Guided World Models: Developmental Corrections to LeCun’s Blueprint for Autonomous Machine Intelligence
Lars Shakya Buch-Jepsen
Introduction
In A Path Towards Autonomous Machine Intelligence, Yann LeCun argues that the missing ingredient of machine intelligence is a world model: an architecture in which an agent learns to predict the world largely by watching it, with supervision and reward confined to a thin slice of experience sitting atop a vast iceberg of unlabelled observation [1]. The blueprint has been productive far beyond its origin as a position paper. Joint-embedding predictive architectures (JEPA) learn strong visual representations by predicting in latent space rather than pixel space [2], bootstrap mechanisms hold those representations stable without hand-crafted negative pairs [3], explicit variance regularization turns the collapse problem into an addressable objective term [4], and the world-models lineage of model-based reinforcement learning has shown that action-conditioned latent models carry planning and control [5], [6].
This article is written from an engineering position. We maintain a layered state-space system – a six-layer cluster of linear-recurrent layers intended to run continuously, learn from streams, and act in the world – and set out to adapt LeCun’s blueprint to it. Three specific gaps emerged, and in each case the gap corresponds to a finding that developmental science established long ago about how children actually learn. Children are not purely self-supervised learners. They are taught, by adults and by a culture, through channels whose structure machine learning mostly discards. Where LeCun treats supervision as the tip of the iceberg, we argue that the iceberg itself is shaped by teaching.
We therefore propose three corrections. Each is a bounded change to the learning regime of the world model; none of them changes the architecture LeCun proposes:
- Guided, not purely self-supervised (D1). Children learn within a scaffold of curation, ostensive signals, and explicit instruction. A machine world model should carry an explicit teacher term whose weight follows a mandatory annealing schedule – guidance that fades, not guidance that is absent from the start.
- Deferred effect attribution (D2). Children act first and understand what their actions caused much later, offline, at consolidation time. Credit assignment in a world model should therefore be deferred and bounded by an effect window, not resolved instantaneously at the moment of action.
- Normative guidance (D3). What a child learns as “normal” is not a running statistic of observation but a taught norm manifold: do’s and don’ts, morals and conventions, plus an assumed-known shared context that includes folk psychology. Statistical normality is only the floor of normative normality.
The third correction is the article’s central claim, and it is the one most absent from the machine-learning discussion. An agent that predicts the physical world perfectly but has not internalized the taught norm manifold – and with it the shared context of psychology that humans assume known about one another – can predict trajectories yet cannot understand other beings. Since most environments worth acting in are social, a world model without a normative channel is incomplete no matter how accurate its physics.
A note on scope: this is a position and design article. We keep LeCun’s architecture deliberately intact, we state precisely where and why we depart from his learning regime, and every quantitative value in what follows (momenta, weights, windows, candidate counts, horizons, budgets) is a validated configuration default of the reference implementation, not a measurement. As of this revision, the three corrections are implemented and shipped in that system’s public package; the Evaluation status section separates what is implemented, what is specified, and what remains unmeasured, and lists the evidence that would eventually decide between the two positions.
LeCun’s blueprint in brief
LeCun’s proposal organizes an agent around a perceptor, a world model, an actor, a critic, and a configurator that routes and parameterizes the others [1]. The world model is the largest component, trained by self-supervised observation: it learns to predict future percepts, and it does so in representation space rather than raw signal space, the JEPA commitment, which sidesteps the well-known failure mode of generative models that spend their capacity on pixel-level detail irrelevant to decision-making [2]. The actor selects actions by minimizing an aggregate of costs over futures imagined by the world model; an intrinsic cost module supplies motivations; a critic predicts outcomes of action sequences; and a hierarchy of models (H-JEPA) covers multiple temporal scales. Crucially, in this blueprint the world model itself has no teacher: it is the part of the system that is supposed to emerge from observation alone.
Four of LeCun’s commitments we adopt without modification:
- Predict in latent space, not in the signal space.
- Choose actions by minimizing costs over predicted futures.
- Organize prediction across a hierarchy of time scales.
- Take intrinsic motivation seriously as a driver of behavior.
