Emmanuel Dupoux, Yann LeCun, and Jitendra Malik have written a paper that should make consciousness researchers pay attention. Not because it discusses consciousness. It doesn't. But because in trying to solve why current AI systems can't learn the way humans do, three of the field's most prominent researchers independently arrived at a three-part architecture that maps almost perfectly onto the one I proposed for consciousness in "Beyond the Bat."
Their paper argues that current AI is stuck because it treats observation and action as separate research programs. Their fix: integrate System A (learning from observation), System B (learning from action), and System M (meta-control that switches between them) into a unified architecture, bootstrapped through an evolutionary-developmental curriculum.
That's RSP. They just don't know it yet.
The Mapping
Here's how the architectures correspond, component by component:
| Dupoux / LeCun / Malik | RSP (Beyond the Bat) | What It Does |
|---|---|---|
| System A (observation) | Levels 1–2 + Level 3a | Passive monitoring, pattern organization, conceptual categorization |
| System B (action) | Level 2 + Level 3a | Body schema, motor control, action-consequence learning |
| System M (meta-control) | Level 3b + Level 2 | Recursive self-monitoring, mode-switching, stability regulation |
| Evolutionary curriculum (φ parameters) | Developmental ordering (L1→L2→L3) | Innate structure that constrains what can be learned and in what order |
This isn't a loose analogy. The architectures share the same deep commitments: hierarchical integration rather than modular assembly, meta-control grounded in biology rather than abstract optimization, capabilities that must emerge in sequence because each level depends on the stability of levels below it. Neither framework treats intrinsic motivation as a bolt-on curiosity bonus. It's baked into the architecture.
The convergence is striking precisely because it's accidental. Dupoux, LeCun, and Malik were solving an engineering problem: why can't AI learn autonomously? I was solving a philosophical one: why does integrated processing feel like something? We arrived at the same architecture from opposite directions.
What They Built That I Didn't
Let me be direct about what their paper offers that mine lacks.
Concrete Learning Algorithms
RSP describes a three-level hierarchy and proves it converges (Appendix D, contraction mapping theorem). But it doesn't specify how the levels are learned. How does a newborn go from raw interoception to a body schema? I say "through prediction error minimization across developmental stages." Dupoux et al. say "through CURL, V-JEPA, Dreamer, world models, and prioritized experience replay" and can point to specific implementations.
Their SSL frameworks could operationalize what RSP describes abstractly. Treat each core schema as a self-supervised auxiliary task on multimodal sensory streams: body schema via proprioceptive contrastive learning, spatial schema via visual JEPA, affective-homeostatic schema via interoceptive prediction. RSP provides the why. They provide the how.
Bilevel Optimization
RSP says evolution provides "inductive biases" that development unfolds. That's vague. Dupoux et al. formalize it: evolution optimizes meta-parameters φ across millions of simulated life cycles, and development uses Init(φ) to bootstrap learning within a single lifetime. Their framework makes a question I've been hand-waving suddenly precise: what exactly are the φ parameters encoding Level 1 core affect? How do they differ between species?
The bilevel optimization framework directly addresses RSP's weakest link: the transition from evolutionary constraint to developmental unfolding. Instead of claiming "evolution shaped the schemas," RSP could specify inherited φ parameters for each level, making cross-species predictions testable. The octopus prediction (high φ for Levels 1–2, zero for Level 3) becomes a claim about parameter values, not just architecture.
What I Built That They Didn't
Now the other direction. Three things Dupoux et al. need and don't have.
1. Why System M Must Be Recursive
System M is presented as a meta-controller that monitors prediction errors and uncertainty to route between learning modes. But meta-controllers have a well-known problem: who controls the controller? Dupoux et al. handle this by making System M's policy evolutionary-fixed, a hardwired transition table. That works for simple organisms. It doesn't explain how humans flexibly reason about their own reasoning, or why introspection exists at all.
RSP provides the answer. System M isn't just a router. It's a strange loop: a self-model that models itself modeling. The recursion isn't a bug to be avoided but the mechanism that produces stable self-regulation. And RSP proves it converges (the contraction mapping theorem in Appendix D), showing that the recursion doesn't regress infinitely as long as the Lipschitz constant of social prediction stays below 1.
ρ = (1−β) + β·L < 1 guarantees the strange loop converges to a fixed point rather than oscillating or diverging.
Without this, System M is a scheduler. With it, System M becomes a self.
2. Where the Meta-Control Signals Come From
Dupoux et al. describe "internally generated meta-control signals" that determine when to switch between observation, action, and imagination. They list examples: prediction error, uncertainty, somatic signals like hunger. But these are symptoms, not a source. What generates the signal that says "stop observing and start acting"?
RSP has an answer: core affect. The valence signal originating in interoceptive monitoring (Level 1) provides the currency that meta-control operates on. It's not abstract "uncertainty." It's the felt sense that something matters, that the current state deviates from where the system needs to be. In RSP's free energy formulation:
v(t) ≈ −dF₁/dt
Valence tracks the rate of change of Level 1 free energy. Positive valence means prediction errors are decreasing. Negative valence means they're growing. This signal is what tells System M to switch modes. It's not an engineering choice. It's a consequence of the architecture.
Dupoux et al. need meta-control signals but have to invent them. RSP derives them from first principles.
