Small 4B model trained with reinforcement learning reportedly outperforms GPT-6 Astra at flying a Tesla Roadster

A post on SkyPilot's blog describes post-training a 4-billion-parameter model using reinforcement learning to control a simulated Tesla Roadster flight. The write-up reportedly claims the model performs 21% better than GPT-6 Astra on this task. The source material available is limited to the blog link and a Hacker News thread with 3 points and no comments, so the specific methodology and benchmark details could not be verified. The result is presented as an example of applying RL post-training to a small model for a niche control task.