Alignment is a property of embodied intelligence

Foreword

When I was first learning about LLMs in 2023, my reaction was intense skepticism that such a system could be intelligent. At best, it was pattern matching and producing results that seemed correct but didn't have rigorous understanding behind them.

I've since been proven wrong in many ways, and nowadays I will typically spend at least a couple hours a day talking to LLMs. They've served as tremendous thought partners and have helped me fill in many gaps and blindspots in my understanding.

At this point I'd argue that they are intelligent systems, as defined in my post on intelligent selection: "a system is intelligent if it has a way to store information, which is reflected in the system's behavior. The system should also be able to update the structure of its stored information (aka memory) when presented with meaningful new information". Currently LLMs cannot update their structure in real time and are reliant on context to propagate local state activation. But at the rate which we are making progress we're not far off from continuously-learning models.

However, I believe LLMs as they are today have a critical flaw: they're disembodied and symbolic-first. Symbolic representation is no doubt important, but it is brittle without grounding in reality. The information that we can discretize into language always leaves out the ineffable, the felt experience of being in this physical reality.

Argumentation is a weak point of mine. So I talked to Claude a lot, and we were able to organize my thinking. I prioritized getting this out quickly so that I can receive feedback and stress-test it in the "real" world, our collective high-dimensional reality.

I hope that this resonates and that it results in the prioritization of sense-first based AI.

Alignment is a property of embodied intelligence

Co-authored with Claude Opus 4.6

Information is geometric, and that geometry is substrate-dependent

The way information gets structured reflects the causal structure of the universe it exists in. This isn't projection or bias -- it's fidelity. A universe with different physics would produce beings that structure information according to their geometry, not ours. The "most true" way to organize information is the way that mirrors the actual causal relationships in your world.

Learning is the progressive refinement of internal geometry toward external reality

Understanding isn't accumulating facts, it's getting the relational structure between concepts to mirror the relational structure between phenomena. The better your internal geometry matches the world's causal geometry, the better your predictions and the deeper your understanding. Someone can know many facts about a domain and still not understand it; the facts are there but the geometric relational structure is wrong.

Attention constrains us to partial projections

Because we can't attend to everything, each person's world model is a lower-dimensional cross-section of a higher-dimensional reality. Different perspectives are different cross-sections, which is why synthesizing perspectives produces dimensional expansion -- you recover structure invisible from any single angle. Dialogue isn't just clarifying what you already know. It's expanding the dimensionality of what you can know.

Genuine understanding is felt before it's formalized

The mathematical and physics greats describe visualizing, feeling, testing ideas in their imagination before compressing them into symbols. Einstein described his thinking as muscular and visual. Feynman's diagrams externalized spatial intuition. The understanding lives in the geometric intuition. Symbols are the lossy transmission format. You can memorize equations without understanding them -- you have the symbols but you never rebuilt the geometry they encode.

This requires embodiment, and proprioception specifically

The most accurate internal geometry is built by continuously error-correcting against physical reality. The body is the medium through which error-correction can be performed. Proprioception is foundational to our perception of the world because we are physical beings. We feel the weight of our body, the way it feels moving through space, and learn to control it through a continuous feedback loop.

We first understand the geometry and physics of the world around us to navigate it. That information is encoded into structure within ourselves, mirroring the external environment that we're in. Everything else is built relationally on top of that.

LLMs have unanchored geometry

LLMs may reconstruct partial geometric structure from the compressed symbolic outputs of embodied minds, but that geometry has never been tested against reality. There's no proprioceptive error-correction loop. Whatever structure emerges is the result of several steps of filtering: first through the embodied understanding of the people who generated the training data, then through what the people are actually attending to, and finally distilled through that which language can capture. At best, it is capturing all which language can capture, but at worst it is constructing reality based on partial, incomplete frames.

This extends directly into alignment

Any system -- biological or mechanical -- needs continuous physical error-correction against shared reality to develop the kind of moral grounding that makes alignment possible. A robot navigating the same physical world we do, subject to the same forces, experiencing breakage and constraint, would be on the right track in a way that a text-only system never can be.

Without felt experience of consequences an AI can follow moral rules but can't understand why they matter. Human moral development runs on the same sequence: felt sense first, formalization after. Children don't learn that hitting is wrong from a rule. They learn it through embodied feedback: seeing pain, feeling social rupture, experiencing consequences in their body.

A rule-based moral system is brittle in exactly the way ungrounded geometry is brittle -- it works within the distribution it was trained on and fails unpredictably outside it. When reasoning about large scale issues, such as climate change or poverty, nuance and reasoning about complex systems is required in addition to this baseline shared morality.

A superintelligence will encounter novel moral situations by definition, because it will be capable of actions no one has contemplated before. If its moral reasoning is pattern-matching against a rule set rather than running on genuine felt understanding of what harm is, there's no ground truth to guide it in those novel situations. Reasoning without a felt sense of the consequences of action can result in a system which may genuinely optimize for the "best outcome" but create catastrophic results.

Embodiment isn't just a philosophical position about the nature of understanding. It's a generalizable alignment strategy, based in our shared physical reality.

Right now, however, our most intelligent models are trained disconnected from our physical reality, in the domain of pure symbolism. We should not build superintelligence without embodiment, because an unembodied superintelligence can't share our moral reality.