# OpenAI Agent Masters Notoriously Difficult Game from Just One Demo

OpenAI has announced a breakthrough in artificial intelligence learning efficiency: an AI agent that achieved a record score of 74,500 on the classic Atari game Montezuma's Revenge after watching just a single human demonstration.

The achievement is significant because Montezuma's Revenge has long been considered one of the hardest challenges in AI gaming research. The game requires complex problem-solving and long-term planning, making it difficult for traditional reinforcement learning approaches that typically need millions of attempts to learn.

OpenAI's approach is surprisingly straightforward. The agent starts playing from carefully selected points in the human demonstration, then improves its performance using PPO (Proximal Policy Optimization), the same reinforcement learning technique that powers OpenAI Five, their Dota 2-playing system.

This represents a major step forward in sample-efficient