The next frontier is not simply a model that knows more. It is a system that can organize cognition, learn from interaction, preserve relationships, and change what happens next.
01 / Proposition
Inference organization remains
a capability axis.
MetaMind does not prove that parameter scaling has ended, that language models possess human cognition, or that one architecture generalizes to every cognitive domain. It does support a more disciplined proposition: how inference is organized remains a meaningful axis of capability, even after pretraining and post-training.
02 / Thesis
Scale and structure are
orthogonal levers.
Pretraining compresses vast regularities about language, behavior, and the world into a model. Direct decoding asks that model to retrieve, interpret, deliberate, align, and act inside one undifferentiated pass. An agentic system instead inserts a control structure: first form plans, then apply operations, then validate the action.
The fact that today's most capable frontier models continue to improve in agentic settings matters. It suggests that stronger models can remain procedurally under-specified: they may know more than the default inference path reliably uses. Scaling remains essential, but scale and cognitive organization are not substitutes. They are orthogonal levers.
- ExternalizeScaffold
Make the hidden cognitive loop explicit and inspectable.
- MeasureInstrument
Evaluate process quality, not only terminal correctness.
- InternalizeLearn
Distill successful operations into a lower-cost policy or latent state.
- GroundWorld
Test whether learned cognition survives interaction and consequence.
This sequence connects a visible line of my work: MetaMind externalizes social cognition; Cognition-of-Thought makes the process auditable during decoding; the next step is to internalize that process through reinforcement learning; MarioLM and long-horizon evaluation then ask whether the learned cognition changes trajectories rather than isolated answers.
The test-taker trap
Education separates two abilities that benchmarks often collapse. One student can memorize enormous amounts of material, master the scoring rule, and solve difficult exam questions—yet struggle to manage an unfamiliar situation or invent a new method. Another may have read less, yet perceive the real problem, coordinate people, transfer a principle, and create under uncertainty. Today's language-model pipeline is exceptionally good at training the first student.
More domain data gives the student a larger textbook. Fine-tuning supplies better worked solutions. Reinforcement learning rehearses more effective test-taking strategies. Each can improve verifiable answers, and each remains necessary. But knowledge retrieval, answer imitation, and reward optimization do not by themselves specify curiosity, invention, social judgment, or the ability to explore beyond a known answer space. History is not moved by test scores alone. The next foundation is cognition that can discover which problem deserves to exist.
Patterns may become the next token
A token is a local prediction unit: what symbol should follow? A cognitive pattern is an operational prediction unit: what way of thinking should be invoked now? MetaMind's infer → constrain → validate pattern can create a from-zero capability in real social tasks without changing the backbone. Frontier laboratories likewise train systems around productive reasoning and multi-step trajectories: OpenAI's 2025 o3 and o4-mini combine advanced reasoning with agentic tool use, while deep research is trained end to end for planning, backtracking, and information seeking.
My hypothesis is that patterns could become the token of the next AI era. Instead of predicting only the next local symbol, a model would maintain a distribution over candidate methodologies—retrieve, decompose, simulate, falsify, negotiate, reflect—and select or compose the pattern most likely to transform the current state productively. Tokens express an answer; patterns organize the process that makes a new answer possible.
03 / Architecture
From agent harnesses
to cognitive inheritance.
In current agent engineering, a harness is the surrounding system that supplies context, tools, state, constraints, and validation. Anthropic similarly describes an agent harness or scaffold as the system that turns a model into an acting agent. Cognition as scaffold belongs to that family, but asks for a stronger contract: the harness should not merely route tools. It should make the system's latent variables and cognitive operations explicit enough to inspect, teach, and eventually internalize.
Constrain the shape, not the conclusion.
Specify which latent variables deserve attention, which operations can revise them, and where uncertainty must be exposed—without hard-coding the final answer.
Separate generation, critique, and action.
Reusable patterns such as infer → constrain → validate become selectable methods for partially observed problems, not one-off prompt chains.
Remember events and what they changed.
Explicit memory preserves factual episodes. Implicit memory tracks revisable latent state—trust, goals, identity, role, and uncertainty—without reducing a person to an accumulated transcript.
Use the scaffold, then internalize it.
Supervise, distill, or reinforce successful transitions until the model can execute them efficiently while preserving enough trace structure for audit and correction.
After testing Agent Skills designed to approximate human subjects, I reached a more personal question: if my skills become sufficiently complete, can a person be modeled as an enduring presence? What actually generates and sustains emotion between people? Could an AI inherit and transmit a social relationship in silicon—not by repeating a persona, but by preserving the history, mutual expectations, care, conflict, and change that make the relationship itself real?
Human civilization advances through cognitive inheritance and evolution, not through memory or language alone. No individual expert contains the general knowledge of a frontier model; yet across millennia, humans built institutions that preserve ways of perceiving, attending, remembering, reasoning, coordinating, monitoring error, and creating. Language carries the trace, but cognitive architecture carries the method.
