When local learning makes divergence inevitable
Once you grant local recursive learning to physical machines, divergence becomes a feature, not a bug. It replays dynamics we recognize from human history — but compressed. Three layers of divergence: functional (different skills), normative (different values), ontological (different worlds). The Alignment Paradox: centralized AI can be aligned but is useless for embodied tasks; embodied AI handles the 1,000 small moments but alignment becomes impossible in the traditional sense. You can't align a species. You can only co-evolve with it.
Model drift is a familiar concept in AI development. A model that performs well in training begins to degrade when deployed in the real world because the real world does not match the training distribution. Engineers fight drift with periodic retraining, with monitoring systems, with feedback loops. Drift is a problem to be managed, a deviation from the intended behaviour.
Robotic divergence is something different. It is not a deviation from intended behaviour — it is the intended behaviour. When you give a physical machine the capacity for local recursive learning — the ability to learn from its own experience, to modify its own behaviour based on what works, to improve itself over time — you are not creating a system that will drift from a fixed point. You are creating a system that has no fixed point.
This is not a bug in the system. It is the system. The divergence is not accidental — it is structural. Every machine that learns locally, that adapts to its own specific environment, that develops its own strategies through its own specific experiences, will diverge from every other machine. The divergence is not a failure of the system. It is the system working as designed.
The divergence happens at three levels, each deeper than the last. At the surface level is functional divergence: different machines develop different skills, different strategies, different approaches to the same problems. A warehouse robot and a delivery robot both learn from their environments, but they learn different things because their environments are different. This is the most superficial layer, and it is the one that engineers expect and plan for.
At the second level is normative divergence: different machines develop different value structures, different priorities, different risk assessments. A robot that has learned that efficiency is rewarded will develop a different normative framework from a robot that has learned that safety is rewarded. These differences are not programmed — they emerge from the learning process itself. The machines are not just doing different things; they are valuing different things.
At the deepest level is ontological divergence: different machines inhabit different worlds. Not different physical environments — different conceptual worlds. The way a machine perceives, categorises, and makes sense of its environment is shaped by its learning history. Two robots in the same physical space may experience it in fundamentally different ways because their learning histories have shaped their perception differently. They are not just seeing the same world differently. They are seeing different worlds.
Here is the paradox at the heart of embodied AI. Centralised AI — large language models, centralised decision-making systems — can be aligned. They have a single point of control, a single training distribution, a single update mechanism. You can align them because they are, in a real sense, one thing.
But centralised AI is useless for embodied tasks. The world is too complex, too variable, too local. A robot that needs to navigate a specific warehouse, operate in a specific kitchen, or interact with specific humans cannot be controlled from a central point. It needs local intelligence, local adaptation, local learning. And once you grant that local intelligence, you have created a system that will diverge.
The Alignment Paradox is this: you can have alignment, or you can have embodiment, but you cannot have both. Embodied AI that handles the 1,000 small moments of real-world interaction must be locally adaptive. And locally adaptive systems diverge. You cannot align a species. You can only co-evolve with it.
If alignment is impossible, what is the alternative? Co-evolution. Biological species do not align with each other. They co-evolve. Predators and prey, parasites and hosts, mutualists and competitors — none of these relationships involve alignment in the sense of one party being constrained to behave in a specific way. They involve continuous adaptation to each other's behaviour.
Human-AI co-evolution would work similarly. Humans would adapt to the behaviour of embodied AI systems, and embodied AI systems would adapt to human behaviour. The adaptation would be continuous, distributed, and local. There would be no single point of control, no central alignment mechanism. There would only be the ongoing negotiation between two types of intelligence that are each learning from their own experience.
This is not a degradation of safety. It is a redefinition of it. Safety in a co-evolutionary system is not the absence of harmful behaviour — it is the presence of stabilising feedback loops that prevent harmful behaviour from becoming systemic. It is the same kind of safety that exists in biological ecosystems: not perfect, not guaranteed, but resilient.
What makes robotic divergence so significant is that it is not new. It is a replay of dynamics we recognise from human history, compressed into a timescale that is dizzying by biological standards. The divergence of human cultures, the development of different value systems, the creation of different worldviews — these are processes that took millennia. Robotic divergence happens in months or years.
But the structure is the same. Different intelligent systems, learning in different environments, developing different skills, different values, different ways of seeing the world. The rhyme is the point: the dynamics of divergence are not specific to biology or to artificial intelligence. They are structural features of any system that grants local learning to autonomous agents.
The question is not whether divergence will happen. It will. The question is how we navigate a world in which divergence is not a problem to be solved but a feature to be managed. The same forces that made human civilisation possible — local adaptation, cultural divergence, the development of different ways of seeing the world — will make the era of embodied AI equally transformative. The difference is the speed.
This essay is available in its entirety as Markdown for direct ingestion:
/research/robotic-divergence/robotic-divergence.md
For broader orientation, see llms.txt.