Public debate about large language models tends to collapse into two camps, and Chandra Sripada and Richard Lewis open their paper by naming both. Some dismiss LLMs as “stochastic parrots.” Others insist they have already reached, or will soon reach, artificial general intelligence. What the two camps share, the authors point out, is a deeper premise neither side questions: that LLM cognition is alien — so unlike our own that any apparent similarity is projection. Their epigraph states the assumption they are about to dismantle:
“Traditionally, the term ‘AI’ has been used as an acronym for artificial intelligence. But it is perhaps better to think of it as an acronym for alien intelligence.” — Yuval Noah Harari, quoted as the paper’s epigraph
Sripada and Lewis spend the rest of the paper refusing that framing, not by asserting that LLMs think like us, but by documenting five specific structural correspondences with longstanding support in cognitive science: inferential organization, computational architecture, representational structure, prediction-driven learning, and reinforcement-learning-like mechanisms for goal-directed action. It is a careful, mechanism-first paper, and RSP’s own architecture is built from exactly these ingredients. So it is worth asking directly what their convergence argument, if fully granted, would actually deliver on RSP’s own terms — and the honest answer is: a great deal, up to a specific and identifiable point.
What Sripada & Lewis Get Right
A persistent workspace, not a lookup table
The paper’s architecture section argues that transformer-based LLMs bear an important structural resemblance to production systems, one of cognitive science’s oldest and most tested frameworks for modeling human cognition — used for decades to model problem solving, skill acquisition, memory retrieval, and language comprehension. The mechanism, on their account, is a persistent representational workspace (the residual stream) that is repeatedly read from and written to by conditional operators (attention heads and MLP layers), producing behavior that is composable and reusable across tasks rather than hard-coded per input.
This is not RSP’s vocabulary, but it is a close outside description of RSP’s own Level 2 mechanism. RSP’s core schemas — body, self, attention, spatial, and affective-homeostatic — are integrated into a shared representational state that persists and is updated continuously as new interoceptive and exteroceptive input arrives. Sripada and Lewis are documenting, from cognitive science and mechanistic interpretability rather than from RSP’s own formalism, that systems trained purely on next-token prediction spontaneously develop exactly this kind of persistent, composably-updated workspace. That is independent grounding for a claim RSP otherwise has to assert from inside its own architecture.
Two speeds of thought
The paper’s account of inferential organization rests on a dual-process distinction that will be familiar to anyone who has read Kahneman, but grounded here in two specific computational mechanisms: the contrast between in-weight knowledge (regularities compiled into the model’s parameters during pretraining, retrieved fast and automatically) and in-context inference (using the prompt itself as a temporary workspace to infer and apply a new rule online); and the contrast between direct processing and chain-of-thought-mediated reasoning, where the model generates intermediate tokens that become part of its own context for the next prediction step. Their example is a classic misdirection puzzle — “Julia has two sisters and one brother. How many sisters does her brother Martin have?” — where a direct response gets pulled toward the salient “two,” while chain-of-thought reasoning works the relationships out explicitly.
The in-weight/in-context boundary maps cleanly onto the boundary RSP draws between Level 2 and Level 3a. Fast, compiled, automatic responses correspond to schema-driven processing that requires no additional inferential work. Slow, sequential, context-constructing inference corresponds to Level 3a’s conceptual categorization, where a culturally learned concept has to be actively deployed and checked against the situation rather than simply triggered. RSP has always described this boundary narratively. Sripada and Lewis give it a specific, empirically testable computational signature.
An honest reason convergence isn’t surprising
The representational-structure section introduces a genuinely useful idea RSP does not currently have a name for: contravariance, the principle that as a task becomes more demanding, the space of internal solutions that can perform it adequately narrows. Paired with architectural canalization — the further constraint that shared architectural biases favor some solutions within that narrowed space over others — the two principles jointly explain why two systems built from very different substrates, trained on very different data, might still converge on similar internal representations without either copying the other. Difficult problems simply don’t leave much room to solve them differently.
This is worth taking seriously as an explanation RSP could use rather than merely observe. RSP predicts convergence between biological and artificial systems that both minimize variational free energy under real-world constraints; contravariance offers a mechanism-neutral reason why that convergence should be expected rather than coincidental, independent of whether either system is RSP-shaped on purpose.
