Breakout test: 9B decision model beats 27B by reacting faster
Two Jev-style decision models output left/stay/right probabilities each tick; the 9B decides 15.6 times a second and makes 92 saves, while the slower 27B lets the paddle drift past the ball.
both these are cloudflare's jev like decision models and the 9b beats the 27b at breakout, and it wasn't close > clef flash 9b: 56 ms an answer, decides 15.6 times a second, 92 saves, 5 balls lost > clef 27b: 176 ms an answer, decides 5 times a second, 25 saves, 32 balls lost > both at q4, 3 minutes each they read the game as one line of numbers every tick and answer left, stay or right with a probability each and the game never waits for them watch the paddles. between two answers the 9b's paddle drifts about 42 px, a third of its width, so the next answer still catches the ball. the 27b's drifts 128 px, more than its own width, so it swings past the ball before the next answer lands i checked whether the 27b is just worse at the game. a script that always moves the right way, held back by the same delays, drops from 25.8 saves per ball at 56 ms to 1.4 at 176 ms, the same collapse with perfect judgment the part that surprised me is when they see the game as pixels instead of numbers, the 27b points at the ball 96.6% of the time against the 9b's 68.6%, it's the better reader, and it still loses because it's late. so for anything live, an agent loop or a game, i'd put the fastest model that's good enough in charge and keep the big one for the calls where being wrong costs more than being slow next i want to freeze the game while each one thinks, so only the judgment counts
This post quotes another post: x.com/sudoingx/status/2107480497559412972