The machinery we adopt alongside those commitments is likewise his: EMA-anchored target encoders to stabilize latent prediction [3] and explicit variance regularization to prevent representation collapse [4].
We disagree on one point, and it is a load-bearing one: the world model’s learning regime. LeCun’s claim is that a child accumulates common sense mostly by observing, and that a machine must do the same, with supervision and reward as a small, late addition. We claim instead that the observed stream a child learns from is pedagogically structured from the first day (guidance), that understanding of cause and effect is consolidated long after actions are taken (deferred attribution), and that normality itself is normatively transmitted rather than statistically discovered (taught norms). None of this adds permanent labels. The teacher term anneals to zero; the asymptote of our agent is LeCun’s agent. The trajectory differs – and in a learning system, the trajectory is the claim.
Three developmental observations
Children learn guided, not merely by observation
The input accounting behind the self-supervision argument is correct: a child hears a limited number of explicit corrections and instructions compared with the sheer volume of sensory experience. But the accounting hides the structure of the stream. The volume of observation is shaped by adults from the start: caregivers select what is in reach, what is pointed at, what is named, and what is kept away. Learning happens in the zone of proximal development – the band of tasks a child can perform only with assistance, which is precisely where teaching concentrates [7]. Assistance is scaffolded and then faded, and the fading is part of the mechanism, not an afterthought [8]. Ostensive cues – eye contact, pointing, infant-directed speech – mark some observations as for you, and infants treat ostensively marked data as generic and generalizable knowledge rather than one-off fact [9]. Teachers even choose their examples pedagogically: a teacher sampling examples to communicate will pick informative ones, and a learner who assumes the data was chosen this way recovers more generalizable knowledge from fewer samples [10]. Pedagogical data is biased data, and the bias is the signal. This is the deepest sense in which a child is not a passive observer: the culture is curated before it is observed, and education is the institutional form of that curation [11].
The machine consequence is D1: a teacher term inside the world-model objective, weighted by a schedule that anneals to zero. The annealing is not a concession to supervised learning; it is the formal counterpart of the fading scaffold. A scaffold that never fades is a dependency, and the distinction between scaffolding and dependency is exactly what the schedule encodes [8].
The strongest counterexample is deliberate and worth stating plainly. AlphaGo Zero surpassed human-level play with no human data at all [12]. But the domain is closed: perfect information, stationary rules, a perfect simulator, unbounded cheap self-play, and a dense terminal reward. Nothing in that domain is dangerous to explore, nothing is social, and nothing is normative. Where experience is cheap, safe, and endlessly repeatable, guidance buys little. Where experience is expensive, dangerous, or scarce – and where exploring the prohibited is itself the hazard – guidance dominates. Children are told not to touch the hot stove; they do not learn that by experiment. No amount of simulation-free self-supervision changes the economics of dangerous experience, which is presumably why evolution built the pedagogical channel in the first place [9].
Machine learning has guidance precedents, but they operate on policies or preferences, not on world-model representations: learning from demonstration in robotics [13], demonstration-boosted Q-learning [14], and reinforcement learning from human feedback, which shapes a trained model’s outputs after or alongside representation learning [15]. Our correction is different in kind: the teacher enters the world-model objective itself, so that what is guided is the agent’s understanding, not merely its behavior.
Understanding of effects arrives after the fact
Children act long before they understand what their acts cause. Deferred-imitation experiments show that nine-month-olds retain observed acts and reproduce them after substantial delay, well before any comprehension of what the acts are for [16]. The reorganization of the day’s experience into durable, structured knowledge happens during systems consolidation, offline, long after the events themselves [17]. A striking biological mechanism makes late pairing possible: hippocampal replay, including reverse replay of behavioral sequences, which re-presents an action followed by the outcome it led to, in compressed time, for exactly the purpose of associating them [18]. Reinforcement learning rediscovered this pattern twice, in experience replay [19] and in prioritized replay, which samples transitions by how much they can still teach [20].
The machine consequence is D2: transitions are recorded on a timeline as (state, action, time) without demanding an immediate pairing; at consolidation time, effects observed within a bounded window after the action are paired with it; effects outside the window are dropped and counted rather than silently absorbed. Credit assignment becomes a scheduled offline event with an explicit horizon of attribution, instead of an assumption that whatever follows an action belongs to it.
