The time when artificial intelligence could only operate with limited human data has reached its end, according to Google DeepMind researchers David Silver and Richard Sutton. The paper Welcome to the Era of Experience presents the authors’ view that AI needs to progress past using static training data and episodic interactions. AI agents will learn in a manner that parallels human learning, according to their proposed future model.
The creators of AlphaZero Silver, together with reinforcement learning pioneer Sutton, express their view that generative AI models like ChatGPT remain limited by human input and evaluation. These models function in separate sessions that have no memory storage and cannot develop long-term objectives. The models face limitations when it comes to independent improvement or exploration.
The researchers propose “streams,” which describes an AI system that interacts with the world through time while learning and adapting through real-world signals, including health metrics and energy consumption and profit data. The approach uses reinforcement learning principles to allow agents to find knowledge that exceeds their human programming.
According to the researchers, this experiential learning process enables capabilities that surpass the current reasoning models, including Gemini and DeepSeek. The authors indicate that enabling AI to operate independently and receive environmental feedback leads toward the development of Artificial General Intelligence (AGI). The authors warn about potential dangers when humans lose their ability to control AI decision-making processes even though they support the concept of long-term autonomy.
Still, the potential is vast. Stream-based AI functions as a health monitoring system, provides educational guidance, and enables scientific discoveries. According to Silver and Sutton, AI will surpass human knowledge through experiential data, which will establish a new period of artificial intelligence that goes beyond human intelligence.
