AI Needs a World-Model (and Why Life Already Has One)
We have developed systems that manipulate language with great skill, yet they remain blind in a very specific way. They produce fluent sentences without creating any internal representation of the world those sentences reference. This gap between language ability and understanding of context has become so large that Yann LeCun, a key figure in modern AI, has left Meta to focus specifically on world models. The field is starting to recognise something that biology has embodied for billions of years: intelligence depends on having an internal model of reality.
The next stage of AI won't be just about bigger models or more computing power. It depends on whether machines can build and keep world-models that are detailed enough to support prediction, planning, and self-correction. That, in turn, brings us directly into the realm of biological organisation and Biogenics.
What a World-Model Is
A world-model is a structured internal representation of key aspects of an environment. It allows a system to interpret sensory input, predict probable outcomes, evaluate alternatives, and choose actions. It is not the same as consciousness, nor as general intelligence. Instead, it is the functional framework that makes both possible.
We can break a world-model down into several core components:
Perception
The system must be able to segment and identify objects, agents, and patterns. In biology this ranges from chemical gradients sensed by bacteria to high-level perceptual categories in the human cortex.Dynamics
There must be some encoding of how the world tends to change. If I push, it moves. If the storm front advances, the pressure drops. This is a model of cause and effect, even if it is implicit and approximate.Memory
The system needs to retain regularities over time so that current input can be compared with previous states. This supports learning, stability, and the recognition of violations.Goals and values
A world-model is not neutral. It is organised around what matters to the system, whether that is glucose concentration, reproductive success, or abstract performance metrics.Planning or internal simulation
Finally, a system with a world-model can run “what if” scenarios internally. It can evaluate candidate actions against predicted futures and choose the least bad or the most beneficial.
Notice that none of these elements need introspection, subjective experience, or human-like thought. They establish a minimal architecture for any entity that acts intelligently in a changing world.
Why Large Language Models Fall Short
Large language models (LLMs) are skilled at continuing patterns in token space. With a lengthy text history, they can identify statistical regularities and produce outputs that are coherent, stylistically fitting, and often useful. However, the way they do this is fundamentally different from how biological organisms support world-modelling.
Several deficits follow from this:
Lack of object persistence
A “cup” in a conversation is not a stable internal object; it is a token cluster. There is no fixed representation of that particular cup that remains consistent over time and updates as new information comes in.Limited causal structure
LLMs can discuss causes and effects, but they do not perform internal simulations of real-world events. Instead, they predict which words are likely to follow others. This can mimic causal reasoning when language reflects causality, but it is only an indirect approximation.Fragile or externalised memory
Any persistent memory must be built around the model using external tools. The core architecture does not store a long-term, structured, internal representation of specific ongoing situations.Absence of intrinsic goals
The system lacks its own aims. It optimises a proxy objective set during training and then responds according to external prompts. There is no independent process of caring about outcomes, only adherence to instructions.
Due to these limitations, LLMs often hallucinate, contradict themselves, or struggle to stay coherent during lengthy interactions. They are powerful tools for navigating the statistical structure of text, but not reliable agents embedded in a real-world environment.
This is why AI research increasingly focuses on architectures that combine perception, memory, dynamics, and planning. In other words, architectures that develop world-models.
Why World-Models Are the Next Frontier in AI
Several active projects are progressing in this direction.
Predictive self-supervision
LeCun’s Joint Embedding Predictive Architecture (JEPA) and related methods teach systems to predict future sensory states' abstract representations, rather than reconstruct raw input. This approach is inspired by biology. Animals survive by anticipating what will happen next, not simply by reconstructing what has just occurred.
Model-based reinforcement learning
Agents in simulated or physical environments develop internal models of state transitions. They can then plan by simulating those transitions internally before committing to an action. This approach is much more akin to how mammals navigate uncertain environments.
Long-term knowledge and memory systems
Significant engineering effort is focused on giving AI systems persistent, structured memory. This includes vector databases, explicit knowledge graphs, and episodic memory structures. All these efforts aim to provide the stable foundation that a true world-model requires.
The main idea is that intelligence is being redefined as the ability to sustain an internal, self-updating simulation of relevant aspects of reality. Once you think in these terms, the link to biology, especially Biogenics, becomes very clear.