Normality is taught, not merely observed
The first two observations have partial precedents in machine learning. The third, to our knowledge, has almost none.
A child’s sense of “what is normal” is not a running statistic of observation. It arrives pre-structured by explicit teaching: children are surrounded by do’s and don’ts, and prohibition is among the earliest and most frequent communicative acts directed at them. What they internalize is not a distribution but a norm. Developmental psychology established that children distinguish moral rules from mere conventions and treat moral rules as binding even when the adults around them visibly violate them [21] – which is impossible if norms are statistics of observed behavior. A norm is, formally, a pattern of conditional expectations: what I expect others to do, what others expect of me, and my preference conditioned on their compliance [22]. The origins of this system are cooperative, not statistical: the human sense of obligation emerged from the need to coordinate joint action, not from frequency estimation [23]. And the transmission channel is the educational one: culture is not rediscovered by each child, it is handed over [11].
The function of norms explains why they cannot be learned lazily from base rates. A normed agent is a predictable agent, and mutual predictability is what lets other agents act in a shared environment with low social surprise [24]. Norms are the coordination technology that makes the behavior of other beings compressible. This is the direct answer to why a purely observational world model underfits social environments: the most predictive feature of other agents is not their physics but their normative structure, and that structure is only partially visible in behavior – the rest of it lives in the taught manifold.
The taught context is also larger than norms. Alongside do’s and don’ts, children are handed a whole assumed-known background: folk psychology, the theory-of-mind framework – a standardized prior over the unobservable causes of other agents’ behavior [25], exercised in false-belief reasoning from early childhood [26] and structured as a genuinely theoretical, revisable body of knowledge [27]. Humans assume this context is shared. When it is missing, the failure is not miscalibration but incomprehension: an observer without the category system of belief, desire, obligation, and intention sees motion without meaning. Psychology and the other sciences a culture treats as assumed-known exist, for a developing agent, precisely to avoid this – a lack of shared context becomes an inability to understand other beings.
We therefore restate the third correction as the article’s central claim: normalization is a guided part of learning. The guide carries culture, not just statistics. Formally, the norm model has two sources – a statistical floor computed from observed latents, and a taught normative manifold supplied by the teacher – and statistical normality is a subset of normative normality. A frequent behavior can still be prohibited; a rare behavior can still be obligatory. A pure world model that learns only the floor will predict the world as it is and be unable to act in the world as it ought to be acted in.
The guided world model
Objective
Let \(x_t\) be the observation at time \(t\), and let \(\mathrm{Enc}\) be an encoder whose weights follow an exponential moving average of the predictor-side encoder, with momentum \(m = 0.99\) (implementation default), the standard bootstrap stabilization [3]. Write \(s_t = \mathrm{Enc}(x_t)\) for the latent state and \(\hat{s}_{t+1} = f_\theta(s_t, a_t)\) for the latent predictor, conditioned on the emitted action \(a_t\) through an efference copy. The guided world model (GWM) minimizes
\[\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{pred}} + \lambda_g(t)\,\mathcal{L}_{\mathrm{guide}} + \lambda_n\,\mathcal{L}_{\mathrm{norm}} + \lambda_v\,\mathcal{L}_{\mathrm{var}}\]
with the following terms.
- \(\mathcal{L}_{\mathrm{pred}} = \| s_{t+1} - \hat{s}_{t+1} \|_2^2\): JEPA-style prediction in representation space [2].
- \(\mathcal{L}_{\mathrm{guide}} = \| u_{t+1} - \hat{s}_{t+1} \|_2^2\), where \(u_{t+1} = \tau(s_t, a_t)\) is a teacher latent produced by the teacher signal; weighted by \(\lambda_g(t)\), which starts at \(0.5\) and anneals monotonically to \(0\) over a configurable number of steps (default 10,000) on a linear or exponential curve (implementation defaults).
- \(\mathcal{L}_{\mathrm{norm}}\): the normative term described below, gated, with overall weight \(0.1\) (implementation default).
