Richard Sutton argues that LLMs are primarily imitation-based systems that mimic human text. He contends they lack genuine world models, goals, and the ability to learn from experience, which are essential components of true intelligence.
Sutton posits that reinforcement learning, with its focus on agents, goals (rewards), and learning from interaction with an environment, is the foundational paradigm for building intelligent systems. He contrasts this with the supervised, imitation-based learning of LLMs.
Sutton's famous essay is discussed in the context of LLMs. He argues that while LLMs leverage compute, their reliance on a finite corpus of human knowledge means they will likely be superseded by methods that learn purely from experience and computation, repeating a historical pattern in AI.
The conversation emphasizes the need for AI to learn continuously from its interaction with the world, a capability inherent in animals but largely absent in current AI systems which have distinct training and deployment phases. This is identified as a key missing piece for AGI.
Sutton presents his view that digital intelligences will inevitably succeed humans. He frames this not as a catastrophe but as a grand, natural transition in the universe's evolution from an era of biological replication to an era of intelligent design.
Keep pulling the thread on Richard Sutton.