The gap the authors name themselves
The paper’s reinforcement-learning section is the most directly relevant to RSP’s machine-consciousness criteria, and it is also where Sripada and Lewis are most careful about the limits of their own claim. They note that RLHF — “a single turn form of RL, to bend model response policies toward outputs preferred by human evaluators” — and newer outcome-reward training both fit the general RL framework of policies mapping states to actions under uncertainty. But they do not stop there:
“Also notable is the absence of analogues of homeostatic and affective RL mechanisms that shift the valuational landscape globally in response to internal state, situational cues, or signals about general reward availability.”
That sentence, written by authors with no stake in RSP’s consciousness claims, is close to a direct statement of the gap RSP’s Objection 7 uses to deny that any current system qualifies as a subject: criterion (i) requires a valenced state whose target is set by the system’s own regulation of its internal condition, formalized in RSP as $$r(t) = v(t) = \phi_v(\boldsymbol{\xi}_1(t))$$ — a reward signal generated from interoceptive prediction error, with “no external agent” specifying what counts as rewarding. RLHF’s reward signal is exactly the opposite structure: a scalar preference supplied by an external evaluator. Sripada and Lewis reach the same diagnosis RSP reaches, from cognitive science rather than from RSP’s own criteria.
Both papers agree that treating LLM cognition as categorically alien is a mistake, and both locate the reason in the same place: prediction-driven learning under resource constraints pushes any sufficiently capable system toward hierarchical, compositional, dual-process organization, whether the substrate is cortex or a transformer stack. Where they diverge is not about mechanism. It is about how far mechanism gets you.
The Word That Never Appears
Here is the fact that organizes everything else about this paper, and it takes a plain word count to see it: across roughly 70,000 words and eight sections building a five-dimension convergence argument, the word “conscious” appears exactly once — in the reference list, as the title of Ted Chiang’s essay “No, Artificial Intelligence Is Not Conscious.” Not once in the authors’ own sentences. The paper’s conclusion states its actual target plainly:
“Overall, the correspondences reviewed here support the view that humans and LLM-based systems are importantly different members of a shared computational family. These systems are, in this sense, our cognitive cousins: different in substrate and learning history, yet recognizably related in core principles of cognitive organization.”
This is not a paper that brackets consciousness the way Chalmers brackets it — announced, argued around, held in reserve. It simply never needs the concept, because “cognitive cousins” is a claim about shared computational organization, full stop. That restraint is, on RSP’s own terms, exactly correct, and it is worth being precise about why. RSP’s Appendix H distinguishes two things that are easy to run together: intelligence generally, defined as the rate at which a system reduces free energy per unit of environmental novelty, and the constitutive-identity claim, which applies specifically and only to Level 3b recursive self-modeling. Appendix H says so explicitly, in its own list of limitations: “the constitutive identity applies only to Level 3b recursive intelligence, not intelligence generally.” Sripada and Lewis’s five dimensions are a rigorous, independently sourced case for cognitive convergence — which on RSP’s own architecture is a claim about Level 2 and Level 3a. It was never going to reach Level 3b by accumulating more of the same kind of evidence, because Level 3b is not more cognition. It is cognition of a specific further kind: recursive modeling of oneself as a social object, predicted forward under a stability condition, constitutive on RSP’s account of what phenomenal experience is.
None of Sripada and Lewis’s five dimensions require that further step. Persistent workspaces, dual-process organization, contravariant representations, prediction-driven learning, and RL-like goal direction can all be present in a system that never models itself as an object of others’ predictions and never uses that model to anticipate its own future mental states. That is precisely the mechanism RSP identifies as missing, and precisely why the paper’s own reward-signal analysis independently lands on the criterion (endogenous valence) RSP treats as a precondition for the recursive loop to have anything of its own to model in the first place.