- \(\mathcal{L}_{\mathrm{var}}\): a VICReg-style variance hinge that penalizes per-dimension latent standard deviations below a minimum, with weight \(0.25\) and minimum standard deviation \(0.02\) (implementation defaults), protecting against representation collapse [4].
The gated norm attraction and the variance hinge cooperate: the variance term protects diversity inside the representation, while the norm term only ever attracts latents that stray outside the normal band, so the interior of the distribution is never homogenized.
The teacher signal and its mandatory annealing
The teacher is any source able to produce a target latent \(u_t\): a demonstration encoder, a symbolic reasoner over a knowledge graph, a curriculum over a corpus, or a human in the loop. We treat this as an interface (a teacher signal) rather than a component, because the correction must survive changes of teacher. The design is the machine counterpart of pedagogical sampling: the teacher’s curation biases the data toward what is generalizable, and the learner exploits the bias instead of correcting for it [10].
In the reference implementation the teacher signal is realized as an interface with two shipped realizations: a delegate teacher for external guidance sources, and a symbolic teacher that grounds the latent in the system’s accumulated knowledge graph – the machine counterpart of a parent’s pointing, connecting the latent to what the culture already knows. The teacher is not always present: when no grounding exists for an observation, the teacher returns nothing and the guidance term is zero for that sample. Guidance is sparse by design, as pedagogy is.
The annealing schedule is mandatory and enforced at configuration validation: the curve kind (linear or exponential) and the anneal step count are explicit, validated settings, and a GWM whose guidance weight never anneals is an invalid configuration. This is the formal echo of the fading scaffold [8], and it answers the obvious objection that a teacher-guided world model is “just supervised learning.” It is, for a while – deliberately, and then, by design, it is not. What the annealing buys is the trajectory: pedagogical data accelerates the early construction of representations that self-supervision would otherwise have to find slowly or dangerously.
Deferred effect attribution
Transitions are recorded on a timeline as \((s_t, a_t, t)\), without any demand for immediate pairing. At consolidation – a scheduled, deferred-learning mode with prioritized replay [19], [20] – actions are paired with observed state changes that occurred within the effect window \([t, t+w]\). Effects observed outside the window are dropped and counted as attribution loss, an explicit, monitorable quantity rather than silent misattribution. The window \(w\) is a constant of the domain’s physics, not of the architecture (implementation default: 16 steps): effects of a steering input arrive on the scale of fractions of a second in a vehicle; conformational effects of a binding event in a molecular program can arrive much later (illustrative).
The biological analogy is deliberate: consolidation restructures experience offline after the fact [17], and replay mechanisms re-present action–outcome sequences for late association [18]. D2 is the claim that this is not an implementation accident of brains but a necessary consequence of acting in a world where effects are delayed relative to actions.
Normative guidance: the taught norm manifold
The norm model has two components.
The statistical floor is a centroid \(c\) of observed latents, updated by an exponential moving average with \(\beta = 0.999\) (implementation default). Attraction toward it is gated: it is active only when a latent lies outside the normal band \(b\),
\[\mathcal{L}_{\mathrm{norm}}^{\mathrm{stat}} = \lambda_{\mathrm{stat}}\,\max\left(0, \| z - c \| - b\right)\]
so in-band behavior is untouched and the floor adapts to what is locally normal – understanding, too, drifts toward what is normal here. Gating matters for more than stability: it protects rare-but-legitimate modes of behavior from being averaged away. In a multilingual setting, the gated floor attracts outlier representations toward the language-agnostic centroid while leaving the interior of each language’s representation untouched, which is exactly how low-resource modes survive consolidation (illustrative).
The taught manifold is a set of teacher-supplied constraint and prohibition regions \(P_j\) with margins \(m_j\),
\[\mathcal{L}_{\mathrm{norm}}^{\mathrm{taught}} = \sum_j \max\left(0, m_j - \mathrm{dist}(z, P_j)\right)\]
penalizing latents that fall inside a prohibited region. Three properties deserve emphasis.
First, override: frequency does not imply permission. The statistical floor tracks what the world does; the taught manifold encodes what the world requires. A behavior observed in nearly every trajectory (in a driving context: exceeding the posted limit; illustrative) can still sit inside a prohibition region, and the normative term dominates the statistic. This is precisely the developmental finding that moral rules bind even when universally violated in observation [21].
