In a recent interview with Dwarkesh, Richard S Sutton laid out an argument against LLMs’ efficacy. Let’s first try to understand Richard S Sutton’s influential essay titled “Bitter Lesson.”
A key historical observation from the essay: general-purpose methods that leverage massive computation, like search and learning, ultimately outperform methods that rely on incorporating human knowledge.
The central idea is that AI research has seen a long-standing conflict between two types of approaches:
• “Weak Methods” (The Winners): These are simple, general-purpose methods based on principles like search and learning. They are called “weak” because they don’t start with specialized human knowledge. Their power comes from their ability to scale with computation. The truly scalable method, according to Sutton, is learning directly from experience, trying things and seeing what happens, rather than from data provided by people.
• “Strong Methods” (The Losers): These methods involve imbuing an AI system with human knowledge and expertise. In the past, this referred to symbolic methods. Still, Sutton argues that large language models (LLMs) fit this category because they are trained on a massive corpus of human-generated text.
The “bitter” part of the lesson is that while researchers often feel good about building their own knowledge into systems, this approach has consistently been “superseded” by the simpler, more scalable computational methods in the long run.
So, what is Sutton's solution?
Sutton argues that the very essence of intelligence is the ability to achieve goals. Reinforcement learning is built around this principle. The primary purpose of an RL agent is to understand its world to achieve goals. In this view, intelligence is the “computational part of the ability to achieve goals”. An action is considered “right” or “good” if it leads to maximizing a reward signal, which provides a clear “ground truth.”
In contrast, Sutton characterizes LLMs as fundamentally about mimicking what people say or do. Their training objective, next-token prediction, is not a substantive goal that changes the external world. Because they lack an external goal, there is no inherent definition of a “right thing to say,” and therefore no ground truth to learn from during their operational life.
Sutton sees LLMs as a mixed case. They scale with massive computation, which aligns with the lesson, but they are also fundamentally a way of ingesting vast amounts of human knowledge. He predicts that this reliance on human-generated data will be their ultimate limitation and will be superseded by systems that learn purely from experience, making it another instance of the bitter lesson. He is skeptical that LLMs are a good starting point, as historically, approaches heavily reliant on human knowledge have been “eaten” by more scalable, experience-based methods.
So, what does this mean for AI investments?
I believe that the next investment wave in AI will be in sensor-based learning of models by exposing them to external stimuli. For example, to make a human-like AI with near-environment awareness, we might want to embed temperature and light sensors (maybe just LiDAR) and make the model experience different situations alongside a human being, which becomes a training paradigm. This solves for making models and training more generalized, but according to Sutton, this will also not solve the problem of not having an inherent goal. The problem boils down to one of consciousness and will; how can we embed will into the models so that they go out and seek the ultimate truth?