The 20-Watt Intelligence: Richard Sutton’s Oak Lab and the Quest to Rebuild AI

In the high-stakes theater of artificial intelligence, where trillion-dollar valuations are built on the back of massive data centers and "brute force" scaling, a fundamental schism is forming. At the center of this intellectual rebellion is Richard Sutton, a man whose contributions to the field are so foundational that he was recently awarded the 2024 Turing Award—the "Nobel Prize of Computing"—for his role in founding modern reinforcement learning (RL).

On Monday, Sutton sent shockwaves through the industry by announcing his departure from John Carmack’s startup, Keen Technologies, to launch a new venture: Oak Lab. Alongside his long-time collaborator Khurram Javed, Sutton is embarking on a mission that is as much a critique of the current AI status quo as it is a roadmap for the future. His goal? To move away from the "weak and inefficient" deep-learning methods that dominate the industry and build a trillion-parameter agent capable of learning in real-time on just 20 watts of power—the approximate energy consumption of a human brain.

Main Facts: A Departure from the "Frozen" Model Paradigm

The announcement, delivered via X (formerly Twitter), was characteristically blunt. Sutton praised his time with legendary programmer John Carmack but noted that he and Javed had "broken away to start our own startup" to pursue "a slightly different path toward understanding intelligence."

The core of Sutton’s argument rests on a distinction that the current AI boom has largely ignored: the difference between learning from curated datasets and learning from experience. Today’s Large Language Models (LLMs), such as GPT-4 or Claude, are trained in a two-step process. First, they are fed massive, "frozen" snapshots of the internet—cleaned, filtered, and curated by humans. Once training is complete, the model is essentially static; it does not "learn" from its interactions with users in a meaningful, structural way. It is a library, not a mind.

Oak Lab proposes a radical alternative. Sutton argues that true intelligence must be created and maintained through run-time experience. This means an AI should learn one step at a time, continuously, from a stream of raw sensory data, without the need for human-prepped datasets or the massive "re-training" cycles that currently cost tech giants billions of dollars.

Chronology: From the Foundations of RL to the "Bitter Lesson"

To understand why Sutton’s move matters, one must look at the trajectory of his career. For decades, Sutton has been the primary advocate for the idea that AI should be built on the principles of trial and error—the essence of reinforcement learning.

• The Early Years and the Textbook: In the 1980s and 90s, Sutton, along with Andrew Barto, authored Reinforcement Learning: An Introduction. This text became the "bible" for researchers, laying the groundwork for how machines could learn to achieve goals by interacting with an environment.
• The DeepMind Influence: Sutton’s work was the direct precursor to the breakthroughs at Google DeepMind. The algorithms that powered AlphaGo—the first AI to defeat a world champion at the game of Go—were built on the RL principles Sutton pioneered.
• The Bitter Lesson (2019): Perhaps Sutton’s most famous contribution to modern discourse is his 2019 essay, The Bitter Lesson. In it, he argued that the history of AI research shows that "general methods that leverage computation are the most effective, and by a large margin." He warned researchers against trying to "build in" human knowledge, arguing that more compute and simpler learning algorithms would always win in the end.
• The Keen Technologies Era (2022-2024): In 2022, Sutton joined forces with John Carmack, the visionary behind Doom and Quake, who had pivoted to "Artificial General Intelligence" (AGI). Their partnership was seen as a "dream team" of engineering and theory. However, Sutton’s departure for Oak Lab suggests that even within the world of AGI research, there are fundamental disagreements on the "how."

Supporting Data: Why Current Models are "Weak and Inefficient"

The technical justification for Oak Lab is laid out in the lab’s inaugural research post, authored by Sutton and Javed. They identify a systemic flaw in the way modern AI is optimized: the reliance on Stochastic Gradient Descent (SGD) and its variants like Adam.

The Noise vs. Signal Problem

In a standard training environment, an optimizer like SGD processes data that has been carefully cleaned. However, in the real world—the "run-time experience" Sutton favors—data is messy. It is a mix of predictable signals and pure noise.

According to Oak Lab’s research, standard optimizers cannot distinguish between the two. When an error occurs, SGD "spreads the blame" across all its parameters. Over time, this causes the model to "quietly absorb the noise," leading to inefficiencies and a loss of plasticity—the ability to learn new things without forgetting old ones.

