Research
Eight strands, one question: how coherent behaviour arises in a collective that nobody is directing — and what it takes to steer it once it has.
Artificial life
Symbolic reasoning in embodied agents
Can a population evolve attention itself, with nobody telling it what "different" means?
Ten clusters sit on a ring, each holding five objects built from shape (four kinds), colour (four kinds) and size (two kinds). In every cluster exactly one object differs from the rest along one of those three attributes. Agents wander between clusters and have to approach and eat it — the clip's camera holds three clusters in frame at once.
The brain reading each scene is a tiny generic attention network: one token per object, a few self-attention heads, a residual connection, a small scoring MLP — about a thousand numbers, with nothing in the architecture that knows what "odd one out" means. Those thousand numbers are the genome. There is no training step and no fitness function that scores an agent on accuracy: an agent only gains energy by actually eating the odd object, loses it by eating the wrong one, and lives, reproduces or starves on that energy alone. Whatever skill at the task exists has to pay for itself in survival, not get handed down by a reward signal that already knows the answer. A "neutral" control, run alongside, decouples the payoff from the agent's actual choice, to tell real selection apart from drift.
The clip contrasts an early population against one two million steps later, same seed. Early on, agents aim close to chance. Later, they aim at the odd object reliably — and the attention weights show a mechanism: objects sharing the value of whichever attribute actually varies start attending to each other, a "who looks like me" pattern that only makes sense once the rule has been found.
A direct descendant of the Polyworld lineage — embodied, foraging, selection-pressured populations — pointed at a symbolic/relational task instead of raw survival.
Cultural evolution and machine learning
Rogers' paradox in language models
If copying is cheaper than learning, who is left to learn?
Cultural evolution has a result that reads like a warning. Social learning is cheap: copy someone who already solved the problem and skip the cost of solving it yourself. So copying spreads. But Rogers' paradox is that a population full of copiers is no better off on average than one with no copying at all — copied information is only as good as the moment it was acquired, and in a world that keeps changing it goes stale. The benefit of copying depends on somebody still paying to learn directly.
The same accounting now applies to machines. Language models trained on the output of other language models lose knowledge across generations, and they lose the rare things first — the tail goes before the common case does. That failure is usually discussed as a data-quality problem. Read through cultural evolution it is something more specific: model collapse is Rogers' paradox, with model weights in the place of a population and training corpora in the place of culture.
A 2026 preprint makes that reading quantitative on small language models, and asks the question the paradox forces: who pays for the fresh data? Copying is the cheap move for any individual model, which is exactly why it scales — and exactly why the ecosystem cannot run on it alone.
Language-driven control of living systems
Prompt-to-Intervention
Can you tell a cellular collective what to do, in ordinary English, and have it comply?
Prompt-to-Intervention (P2I) is an architecture that maps a free-form sentence onto a bioelectric intervention on a collective of cells. A sentence embedding conditions a multi-head convolutional policy network, whose parameters are found by CMA-ES. The crucial move is the reward: rather than hand-specifying a target pattern, a vision–language model reads the resulting behaviour and scores how well it matches the prompt. Nothing about the target has to be written down in advance.
The system generalises past its training prompts, which is the part that matters. Instructions it has never seen still land on the intended behaviour more often than not — evidence that what is being learned is a mapping between semantic space and dynamical space, not a lookup table.
The current work takes this off the simulator. Xenobots — self-assembling constructs built from living cells — are the target, and the question is whether an ordinary English instruction can elicit a prescribed change in how one moves. The clip above is one such result: a biobot’s own archival trajectory, before and after being told, in plain English, to move faster, slower, or stop. Live assays are underway at Tufts. A patent application covering the general approach has been filed through the University of Vermont, and a manuscript describing the method is under review — so the mechanism itself stays undescribed here until that process is done.
Simulated tissue engineering
Did the sentence do the work?
A network can score well while ignoring its own prompt entirely — only a control tells the two apart.
Every tissue state on this page is a real run of BETSE (the BioElectric Tissue Simulation Engine), Pietak and Levin’s open-source bioelectricity simulator: ion-channel and gap-junction activity solved by finite volume methods, not an illustration of what a cell collective might do.