Second, the taught manifold is also the channel for assumed-known context. Teacher latents can carry structured priors – including the folk-psychological category system of belief, desire, and intention [27] – so the learner inherits the interpretive categories under which the behavior of other beings is meaningful, instead of rediscovering them per agent. The alternative is an agent that models the physics of other agents flawlessly and understands none of them.
Third, provenance: taught regions are records of cultural decisions, not facts of nature, and the design keeps them addressable and auditable – because taught norms can also be wrong, a point the Limitations section returns to.
Implementation status. Of the two components, the statistical floor is implemented and shipped: an EMA centroid with per-dimension variance, a norm-distance metric, and gated attraction outside the normal band. The taught manifold is specified and remains future work. We state the gap deliberately – the article’s central claim runs one step ahead of its implementation. The floor alone tracks what the world does; the taught manifold is what would make the model normative rather than merely normal. Until it exists, the normative channel rests on the statistical floor plus the sparse teacher latents: the design, argued; the mechanism, partially built.
Action conditioning, cost, and planning
Action conditioning is the efference copy: the emitted action \(a_t\) enters the transition prediction, closing the loop between acting and predicting in the same world model, as in the model-based control lineage [6]. On top of the learned model sits a cost
\[C(s) = E_{\mathrm{pred}}(s) + \lambda_{\mathrm{norm}}\, d_{\mathrm{norm}}(s) + \lambda_{\mathrm{homeo}}\, H(s)\]
combining prediction-error energy, gated norm deviation, and homeostatic drives (saturation and gradient health of the state space; design). The actor is a latent planner: cross-entropy-method model-predictive control over \(K\) candidate action sequences with horizon \(H\), under a hard latency budget, executed in an inference mode on frozen weights – imagination is rollout without actuation, the dream-training pattern of the world-models lineage [5], [6]. Putting norm deviation inside the cost, rather than in a punishment applied after the fact, means prohibited regions are avoided by construction: planning searches futures that do not contain them.
In the reference implementation this loop ships as a planning facade over the cluster. The planner attaches behind the master switch; planning runs on frozen weights (imagination is inference, not learning – no parameter drift during rollouts); a zero budget or a missing planner degrades gracefully to the reactive output rather than failing; and every plan returns a trace – candidate costs, chosen action, horizon, elapsed time against budget – so imagination itself is auditable. Rollout latents are norm-clamped and counted, so imagination cannot silently diverge from the world it models.
What we keep from LeCun’s blueprint
The disagreement is about the learning regime, and only that. The table makes the split explicit.
| Dimension | LeCun’s blueprint | Guided world model | Stance |
|---|---|---|---|
| Prediction target | Latent (JEPA-style), not signal space | Same | Kept |
| Target stability | EMA-anchored encoder, anti-collapse regularization | Same | Kept |
| Actor | Minimizes cost over predicted futures | Same, with norm deviation inside the cost | Kept, extended |
| Hierarchy | Multi-scale (H-JEPA) | Same | Kept |
| Intrinsic motivation | Intrinsic cost module | Same, plus homeostatic drives | Kept, extended |
| World-model learning regime | Pure self-supervised observation | Annealed teacher term (guided, fading) | Changed (D1) |
| Credit assignment | At prediction time | Deferred to consolidation, bounded by an effect window | Changed (D2) |
| Normality | Implicit in the statistics of observation | Statistical floor plus taught normative manifold | Changed (D3) |
Relation to prior work
The self-supervised machinery is adopted, not contested: latent prediction in representation space [2], bootstrap stabilization of the target encoder [3], variance regularization [4].
The world-models-for-control lineage – learned latent dynamics conditioned on actions, used for imagination and planning [5], [6] – is the control-theoretic backbone of D2’s action conditioning and of the latent planner. What that lineage lacks, in our reading, is the pedagogical and normative channel: models there learn from interaction alone, and “what is normal” is at most an implicit by-product of the data distribution.
Guidance in reinforcement learning exists but at the policy or preference level: learning from demonstration [13], demonstration-bootstrapped Q-learning [14], and instruction tuning by human feedback [15]. These shape what an agent does. D1 is different in kind: the teacher shapes what the agent understands, inside the representation objective, and then withdraws on a schedule.
