Pong
A classic arcade game that simulates table tennis; the player controls a paddle to bounce the ball back.
Timeline (last 90 days)
In experiment 1, the researchers built a multilayer perceptron with 500 inputs and 5 outputs, trained on four games—Breakout, Pong, Pac-Man, and Freeway—and the LLM, relying solely on the network architecture and raw weights (no tools, no network access), correctly identified all of the corresponding games.
In the experiment, Qwen 3.8 27B reached 75% recognition confidence for Pong.
In experiment 1, Astra was given only the raw weights of four game networks (Breakout, Pong, Pac-Man, Freeway) with no hint of which games they were, and correctly named all of them.
In the MIT experiment, four tiny policy networks played grid-simplified versions of Breakout, Pong, Pac-Man and Freeway respectively; each network had 17,253 weights, and Astra got all four right, with confidence between 96% and 99%.
In the Pong-related tests, the run time exceeded one hour; on the Freeway task, Qwen simply failed.
In the game-identification experiment, Astra correctly identified all four small networks (Breakout, Pong, Pac-Man, Freeway) using only the raw weight-number text of each.
A developer's locally built AI project called Aura, running on a Mac, completed a full end-to-end task from code debugging to game control using a single natural language instruction, and defeated the computer opponent 5 to 2.
The complete instruction the developer gave to Aura was: Pong is broken, fix it, then play against the computer until you win.
All Facts (8)
In experiment 1, the researchers built a multilayer perceptron with 500 inputs and 5 outputs, trained on four games—Breakout, Pong, Pac-Man, and Freeway—and the LLM, relying solely on the network architecture and raw weights (no tools, no network access), correctly identified all of the corresponding games.
50%UnverifiedIn the experiment, Qwen 3.8 27B reached 75% recognition confidence for Pong.
50%UnverifiedIn experiment 1, Astra was given only the raw weights of four game networks (Breakout, Pong, Pac-Man, Freeway) with no hint of which games they were, and correctly named all of them.
50%UnverifiedIn the MIT experiment, four tiny policy networks played grid-simplified versions of Breakout, Pong, Pac-Man and Freeway respectively; each network had 17,253 weights, and Astra got all four right, with confidence between 96% and 99%.
50%UnverifiedThe complete instruction the developer gave to Aura was: Pong is broken, fix it, then play against the computer until you win.
50%UnverifiedIn the Pong-related tests, the run time exceeded one hour; on the Freeway task, Qwen simply failed.
50%UnverifiedIn the game-identification experiment, Astra correctly identified all four small networks (Breakout, Pong, Pac-Man, Freeway) using only the raw weight-number text of each.
50%UnverifiedA developer's locally built AI project called Aura, running on a Mac, completed a full end-to-end task from code debugging to game control using a single natural language instruction, and defeated the computer opponent 5 to 2.
50%