The Future of AI: Less Predictable, More Powerful

At the NeurIPS conference in Vancouver, former OpenAI chief scientist Ilya Sutskever made a striking prediction: as AI systems gain reasoning capabilities, they will become far less predictable.

Sutskever, who co-founded Safe Superintelligence Inc after his time at OpenAI, discussed the limitations of current AI models, particularly the constraints of data availability. He emphasized that while computational power is increasing, the data used to train these models is finite, as "we have but one internet."

Despite these challenges, Sutskever outlined potential solutions, such as AI generating its own data or evaluating multiple answers to improve accuracy. He also touched on the development of AI agents with deeper understanding and self-awareness, capable of reasoning like humans.

However, this reasoning power comes with a caveat: the more AI reasons, the more unpredictable it becomes. Sutskever illustrated this point with examples like AlphaGo's unexpected moves in the game of Go, and the unpredictability of top chess AIs to human players.

As AI continues to evolve, Sutskever believes it will be "radically different," ushering in an era of superintelligent machines that, while powerful, may remain enigmatic to their creators.