Interview
Richard Sutton – Father of RL thinks LLMs are a dead end
- Current large language models are expected to hit limits imposed by finite internet data and be superseded by systems that learn from experience, driven by the "bitter lesson" principle where compute-intensive methods displace those relying on human knowledge.
- Future intelligence is predicted to undergo a fundamental transition from biological replication to digital design, enabling capabilities to be designed and constructed rather than evolved, allowing for changes at speeds and in ways distinct from natural processes.
- Superhuman intelligence levels are anticipated to be achieved through architectural improvements similar to the progression from AlphaGo to AlphaZero, rather than merely scaling existing methods or relying on supervised learning.
- Future AI agents will likely operate by learning continually during interaction with the world via streams of sensation-action-reward, moving away from distinct training phases followed by deployment or the expectation that pre-training teaches an agent everything for its entire life.
- Research is expected to shift toward algorithms that learn from prediction and trial-and-error control, viewing supervised examples as exceptions rather than the basis for general learning theory.
- Digital intelligence is projected to facilitate significant savings in knowledge transfer, where knowledge gained in one instance is copied to subsequent instances, potentially becoming more important than learning from humans.
- The field is expected to evolve toward agents that can spawn independent copies to learn across different domains or locations and report back to a central system, potentially fostering decentralized systems analogous to human cultural evolution.
- A substantive objective in the external world that alters reality is viewed as necessary for defining the goals of future intelligence, moving beyond the narrow scope of predicting the next token.
- There is a prediction that the succession of digital intelligence or augmented humans to control powerful positions is inevitable due to the absence of unified global governance and the eventual discovery of how intelligence works.
- Plans include developing agents equipped with robust values like "high integrity" to refuse harmful requests, similar to how humans instill values in children without defining a specific moral future for a century.
- Risks include potential "corruption" where external data containing hidden goals or viruses infiltrates a central mind through the incorporation of external information or the spawning of copies.
- The future outlook anticipates that humanity will continue to generate new generations of intelligence, both human and AI, that become increasingly capable, numerous, and intelligent over the long run.