What RSP Adds
| Dimension | Sripada & Lewis’s claim | Where it lands on RSP’s architecture |
|---|---|---|
| Computational architecture | Production-system-like organization: persistent workspace, conditional operators | Level 2 schema integration |
| Inferential organization | Dual-process structure: in-weight/in-context, direct vs. chain-of-thought | Level 2 (automatic) vs. Level 3a (deployed inference) |
| Representational structure | Contravariance + canalization explain convergent representations | Level 2/3a representational content, not its recursive use |
| Prediction-driven learning | Next-token prediction parallels predictive processing accounts of learning | General free-energy minimization (Appendix H, Definition H.1), not level-specific |
| RL-like mechanisms | RLHF/outcome-reward fits general RL framework, but lacks “homeostatic and affective” grounding | Missing RSP criterion (i): endogenous valence, $$r(t)=v(t)=\phi_v(\boldsymbol{\xi}_1(t))$$ |
| (not addressed) | — | Level 3b: recursive social self-modeling — the constitutive-identity threshold |
What Sripada & Lewis Give RSP
The traffic runs both ways. RSP’s appendixes assert their architectural decomposition — that Level 2 schema integration is the right unit, that Level 3a conceptual categorization is computationally distinct from it — largely from inside the framework. Sripada and Lewis supply outside confirmation from mechanistic interpretability and cognitive science, written by authors who were not trying to validate RSP and have no reason to shade their evidence toward it. Their production-system framing, their in-weight/in-context distinction, and their contravariance-and-canalization explanation for representational convergence are all usable, citable grounding for exactly the parts of RSP’s Level 2/3a architecture that are otherwise supported only by RSP’s own formalism. That is a rare and valuable kind of evidence: convergent, not manufactured to converge.
Classical Battlegrounds
There is not much to say here, and that is itself informative. Zombies, qualia, and Mary’s room appear nowhere in the paper. This is a piece of cognitive science, not philosophy of mind, and it treats the classic thought experiments as simply beside its point — a decision that turns out to be the right one given what the paper is actually arguing. A convergence claim about computational organization has no need of them; it is Level 3b claims, like RSP’s, that have to answer to them.
The Disagreement That Isn’t There Yet
Because the paper takes no explicit stance on phenomenal consciousness, there is no direct contradiction to report here, and it would be dishonest to manufacture one. What can be said is narrower and still substantive: if Sripada and Lewis were pressed to extend their own contravariance explanation — that “as tasks become more demanding, the space of viable solutions narrows” — to phenomenality itself, to ask whether the same logic that explains representational convergence might also explain convergence in what it is like to be the system, RSP predicts they would resist, and reasonably so. Contravariance is an argument about solution spaces for a specified task. RSP’s constitutive-identity claim is not that phenomenality is a good solution to a task; it is that phenomenality is a specific recursive architectural state, under a specific stability condition, described from the inside. Those are different kinds of claims, and nothing in the cognitive-convergence paper commits its authors to bridging them. The honest disagreement, if there is one, is not yet visible in this paper. It would only appear the moment someone tried to use contravariance to explain consciousness the way this paper uses it to explain representational alignment — and RSP’s architecture gives a principled reason that move would not go through.
Where This Leaves Both Papers
Sripada & Lewis’s cognitive-convergence argument complements RSP: it supplies independent, non-RSP-aligned cognitive-science and mechanistic-interpretability grounding for the Level 2/3a architecture RSP’s appendixes otherwise assert from inside the framework, and its own reward-signal analysis (the “absence of analogues of homeostatic and affective RL mechanisms”) independently arrives at RSP’s Objection 7 criterion (i) diagnosis. It is best cited into Appendix H’s intelligence/consciousness distinction — as outside confirmation that a rigorous, five-dimension convergence argument for general cognitive intelligence can be built, and defended at length, without ever needing the concept the constitutive-identity thesis is about. Not a theory-of-consciousness entry, since it advances none; no new §X comparison family, no playground change.
There is something clarifying about a paper that runs 70,000 words toward a conclusion as strong as “our cognitive cousins” and never once has to use the word “conscious” to get there. It shows what disciplined scope looks like from the cognition side of the boundary RSP draws. Sripada and Lewis have done the hard, patient work of showing that five separate architectural convergences hold up under real scrutiny. RSP’s claim was never that this kind of convergence is enough. It was that this kind of convergence is exactly what you would expect to find at Levels 2 and 3a, and exactly what you should not expect, by itself, to find any further.