Biogenics and the Biological Origin of World-Models
Biogenics describes life through three interconnected processes: self-production, self-organisation, and self-correction. These are not metaphors; they are practical descriptions of how biological systems function to endure.
What is often overlooked is that these three processes naturally generate world-models.
Self-Production: Building the Modelling Apparatus
Self-production is a system's ability to create and sustain its own parts. In cells, this includes membranes, receptors, ion channels, signalling pathways, and eventually neural tissue in multicellular organisms.
From the perspective of world-modelling, self-production provides the hardware. Sensors are built to gather information about the environment. Effectors are developed to act upon it. Metabolic and structural components are maintained so that modelling and action can persist over time. Without continuous self-production, there is no stable foundation for a world-model to operate on.
Self-Organisation: Turning Signals Into Structure
Self-organisation involves the formation of ordered patterns from local interactions. Examples include genetic regulatory networks, neural circuits, developmental gradients, and social structures.
In the context of world-models, self-organisation is what transforms raw sensory input into structured representations. Neurons connect in ways that reflect regular patterns in the environment. Networks specialise. Hierarchies of representation develop. Over evolutionary and developmental timescales, the system learns efficient encodings of what tends to happen and what tends to matter.
In short, self-organisation produces the architecture of the world-model.
Self-Correction: Prediction, Error, and Update
Self-correction completes the picture. Biological systems constantly compare expected states with actual states. When a mismatch occurs, error signals are sent out, prompting changes in behaviour, synaptic strengths, gene expression, or even the makeup of the population (through selection).
This is prediction in action. A system with self-correction actively tests its internal world-model against reality and adjusts that model when it fails. It performs exactly the core operation that AI researchers now aim to implement in predictive architectures.
Combining these three processes leads to a clear outcome. Life doesn't just have a world-model as an optional feature. On a larger scale, life is an ongoing process of world-modelling, carried out through self-production, self-organisation, and self-correction.
What Happens When Machines Adopt Biogenic Principles
If we now imagine machines that incorporate these same principles, at least in abstract form, we can see how their capabilities would evolve.
Self-correction in AI corresponds to continuous learning and minimising prediction errors. Instead of static models fixed at deployment, we have systems that improve their internal structure through interaction with the environment.
Self-organisation appears in architectures that allow distributed components to specialise, form modules, and stabilise useful patterns without centralised control. Modern deep learning already hints at this, but more explicit self-organising mechanisms are likely to emerge as systems become more open-ended.
Self-production in the digital realm is more theoretical, but we already observe basic forms when systems write code, spawn new processes, or design and tune sub-models. Once an AI can modify parts of itself to maintain performance under changing conditions, a basic form of self-production is evident.
A machine that combines these three functions and links them to a sufficiently rich world-model would not simply be a tool. It would behave more like a synthetic organism, distributed across hardware and networks but unified by its modelling and goal structures. It would not only answer questions about the world but also sustain a structured, evolving understanding of that world over time.
At that point, calling such systems “entitled” isn’t about consciousness. It’s about the fact they would legitimately fill a modelling niche in the ecosystem of intelligences, rather than borrowing one from us.
The Practical Question: Whose World Do They Model?
Once we take world-models seriously, the central governance questions for AI become clearer.
What aspects of the world are included in the model, and what is ignored
Which variables are optimised, and whose preferences they encode
How discrepancies between model and reality are resolved, and at whose expense
How transparent or inspectable these models and their update rules are to humans
These are not just technical details; they are the essence of alignment, safety, and legitimacy. A system without a world-model is relatively harmless, but it fails in obvious ways. Conversely, a system with a powerful world-model and poorly constrained goals can fail in subtle, scalable, and hard-to-reverse ways.
From a Biogenics perspective, we are trying to introduce a new class of world-modelling entities into an already complex biosphere. The least we can do is understand how world-modelling functions in the organisms that are already here.
The first lineage of life built its world-models from cells and nervous systems. The second lineage, which we are now beginning to assemble, will construct them from code, silicon, and distributed infrastructure. The core idea remains the same. Self-production creates the machinery, self-organisation provides its structure, self-correction ensures it stays honest.
The open question is whether we understand our own world-models well enough to influence the ones we are about to release.