toks

中文参考

我让 Jev 和一个 LLM 玩 Doom，但加了一个变化：尝试多个未来，然后从最好的结果继续。Jev 选择动作；自定义 harness 使用 @microsandbox，从同一检查点并行运行多个尝试；LLM 复盘发生的事情，并指导下一步尝试。

原帖全文 / EN

i got jev + an llm playing doom, with a twist: try several futures, then keep playing from the best outcome. jev picks the moves. a custom harness uses @microsandbox to run attempts in parallel from the same checkpoint. the llm reviews what happened and guides what to try next.