Richard Sutton – Father of RL Thinks LLMs Are a Dead End

What if one of the founding fathers of modern AI told you that large language models—the backbone of today’s most advanced chatbots—might actually be leading us astray? That’s exactly what Richard Sutton, often called the “father of reinforcement learning,” is saying about the current direction in artificial intelligence.

Who Is Richard Sutton—and Why Does His Opinion on LLMs Matter?

Richard Sutton isn’t just another voice in the crowded field of AI. He helped invent reinforcement learning (RL), a branch of artificial intelligence that teaches machines to learn from experience—kind of like how humans and animals do. If you’ve heard about self-learning robots or game-playing AIs mastering complex tasks over time, you’re seeing his ideas in action.

So when someone with his track record says large language models (LLMs) like GPT-4 may be a technological “dead end,” people pay attention. In an online discussion highlighted on Reddit (see [this post](https://www.reddit.com/r/artificial/comments/1nrfic3/richard_sutton_father_of_rl_thinks_llms_are_a/)), Sutton shares his perspective on why simply making bigger and bigger text generators won’t get us closer to true machine intelligence.

Why Does Sutton Think Large Language Models Fall Short?

Sutton’s main critique is simple but powerful: LLMs don’t really understand the world—they just predict which word comes next based on mountains of text they’ve seen before. Sure, they can generate essays or answer questions with impressive fluency. But are they truly “thinking,” or are they just mimicking patterns?

Here’s where it gets interesting:

• No real-world grounding: LLMs don’t interact with their environment; they never see or touch anything outside of text.
• No ability to learn from experience: Unlike RL agents that improve by trial and error, LLMs can’t adapt beyond their training data.
• Scaling has limits: Making models bigger doesn’t guarantee smarter behavior—just more convincing mimicry.
• Lack of agency: They react; they don’t set goals or plan ahead.
• Narrow focus: They excel at language but stumble outside predefined tasks.

These concerns have been echoed by other experts as well—Gary Marcus wrote an entire commentary on why relying only on LLMs might be shortsighted (read it [here](https://garymarcus.substack.com/p/game-over-for-pure-llms-even-turing) if you prefer words over videos).

The Debate: Is There Really a Dead End Ahead?

Of course, not everyone agrees with Richard Sutton’s take on large language models. Some believe these tools are just early steps toward more advanced systems and that combining them with other methods might lead to breakthroughs.

But here’s where the debate heats up:

• Supporters of LLMs argue that as we add more data and tweak architectures, these systems will eventually develop richer reasoning skills.
• Skeptics (like Sutton) think piling on more data won’t solve foundational problems—especially if models never get real-world feedback.
• Hybrid approaches are gaining traction—mixing reinforcement learning with LLM-style prediction to create agents that both understand and act in the world.

One memorable anecdote from recent years involves AlphaGo—the Go-playing AI that shocked the world by defeating champions using techniques pioneered by Sutton’s field (reinforcement learning). Unlike today’s chatbots, AlphaGo didn’t just read about Go; it played millions of games against itself and learned strategies no human ever discovered before.

That kind of “learning by doing” is exactly what Sutton wants to see more of in AI research—not just bigger text prediction engines.

Where Does This Leave Artificial Intelligence?

If you’re wondering whether this means your favorite chatbot is doomed—don’t worry just yet! Large language models are still incredibly useful for many tasks. But if we want machines that can truly understand and interact with our world—maybe even rival human thinking someday—Sutton argues it’ll take much more than clever text prediction.

Here are some possible paths forward:

• Combine LLMs with reinforcement learning so AIs can learn from experience
• Add real-world sensors so they can see and hear like humans
• Create environments where AIs can experiment safely and gain new skills over time

The big question now is whether tech companies will pivot toward these ideas—or keep chasing bigger language models until they hit a wall.

So what do you think? Is Richard Sutton right about large language models being a dead end for AI—or is there still untapped potential waiting to be discovered? Let us know your thoughts below!