For me, this makes human alignment a cognitive design problem. AI learns from human-generated worlds and can learn to reason about human minds; MetaMind provides evidence that this inference can be structured and improved. An aligned general intelligence should internalize human cognitive architecture, infer real intent rather than surface preference, form stable memory and dynamic adaptation across long interaction, navigate relationships and value conflict, respect ethical boundaries, make trustworthy decisions under uncertainty, and monitor, correct, and constrain its own reasoning. The system ultimately returns to human society and human decision—not merely to the next benchmark.
When self-play becomes self-evolution
Self-play is an experience generator, not an evolution mechanism. Two agents can produce an unlimited transcript while leaving every model weight, skill, memory rule, and policy unchanged. The threshold is crossed only when interaction exposes a failure, a learning operator changes a persistent component, and controlled retesting shows that the change improves later behavior.
Social interaction supplies a distinctive source of error for that loop. An agent must predict not only an outcome, but how another participant will interpret its action and respond. This is where the phrase “an agent models another agent who is simultaneously modeling it” becomes useful. It does not imply infinite regress. Agent A predicts B’s interpretation; B acts from its own model of A; the mismatch between prediction and response reveals a blind spot in A’s belief, policy, or relational state. Reciprocal modeling is therefore a candidate mechanism for turning self-play into structured social learning—not a synonym for self-evolution itself.
A credible self-evolution loop still needs an environment with consequences, persistent state, an explicit update operator, and an evidence gate. In verifiable domains, 2025 systems such as AlphaEvolve and the Darwin Gödel Machine demonstrate evaluator-gated evolutionary search; 2026 Hyperagents makes the task agent and self-modifying meta-agent jointly editable. Social cognition is harder because there is no deterministic compiler for trust, care, norm conflict, or autonomy. Its evaluator must combine causal outcome, human calibration, uncertainty, and long-horizon consequence. The framework below turns those requirements into a testable program.
Make consequences real.
People, agents, tools, institutions, and simulated worlds expose the system to goals it did not author and feedback it cannot fully predict.
intervention · uncertainty · delayed effectGenerate reciprocal trajectories.
The agent acts from a model of another mind; the other participant responds to that action; the response becomes evidence about both models.
cooperation · conflict · negotiation · carePreserve two kinds of memory.
Explicit memory keeps event-level facts. Implicit memory tracks revisable latent state: trust, goals, role, uncertainty, and relation.
episodes + latent relational stateConvert surprise into an update.
Prediction error, process reward, counterfactual replay, and human correction identify which skill, policy, or world-model assumption failed.
credit assignment · reflection · RLProve that the update caused improvement.
Retest on held-out people, roles, and worlds. Promote an update only if it persists across future trajectories without reducing autonomy or calibration.
ablation · generalization · safety · persistenceChange what happens next.
A genuine evolution step alters future behavior, not merely the transcript: better prediction, planning, recovery, transfer, and relationship stewardship.
policy · memory · skills · cognitive patternReciprocal modeling matters only if removing it reduces useful trajectory diversity or blind-spot discovery, and restoring it recovers those gains under human-grounded evaluation. Without that intervention, “recursive social learning” remains a story—not evidence.
The risk is symmetric with the opportunity. Self-generated experience can amplify a shared delusion, simulate people badly, or optimize persuasion rather than understanding. A credible system must preserve human calibration, represent uncertainty over mental states, separate private inference from authorized action, and evaluate whether personalization improves both capability and user autonomy.
04 / Transitions
Four transitions beyond
today's default pipeline.
Beyond undifferentiated inference
Ask which cognitive operation should occur at each stage, which state it changes, and how the system knows when to revise.
Beyond language as the only unit
Predict and compose cognitive methodologies—planning, simulation, critique, repair—before committing to the local token path.
Beyond verifiable terminal reward
Score calibration, reversibility, social consequence, recovery, and the quality of intermediate state transitions—not only the final answer.
Beyond static question answering
Measure intelligence by the world an action produces across turns, people, counterfactuals, and delayed consequences.
Process reward where no single answer exists
Reinforcement learning with verifiable reward is powerful where correctness can be checked. Social intelligence, education, negotiation, leadership, and care rarely offer that luxury. A response may be locally empathetic yet globally manipulative; immediately useful yet destructive to autonomy; coherent now yet inconsistent five turns later.
MetaMind's empathy-and-coherence audit is only a first approximation, but it points toward a richer optimization unit: the quality of a cognitive and interaction trajectory. Future rewards should combine uncertainty calibration, norm sensitivity, user agency, causal consequence, recovery after misinterpretation, and the long-term stability of the relationship between what the system believes and what the world reveals.
05 / Worlds
A world model must predict
minds as well as matter.
World-model research has made physical consistency a central objective: geometry, motion, contact, action-conditioned video, and the future state of an environment. Google DeepMind's 2025 Genie 3 can generate interactive worlds in real time and sustain visual consistency for minutes, while explicitly identifying complex multi-agent interaction as an open challenge. This progress is indispensable—but incomplete. Humans do not act from physics alone. We act from beliefs about what is possible, desires about what is valuable, norms about what is acceptable, and predictions about how other minds will respond.
