News · 2026-07-13

Richard Sutton's Oak Lab bets against frozen models: a trillion-parameter agent on 20 watts

Richard Sutton - the Turing Award winner who co-wrote the field's foundational reinforcement-learning textbook and authored the influential 'bitter lesson' essay - has launched Oak Lab with an explicit and deliberately audacious north star: a trillion-parameter agent that learns and plans in real time on about 20 watts of energy, the rough power budget of a human brain. The bet underneath the goal is a direct challenge to the dominant paradigm: Sutton is wagering on continual, experiential learning over the static pre-train-then-freeze approach that produced today's large language models.

Key facts

• Richard Sutton announced Oak Lab, reported by The Decoder.
• The stated goal: a trillion-parameter agent that learns and plans in real time on roughly 20 watts.
• The approach centers on experiential/continual learning rather than static pre-training.
• It is a mission statement and research direction, not a current result.

To understand why this is a philosophical shot across the bow, you need the contrast. Today's LLMs are trained in one enormous, expensive pass over a fixed dataset, and then the weights are frozen. When you chat with one, it is not learning from you; it is running a fixed function. Everything it 'knows' was baked in at training time, and updating that knowledge means another training run. This works astonishingly well, but it is nothing like how a human or animal learns. We learn continuously, from our own experience, adjusting as we go, without stopping to retrain from scratch. For more on this distinction, see our lessons on training vs inference and reinforcement learning post-training.

Sutton's whole career argues that the second way - learning from experience, in the loop, continually - is the more powerful path in the long run, even if the frozen-model approach is winning right now. The 20-watt framing is the provocation that makes the point vivid. Human brains do open-ended, continual learning and sophisticated planning on the energy of a dim light bulb. Today's frontier models, by contrast, consume staggering amounts of power both to train and to run, and still cannot learn from their own experience once deployed. Setting 20 watts as the target is Sutton's way of saying the current paradigm is not just philosophically incomplete but wildly energy-inefficient compared to the one existence proof we have of general intelligence.

The analogy: a frozen LLM is like a brilliant scholar who read every book ever written up to a cutoff date, then suffered total amnesia for anything new - endlessly knowledgeable, incapable of learning from today. Sutton wants to build the opposite: an agent that might know less at any given moment but never stops learning, accumulating competence from its own ongoing experience the way a person accumulates skill.

The essential caveat, which Sutton would be the first to insist on, is that this is a goal, not a result. There is no 20-watt trillion-parameter continual learner today, and building one runs headlong into the hardest open problems in the field: catastrophic forgetting (learning new things tends to erase old ones), stability, and the sheer difficulty of making continual learning work at scale without the system degrading. The 20-watt number is an aspiration, a north star to orient research, not a spec sheet.

Why it matters: the weight of Sutton's credentials makes this more than one more lab launch. The 'bitter lesson' - his argument that general methods leveraging computation beat clever hand-engineered ones - has shaped how the field thinks about scaling. When the person who wrote it stakes a new lab on the claim that the scaling-the-frozen-model paradigm is a detour from real intelligence, the field listens, even from those who think he is wrong. It sharpens the central open question of the moment: is the path to more capable AI more of the same (bigger frozen models), or something architecturally different (agents that learn as they live)? Oak Lab is a bet on the second answer.

Primary source, verified: read the paper →

Ground Truth. (2026, July 13). Richard Sutton's Oak Lab bets against frozen models: a trillion-parameter agent on 20 watts. Ground Truth. https://groundtruth.day/news/richard-sutton-oak-lab-20-watt-continual-learning.html

@misc{groundtruth:richard-sutton-oak-lab-20-watt-continual-learning,
  title  = {Richard Sutton's Oak Lab bets against frozen models: a trillion-parameter agent on 20 watts},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/richard-sutton-oak-lab-20-watt-continual-learning.html}
}

Topics: reinforcement-learning · continual-learning · richard-sutton · agents · research-labs