Computer use on Jev at about 90ms a decision
Milind S is driving a Mac with Jev at roughly 90ms per decision, no screenshots and no LLM in the loop. Meanwhile Zachi killed pg-jev after 3 billion tokens of testing, because query speed falls over above 50k rows.
Built with it
Milind S drives a Mac with Jev at about 90ms per decision
Milind S built a computer use loop where Jev picks which UI element to click, with no screenshots and no LLM. A local CoreML model segments every button on screen, on-device OCR reads the labels, and that text is all Jev gets, at about 90ms per decision. ThePrimeagen described a similar probabilistic setup built on Cloudflare's decision API.
~90ms per decision. Faster than any LLM computer use I've tried.
Drew Breunig sketched a DSPy model router on Jev in 15 minutes
Drew Breunig built a model router with DSPy and Jev that picks the model from the initial prompt, sketched in about 15 minutes. Harrison Chase said LangChain ships this shape as ModelRouterMiddleware, and that Jev is cheap enough to re-pick the model after each tool result with a custom hook. Breunig said he was surprised at how well it works.
jev is cheap enough that you can also re-pick after each tool result with a custom hook
The case against
Zachi ends pg-jev: query speed breaks above 50k rows
Zachi called the newest pg-jev release its last, saying the thing will not work as expected. Under 50k rows it is fine, but above that query speed is the bottleneck, a verdict he reached after spending about 3 billion tokens on benchmarks and testing.
Under 50k rows its fine, but for everything above that, query-speed is a bottleneck here.