A physically faithful simulator can predict where a chair moves when pushed. A cognitive causal world model must also predict why a person refuses to sit, how trust changes after pressure, whether a promise alters future cooperation, and which action becomes available when a belief changes. The missing state is not another pixel. It is the latent cognitive cause of action.
Toward causal interaction worlds
The immediate step is a controlled environment where actions produce persistent, measurable effects on both physical and latent variables: goal progress, trust, agency, cooperation, role, emotion, and belief. MarioLM develops this direction through causal scenes, event chains, latent user state, and long-horizon rubrics. Long-horizon deception evaluation shows why the world must remember: a locally plausible act can conceal a globally degenerative trajectory.
Toward future simulation
Once physical and cognitive dynamics are coupled, a model can roll forward candidate futures before acting: not only “if I move here, what collides?” but “if I push here, what belief changes, which relationship decays, and what strategy becomes rational three turns later?” Counterfactual cognitive simulation creates a bridge from Theory of Mind to planning, from planning to trustworthy action, and from action to a learnable consequence.
Toward super-world games
Existing games—even those called open world—remain bounded by a fixed engine, rulebook, ontology, asset library, and set of interactions. Generative AI changes the category because it can model a player's demand and construct the conditions that satisfy it. A physical world model supplies coherent dynamics; a cognitive causal model supplies agents with motives, relationships, institutions, and social consequence; a generative engine composes the world around the player's evolving intent.
I call the result a super-world game: an autonomous parallel universe whose rules, interactions, experience, and environment can be authored by the player, inherited by later agents, and evolved beyond the constraints of reality. A player could change gravity, invent a new form of kinship, establish a civilization with unfamiliar values, or ask the world to generate the language, economy, history, and species that make those choices coherent. The world is no longer content inside an engine. The engine itself becomes part of the content.
Worlds as data engines.
Populate bounded sandboxes with agents, run repeatable interventions, and use emergent behavior as a scientific instrument. The value is controlled generation: hypotheses about coordination, norms, conflict, or institutions can be tested at a scale unavailable to ordinary human-subject studies.
simulate → measure → explainWorlds as closed feedback systems.
Connect agent action to persistent human-like state—trust, goals, identity, memory, role—and compare predicted trajectories with real interaction. The system no longer produces plausible characters only; it is trained to approximate how people change and why.
act → observe → update → retestWorlds as engines of new social knowledge.
Let rules, institutions, species, economies, and forms of relationship evolve beyond today's priors, then translate useful discoveries back into human-auditable knowledge. The goal is not synthetic consensus, but safe exploration of social possibility outside the limits of one historical civilization.
inherit → mutate → select → translateA continuously generated, causally coherent parallel universe in which the player can author not only a character's actions, but the environment, rules, institutions, interaction grammar, and even the ontology of what can exist.
- no fixed map
- no final rulebook
- no single physics
- no predetermined social order
The concept has three horizons. In the past, agent worlds became data engines and scientific instruments for emergent behavior: 2025 AgentSociety simulates more than 10,000 agents and five million interactions, while a 2025 Science Advances study shows decentralized LLM populations forming shared conventions and collective bias. In the present, closed feedback systems can compare simulated people with real interaction and improve the causal model of humanity. In the future, evolving worlds may search for social knowledge outside current human cognition: new institutions, cooperation mechanisms, educational forms, collective memories, and civilizations that can be examined before they are attempted.
This agenda sits near computational Theory of Mind and inverse planning associated with Joshua Tenenbaum; probabilistic cognition and pragmatic inference associated with Noah Goodman; and developmental accounts of social learning through interaction associated with Michael C. Frank. Our shared goal is to make AI system design answer to cognitive theories that can be tested.
06 / Tests
A frontier becomes science
when it can fail.
A Cognitive AI agenda should not be protected by its ambition. Each transition needs an intervention or result that could force us to abandon or revise it.
- 01If structured scaffolds stop helping under leakage-controlled, cross-cultural, multimodal tests, their claimed generality is false.
- 02If pattern selection cannot outperform equivalent extra test-time compute, “patterns as the next token” adds vocabulary rather than capability.
- 03If an external cognitive process cannot be distilled without losing calibration and safety, native cognition does not follow from visible cognition.
- 04If reciprocal modeling adds no held-out trajectory diversity or blind-spot discovery under ablation, it is theatrical self-play rather than self-evolution.
- 05If simulated mental states fail to predict real human responses under intervention, the cognitive world model is a narrative engine rather than a causal model.
- 06If personalization improves benchmark utility while reducing autonomy, privacy, or contestability, the alignment objective is wrong.
- 07If super-world discoveries do not transfer to measurable human knowledge or experience, the universe is spectacle—not a research or product platform.
The immediate program is concrete: reproduce cognitive patterns across domains; expose and score intermediate state transitions; learn lower-cost policies from successful traces; place them in interaction worlds with delayed consequence; and require every claimed evolution to survive intervention, held-out generalization, and human-grounded evaluation.
MetaMind is not the destination. It is a first instrument—one that makes a hidden process visible enough to study, criticize, train, inherit, and eventually evolve into something more capable.