I Made This AI Play 2048 (it got crushed 108,474 times)

Ben is an AI cube. He stands on four arrow pads in front of a giant 2048 screen: one hop, one move. Above him hangs a press with 16 notches, one for every empty cell on the board. Every new tile drops it a notch. Every merge lifts it back up. When the board is full and nothing can merge, it falls.

Nobody told Ben the rules or a single strategy. He gets flattened again and again, and every crush teaches him a little. 200,000 games.

How far will he get: 2048? 8192? 16384? Write your guess in the comments. The answer is at 5:19.

What you’ll see:
• An AI learning 2048 from zero, with no strategy told
• Ben’s choice on screen: the real score of each of his four moves, every move
• Inside his head: the four futures he imagines before every move
• How he finds the famous corner trick by himself (10% → 53% of his moves)
• Game 108,475: eleven moves with the press on the last notch

Nothing is staged. Ben is a real learning agent (TD learning with an n-tuple network). I don’t control his decisions; every move, merge, crush and number on screen comes from the recorded training games. The press is the show’s way to make empty cells visible: its height is the number of empty cells on the board.

I taught an AI 2048 from scratch. It made 108,474 mistakes before finally figuring out how to reach the highest tile.

Ben started with zero knowledge of the game rules, no understanding of tile merging, and no concept of the goal. He had to figure out how to increase tile values through pure trial and error. The video tracks his progress as he moves from random inputs to strategic planning, showing how an AI builds a mental model of a system that it has never seen before.

This is a look at how reinforcement learning allows a system to master complex logic. By the end, Ben achieves a score of 108,475 and 91% completion of the board. Nothing is staged. Ben is a real machine learning agent, and every move you see comes from the recorded training games where he slowly optimized his decision-making process. This artificial intelligence experiment provides a clear view of how an AI learns to optimize a board without being told how to win.

Every episode, Ben gets a harder challenge. Subscribe to AI ArenaLAB and see what he learns next.

0:00 Crushed, again and again
0:12 2048 with a press over his head
0:45 Game one
1:22 Inside his head: four futures
1:48 His first 2048
2:24 Why he still lost
2:36 Thousands of games
2:46 Where he keeps his biggest tile
3:12 One cell from the end
3:48 The corner trick
4:08 Game 108,475
4:44 The last notch
5:19 The answer

Watch the latest AI ArenaLAB experiments:
I made this AI cross a busy road:
I gave this AI only 3 memories:
An AI plays Pac-Man without the rules:

What happens when an AI forgets everything after every mistake?

Every episode, Ben gets a harder challenge. Subscribe to AI ArenaLAB and see what he learns next.

#AI #2048 #MachineLearning

source