Sutton's Philosophical Paradox: When the Father of Reinforcement Learning Opposes LLMs, He Breaks His Own Two Iron Laws
> Seven pages of prose, zero experiments, zero benchmarks, not a single new algorithm. Richard Sutton—2024 Turing Award co-winner and the father of reinforcement learning—posted a philosophical position paper, *Toward Enactive Artificial Intelligence* (arXiv:2605.24238), in May 2026, laying philosophical groundwork for the entire anti-LLM路线. Then the whole world misread it. > > Worse: Sequoia, NVIDIA, and Google have pooled $1.1 billion into David Silver's Ineffable Intelligence at a $5.1 billion valuation, betting everything on "Sutton being right." The problem is—the foundation itself is cracked.
---
1. What Enactive Actually Means: Not Generative, But Its Opposite
Sutton and Rafiee's core argument: AI needs an enactive approach to perception and cognition. The enactive view treats perception as active, skillful engagement with the world—agents perceive through acting, and understand the world by grasping how their actions shape experience.
This contrasts with the classical view in which perception is a passive internal process: the brain receives sensory input, processes information, and constructs internal representations. The enactive view says: perception is not receiving the world, but generating perceptibility through coupling with the world.
The paper proposes four key concepts:
- Experience: not passive data reception, but active, skillful interaction
- Action-Perception Inseparability: perception and action unfold simultaneously, not sequentially
- Autonomy: normativity arises from the agent's own organization, not external imposition
- Embodiment: the body is the condition that makes perception possible, not an optional add-on
- They have internal representations (vision-language encoders)
- They have external reward/imitation signals
- Their bodies are interfaces for executing pre-computed policies
- "Intelligence Without Representation" (1991): against internal representations; the world is its own model
- Subsumption Architecture: layered behavior architecture with perception directly coupled to action
- Situatedness and Embodiment: real intelligence is situated in the world
- After 2012, deep learning crushed coupling-based behavioral methods in vision and language
- Brooks later conceded his approach couldn't handle abstract reasoning, language, or long-horizon planning
- Today's robotics uses deep learning + RL hybrids, not pure behaviorism
- Table 1: $5.1B valuation, zero product, zero revenue. No breakthrough by 2028 means a down round or no exit
- Table 2: academic reputation. If enactive is proven philosophically self-contradictory, Sutton's late-career legacy suffers
- Table 3: existing technical debt and scale effects. LLMs already talk, VLAs already act; the question is integration and reliability
- Sutton & Rafiee, "Toward Enactive Artificial Intelligence", arXiv:2605.24238, 2026-05-22
- Sutton, "The Bitter Lesson", 2019
- Sutton, "Reward Hypothesis", 2004
- Brooks, "Intelligence Without Representation", 1991
- Silver & Sutton, "Welcome to the Era of Experience", 2025
- Ineffable Intelligence funding details: TechCrunch, SiliconANGLE, Caproasia, 2026-04-27
Key clarification: enactive is not generative AI—quite the opposite. Generative AI builds internal representations and then outputs; enactive says the agent need not have internal representations, because the world itself is its model. Generative AI is "perceive first, then act"; enactive is "perceiving and acting simultaneously."
---
2. First Pillar Crumbles: The Reward Hypothesis vs. Autonomy
Sutton's Reward Hypothesis (2004) is the cornerstone of RL:
> "All of what we mean by goals and purposes can be well thought of as maximization of the expected value of the cumulative sum of a received scalar signal (reward)." > > — Sutton, 2004
That means: all goals can be described as maximization of expected cumulative scalar reward. An agent's goals, purposes, and normativity all come from an externally defined reward function.
But in the 2026 enactive paper, Sutton writes in plain text:
> "RL marks a significant shift in that the agent evaluates its behavior through experience... However, evaluation criteria remain externally defined through the reward function." > > "Several lines of work have sought to relax dependence on externally specified criteria... Nevertheless, full autonomy in the enactive sense, where normativity arises from the agent's own organization, remains unrealized."
The contradiction exposed: enactive's core requirement—autonomy—demands normativity arise from the agent's own self-maintaining organization, not from external imposition. But RL's reward function is precisely an externally imposed scalar signal. Sutton demolished the temple he himself built.
