RL-post-trained 4B model flies Tesla Roadster, beats GPT-6 Astra and other frontier LLMs
skypilot_org · x · 2026-10-23
After Tesla's flying Roadster demo was delayed two weeks "due to high winds," the SkyPilot team built an RL environment letting LLMs fly it. Result: GPT-6 Astra, DeepSeek, Muse Spark Open, Beam, and Nemotron 3.5 Super VL all landed 0% of kickflips in real time—even with unlimited thinking time per decision.
They then post-trained a Qwen 4B model with RL—first learning from teachers, then practicing itself—using SkyRL and vLLM, orchestrated by SkyPilot's Job Group. That small model outperformed the frontier LLMs, even flying in a storm. A fun yet hardcore demo that small models + domain RL can beat general frontier models.
More from Fun
- 3 devs spent 2 years building an AI-assisted roguelike horror card game, demo on Steam — AlvinhoGames_ · 2026-10-23
- 'Woken by the token reset': developers joke about living on AI usage quotas — burhop · 2026-10-23
- AI-generated movie poster 'The Chase' crams a wolf, a plane and muscle cars into one chase — Aggravating-Sir-115 · 2026-10-23
- Reporter Loops an AI Agent to Hoard Shopping Cart Holds for Hours — lfschiavo · 2026-10-23
- Dating app rejection comes back with only an 'InvalidDateException' error message — gregmushen · 2026-10-23
- AI circle joke: mathematicians should teach 'ethics of matrix multiplication' for Anthropic's ToS — kuza55 · 2026-10-23