Artificial intelligence is rapidly evolving, and recent breakthroughs are challenging conventional approaches to AI development. A new paper from OpenAI reveals something incredible about the nature of AI. Instead of teaching AI sophisticated strategies, a better approach may be to allow it to learn on its own.
The "You Shall Not Pass" Game
In the "You Shall Not Pass" game, a red agent tries to prevent a blue character from crossing a line. When two regular AIs play, sometimes the red agent wins, and sometimes the blue agent gets through. However, when a "hacker adversarial agent" that does absolutely nothing is introduced, it reprograms its opponent to make mistakes and behave randomly. Teaching AI strategies might prevent it from discovering the best ones, including strategies humans would never find.
Generalist vs. Specialist AI
The idea that teaching AI strategies might prevent it from discovering the best ones can be further illustrated when comparing specialist and generalist AIs. Imagine training an AI to master one specific game. If you want to play a different game, you need a different AI. However, something surprising happens when you train one AI to play many games.
Consider a muscular wrestler who has wrestled his whole life (specialist AI) versus a scrawny guy who dabbles in wrestling, swimming, and 20 other sports (generalist AI). Who wins? The specialist should win, but the generalist can beat a specialist that has played one particular game a ton.
OpenAI applied this concept to programming. Their specialist system was shown handcrafted, human-taught data and strategies to excel at one thing and performed well enough to win a gold medal under relaxed conditions. However, a smarter, generalist agent with no specialized knowledge learned on its own and beat the specialist. This is because the generalist AI learns something in one task and can apply it to another.
The Implications
These findings suggest that artificial general intelligence is a possibility. This could lead to many benefits, such as designing new drugs to defeat previously untreatable diseases and giving a personalized teacher to every child. The o3 AI now ranks among the best human programmers in the world. To achieve intelligence, the focus should be on simple algorithms and lots of computing power.
Comments