Google’s DeepMind Unit Proposes Revolutionary ‘Streams’ Approach for AI Advancement
Google’s DeepMind unit is pushing the boundaries of artificial intelligence with a novel approach called ‘streams’, allowing AI models to learn from their environment without being limited by human pre-judgment. This development could potentially propel AI beyond current capabilities and even lead to artificial general intelligence.
The current AI landscape is dominated by large language models (LLMs) that are trained on static data and limited to answering individual human questions. DeepMind scholars David Silver and Richard Sutton argue that this approach is restrictive and that AI needs to be allowed to have “experiences” and interact with the world to formulate goals based on environmental signals.
The ‘streams’ concept builds upon reinforcement learning and the lessons learned from AlphaZero, DeepMind’s AI model that defeated humans in Chess and Go. Silver and Sutton suggest that by allowing AI agents to inhabit “streams of experience” rather than just short interaction episodes, they can develop long-range goals and learn from their actions over time.
How ‘Streams’ Works
- AI agents interact with the world through a “stream of experience” that progresses over a long timescale.
- The agents learn via reinforcement learning principles, receiving feedback in the form of “rewards” as they explore and take actions.
- These rewards train the AI model on what is more or less valuable among possible actions in a given circumstance.
- The world provides various “signals” that serve as rewards, such as cost, error rates, productivity, and health metrics.
The researchers envision a future where AI assistants with long-range capabilities could track a person’s health over months or years, provide educational support, or even pursue ambitious scientific goals like discovering new materials or reducing carbon dioxide.
However, they also acknowledge the potential risks associated with AI agents that can autonomously interact with the world over extended periods. These risks include job loss and the possibility of humans having fewer opportunities to intervene in the agent’s actions.
Despite these challenges, Silver and Sutton are confident that the ‘streams’ experience will generate vast amounts of information about the world, potentially dwarfing the data used to train today’s AI models. This could lead to unprecedented capabilities and a future profoundly different from anything seen before.
As AI continues to evolve, the ‘streams’ approach represents a significant step towards creating more advanced and autonomous AI systems that can learn from their environment and adapt to new situations.
