2026-05-07

I’m building bichess, an open-source platform for Fog of War chess. In Fog of War, each player only sees the squares their own pieces could legally move to; everything else is hidden. Surveying the existing landscape this week made the strategic frame land sharper than it had before: the question isn’t whether a niche is empty, it’s which kind of empty.

The endpoints are occupied. CMU’s Obscuro paper, revised in February, claims the first superhuman AI for the variant. It beats prior SOTA and the world’s best human players. Closed-source, no public play surface. That’s the research ceiling, and it confirms the variant has the strategic depth to draw a top imperfect-information AI lab.

The other endpoint is chess.com. Their Fog of War mode runs an annual community championship: $2,500 purse in 2024, 3,844 peak viewers, 11,559 hours watched. Web-only, no engine, no postgame reveal, no replay polish. That’s the casual floor: a real audience that already shows up, served by a thin product.

The middle is empty. The most-starred open-source Fog of War engine on GitHub has 7 stars and is inactive. Lichess won’t fill the gap. When users petitioned for the variant, a maintainer’s reply was “It’s difficult to add (I spent a day on it and didn’t get anywhere),” and lichess has stated they have no plans for new variants. Whatever lands in the middle won’t be displaced by either endpoint moving in.

The trap with empty-niche thinking is treating all empty markets as opportunity. Most empty markets are empty because nobody wants them: no research interest, no audience, no casual surface. Building into one of those is just lonely. The diagnostic that separates “real territory” from “wasteland” is whether the endpoints are occupied. A closed research ceiling above proves the depth. A thin casual floor below proves the audience. The middle is only worth building when both endpoints are present and neither is shipping the open and accessible version.

The same shape shows up outside chess. Imperfect-information game AI generally: Libratus solved poker without producing public poker bots; AlphaStar won at StarCraft without a surface to play it. Music generation: academic state of the art above, consumer apps with the knobs hidden below, thin middle. Robotics simulation: heavy research stacks above, toy hobbyist projects below, little in between.

If a niche looks empty, find the endpoints first. A closed research ceiling and a thin casual floor are the structural permission slip. Without them, the emptiness is the answer to a question nobody asked.