When Do Digital Twins Cross Into Knowledge?

Large Language Models process tokenized sequences and have become the default substrate for many digital twin implementations, a choice that reflects institutional momentum. The limitations are plain. Systems built on autoregressive next-token prediction perform well at interpolating within symbolic distributions, and they stop there. Modeling causality in continuous physical domains (material stress propagating through a turbine blade, fluid dynamics inside a pressurized pipeline) calls for something structurally different.

First, computational costs compound the problem. Real-time sensor fusion from industrial deployments (a modern automotive assembly line may carry several thousand sensors, a large petrochemical facility tens of thousands) generates data volumes where continuous model updating is needed. Labeled training data for such environments is expensive to produce and perpetually incomplete and healthcare and aerospace applications face an additional constraint: regulators and engineers want explanations traceable to physical mechanisms. Statistical performance alone satisfies neither group.

Yann LeCun has stated the core objection plainly and repeatedly, and the argument deserves engagement on its merits. Autoregressive language models trained on next-token prediction acquire symbolic interpolation ability. Object persistence, intervention consequences, and long-horizon reasoning of the kind physical planning demands remain outside what the architecture produces today. Biological intelligence requires orders of magnitude less training data to acquire comparable world understanding (and in most physical domains, superior understanding).

The alternative LeCun and collaborators have pursued, something called Joint Embedding Predictive Architecture (JEPA), learns abstract representations in latent space. A predictor network forecasts future embeddings from current states. Then, hierarchical variants stack these operations across multiple temporal and spatial scales. Planning computations remain tractable and uncertainty is represented explicitly. Video-based implementations have shown measurable capability in motion understanding and action anticipation, domains where physical grounding is a must and where the shortfalls of purely symbolic approaches are most visible.

The business implications are concrete. Manufacturing operations that run forward simulations grounded in learned physical dynamics reduce the number of physical prototyping iterations. Healthcare systems that anticipate physiological responses without exhaustive per-patient labeled datasets lower both cost and latency. Edge deployment, where inference must complete within milliseconds and bandwidth is constrained, penalizes architectures that carry large computational budgets. These are just engineering requirements.

In January 2026, LeCun co-founded Advanced Machine Intelligence Labs (acronym AMI also means "friend" in French), securing 1.03 billion dollars in seed funding at a 3.5 billion dollar valuation. Think about it. The participation of major technology funds and industrial investors signals that institutional capital has reached conclusions consistent with the technical arguments. The stated focus (physical-world understanding, persistent memory, safe planning) maps directly onto the deficiencies which current architectures leave unresolved regardless of additional scale.

Now, a separate and insufficiently examined problem concerns training data provenance. Where the data comes from. Corpora assembled from digitized text are English-dominant and reflect the conceptual thinking of a small number of societies so Digital Twins modeling urban systems, labor dynamics, or ecological processes will reproduce these biases in their outputs. In fact, the error is technical as much as it is ethical.

A simulation of land use encoding extractive resource systems will generate predictions that diverge systematically from actual behavior in communities organized around relational stewardship principles. In the Pacific Islands for instance, our indigenous communities cultivate knowledge of environmental interconnection that differ structurally from the assumptions embedded in most available training data. DTs deployed in these contexts without corresponding adjustments will produce errors that are difficult to detect and expensive to correct. We know a bit about MTPE.

The physics, by contrast, is invariant. Microgravity behaves identically regardless of the cultural background of the engineers who modeled it. Radiation flux follows known laws on a Martian surface or in low Earth orbit as reliably as in a terrestrial lab. World models trained on sensor streams from varied environments acquire latent representations that transfer across physical contexts precisely because the underlying dynamics remain constant. Moreover, cultural and institutional modules can be layered above this invariant physical core to enable adaptation while retaining the predictive fidelity that physical grounding provides.

Hybrid configurations that combine physics-based simulation engines with predictive embedding models is the logical integration pathway. The physics engine constrains the solution space, the learned model handles high-dimensional state estimation and forward prediction that analytical methods cannot tractably compute. Then, edge hardware enables real-time synchronization. Latent-space prediction reduces bandwidth consumption by representing states compactly. Organizations that have piloted these configurations report improved alignment between model outputs and observed physical behavior, a prerequisite for operational trust that purely data-driven approaches have consistently failed to achieve.

In light of these observations, the transition calls for investment that is unglamorous but substantial. Sensor quality sets the upper bound on what world models can learn, while systems trained on degraded or poorly calibrated inputs will produce degraded representations, regardless of sophistication elsewhere in the pipeline. Standardized protocols for multimodal data ingestion accelerate iteration cycles and workforce development must bridge the gap between domain expertise (metallurgy, fluid mechanics, physiology) and the model interpretation skills these systems need. These are slow, expensive, and inescapable steps.

Now, Pacific Island epistemologies supply one instructive case. Communities across the region cultivate relational approaches to their environment and collective stewardship. DTs that incorporate analogous structural assumptions generate simulations that connect more directly with local stakeholders and with the actual causal relationships governing the systems being modeled. Global businesses expanding into these markets achieve greater predictive precision when they consult specialists who translate systems embedded in local knowledge. Yes, interfaces and training data that respect place-based ontologies produce better models.

We can say the same logic applies to orbital and planetary habitats. Microgravity, radiation exposure, and severe resource constraints reshape daily practices in ways that terrestrial training data cannot capture. DTs for orbital stations or planetary bases must maintain fidelity to human decision-making under these conditions while the underlying physics supplies the invariant foundation. Recently, heritage programs in spacecraft simulation have taught that physics-first approaches enable reliable prediction in extreme environments. The extension to long-duration crewed missions follows the same principle: Learn the physical dynamics from sensor streams, layer human behavioral and cultural modules above that foundation, and the resulting system transfers across contexts where direct training data is absent.

The trajectory, given adequate technical and organizational investment, points toward simulation systems that serve as predictive instruments with genuine operational utility. They will anticipate equipment failures before sensor readings breach alarm thresholds, simulate intervention consequences before irreversible commitments are made, and support iterative design at a pace that physical prototyping cannot match. Executives evaluating technology roadmaps will find that organizations which piloted physics-grounded world models earliest are accumulating advantages where better models produce better decisions, which generate better operational data, which improve subsequent model iterations.

Whether most organizations will make the investments before competitors who already have done so is, like most questions about institutional behavior, considerably less certain than the underlying physics. The technical direction is established. The 1.03 billion dollars committed to AMI Labs in a single seed round only confirms that some institutional actors have concluded the window for early positioning is finite. Organizations that treat sensor infrastructure, multimodal data protocols, and culturally calibrated training corpora as engineering priorities will find the gap between simulation and physical reality narrowing at a pace their competitors will struggle to replicate.