Video unavailable. Open the source below.
NewJEV-Harness²
Google Research's Harness² wired into JEV 1.14 so the harness rewrites prompts, builds tools and distils experience without touching weights.
Video unavailable. Open the source below.
NewGoogle Research's Harness² wired into JEV 1.14 so the harness rewrites prompts, builds tools and distils experience without touching weights.
Limited inference from the post and clip: Jev supplies the model decisions inside a harness that rewrites its own prompts, builds its own tools and distils experience between rounds. No weights are changed. The harness code and benchmark numbers are not in the saved post.
It tests the claim that the harness around a decision model, not the model itself, carries most of the gain. The post reports gains on vision benchmarks; the runs and evaluation setup are not public, so the size of the effect stays unverified.