Richard Sutton — 2024 Turing Award co-winner and the father of reinforcement learning — has formally announced his departure from Keen Technologies, founding Oak Lab with his student Khurram Javed.
The goal is explicit: build a trillion-parameter agent that learns and plans in real time on a machine drawing just 20 watts. Twenty watts — the human brain’s power budget. Today’s most advanced AI models consume megawatts for a single training run.
On its face the news is small — a 69-year-old professor starting a company sounds like academic idealism’s last act. But set Sutton’s four-decade intellectual trajectory against what he is challenging, and this is not a heartwarming “old scientist’s final venture.” It is a frontal challenge to the AI industry’s most core assumption.
What Sutton Is Challenging
He stated it bluntly in his announcement:
“We believe current deep learning methods are weak and inefficient — what’s needed is not patching but fundamentally new ideas and a complete redesign.”
Note who is speaking. Not a marginal scholar or an ousted dissenter. Richard Sutton founded modern reinforcement learning, invented temporal-difference learning, and co-wrote Reinforcement Learning: An Introduction with his advisor Andrew Barto — the field’s standard textbook for four decades. His students include David Silver, core designer of AlphaGo, and Doina Precup, head of DeepMind Montreal. From 2017 to 2023 he was a Distinguished Research Scientist at DeepMind.
In 2019 he wrote The Bitter Lesson, among the most-cited essays in the field: general methods that leverage computation ultimately beat methods built on hand-encoded human knowledge. The essay was widely read as an endorsement of the Scaling Law — Sutton himself said it: bigger models, more data, more compute are the path to general intelligence.
But Sutton has increasingly realized his essay was misread. The Bitter Lesson never said “pile on compute and you’re done” — it said “stop hard-coding human knowledge into systems; let them learn from experience.” The trouble is that mainstream deep learning still depends heavily on human-produced data: models learn from text humans wrote, photos humans took, labels humans applied. It is not learning from its own actions; it is imitating the traces of humanity’s past. In 2025, Sutton and David Silver jointly proposed the “Era of Experience,” stating plainly that AI is shifting from depending on human data to depending on experience generated by agents interacting with their environments. Oak Lab is that thesis incorporated.
Offline Versus Real-Time Learning: Two Philosophies of AI
To understand why Sutton left, understand the split between the two learning paradigms. The current mainstream — offline pretraining — spends months and hundreds of millions of dollars on one-shot training over vast text corpora. When training ends, parameters largely freeze; the model ships. Even conversing with billions of users daily, almost none of that interaction converts into new capability. The model reuses what it learned at graduation, or briefly holds information in its context window. It is like someone who finished college and never learned again — every answer built on graduation-day knowledge.
Sutton’s target paradigm — real-time online learning — is an agent perceiving the environment, acting, and adjusting its own behavior from results. Every new experience produces learning simultaneously; no long gap before concentrated retraining. Sutton’s own words: “Every moment an AI runs should be a moment of learning.”
Oak Lab is built around an architecture called OaK (Options and Knowledge). The core idea: let the agent discover temporally extended abstract structures from its own experience and convert them into verifiable, plannable, reusable “skills.” Example: a robot fetches water from the kitchen for the first time — recognizing the room, avoiding obstacles, picking up the cup, opening the tap. Traditional AI treats the whole sequence as one decision task. The OaK architecture lets the agent decompose its practice into high-level skills — “go to the kitchen,” “pick up the cup,” “fill with water” — and, next time a similar goal arises, directly recall the existing skills and adapt them to the environment. This “temporal abstraction” — condensing scattered actions into reusable skills — is a core capability of human intelligence, and today’s AI has almost none of it.
More radical still: Oak Lab’s learning phase neither stores historical data nor replays past experience. They envision batch-size-one real-time learning — each new experience immediately becoming one update. Combined with event-driven neural networks, the required computation and energy could drop by orders of magnitude. Hence the seemingly extreme long-term target: trillion parameters, real-time learning, 20 watts.
Three Paradigms: a Frame for AI’s Learning Eras
Placed in a larger coordinate system, AI’s learning paradigms fall into three eras. Paradigm one: the Data Era — dependent on human-labeled datasets; models learn humanity’s existing knowledge. Early machine learning and supervised learning’s golden age. Limit: the ceiling comes from data quality and scale; models cannot exceed the boundary of human knowledge. Paradigm two: the Compute Era — large models plus Scaling Law, extracting statistical regularities from oceans of data with extreme compute. GPT, Claude, Gemini. Limit: training is one-shot; the model cannot keep learning after deployment; training costs run to megawatts; data eventually runs out. Paradigm three: the Experience Era — agents generating new experience continuously at runtime and learning every moment. Representative: Oak Lab, if it succeeds. Potential: continuous adaptation to changing environments; energy consumption approaching biological intelligence; independence from human data, opening knowledge territory humans never reached.
Virtually all the industry’s resources and attention sit in paradigm two. Sutton is betting that paradigm two’s ceiling is near, and paradigm three is the true road to AGI.
What the Frame Explains
The three-paradigm frame clarifies much of the current industry. Why did RL suddenly become central? DeepSeek’s R1 surprised not because the model was bigger but because RL let the model learn to “think” during inference. OpenAI’s o-series and Google’s reasoning models are all probing toward learning from experience. Why are embodied AI and agents hot? If AI can only answer from offline data, it will never perform tasks in the physical world or in complex workflows. The agent boom is the industry collectively searching for the bridge from paradigm two to paradigm three. Why will energy become AI’s ultimate bottleneck? Paradigm two’s energy consumption grows exponentially — Sam Altman’s line that “the bottleneck isn’t model capability, it’s energy” is correct inside paradigm two. But if paradigm three holds — a trillion-parameter agent at 20 watts — the underlying logic of the compute economy gets rewritten. Why is “edge AI” more than a buzzword? If AI’s final form is a real-time learning agent, it must run close to where it acts — on robots, in cars, in phones. Today’s cloud-inference architecture is transitional.
An interesting footnote: from 2017 to 2023 Sutton led DeepMind’s Edmonton team. In 2023 DeepMind shut the lab, and Sutton joined Carmack’s Keen Technologies. Less than three years later, he left again. This is not an easy man to work with — not because of temperament, but because his thinking keeps walking outside the mainstream.
What You Can Do
If you are a technical decision-maker at a model company — do not file “real-time learning” under five-years-away. Start reserving continual-learning interfaces in your model architectures now. Every delivery model built on pretraining-plus-fine-tuning will eventually be displaced by continual learning. If you are an AI founder — watch Sutton’s direction seriously. Not because he will certainly succeed, but because the direction he names — real-time learning, low power, high autonomy — is the next growth pole of AI infrastructure. A trillion-parameter model at 20 watts, if achieved, redefines the entire edge-computing market. If you are an AI researcher — OaK’s “temporal abstraction” and “batch-size-one real-time learning” are two directions worth mining. Whether Oak Lab hits its targets or not, the exploration itself carries value. If you are simply an AI user — remembering one Sutton sentence is enough: “Real intelligence is not learned from data; it grows from action.” It is his startup creed, and also a simple yardstick for evaluating any AI product: can it learn from experience? Or does it only repeat the answers it memorized?
References: Richard Sutton on X · QbitAI reporting (via 36Kr)