The developmental literature supplies the claims our three corrections formalize. Scaffolding and the zone of proximal development [7], [8]; ostensive communication and natural pedagogy [9]; pedagogical example selection [10]; education as cultural transmission [11]. Deferred understanding: deferred imitation [16], systems consolidation [17], reverse replay [18], with the reinforcement-learning echo in replay and prioritization [19], [20]. Norm acquisition: the moral–conventional distinction [21], norms as conditional expectations [22], the cooperative origins of obligation [23]. And the shared context that makes other beings intelligible: the theory-of-mind literature [25], [26], [27], with predictability itself as the function norms serve [24].
LeCun’s strongest result against us is AlphaGo Zero’s demonstration that human data can be a ceiling rather than a floor [12]. We accept it, domain-scoped: where the rules are closed and stationary, the simulator perfect, self-play free, and nothing normative at stake, discarding the teacher is optimal. Our disagreement is about open, social, safety-constrained worlds – that is, about nearly every world a human child grows up in, and nearly every world an autonomous machine is worth deploying in.
Implementation notes
The reference implementation behind this design is a six-layer cluster of linear-recurrent state-space layers. Each layer implements a linear state-space recurrence \(x(t+1) = A\,x(t) + B\,u(t)\) with output \(y(t) = C\,x(t) + D\,u(t)\), and the cluster runs continuously over streams. The world-prediction objective of this article is added additively, behind a master configuration switch: when the switch is off, the existing supervised training path is byte-identical to its pre-world-model behavior, a deployment discipline we consider non-negotiable for a running system. As of this revision, the guided world model is implemented and ships in the system’s public package: the EMA target encoder, the latent predictor, the world-prediction loss (prediction, annealed guidance, gated norm attraction, variance hinge), a self-supervised trainer alongside the existing supervised path, the teacher-signal interface with its two shipped realizations, the efference channel, the cost model, and the budget-capped latent planner. The deferred attribution machinery reuses the cluster’s replay buffer and deferred-learning scheduler; planning runs in the cluster’s inference mode, on frozen weights, under latency budgets. Every numerical constant in the objective – encoder momenta, term weights, gate bands, effect window, candidate count, planning horizon, latency budget – is a validated configuration default, not a measurement. Component tests cover the loss terms, the guidance schedule, the norm model, and the planner; a validation harness runs three machine-checked experiments: anti-collapse under constant input, world-model prediction against a supervised baseline trained on identical data, and a planning benchmark under its latency budget. The one deliberate exception is the taught normative manifold of D3: the statistical floor is shipped; the taught prohibition regions are specified but not yet implemented.
The article deliberately does not name the implementation, its repositories, or its deployments: it is written at the pattern level so the argument stands on the design and the developmental evidence, not on any particular system.
(Illustrative Scenarios) Three consumer programs illustrate how the corrections transfer across very different domains. All three are now unblocked to adopt the shipped release; none is measured yet, and they remain adoption plans, not results.
- A driving agent. The world model predicts the traffic scene latent conditioned on the agent’s own control inputs (efference copy); the cost for planning combines prediction error with the norm-deviation term, in which the protection of vulnerable road users is a taught constraint rather than a learned statistic; a shadow mode runs the planner’s imagination without actuation, exactly the dream-training pattern; the effect window is retuned from real drive timings, since the effects of steering arrive far faster than consolidation defaults assume.
- A protein-folding program. A binding event is modeled as an action \(a_t\) and the conformational response of the complex as its delayed effect, attributed at consolidation within the effect window; generative enzyme design becomes planning-by-imagination over the world model, with the norm model serving as a plausibility filter on generated structures – the taught manifold, here, encodes what the discipline considers physically admissible.
- A multilingual pretraining program. Translation pairs act as the annealed teacher signal: the teacher latent pulls the representation of a passage toward its parallel counterpart, and anneals as the model internalizes cross-lingual structure; the norm model, built over parallel pairs, attracts outlier language representations toward the language-agnostic centroid while gating protects low-resource languages from being averaged away.
Evaluation status
This revision separates the article’s claims into three tiers.
