Today I came across ARC-AGI-3 tests for AI models. That is, if you think that AI can do anything for you, consider that you have already earned a large part of the $850,000 prize fund – just submit your beautiful code, and let it be the best. But it doesn’t work that way. Humans handle 100% of tasks, while programs have not yet surpassed the 2.5-3% mark. What are those tasks?
The program is given not a textual task but a mini-game. There are no instructions or descriptions of the rules. At each step, the program receives a “frame” — a JSON object with the current state of the field. It is a grid of up to 64×64 cells, where the cells are encoded with numbers from 0 to 15 (representing different colors or types of blocks). The agent has a standardized interface (actions 1–7), including basic steps and the ability to specify specific coordinates (X, Y). But AI does not know beforehand what these actions do! In one task “Action 1” might shift a block to the right, in another repaint it, in a third turn on gravity. To understand the logic, AI must actively “poke with a stick” at the environment. It performs an experimental action, the environment reacts and returns a new frame. The agent must analyze the changes, update its theory about the physics of this particular world, and make the next deliberate step. The game can end with the status WIN, GAME_OVER, or continue further.
Again – humans score 100%, machines – 2-3%. You can try it at the ARC PRIZE website, there are sample tasks available.
Hi, I'm Rauf Aliev. 