Rich Sutton posits that the need to specify "continual learning" is a recent, strange development. Historically, all learning was assumed to be ongoing. The current AI paradigm of training static models is an aberration from this natural, common-sense view of intelligence.
The essence of Sutton's "Bitter Lesson" is a warning against the temptation to embed human knowledge into AI systems. Instead, AI progress is driven by general methods like search and learning that scale with computation, a lesson learned over decades of research.
LLMs initially validate the Bitter Lesson by scaling immensely with computation on internet data. However, they also illustrate its warning: once they exhaust this finite human-generated dataset, their reliance on prior knowledge becomes a bottleneck, limiting further learning from direct experience.
According to the "Big World Hypothesis," the real world is infinitely more complex than any simulation. Therefore, synthetic data is a flawed approach because it's a small, incorrect approximation that is ultimately bottlenecked by the human expertise required to create and validate it.
Rich Sutton argues that LLM-based assistants are not "experiential learners." Despite in-context learning, their fundamental weights never change after deployment. This prevents them from truly adapting, forming new concepts, or updating their core understanding based on user interactions.
The "Big World Hypothesis" posits the world's complexity vastly exceeds any AI's capacity. Consequently, AI must rely on approximations. Continual learning is essential because these approximations must be constantly tuned to the specific, immediate context the AI finds itself in.
To enable continual learning without destroying past knowledge (catastrophic forgetting), Rich Sutton suggests a specific algorithmic fix. The solution involves meta-learning a unique learning rate (step size) for every individual weight in the network, ensuring updates are precise and non-destructive.
Sutton contends that focusing on supervised learning is a mistake when trying to build AGI. Animals, including humans, do not learn core skills like walking or seeing via labeled examples. True intelligence is rooted in experiential, goal-driven learning, making school-style supervision an irrelevant model for core AI.
A critical gap in AI is that systems cannot yet learn a model of their world from raw experience and then use that model for high-level planning. Successes like AlphaGo relied on a pre-defined model (the game rules), but true intelligence requires discovering these models independently.