Google’s DeepMind unit has proposed a new approach to artificial intelligence, termed “streams,” which allows AI models to learn from their environment through experiences rather than relying solely on human-generated data. In a recent paper, researchers David Silver and Richard Sutton argue that traditional AI development is limited by static data and human biases, which restrict the models’ ability to self-discover knowledge. They suggest that empowering AI with experiential learning will unlock incredible new capabilities. The “streams” approach draws on reinforcement learning principles, allowing AI agents to interact with the world and adapt over time, similar to human learning. This method contrasts with the current focus on short, isolated interactions typical of large language models. By connecting AI agents to rich action and observation spaces, the researchers believe these models can derive valuable feedback from various environmental signals. They envision a future where AI agents operate autonomously, continuously updating their understanding based on real-world interactions. While human-defined goals will still be important, the agents will be able to explore and learn independently, enhancing their performance and adaptability. Overall, this shift could significantly advance the capabilities of AI systems.