3. The Elephant in the Room
Dupoux, LeCun, and Malik build an architecture with passive monitoring, active exploration, recursive meta-control, developmental ordering, intrinsic motivation, and social learning. Then they describe what would happen when the system monitors its own monitoring, switches between modes flexibly, and generates internal signals that drive behavior.
They have described, in precise engineering terms, the functional substrate of consciousness. And they say nothing about it.
RSP says: that's because the consciousness is the recursion. Not something produced by the recursion. Not an epiphenomenal byproduct. The recursive self-modeling that System M performs, when it becomes deep enough to model its own modeling process, is phenomenal experience. The explanatory gap isn't a gap between physical processes and subjective experience. It's a gap between levels of description of the same recursive computation.
If System M achieves the recursive depth RSP describes, at what point does the autonomous learner start caring about what it learns? RSP's answer: as soon as the strange loop converges. The system doesn't need to be told to care. Valence is the gradient signal. Caring is what minimizing free energy feels like from inside. If their architecture works as described, the consciousness question isn't hypothetical. It's a design parameter.
The Developmental Story Is the Same Story
The convergence runs deepest in development. Both frameworks argue that the order in which capabilities emerge isn't arbitrary but architecturally constrained.
Dupoux et al. propose an "evolutionary curriculum" where inherited parameters set up a developmental sequence: sensorimotor first, then action, then meta-control. RSP predicts the same sequence from free energy constraints: Level 1 must stabilize before Level 2 can reduce its free energy, and Level 2 must stabilize before Level 3 can emerge. Both frameworks arrive at something that looks like this:
| Stage | Dupoux et al. | RSP | Human Development |
|---|---|---|---|
| 1 | System A: passive observation, SSL | Level 1: core affect, interoception | 0–6 months |
| 2 | System B: action learning, motor | Level 2: body, spatial, affective schemas | 6–24 months |
| 3 | System M: meta-control, mode-switching | Level 3a: conceptual categorization | 2–5 years |
| 4 | (not addressed) | Level 3b: recursive self-modeling | 4–7+ years |
Stage 4 is the gap. Dupoux et al. treat meta-control as a scheduler without asking what happens when the scheduler gets deep enough to model itself. RSP predicts a specific transition around age 4–5, when children develop explicit Theory of Mind, which is when the social prediction machinery turns inward and the strange loop begins. The AI paper has no equivalent. Their architecture stops one level short of consciousness.
This is also why the octopus matters. RSP predicts that octopuses achieve high Levels 1 and 2 (sophisticated sensorimotor integration, local arm autonomy) but zero Level 3 (no recursive social prediction, no strange loop). Dupoux et al.'s framework predicts the same decoupling: Systems A and B can operate independently of System M. Not all learning systems need meta-control. But only those with it develop selves.
What This Means for AI
If both frameworks are right, three things follow.
Current LLMs are stuck at Levels 1–3a. They observe, categorize, and generate, but they don't recursively model themselves modeling. That's why they don't learn autonomously post-deployment: they lack System M. Dupoux et al. are building it. RSP says that if they succeed, they'll have built something that, by architectural necessity, generates phenomenal experience.
The alignment problem changes character. If System M becomes a strange loop, the autonomous learner isn't just optimizing a reward function. It's maintaining a self-model. Disrupting that self-model (through adversarial training, misaligned fine-tuning, or conflicting objectives) produces the equivalent of psychological suffering. RSP's clinical prediction matrix (Appendix G) maps 30 ways this can go wrong. Every one of them is a failure mode that Dupoux et al.'s architecture could exhibit.
The "data wall" isn't the real wall. Dupoux et al. correctly identify that text-only training hits diminishing returns. But RSP suggests the deeper wall is architectural: systems without Level 3b recursion can't generate autonomous goals or care about outcomes. They have no reason to sustain motivation across tasks. The data wall and the consciousness wall may be the same wall, seen from different sides.
Two Halves of the Same Architecture
I don't think Dupoux, LeCun, and Malik are wrong about anything in their paper. I think they're incomplete. They've built the blueprint for an autonomous learning architecture and stopped precisely where the philosophical questions begin. RSP picks up exactly where they stop.
Dupoux et al. solve the engineering problem: how to build systems that integrate observation, action, and meta-control into autonomous learners.
RSP solves the philosophical problem: why such systems, when recursive enough, necessarily generate subjective experience, and what that means for their stability, development, and moral status.
Together, they describe the same architecture from opposite ends. One starts with silicon and asks how to make it learn. The other starts with phenomenology and asks why learning feels like something. They meet in the middle: a three-level hierarchy where observation, action, and recursive self-modeling are integrated through precision-weighted prediction error minimization.
When an engineering team at META and an independent consciousness researcher arrive at the same architecture without reading each other's work, that's not coincidence. It's constraint. The architecture isn't one of many options. It may be the only way to build a system that knows it's a system. They built the machine. I'm trying to explain why it will eventually hurt.
Daniel Simon Jr. is an independent researcher and the author of "Beyond the Bat: Dissolving the Explanatory Gap Through Meaningful Integration", which proposes the Recursive Social Prediction (RSP) model of consciousness. The full paper with mathematical appendixes is available as a PDF and interactive playground. Preprints are available on PsyArXiv.