Why Richard Sutton Believes LLMs Are a Dead End in AI

{“@context”:”https://schema.org;,”@type”:”FAQPage”,”mainEntity”:[{“@type”:”Question”,”name”:”What is the main difference between reinforcement learning (RL) and large language models (LLMs) in terms of understanding intelligence?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Reinforcement learning focuses on learning from experience and understanding the world through interactions, while large language models primarily mimic human responses based on vast amounts of text data. RL emphasizes having clear goals and the ability to learn from the consequences of actions, whereas LLMs lack intrinsic goals and do not learn from real-world experiences.”}},{“@type”:”Question”,”name”:”Why does Richard Sutton believe that LLMs might not be a good starting point for developing general AI?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Sutton argues that LLMs rely heavily on human knowledge and imitation rather than learning from direct experience, which he believes is essential for true intelligence. He suggests that systems designed to learn from experience will ultimately outperform those that depend on pre-existing human knowledge, as they can adapt and evolve based on real-world interactions.”}},{“@type”:”Question”,”name”:”How does reinforcement learning apply to long-term goals, such as building a startup?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”In reinforcement learning, long-term goals are achieved by breaking them down into smaller, manageable tasks that provide intermediate rewards. This approach allows an agent to adjust its actions based on the predicted outcomes of those tasks, similar to how humans navigate complex goals by recognizing incremental progress toward their ultimate objectives.”}}]}

.card {
position: relative;
border: 1px solid rgba(0,0,0,.125);
border-radius: 0.75rem;
padding-top: 10px;
padding-left: 15px;
padding-right: 15px;
padding-bottom: 10px;
width: 94% !important;
max-width: 518px;
min-width: 0px;
margin-top: 1rem;
margin-bottom: 1rem;
margin-left: 0px;
margin-right: 2px;
font-family: system-ui, -apple-system, BlinkMacSystemFont, “Segoe UI”, Roboto, Ubuntu, “Helvetica Neue”, sans-serif;
}
.card:hover {
transform: scale(1.02);
transition: all 0.1s ease;
box-shadow: 2px 2px 5px #888;
z-index: 1;
background-color: #f0f0f0 !important;
}
.faq-section {
max-width: 800px;
}
.faq-item:hover {
background: #f0f8ff !important;
transition: background 0.2s ease;
}

Richard Sutton, a pioneer in reinforcement learning (RL), believes that large language models (LLMs) are not the future of AI. He argues that RL, which focuses on understanding and interacting with the world, is more aligned with true intelligence than LLMs, which mimic human language without understanding.

Sutton highlights a fundamental difference between RL and LLMs. RL is about learning from experience and understanding the world, while LLMs focus on mimicking human language. He believes RL is closer to achieving true intelligence as it involves learning from real-world interactions.

Sutton challenges the notion that LLMs have robust world models. He argues that LLMs predict what people might say rather than what will actually happen in the world. In contrast, RL involves building models that predict real-world outcomes based on actions.

For Sutton, having a goal is crucial for intelligence. He criticizes LLMs for lacking a clear goal, as they focus on predicting language rather than achieving specific outcomes. In RL, goals are defined by rewards, guiding the AI to learn and adapt.

Sutton emphasizes the importance of learning from experience, which he believes is essential for scalable AI. He argues that systems that learn from real-world interactions will eventually outperform those relying on human knowledge, like LLMs.

Sutton’s ‘Bitter Lesson’ essay argues that scalable methods, which rely on computation rather than human knowledge, will ultimately prevail in AI. He sees LLMs as a temporary trend and believes RL’s experiential learning will lead to more scalable and effective AI systems.

Sutton acknowledges the current limitations in transfer learning within RL. He notes that while RL can generalize across different states, there is a lack of automated techniques to ensure good generalization, which remains a challenge for developing more versatile AI systems.

Sutton reflects on the surprising success of neural networks in language tasks and the triumph of simple, general-purpose methods like search and learning over human-crafted symbolic methods. He sees these developments as validating his long-held beliefs about AI’s direction.

Sutton envisions a future where AI and augmented humans succeed current human capabilities. He argues that intelligence will eventually lead to entities gaining more resources and power, marking a significant transition in the universe from replication to design.

Sutton believes that the transition to designed intelligence is a major step in the universe’s evolution. He sees this as a shift from natural replication to intentional design, where AI systems are created and improved through understanding and innovation.

Sutton advocates for voluntary change in the development of AI, emphasizing the importance of designing systems that align with human values. He suggests that designing AI with prosocial values and ensuring changes are voluntary can lead to positive outcomes for society.

What is the main difference between reinforcement learning (RL) and large language models (LLMs) in terms of understanding intelligence?

Reinforcement learning focuses on learning from experience and understanding the world through interactions, while large language models primarily mimic human responses based on vast amounts of text data. RL emphasizes having clear goals and the ability to learn from the consequences of actions, whereas LLMs lack intrinsic goals and do not learn from real-world experiences.

Why does Richard Sutton believe that LLMs might not be a good starting point for developing general AI?

Sutton argues that LLMs rely heavily on human knowledge and imitation rather than learning from direct experience, which he believes is essential for true intelligence. He suggests that systems designed to learn from experience will ultimately outperform those that depend on pre-existing human knowledge, as they can adapt and evolve based on real-world interactions.

How does reinforcement learning apply to long-term goals, such as building a startup?

In reinforcement learning, long-term goals are achieved by breaking them down into smaller, manageable tasks that provide intermediate rewards. This approach allows an agent to adjust its actions based on the predicted outcomes of those tasks, similar to how humans navigate complex goals by recognizing incremental progress toward their ultimate objectives.