A question runs underneath the P2I work above, and it is easy to skip past: when a system like this succeeds, did the sentence do the work? A network can score well while ignoring its input entirely. Six earlier single-cell controllers in this project did exactly that — each emitted a fixed intervention and never read its prompt — and nothing in a training curve gives it away, because a dead language channel looks like a model that is still learning.
Three controls settle it, and they have to run beside the training rather than after it: retrain on shuffled sentence-to-target pairings, feed an empty prompt at test time, and hold back paraphrases that were never trained on. Run across twenty sentences and sixty simulations of a bioelectric tissue, the trained network does read the sentence, and the evidence is stronger than a drop in score. Handed a different target’s sentence it does not degrade — it builds the opposite voltage pattern, every time, on sentence pairs that differ by one phrase.
The same three controls, run on an unrelated substrate — a single driven oscillator whose behaviour is known in closed form, so the best possible intervention can be written down rather than searched for — locate where the limit actually is, and it is not in the network. Sentence encoders place opposites close together: “go up” and “go down” sit at cosine 0.78, while “go up” and “head north”, which mean the same thing, sit at 0.32. The controller therefore picks up the axis and loses the sign. On the tissue the same bottleneck shows up a second way: the five phrasings of each request rank in exact inverse order of how much they resemble one another, worst where they share a template.
Two independent measurements on two substrates, pointing at the same component. It makes the next move a measurement rather than a guess: more training will not recover a distinction the encoder never encoded.
Embodied dynamical agents
Language-driven control of embodied agents
Does the same trick work on something that has to move through a world?
Convince-not-control extends here from a bioelectric field acting on a cluster of cells to a driven Kuramoto oscillator whose coupling and forcing terms act on a Braitenberg-style vehicle. A reactive coach re-issues its instruction as the vehicle’s own state drifts, which is closer to how the cellular version actually has to work than a single open-loop prompt is.
Skills compose zero-shot: a vehicle trained to move right and one trained to avoid a wall combine into one that does both, without retraining on the combination. Tested first on 2D maze and lava-crossing tasks, where a wrong correction ends the episode outright rather than merely costing some reward.
This is the rehearsal track for a physical Braitenberg vehicle: the simulated substrate here stands in for one where the dynamics cannot be reset between attempts.
Artificial life
Evolutionary transitions in individuality
How does a population of individuals become one individual?
VitaNova is a simulation in which agents with evolving neural controllers self-organise into flocks — and then the flocks begin to reproduce as units, dividing to found daughter flocks. Role specialisation appears inside them without being selected for directly.
The clip above is one recorded life cycle, not an illustration: a collective disperses, re-aggregates into a ring, then genuinely splits in two — the census caught this fission at step 2850 of the run (seed 7) — and the two daughter collectives are shaded separately from that point on.
The point is not that the outcome is surprising; multicellularity and eusociality both happened. The point is that natural selection and self-organisation had to work together to produce it, and the model lets you take the two apart and watch what each contributes.
Computational intelligence
Learning, evolution, and the Baldwin effect
What does an individual’s ability to learn do to the population it belongs to?
My doctoral work concerned evolving self-taught neural networks: agents that generate their own training signal rather than receiving one from the environment. Across dynamic and rugged fitness landscapes, a consistent asymmetry appears — self-teaching buys robustness when the environment moves quickly, while social learning buys speed when it is comparatively stable.
This is the Baldwin effect examined as an engineering question rather than a historical one: not whether it happened, but under what conditions learning is worth its cost to a population.
The same accounting now applies to machines, and that question has become its own line of work — see Rogers' paradox in language models, above.
Dynamical systems for markets
Phase transitions in price
The same transition question, asked of a different collective.
A market reorganises without anyone deciding that it should — which makes it, formally, the same kind of object as the collectives above. I treat it as a dynamical system with identifiable regimes, so that the question stops being “is this reading high or low” and becomes “which regime is this, and is it building or breaking”.
Working this way makes several timeframes resolvable as one object rather than several charts to reconcile, and makes the set of possible turning-point events enumerable instead of interpretive — including the transition that starts and then fails, which is the common case and the one most approaches discard as noise.
This strand is commercial rather than academic: the method is proprietary and unpublished. It is built and running as Theta.