Fair point. Frozen weights only rule out overfitting through gradients. The harness itself (prompts, tools, lessons) can still overfit the bench. So Propose and Probe only ever see a dev split. Each adopted version gets scored once on a held-out set of new images it never touched, and that's the number we'll report. Lessons are kept as general cue → action rules (low_light → boost_contrast), not per-image fixes. Held-out results ship with the repo.

这就是开源的意义!希望简中互联网开源仓库氛围可以因为智能硬件的浪潮变好

The UI pipeline in this video will be open-sourced together with the repo.

在做手势控制

直接搬运？转发都不转的吗

Iterating fast, open source soon. Follow along🏊

@bot 在深圳给我找到这种农场

看到没笑的这次私信我我v你100

是她是她还是她

压声第一人

我马上搓apple watch里，不和硬件绑死了

Of course, that includes the data for the entire benchmark, which will be stored in the repository.

感谢转发，后续Jev-Harness2完全开源

忘记 Siri 吧。 我们在重构人与 iPhone 的交互方式。 如果一个 Agent 真的能替你操作手机，它本身就不应该再有一个复杂的界面。 Part 1：语音交互与意图理解。 倒计时开始👀

那就用我仓库的gbot，直接cli化让语音调用grok bot，速度提升极为明显

技术内容流量是低呀哈哈

这个炼化的poteto太像我的一个前女友了，实在太绷不住了

Tibo Why am I burning through my quota 100x faster than with Claude?

看了不笑的给我发私信我v你50

We plugged Google Research’s Harness² into JEV 1.14. Zero weights touched, yet driving massive gains on vision benchmarks round after round. Under the hood, the system: ⚡️ Rewrites its own prompts 🛠️ Builds its own tools 🧠 Distills its own experience xitter.cf/openmys/status/2112024…