返回探索

Jev 实验心得:需要明确提示、成本极低,用 Astra 调优

用户分享使用 Jev 的实验发现:需要像 GPT-4 时代模型那样显式提示,成本非常低,并使用 Astra 进行了约8小时的调优。

A few learnings from this experiment: - Jev feels like a throwback to GPT-4-level "dumb", where you need to prompt a lot of things explicitly. (e.g. 'kill when you are fully isolated for the last 5 seconds') - Insanely cheap $: Let Astra run with it to tune for like 8h, and it

· 1 次赞在 X 打开