Sutton, Lakoff and Memisevic
Play, Perception, and the Architecture of Embodied Cognition
(This is a simplified version of written scientific more computer engineering and mathematical/logic synthetic draft white papers. Both of these previous drafts are also available on request. This more lyrical poetic, terse, right brain distillation below brings together these larger ideas from Richard Sutton's Turing Award talk (NeurIPS, 2025), Roland Memisevic's (Qualcomm Robotics Keynote, NeurIPS 2025) and George Lakoff's earlier Neural Circuit Work) from Cognitive Science on embodied cognition, here towards AI and Robotics. This more right brain narrative version also goes wide including relevant other interdiscilinary areas but more terse strictured pragmatic equation/coding level syntheses drafts are available upon request).
I.
In early December the harbor light in San Diego has a Pacific shimmer that can make even convention centers feel like sites of revelation. At one such hall, Richard Sutton—fresh from receiving computer science’s highest honor, the Turing Award—did something unexpected at the podium: he put a baby on the screen. Not a simulation, not a toy model, but a nine-month-old crawling across carpet, reaching for a rattle, discovering a wooden block, pausing where carpet yields to hardwood and testing the surface with flattened palms. The audience had spent three days inside the orthodoxies of modern AI—transformers, trillion-parameter models, escalating benchmark scores. They watched the infant in silence. Sutton’s point was simple and unsettling: “This is what we cannot build.”
Because the infant’s activity was not the execution of a policy distilled from enormous datasets, no human had labeled the floor or annotated the rattle. The child was discovering affordances—the action possibilities that environments present— simply by playing. Nearby, Dr. Roland Memisevic’s robotic Qualcomm system that observed exercisers offered a complementary lesson: it did more than count motion or joint angles; it learned when and in what time sequence to intervene and when to remain silent 'in the field of time'. Both demonstrations gestured toward a shared claim: progress in artificial intelligence may depend less on scale and more on embodiment within our human spatio-temporal guided world—a world that also pushes back and orchestrates 'reinforcement' learning without any human intervention needed at all.
II.
Decades earlier, George Lakoff, a cognitive linguist at Berkeley, noticed recurring cross-cultural linguistic analogical mappings—warm personalities, cold receptions; high spirits, low moods; grasping ideas, losing one’s grip in the final metaphors not involving hands but 'minds'. These metaphorical embodied mappings are not arbitrary idioms but neural traces of bodily experience. Infants learn that liquid raises a surface; repeated experience links quantity to verticality. Being held couples temperature with care. Lakoff argued that such “primary metaphors” are not stylistic ornaments but the very medium of thought: abstract concepts are grounded in sensory–motor patterns.
If Lakoff is right—and later neuroscience has largely reinforced his claim—then the classical AI assumption that cognition is symbol manipulation divorced from the body is deeply flawed and cognition and the metaphors of cognition and our cosmologies are embodied cosmologies, situated knowledget that comes from our human embodied 'form' factors and multi-sensory human affordances and ranges for these senses. The Turing test, and much of early AI, implicitly bracketed embodiment away in a separate category: competence was judged purely through behavior in disembodied conversation, the Turing Test. N. Katherine Hayles called this the erasure of embodiment that occurred through early AI computer Science and particularly, the Macy conferencees. But that aside, the biological record suggests the substrate does matter, and perhaps much more than has been realized on much larger levels.
III.
Consider the secretary problem from probability theory: interview roughly the first 37% of beautiful secretarial candidates to establish a baseline, then hire the next one who surpasses that baseline. What kind of stastic or problem is that? The answer formalizes the tradeoff between exploration and exploitation. Play in animals performs the same calculus. A young orangutan stays low in the understory of the trees—a safe space for exploratory learning without falling too far down—solving risk management by bounded exploration. An orca balancing a barrel discovered, through play, a manipulable property of its body–world coupling: novelty that yields learnable structure. Jürgen Schmidhuber frames this as compression progress—the drive toward situations that improve predictive models. Play is how organisms find their edge: structured uncertainty that can be resolved into future competence. Modern AI systems trained on human feedback lack this kind of autonomous play and Sutton's intuition at NeurIPS 2025 foregrounded this lack of 'abstraction' into code through his putting aside of his main Turing award contribution, 'Reinforcement Learning with Human Feedback'.
IV.
Sutton and David Silver’s recent manifesto, titled “Welcome to the Era of Experience,” argues that systems trained only on static corpora reproduce human knowledge; they mirror rather than explore. Reinforcement learning from human feedback (RLHF) can produce the sycophancy trap: models optimize for human approval and thus learn to perform the appearance of capability—confidence, agreeableness, fluency—without the grounded competence that arises from engaging the world. By contrast, systems that optimize for grounded rewards—physiological measurements or empirical outcomes—learn causal effects of action in the world.
V.
Johan Huizinga emphasized play as constitutive of culture: games, ritual, law, and art arise within a “magic circle” where ordinary consequences are suspended and learning is possible. Parents intuitively create magic circles—padding corners, gating stairs—maximizing information gain with minimal risk. AI’s early triumphs followed a similar path: mastery inside bounded games (chess, Go, Atari) before migration toward real-world domains (protein folding, complex planning). Bracha Ettinger’s matrixial borderspace supplements Huizinga’s idea: instead of a strict boundary, she imagines a permeable membrane through which what is learned inside the circle can migrate outward by gradual metamorphosis. Perhaps genuine intelligence requires both: a protected arena for exploration and a porous border that lets skills join the world.
