Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra
Researchers trained a 4‑billion‑parameter language model using reinforcement learning to generate flight plans for a Tesla Roadster. The resulting model outperformed the GPT‑6 Astra by 21 % in simulated flight performance metrics. The work demonstrates that smaller, fine‑tuned models can surpass larger baselines when tailored to specific tasks. The study was shared on Hacker News in the field of autonomous vehicle navigation.