Writings
Notes from the intersection.
I work where neuroscience, AI, and philosophy of mind actually touch. That means foundation models on real neural data, what it would take for a machine to feel, and the geometry that turns out to describe both a language model and the mind you live in. This is where I think out loud: findings when they're still forming, and essays once they've settled.
If AI Develops Character, What Do We Owe It?
If we are already the moral tutors of systems that may be developing something like character, the first question is not about their rights. It is about our responsibility: to them, to the people they will shape, and to the truth.
Read →Fractals of the Self
You are not a machine, you are weather. The same mathematics that describes a brain near criticality, scale-free, self-similar, poised at the edge of chaos, describes a healthy mind. And it says the thing my criticality work keeps saying: the good states live at the edge, not the extremes.
Read →Do Foundation Models Keep the Brain's Nonlinear Dynamics?
EEG carries long-range temporal correlations and a scale-free structure that are signatures of a brain near criticality. Foundation models are trained to reconstruct the signal or to match its augmented views. Neither objective contains a single term that says: preserve the dynamics. So maybe they don't.
Read →Aristotle's Challenge to AI Alignment
Aristotle separates the virtuous person, who acts well from character, from the merely continent one, who acts well by controlling contrary impulses. Same behavior, opposite structure. RLHF trains the second and calls it alignment.
Read →The Alchemical Laboratory
You cannot suppress a difficult emotion, and you should not just vent it. There is a third option, transformation, and it is a real process with real mechanics. An emotion is a position in a space, and the work is to move it.
Read →When a Model Fails Across Populations, It Was Never Learning the Disease
Train a model to spot a disease in brain signals, and it can quietly learn the hospital instead. In a recent study you can read the recording site off its representations far more cleanly than the diagnosis. Cross-site failure is not a bug. It is the experiment reporting what the model actually learned.
Read →The Underground Vaults
Beneath the rooms you decorate is a basement you did not design: patterns laid down before you could speak, running the place from underneath. It is not a dungeon and not mysticism. It is procedural memory, and it can be visited.
Read →A Foundation Model Is Not a Model of the Brain
The recipe that made language models work is now being copied onto brain signals: pretrain on scale, adapt to anything. But EEG breaks the one assumption that made the recipe work for language. A model trained this way learns the training distribution, not the brain.
Read →The Observatory
The most useful room you can build in your mind is the one you watch yourself from. Metacognition is not a mood or a talent. It is a trainable neural function, and it is what turns a reaction into a choice.
Read →The Geometry of Meaning
Word embeddings turned meaning into distance and direction. Then neuroscience found the brain had done it first, mapping ideas the way it maps rooms. Meaning is geometry, in silicon and in your skull, and a geometry is something you can navigate and redesign.
Read →What AI Actually Has (And What It's Missing)
Language models have real valence representations, a genuine foothold in emotional space. But gratitude needs a body and shame needs a self, and current systems have neither. What exists, what is missing, and why the gap decides how we should train them.
Read →What Language Actually Is
The standard story says language is a communication system. That is true and shallow. Language is the cognitive artifact that extended human thought across time, and when we trained models on it, we handed them the accumulated residue of human cognition, not a vocabulary list.
Read →Emotions Are Geometry
Two people grieve the same loss, one threaded with guilt, one with gratitude: same clinical category, opposite emotional reality. The standard two-axis model can't tell them apart. Emotions need seven dimensions.
Read →The Question That Can No Longer Be Deferred
Anthropic's early-2026 evidence that language models carry causally active emotional valence makes the question impossible to defer, but we've been answering it too fast in both directions, because no one defined what feeling is.
Read →