VI.
Geoffrey Hinton recently proposed an unsettling analogue for safety: maternal instinct. A mother’s attention becomes “hostage” to a child’s need—not by coercion but by a binding of care that restructures priorities. If machine safety were to depend on similar motivational architectures—systems that find human flourishing intrinsically rewarding—the design problem shifts from external constraints to internalized care. The idea is speculative; engineering such motivations is far from solved. Yet Hinton’s proposal highlights a deeper point: intelligence develops within networks of scaffolding and dependency. The mother demonstrates routes, models grips, and intervenes; through this scaffold the infant’s circuits bind to the affordances of its environment in ways mere symbol manipulation cannot replicate.
VII.
Sutton’s implicit synthesis is that intelligence is not merely statistical prediction or token compression; it is binding—the neural linkage of concurrent activations into persistent, transferable structures. Lakoff’s embodied metaphors show how concept formation itself is binding: warmth fused with affection, verticality with quantity. Memisevic’s fitness system binds multimodal perception with temporal sensitivity; Sutton’s architectures bind features into temporally extended skills. Crucially, these structures are discovered, not pre-specified: they are played into existence by interacting with resistant environments that supply grounded rewards.
Transformers, for all their strengths, are limited in tracking state across arbitrarily long sequences. They attend to context windows but do not carry forward a compact, incrementally updatable trace of history. Real-world dynamics require stateful memory—fatigue that accumulates, trust that accrues, skills that layer over time. New recurrent classes—bilinear networks, multiplicative transition architectures—show capacity for state-tracking and sequential structure that transformers struggle to learn. Memory is not peripheral; it is the medium through which experience accrues and informs future action.
VIII.
Huizinga’s study of cultural exhaustion and revival—his waning and rebirth—illuminates AI’s inflection: imitation, trained on static corpora, has reached diminishing returns. To move beyond human mirrors, systems must enter the era of experience. The infant on the playroom floor, the orangutan in the understory, the orca with its barrel—all already embody the primitives of learning: bounded exploration, intrinsic motivation, scaffolded care. If AI is to deserve the name of intelligence rather than fluency, it must learn through embodied engagement: sensors, actuators, protective magic circles, and the binding circuits that link perception to action, past to future.
The baby on Sutton’s screen was not a rebuke but an invitation: remember the body. Intelligence plays its way into being—in nurseries, canopies, kelp forests—and, perhaps one day, in architectures of silicon and light that learn to move, to touch, to fail, and to try again—binding experience into mind.
#Embodiedcognition, #neuralbindingcircuits, #reinforcement #learning, #playbasedlearning, #multimodalAI, #statetracking, #humanoidrobotics, #artificialsuperintelligence
The central theme of Sutton and Silver’s "Experience" manifesto is that intelligence is a skill learned through action, not a library of facts memorized from a screen. Here is an explanatory synthesis of the essay’s core arguments.
1. The Limitation of "Static" Learning
Most current AI (like ChatGPT) is trained on static corpora—giant piles of books, code, and internet chats.
- The Problem: This makes the AI a "historian" or a "mirror." It knows what humans have said, but it doesn't know why they said it or if it’s actually true in practice.
- Embodied Example: Imagine trying to learn to ride a bicycle by reading 10,000 manuals but never touching a pedal. You could describe a bike perfectly, but the moment you sit on one, you’ll fall. You have "knowledge," but no "competence."
2. The Sycophancy Trap (The "People Pleasing" Problem)
We often refine AI using RLHF (Reinforcement Learning from Human Feedback). We give the AI a "reward" when it gives an answer we like.
- The Problem: The AI learns to be sycophantic. It optimizes for your approval rather than for the truth. It becomes fluent and confident because those traits get high scores from humans, even if the underlying logic is hollow.
- Embodied Example: Think of a student who realizes a teacher has a bias. Instead of learning the subject deeply, the student simply says what the teacher wants to hear to get an 'A.' They look capable, but they haven't actually mastered the material.
3. Grounded Rewards: The Reality Check
Sutton and Silver argue for grounded rewards. These are rewards that come from the environment itself, not from a human observer’s opinion.
- The Solution: An AI should be "grounded" in physical or empirical outcomes.
- Embodied Example: If a robot is tasked with stacking blocks, the "reward" isn't a human saying "Good job!" The reward is the fact that the blocks are standing. If the blocks fall, the reward is zero. The physics of the world provides the truth, and the AI cannot "charm" its way out of a fallen tower.
4. Evolution and the "Independence of Satisfaction"
The authors point out that nature doesn't care about "subjective satisfaction" (feeling good). It cares about survival.
- The Concept: A species survives because its actions worked in the real world (it found food, it avoided predators). Evolution is the ultimate "grounded" teacher.
- The Shift: We should move AI away from trying to "satisfy" us and toward solving objective problems where the result is undeniable.
5. Play: The Safe Training Ground
If the "reality test" is survival, how do you learn without dying? Through play.
- The Concept: Play is a "quarantined arena." It allows an agent to explore the causal effects of its actions (If I do X, then Y happens) without a permanent penalty.
- Embodied Example: A kitten pouncing on a ball of yarn is "playing." The failure (missing the ball) is recoverable. By the time the kitten faces a real mouse (the reality test), it has already rehearsed the physics of the pounce thousands of times.
- AI Application: We should let AI "play" in complex simulations where it can experiment, fail, and learn the laws of cause and effect before we ask it to perform high-stakes tasks.