More sharply: if enactive requires endogenous normativity, then the Reward Hypothesis—all goals are external reward maximization—is exactly what enactive regards as a fundamental defect. One cannot simultaneously believe "all goals are external reward" and "true autonomy requires endogenous normativity."
---
3. Second Pillar Pierced: The Bitter Lesson vs. Hard-Coding Cognitive Theory
Sutton's 2019 essay *The Bitter Lesson* distilled 70 years of AI research:
> "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective... We have to learn the bitter lesson that building in how we think we think does not work in the long run." > > "...They are not what should be built in, as their complexity is endless; instead we should build in only the meta-methods that can find and capture this arbitrary complexity."
Core of the Bitter Lesson: don't hard-code human theories of cognition (space, objects, multi-agent symmetry) into architectures; build only meta-methods and let computation search and learn.
But what is Sutton's enactive paper doing?
> "We identify and develop four key enactive concepts that we find most relevant to AI: experience, action-perception inseparability, autonomy, and embodiment."
The contradiction exposed: enactive cognitive theory—including Merleau-Ponty's phenomenology of perception and Varela's autopoiesis—is precisely the philosophical version of "how we think we think." Using the four enactive concepts as design principles for AI architectures is exactly the "building in human knowledge" move Sutton himself denounced. The paper's demand that embodiment be a "constitutive principle of cognition" rather than an "external constraint" is a philosophical judgment being hard-coded into an engineering architecture.
---
4. The Only Viable Defense: Enactive as "Meta-Structure"
One defense remains: enactive is not specific cognitive content but a meta-structure—not about what minds contain, but how they are organized. Not hard-coded spatial representations, but an architectural constraint like "perception and action must be inseparable."
This half-works: if enactive is merely a meta-method-level constraint, it might be compatible with the Bitter Lesson. After all, convolutions and certain invariances are architectural constraints too, yet deep learning lets them emerge from data.
But the problems: 1. The paper's criticisms of RL—"evaluation remains externally specified," "action-perception inseparability is not fully realized," "embodiment is treated as an implementation detail"—demand that engineering meet specific philosophical criteria. That's not a meta-method; that's a philosophical checklist. 2. Enactive tradition's core claims (autopoiesis, coupling-constitution) are themselves substantive cognitive theories, not neutral meta-methods. "The body is not an interface but the condition of cognition" is a substantive philosophical assertion.
If Sutton wants this route, he must distinguish in the paper: which enactive insights are meta-structure (Bitter-Lesson-compatible) and which are specific cognitive content (Bitter-Lesson-violating). He does not make this distinction.
---
5. Two Knives from Cognitive Science
5.1 The Scaling Up Problem
Brooks' subsumption architecture was spectacular at insect-level tasks (obstacle avoidance, phototaxis) but fails beyond insect level. The paper itself admits:
> "Brooks' machines have three main problems: They fail if sensor data had to be integrated in nontrivial ways over time... The task of the robot is difficult to change... The finished architecture is hard to understand when too many simple machines interact."
Enactive faces the same scaling problem: if perception-action coupling and autopoietic organization are the basis of intelligence, why would abstract reasoning, language, planning, and counterfactual thinking—capacities that clearly need internal representations—emerge from pure coupling? The paper offers no emergence path from embodied coupling to symbolic reasoning.
5.2 The Coupling-Constitution Fallacy
Adams & Aizawa's classic critique: coupling with the environment does not equal constitution of cognition. Cognitive processes may be coupled to the environment, but that doesn't make the environment part of the cognitive process.
The enactive tradition often commits this fallacy: because perception-action coupling is crucial to cognition, it infers that body and environment "constitute" cognition. But coupling is causal; constitution is ontological. "The body is the condition for perception to be possible" can be read as a coupling claim (conditional) or a constitution claim (the body constitutes cognition). This ambiguity makes enactive's philosophical foundation suspect.
---
6. Engineers Don't Listen to Philosophy: VLAs Are Already Folding Laundry
VLA (Vision-Language-Action) submissions at ICLR 2026 reportedly surged to 164 papers (the exact figure may not be verifiable, but the trend is clear): robotics is solving real problems with Transformer+RL hybrid architectures.
Google DeepMind's RT series, Physical Intelligence's π0, various end-to-end robot policies—these systems do exactly what Sutton's paper criticizes:
But they can fold laundry, plug in cables, twist bottle caps. Engineering reality doesn't wait for philosophical approval.