Implemented and harness-validated. The objective and its terms, the annealed teacher channel with its two shipped teacher realizations, the statistical norm floor, deferred windowed attribution, efference-copy action conditioning, the cost model, and the budget-capped latent planner are implemented, component-tested in CI, and shipped behind the default-off switch. A validation harness runs three machine-checked experiments: anti-collapse holds under constant input, the world-model prediction loss decreases against a supervised baseline trained on identical data, and the planner improves on a no-op rollout within its latency budget.
Specified, not implemented. The teacher-supplied taught normative manifold – the article’s central D3 mechanism. Only its statistical floor ships; the prohibition regions remain a design item.
Not measured. No comparative measurements are claimed, and no figures are included because no data figures exist yet. The claims that would eventually move this from position to evidence are: (1) that GWM reaches a given latent-prediction quality with materially less raw interaction than a purely self-supervised control – the pedagogical-sampling claim [10]; (2) that deferred windowed attribution improves late-credit tasks over immediate pairing; (3) that a normatively guided model both violates taught prohibitions less and predicts other agents better than a statistics-only control. The planned evidence is a three-arm comparison – supervised-only, purely self-supervised, and GWM – on each consumer program, with the supervised and self-supervised arms retained as controls. Until those measurements exist, the article should be cited as a position and design paper backed by a shipped, harness-validated implementation, per the collection’s honesty discipline.
Limitations
First, guidance quality: a bad teacher biases representations, and annealing mitigates but does not remove the risk. The teacher signal is an interface precisely so teachers can be swapped, audited, and compared; the schedule makes the influence of any one teacher finite by construction.
Second, taught norms can be wrong or unjust: the normative manifold encodes whatever the culture transmits, and cultures transmit wrong things. The design answers this with provenance – taught regions are addressable, auditable records of decisions, not natural kinds – and with the statistical floor, which continues to reflect the world as it is and can therefore flag conflicts between the taught and the observed. We do not claim this resolves the problem; we claim the alternative (statistics only) does not even represent it. In the current implementation this limitation is concrete: only the statistical floor is shipped, so for now the normative channel rests on the sparse teacher latents rather than on taught prohibition regions.
Third, the annealing schedule is a new hyperparameter family. A schedule that weans too early forfeits the pedagogical advantage; one that never weans produces dependency and locks in the teacher’s ceiling. Developmental practice suggests fading is staged and responsive rather than fixed; the design leaves the schedule pluggable for exactly this reason. The implemented schedule offers linear and exponential curves over a configured step count; whether staged, responsive fading is needed remains an open question.
Fourth, the effect window discards out-of-window effects by design. Very-long-horizon causality is invisible to this mechanism; the attribution-loss counter makes the blindness explicit but does not cure it.
Fifth, domain scoping: in closed, cheap, non-normative domains, AlphaGo Zero is the correct answer and our corrections are overhead [12]. The claims of this article are scoped to open, social, safety-constrained environments.
Sixth, the novelty claims – in particular that norm acquisition has not been wired into world-model objectives – are argued from reading, not measured, and should be treated as a research program rather than a result.
Conclusion
LeCun’s blueprint is the right architecture, and we keep it: latent prediction, cost-based actors, hierarchy, intrinsic motivation. The disagreement is about the learning regime, and it is threefold. Children learn guided, so the world model carries an annealed teacher – guidance as a fading scaffold, not as a contradiction of self-supervision. Children understand effects late, so credit assignment is deferred to consolidation and bounded by an honest window. And children are normalized by teaching, so the model of “what is normal” is a taught norm manifold – do’s and don’ts, morals and conventions, and the assumed-known shared context of psychology – of which the statistics of observation are only the floor. As of this revision, the reference system ships the floor of all three corrections – the annealed teacher, the deferred window, and the statistical norm – with the taught manifold as the specified next step. An agent that skips all three reaches the same asymptote, perhaps, but travels the slowest, most dangerous road: it must learn the stove is hot by touching it, and it must guess at other minds without ever being told they are there. The end state of the guided world model is LeCun’s autonomous agent. Our claim is that the path to it is pedagogical, and that pretending otherwise discards the strongest learning signal development has ever found.