Decision theorist, formal epistemologist, and philosopher of science and of AI.
I am an Assistant Professor of Philosophy at UNC Chapel Hill, Core Faculty for the UNC PPE Program, and co-director of ∃xProof AI with Aydin Mohseni. I work on the foundations of AI safety and alignment, on the foundations of rational agency, and on the philosophy of science more broadly.
Inquiring together
I think of inquiry as something we do together. We start in different places, keep our minds open, and let shared evidence carry us toward the truth. Bayesian epistemology is our most successful theory of inquiry, and for this reason much of my work proceeds within that framework. Below is a toy model of Bayesian inference.
Rolls
How it works
Research
AI safety and alignment
When does an AI system have beliefs and goals, and how could we measure them? When should we delegate a decision to an AI system whose values may not quite match ours? With Ben Levinstein I have written Radical AI Interpretability, a book on the conceptual foundations of interpretability, forthcoming with Cambridge University Press.
Take delegation. An AI system may know far more than we do and still want something slightly different. I develop models that help us understand when an agent’s capabilities pay for the cost of its misalignment.
How it works
Each shot stands for a decision made under uncertainty, and where it lands is how that decision turns out. The bullseye is the outcome you most want, and the rings score each outcome by your lights. You aim at the bullseye, since you want the right thing, but your shots scatter, since you can’t see the target well. The AI sees it clearly, so its shots are tight, but it aims a little off, at what it wants. After ten shots each, the higher total wins and its shots turn red. The runs are set up so that ten rounds reliably show which choice has the higher expected score: if it is the AI’s, delegating pays. How far off the AI aims varies from run to run, so sometimes you should keep the decision.
Rational agency
I study the foundations of rational agency. For example, how should a rational agent act on the evidence she has? If you know what a better-informed version of yourself would do, should you already do it? I also take notions that look metaphysically loaded, such as chance, causation, or control, and try to naturalize them by showing that they can be recovered from an agent’s own beliefs and desires.
Take chance. The triangle at the top of the page cheats a little, since its hypotheses are chances. De Finetti showed how to do without them. Credences over chances are a way of representing exchangeable credences about observables, such as the outcomes of rolls of a die, not beliefs about something extra in the world. The figure shows this with the die from the top of the page; you can switch it to a coin.
Flips
Credence over a coin’s chance of heads, and the credences about the flips it averages to.How it works
An agent is unsure of a coin’s chance of heads, θ. The curve is her credence over θ, and it sharpens as she conditions on the flips. Averaging the chances by that credence gives her credences about the flips themselves, in the bars. They share the curve’s scale, so her credence in heads next sits directly under the curve’s balance point. Those credences are exchangeable: HT and TH always tie, and so do HHT, HTH, and THH. De Finetti’s theorem runs the other way too: any exchangeable credences about an unending sequence of flips can be written as such an average.
Rolls
One agent’s credence over the three hypotheses, and the credences about the rolls it averages to.How it works
At the top is one agent’s credence over the three hypotheses from the first triangle. Each hypothesis is a die that favours one face: that face comes up half the time, and each of the others a quarter. In the middle is a triangle whose corners are the die’s faces, so a point in it is a chance distribution over a single roll. The three hypotheses are the small rings. Averaging them, weighted by her credences, gives her credence about the next roll: the dot, with its numbers at the corners. Each line from a ring to the dot is as strong as her credence in that hypothesis. With each roll she conditions, the bars shift, and the dot moves with them. At the bottom are her probabilities for three orders of the same rolls, which always tie: her credences about the rolls are exchangeable. De Finetti’s theorem runs the other way too: any exchangeable credences about an unending sequence of rolls can be written as such an average.
Show it with:
Formal philosophy of science
I use formal models to ask what justifies the methods of the sciences, including machine learning, and how scientific and social practices can emerge.
Take language. In Lewis–Skyrms signaling games, meaning emerges from repeated interaction alone, with no prior agreement about what anything means. With Jacob VanDrunen I extend these games so that agents must also learn which features of their environment carry signals in the first place. Natural salience, it turns out, can hold the evolution of signaling back.
How it works
Each run has three, four or five states, with as many signals and acts. Each round, nature picks a state (left). The sender sees it and sends a signal (middle), an arbitrary mark that means nothing to begin with. The receiver sees only the signal and picks an act (right), and the act succeeds when its shape matches the state’s, like a key fitting its lock. Neither starts with any conventions. Each time an act succeeds, both become a little more likely to do what they just did. The thickness of each strand shows how likely that choice is. A dot traces one round in many. When the act succeeds, it swells and the path just taken flashes, since both choices on it are reinforced; when it fails, the act shrinks and shakes. Most rounds are played offscreen. Most runs reach a signaling system, in which each state gets its own signal and the receiver reads it correctly, and then the system lights red. Other runs settle into partial pooling, where two states share a signal, and fade. The more states there are, the more often this happens: about one run in ten with three states, one in four with four, and nearly half with five.
How it works
This is the attention game from Sifting the Signal from the Noise. Each round, the sender sees one of two states (left) and makes one of two marks, which sets one feature (middle). The other two features are set by coin flips (the small coins beside them), so they have nothing to do with the state. The receiver does not know which feature carries the signal. She picks one to look at, reads it, and acts (right), and her act succeeds when its shape matches the state’s. After each success, both become a little more likely to do what they just did, including where she looked: the path just taken flashes and the act swells. After a failure the act shrinks and shakes, and nothing changes. So the strands (how likely each path is) and the outlines (how often she looks at each feature) shift toward the signal. Once she looks at the signal more than at anything else and reads it correctly, the system lights red. In about one run in ten, learning freezes first, usually with the receiver fixed on a noise feature, and the run fades.
Show it with:
Recent and upcoming talks
- Radical AI InterpretabilityMinds, Meaning, and Machines, Hong Kong, November 2026
- Margins of Misalignment: Delegation to AI Beyond the Bayesian IdealizationPPE Society Annual Meeting, New Orleans, November 2026
- Why Are the Social Sciences Hard?PPE Society Annual Meeting, New Orleans, November 2026
- A Bayesian Reduction of CausationPhilosophy of Science Association Biennial, San Diego, November 2026
- Splitting the Atom: Atomless Semantics for CounterfactualsPhilosophy of Science Association Biennial, San Diego, November 2026
- Panel: The Language of AI Safety for ScienceAISciK Workshop, NeurIPS, Atlanta, December 2026
Sapere aude
Knowledge enormous makes a God of me.
Names, deeds, grey legends, dire events, rebellions,
Majesties, sovran voices, agonies,
Creations and destroyings, all at once
Pour into the wide hollows of my brain,
And deify me, as if some blithe wine
Or bright elixir peerless I had drunk,
And so become immortal. . .
— John Keats, Hyperion
Contact