Ilya Sutskever warns that advanced AI reasoning may lead to unpredictability
Ilya Sutskever, former chief scientist at OpenAI, stated that AI systems with advanced reasoning capabilities will become less predictable. He noted that the approach of scaling up data for pre-training AI is reaching its limits, as data growth cannot keep pace with computing power. During a talk at the NeurIPS conference, Sutskever emphasized that while AI has made significant strides, the current method of pre-training will soon change. He suggested that future AI could generate new data and evaluate multiple answers to improve accuracy. Sutskever predicted a future with superintelligent machines that can reason like humans. However, he cautioned that increased reasoning could lead to unpredictable outcomes, similar to surprising moves made by advanced AI in games like Go and chess.