You vs shad0w.
One rifle. Sixty decisions a second.
A first-person duel with Counter-Strike 1.6's rifle rules (a head shot kills, the first shot is accurate only when you stand still, the spray climbs) on a small three-lane desert map with the classic callouts: Long A, Short, Mid Doors, B Tunnels. No enemy on the radar: listen for its footsteps. Your opponent's brain is a shad0w table that learned tactics from an LLM's answers: when to push, sidestep, crouch, jump, fire or reload. Here it runs in your browser, with no server and no network, and makes those three calls every frame in microseconds. The LLM it learned from needs over a second per decision; at this speed it would still be thinking about your first peek.
An LLM taught it. With the game paused.
How hard is it?
You always face the hardest version: its hand reacts in 150 ms, about as fast as the quickest human reactions, and turns fast. Before going live it played 600 rounds against each of two simulated players with the same rules and the same rifle: a skilled player (250 ms reactions) and a very good one (180 ms). They are simulations, not people: a real player also reads sounds, pre-aims angles and learns its habits, so expect your own result to differ.
Same opponent, different brains.
To compare brains with the LLM itself, every brain played against the same scripted opponent with the same, slower hand (not the one you face): 400 rounds each for shad0w and the hand-written rules, 10 each for the LLM itself (every one of its rounds costs real calls). The LLM, with the game paused, decides 10 times a second; its student decides 60. A brain is only as good as its teacher's choices, and a well-tuned hand-written bot is still a strong baseline.