The paper's response is to concede RL has "structural resonance":
> "RL exhibits structural resonance with enactive principles through its emphasis on action, agent-environment interaction, feedback-driven adaptation, and agent-centered evaluation. However, this resonance should not be taken as theoretical equivalence..."
Translation: "Your engineering resembles what I described, but you aren't pure enough." The classic philosophical stance toward engineering—when engineering succeeds, philosophy says "that's not true X." But capital and paper counts flow to the engineers.
---
7. Rodney Brooks Played This Same Script—and Got Crushed by Deep Learning
From 1985–1991, Rodney Brooks proposed a route strikingly similar to Sutton's:
Brooks built Genghis (six-legged walker) and Cog (humanoid social robot), impressive at the time. But then what?
Sutton's enactive paper is philosophically near-isomorphic to "Intelligence Without Representation." If Brooks' pure behavioral route lost to deep learning + representation learning thirty years later, why would the enactive version differ?
The paper's implicit answer: RL is closer to enactive than Brooks' behavioral architectures because RL learns. But as the paper itself admits, RL's "learning" is still optimization under an external reward function—not the autonomous organization-maintenance enactive demands.
---
8. Three Tables, Three Bets: 2028, 2028, 2030
Table 1: David Silver's Ineffable Intelligence
April 2026: David Silver (former DeepMind RL lead, AlphaGo/AlphaZero architect) founded Ineffable Intelligence with a $1.1 billion seed round at a $5.1 billion valuation.
Investors: Sequoia, Lightspeed, NVIDIA, Google, DST Global, Index Ventures, the UK Sovereign AI Fund.
Silver's mission: build a "superlearner"—AI that learns through RL rather than human data, discovering "all knowledge from its own experience." He co-authored the 2025 paper *Welcome to the Era of Experience* with Sutton.
Stakes: $1.1B + NVIDIA GPU priority allocation + Google non-compete waivers The bet: by 2028, the superlearner demonstrates autonomous learning beyond LLMs in open domains
Table 2: Sutton's Enactive Academic Program
Sutton builds the philosophical foundation; Silver handles engineering. But if the paper's contradictions—Reward Hypothesis vs. autonomy, Bitter Lesson vs. hard-coded cognitive theory—cannot be bypassed or resolved in engineering, the theoretical foundation stays cracked. The bet: by 2028, engineering either proves enactive philosophy can evade these contradictions or proves they don't matter in practice
Table 3: LLM + VLA Engineering Route
OpenAI ($85B valuation), Anthropic (Google's $4B investment), xAI, plus Physical Intelligence, Figure, and other robotics companies. Route: LLMs provide world models and reasoning, VLAs provide action, RLHF/RL provides alignment and optimization. The bet: by 2030, general intelligence arrives via "large model + embodied actuators + RL fine-tuning," with no need for Sutton's philosophical revolution
Whose Chips Are Still on the Table
The answer won't be argued out on arXiv. It will be decided in the commercial battlefields of the late 2020s—either Silver's superlearner shows autonomous skills no human ever taught it, or the LLM+VLA engineering mash-up achieves qualitative change through quantitative scale.
---
Conclusion: A Self-Contradictory Blueprint, and a Prescient Warning
Sutton's enactive paper is both a warning (the LLM route genuinely neglects embodied interaction and intrinsic normativity) and a blueprint (one whose foundation is cracked).
The core contradiction: Sutton tries to fill enactive insights with RL, but RL itself—aimed at external reward maximization—is precisely the "externalization of normativity" enactive criticizes. He tries to use philosophy to legitimize the RL route, but the philosophy demands more than RL can provide.
Ultimately this returns to an older question: does AI progress come from philosophical insight or engineering brute force?
Sutton answered clearly in 2019: computational brute force. The 2026 enactive paper seems to say: no, this time we really need philosophy. But that philosophy conflicts with his own two iron laws.
Perhaps the most honest conclusion: the 2026 paper is not the final verdict of the god of reinforcement learning, but the final concern of a 77-year-old (born 1948) about AI's direction. He sees the limits of LLMs, but the philosophical tools he borrows—enactive cognitive science—are incompatible with the RL paradigm he has believed in all his life.
This is not a weakness. It is genuine intellectual tension. And precisely this tension makes this seven-page paper one of the most worth-reading philosophical texts in AI in 2026—not because it provides answers, but because it exposes the problems.
---
References: