AI's Reasoning Power Will Make It Less Predictable, Says Ilya Sutskever

Former OpenAI chief scientist Ilya Sutskever, a leading figure in artificial intelligence, has predicted that AI's reasoning capabilities will make the technology far less predictable, reports Reuters.

Speaking at the NeurIPS conference in Vancouver, Sutskever, who is now co-founder of Safe Superintelligence Inc., argued that the current approach to "pre-training" AI systems with massive datasets is nearing its limit. He noted that while computing power continues to grow, the amount of relevant data available for training is constrained by the finite nature of the internet.

Sutskever suggested several ways to overcome this obstacle, including using AI itself to generate new data and enabling models to evaluate multiple answers before providing a response. Other researchers are exploring the use of real-world data for training.

However, Sutskever's main point focused on the unpredictable nature of future AI systems with advanced reasoning abilities. He drew a comparison to AlphaGo, a system developed by DeepMind that surprised experts with its complex moves during a 2016 victory over Go champion Lee Sedol.

"The more it reasons, the more unpredictable it becomes," Sutskever stated. He further emphasized that even the most sophisticated chess AIs are unpredictable to top human players.

Sutskever envisions a future where AI agents will possess a deeper understanding and self-awareness, reasoning through problems like humans do. He emphasized that this new generation of AI will be "radically different" from current systems.