The NetworkIDBD Solution

To fix this, Sutton is reviving and updating an idea he first proposed decades ago: Incremental Delta-Bar-Delta (IDBD). Oak Lab has introduced a neural version called NetworkIDBD.

Unlike SGD, which updates all weights equally based on error, NetworkIDBD learns to assign credit selectively. It evaluates which signals have been historically predictive and rewards only those. In Oak Lab’s initial tests, this algorithm successfully identified real patterns in environments where SGD became "drowned in junk." This selective credit assignment is the key to "continual learning"—allowing a model to evolve indefinitely without its performance degrading into noise.

The 20-Watt Goal

The most striking figure provided by Oak Lab is the energy target. Current frontier models are trained in data centers that draw megawatts of power—enough to fuel small cities. These models require massive Nvidia H100 clusters just to function.

Oak Lab is aiming for an agent that learns and plans in real-time on 20 watts.

• Human Brain: ~20 Watts.
• Modern AI GPU: ~300-700 Watts (per chip).
• Oak Lab Goal: A 1-trillion parameter agent (comparable in size to GPT-4) running on the power of a lightbulb.

Official Responses: The Contrarian Stance

While the broader AI industry is currently obsessed with "Scaling Laws"—the idea that simply adding more GPUs and more data will lead to AGI—Sutton’s official stance via Oak Lab is one of calculated dissent.

In his announcement, Sutton did not mince words about the current state of the art: "Current deep-learning methods are weak and inefficient, and need not more tweaks, but fundamentally new ideas and a thorough reworking."

This puts Sutton in a small but elite group of "AI skeptics" who believe the LLM path is a dead end for true intelligence.

• Yann LeCun (Chief AI Scientist at Meta): LeCun has famously argued that LLMs do not understand the physical world and has pivoted toward "World Models" (V-JEPA), arguing that machines must learn like infants—by observing and interacting, not by reading text.
• David Silver (Principal Scientist at DeepMind): Silver, a former student of Sutton’s, continues to push the boundaries of RL, betting that "search and learning" are the only two things that matter for superintelligence.

Oak Lab’s position is that the industry is currently "distilling" intelligence from human data rather than "generating" it from experience. Sutton’s departure from Carmack’s Keen Technologies suggests a desire for a "purer" research environment, free from the immediate pressures of building a commercial product, focused instead on the fundamental mechanics of how a mind interacts with time.

Implications: A Shift from "Generative" to "Agentic" AI

The launch of Oak Lab comes at a pivotal moment for the AI industry. The initial "wow factor" of generative AI is beginning to give way to concerns about energy sustainability, data exhaustion, and the limitations of reasoning.

1. The Energy Wall

If the industry continues on its current path, AI energy demand is projected to skyrocket, potentially straining global power grids. Sutton’s focus on 20-watt intelligence offers a potential exit ramp. If Oak Lab can prove that efficiency comes from the algorithm rather than the hardware, the economic landscape of AI shifts from the hands of those who own the most chips to those who own the best ideas.

2. The End of Pre-training?

If Sutton is correct that intelligence is "maintained from run-time experience," the multi-billion dollar "pre-training" phase used by OpenAI and Google might eventually become obsolete. Instead of a "frozen" model that is updated every six months, we would have "agents" that grow smarter every hour they are "alive" in a digital or physical environment.

3. On-Device Intelligence

A 20-watt, trillion-parameter agent would be the holy grail for edge computing. It would allow for true, human-level intelligence to reside on a smartphone or a small drone without needing a connection to a massive server farm. This has profound implications for privacy, robotics, and space exploration.

4. The Rebirth of Reinforcement Learning

For the past three years, RL has taken a backseat to Supervised Learning (LLMs). Sutton’s Oak Lab signals a "return to roots" for the field. By focusing on the "Credit Assignment Problem" and "Continual Learning," Sutton is betting that the next leap in AI won’t come from a bigger building full of GPUs, but from a more elegant understanding of how a single entity learns to navigate the world.

Conclusion

Richard Sutton is betting his storied career and his Turing Award legacy on a contrarian gamble. While the rest of the world is building "bigger brains," he is trying to build a "better way to learn." Whether Oak Lab can deliver on the promise of a 20-watt trillion-parameter agent remains to be seen, but Sutton’s track record suggests that his "Bitter Lessons" are rarely wrong. If he succeeds, the future of AI will look less like a static encyclopedia and more like a living, breathing mind—one that never stops learning from the world around it.