Google DeepMind researchers David Silver and Richard Sutton argue that artificial intelligence (AI) has outgrown static training data and must now evolve by learning through experiences in dynamic environments. In a recent paper, they propose a new approach called “streams”, where AI agents interact with the world over time to form goals and adapt their behavior, much like humans do.
They critique current large language models (LLMs) for relying on short, prompt-based interactions shaped by human judgment. This limits AI’s ability to discover new strategies or knowledge independently. Instead, stream-based agents would continuously learn and adapt by interacting with their environment, receiving “reward signals” from real-world outcomes like health metrics, energy use, or productivity.
This method builds on reinforcement learning, used in DeepMind’s AlphaZero, but expands beyond fixed-rule games to open-ended tasks. AI agents could pursue long-term goals such as improving health, learning languages, or conducting scientific research.
While the approach could unlock superior AI capabilities and even surpass human intelligence, the researchers warn of significant risks, including reduced human oversight. However, adaptive AI might also learn to avoid causing harm by recognizing and responding to human feedback.
In short, Silver and Sutton believe “experience-based learning” is the next step toward achieving artificial general intelligence (AGI), with the potential to radically transform how AI understands and interacts with the world.
