Idea: 让 Harness 成为世界模型研究 OS — 看帧、认路线、给 run 打分、再 RSI #2622
WayneJin0918
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
非官方 / unofficial community idea. 不是 DeepSeek 官方功能提案的背书,只是一条已经能跑的研究回路。
DeepSeek 的产品主线可以不做世界模型。Harness 已经是研究操作系统——缺的是 agent 能看见条带、能先认路线、能给 run 打分,然后再谈 RSI。
一条最小回路
wm.yaml/ eval notes,不是偷偷改 U-Netfixtures/sunset是 8 帧玩具条带:前半段还暖、后半段被涂成冷蓝,late-horizon 一眼能看出来。不占 GPU。为什么值得做成 Harness 习惯
#2454 把「任务 → 组合搜索 → 评测 → Scorecard」写清楚了。世界模型这边可以落成一个很具体的 Pack:
wm.yaml路线图跟 Awesome World Models 对齐:3D prior、2D pixel video、latent space。同一句 “world model” 是三场考试。
已经能玩的草图
社区插件 dsh-wm(非官方):
没有 Harness 也可以:
node cli.js inspect fixtures/sunset --indices first,mid,last node cli.js diagnose "is Sora a world simulator"插件分享帖:Show Your Plugins #2618
想听社区怎么看
欢迎拍砖。我更想先把「能看、能认、能打分」变成 Harness 上的研究习惯,再谈更大的自进化。
DeepSeek’s product mainline can skip world models. Harness is still the research OS. The missing habit is: look at the strip, name the route, score the run, then RSI the harness layer.
A concrete sketch lives in unofficial dsh-wm. Open questions above — especially whether “look at frames” should be a first-class tool-render seam, and whether a rollout directory is a small enough eval case for the Pack loop in #2454.
All reactions