Rich Sutton: AI's 'Weird Field' Needs to Relearn 'Learning'

In a recent wide-ranging discussion, AI pioneer Rich Sutton, widely recognized for his foundational work in reinforcement learning, shared his perspective on the current state of artificial intelligence, particularly the prevalent large language models (LLMs). Sutton, joined by his Oak Lab co-founder Kuram Javeed, argued that the field has developed a somewhat 'weird' focus on concepts like 'continual learning', asserting that all learning is inherently continual and that the field's current trajectory might be overlooking fundamental principles.

The 'Ordinary' Way of Thinking

Sutton pushed back against the notion that his views are radical, stating, "I see it as like I'm thinking the ordinary way. It's just everyone else that's thinking a bit weird." He elaborated that before the recent "AI craziness," the idea of learning that wasn't continual would have been nonsensical. He posits that AI systems, like humans, should naturally learn through action and perception, a process that is continuous rather than compartmentalized.

Recounting his early career dedication to reinforcement learning, even during AI winters, Sutton expressed his conviction stemmed from a fundamental desire to understand the mind. "Learning is a central part of the mind and having a goal is a central part of the mind, central part of intelligence," he explained.

The 'Bitter Lesson' and LLMs

The conversation delved into Sutton's influential essay, "The Bitter Lesson," which posits that methods that scale with computation, rather than human knowledge, ultimately prove more effective in the long run. While acknowledging the incredible progress of LLMs, Sutton sees them as both a positive and negative example of this lesson.

"First, large language models enabled enormous scaling with computation and you could just drink in the internet and scale so much," Sutton noted. However, he cautioned that the reliance on finite internet data could eventually become a bottleneck. He criticized the concept of synthetic data generation as a "big mistake," arguing that the "big world hypothesis" suggests the world is infinitely complex and that true learning comes from direct experience, not human-curated data. "Humans decide what's a good synthetic data and versus a bad synthetic data because I can write a program that can output a lot of synthetic data which would hurt programs right now," he stated, highlighting the human bottleneck in such approaches.

Continuous Learning and the Future of AI

Sutton and Javeed discussed their research agenda at Oak Lab, focusing on achieving true continual learning. Sutton stressed that current LLMs, despite their conversational abilities, do not truly learn when interacting with users as their weights remain static. He advocates for algorithms that "metalearn how to learn", enabling systems to continuously update their models of the world and adapt to new information without "catastrophic forgetting."

The duo also touched on the broader implications of AI development, with Sutton expressing a desire for AI to be self-consistent and capable of maintaining its own coherence. He envisions a future where AI, rather than making humans irrelevant, could make the world "even more exciting and interesting for humans."