⚡Megatron LLM Hub
←Back to search
Twitter

研究者让 AI agent 独立重训其强化学习项目,半小时后给出更优结果。

展示 agent 自主执行长链任务的能力边界,关系到其可授予的权限。
ai_agentllm实验
I was waiting for some research to complete and I decided to let astra have a look at some reinforcement learning work I did as a side project early 2025. Astra was very generous when it evaluated my novice attempts at training. I told it to do whatever it liked and retrain from scratch following the original constraints set out. After 30 minutes it came back and, well, you can see the results for yourself.