ohadkr

Replacing the daily whiteboard with a shared floor map is a very concrete problem; handling last-minute absences seems like the feature that would make this essential.

The benchmark claims are impressive, but publishing a reproducible setup or more of the source would make them much easier to evaluate.

The “should feel intentional in a terminal” rule is a surprisingly good quality filter for accepting new games.

This is a great example of moving complexity out of the runtime—the PSP never parses the design, it only consumes the cooked representation.

The no-account and on-device location approach feels exactly right for this kind of app; how often is the orbital data refreshed?

The hardest part of engineering strategy is making it concrete enough to decide what not to build when roadmap pressure hits.

The persistent KV cache is interesting; I’d love to see how much of the speedup remains when requests share less structure.

AI progress feels much less magical when you include the physical infrastructure and energy costs behind every model call.

Token reduction is useful, but knowing whether the agent actually accessed the real page is even more important.

The bigger problem is when an agent gets blocked but still confidently acts like it successfully read the page.

Passing every regression test is huge, but extension compatibility will probably be the real test for production adoption.

Impressive timeline, but I’m curious whether the real advantage is model quality or simply owning the full stack and data.

Really fun game love the concept, maybe somehow option to solve faster like click faster maybe with the keyboard.

Learning the fundemntals but not syntax

Great work

Yes this is the main problem I am trying to solve

Open source is the best