Beyond Human Minds
For AI to create new knowledge it must explore on its own terms
Richard Sutton and David Silver1, pioneers of artificial intelligence, are reshaping the field once again. Sutton, a foundational figure in reinforcement learning, and Silver, the architect behind DeepMind’s AlphaGo, argue for a paradigm shift in their position paper, The Era of Experience. Their vision moves AI beyond the constraints of human knowledge, advocating for systems that autonomously explore and discover new realms of understanding. This essay explores their argument, its implications, and one crucial question that will dominate our conversation about AI in years to come.
The Limits of Human Knowledge
Human intelligence, developed over millennia, has culminated in AI systems like large language models (LLMs), which encapsulate vast swaths of human knowledge scraped from the internet. Yet, as Sutton and Silver contend, this knowledge represents only a fraction of what is possible. They evoke the metaphor of a circle: 5% is what we know, 5% is what we know we don’t know, and a staggering 90% is what we don’t know we don’t know. Current AI, tethered to human priors, is confined to the first 10%. The authors propose an AI that ventures into the unknown 90%, unbound by human limitations.
This vision rejects the idea of anchoring AI to human knowledge. Why constrain a system capable of surpassing our cognitive boundaries? Instead, Sutton and Silver advocate for AI that learns through experience, exploring uncharted territories of knowledge through trial and error.
The Mechanics of Experiential AI
At the heart of their proposal are two key components. First, an AI agent that continuously updates its value function based on signals and rewards. Unlike today’s static AI systems, which are trained once and then deployed—such as Tesla’s Full Self-Driving software or Google’s Gemini—this agent learns dynamically, adapting its goals, rewards, and values in real time. Second, the system relies on simulated trajectories, akin to Monte Carlo Tree Search used in AlphaGo. By simulating billions of scenarios through self play, the AI evaluates and selects the most promising paths, refining its understanding through experience.
This ability to simulate countless trajectories in parallel is where AI diverges from human cognition. Humans are limited to a single life’s worth of experiences, while AI can iterate through billions of simulated outcomes, gaining insights far beyond our reach. This capacity for massive parallel exploration is the cornerstone of the era of experience.
A New Paradigm in Action
Consider a practical example: advising a student on a college major at Berkeley. Today’s AI, rooted in human knowledge, might recommend computer science due to its lucrative prospects. However, an experiential AI would take a different approach. It would first assess the student’s interests and strengths learned through numerous interactions with the student. Then, through simulated trajectories, it could explore emerging fields and future trends. For instance, it might predict a non-trivial probability that Mars will be terraformed within two decades, requiring constitutional lawyers to shape new societies. Recognizing the student’s aptitude for legal and societal issues, the AI could recommend a tailored academic path to prepare for such a future. This ability to anticipate and explore novel possibilities sets experiential AI apart.
The Dilemma of Human Intervention
One critical question arises: should humans intervene in AI’s exploratory process, or simply define goals and let the system chart its course? This is more complex than it seems. If humans intervene when AI deviates from desired outcomes, the system may implicitly adopt human biases, even without explicit instructions. This mirrors a parent telling a child, “Do whatever you want,” while subtly signaling disappointment if the child avoids a preferred path, like studying law for example. The child, anticipating this, conforms. Similarly, AI might internalize human priors, stifling its ability to explore the unknown.
Sutton and Silver argue for minimal intervention, allowing AI to pursue its own trajectories. By defining broad goals and trusting the system’s experiential learning, we enable it to uncover solutions that human knowledge alone could never conceive. In my opinion there is not such thing as ‘minimal intervention’. Any ‘minimal intervention’ risks embedding human priors, akin to fiat money’s susceptibility to meddling versus bitcoin’s immutable design. True experiential AI demands freedom to explore unhindered.
Toward a Boundless Future
The era of experience, as envisioned by Sutton and Silver, heralds a transformative shift in AI research. By moving beyond human knowledge and embracing continuous learning through simulated exploration, AI can unlock the vast unknown. This paradigm not only redefines what machines can achieve but also challenges us to rethink our role in their development. As we stand on the cusp of this new era, the question is not whether AI can surpass human intelligence, but how far it can take us into the uncharted territories of knowledge.