coding as defined by Karpathy Sensei. Vibe RL obviously involves Vibe Coding, but the key distinctions are: The agent interacts with experiment tools in realtime and makes decisions without much prompting. Proposes concrete hypotheses with verifiable steps to analyze the results. Builds custom interfaces for collaborating with humans.…
saved by
related reading
- Lil'Loglilianweng.github.io
- LLM Powered Autonomous Agents | Lil'Loglilianweng.github.io
- Shipping at Inference-Speed | Peter Steinbergersteipete.me
- As Rocks May Think | Eric Jangevjang.com
- Vibe engineeringsimonwillison.net
- Here’s how I use LLMs to help me write codesimonwillison.net
- How to Vibe Code as a Senior Engineerblog.alexmaccaw.com
- What I learned building an opinionated and minimal coding agentmariozechner.at
- The vibe coder's career path is doomedblog.florianherrengt.com
- State of RL for reasoning LLMs | A. Weersaweers.de
- Non_Interactive – Software & MLnonint.com
- combinatorial reasoning environments for LLMs and RLdjdumpling.github.io