Anthropic has achieved its artificial intelligence completing Pokémon Fire Red without external help, only by analyzing screenshots. The system learned through trial and error, like a human player but without fatigue. However, the thread breaks when you ask it something about politics or health: it responds with hallucinations.
Visual reinforcement learning without human intervention 🎮
Anthropic's model processes each frame of the game as a static image, making movement and combat decisions based on rewards. It does not access RAM or internal cartridge data. It uses convolutional neural networks to identify sprites and text, and a reinforcement algorithm that optimizes the route. Training required thousands of simulated playthroughs, but the result is an agent capable of defeating the Elite Four without knowing the game's rules.
It beats the game, but doesn't know what a political party is 🤖
The same AI that defeats Mewtwo with a leveled-up Charmander goes blank if you ask it to explain the healthcare system. It's like a friend who solves a Rubik's Cube in ten seconds but thinks Finland is in Africa. Very skilled with pixels, but zero common sense. At least it doesn't complain about loading times or ask you to save the game every five minutes.