← BackHacker News front pagewonderskyOct 30, 2026, 23:27 UTC+8Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 AstraSkyPilot 在飞行 Roadster 模拟环境中后训练 Qwen3-4B,风暴翻转越车任务的成功着陆次数比不限思考时间的 GPT-6 Astra 多 21%。Read the original ↗ShareCopy linkShare imageExport MarkdownSee every follow-up