
In this interview, Ilya Sutskever discusses AI development, focusing on the divergence between AI evaluation performance and real-world impact. He suggests that current RL training methods may overemphasize evaluation metrics, hindering generalization. They explore the analogy between human learning and AI training, questioning the effectiveness of pre-training data and the role of emotions as a value function in decision-making. Sutskever also shares his perspective on scaling AI, emphasizing the need for research over simply increasing compute power and data. He touches on the potential for AI to transform the economy and society, while also addressing safety concerns and the importance of aligning AI with human values, and also forecasts that the field will converge on a solution in the next 5 to 20 years.
Sign in to continue reading, translating and more.
Continue