Welcome
I am a philosopher of mind and language working on artificial cognition and normativity. As AI models are increasingly said to understand, believe, reason, and hallucinate, I draw on empirical work in mechanistic interpretability and linguistics to make sense of this extension of cognitive language to AI: I run interpretability experiments on large language models and corpus studies of how language changes under their influence. I am an Ambizione Fellow at the University of Bern and incoming Swiss National Science Foundation Professor at the University of Zurich.
The thread running through my work is the evolution of thought and talk: how concepts expressed by near-universal words first arose, what fits ways of thinking and talking to their circumstances, and how the vocabulary of mind adapts to systems that do what minds do by very different means. In the philosophy of language, I study how cognitive words become conventional for AI, and whether the inferences they invite get calibrated to what the models actually do. In the philosophy of mind, I examine which forms of understanding, concepts, and world models emerge inside neural networks. And in my work on normativity, I investigate the extent to which LLMs are sensitive to different types of normative considerations, and I test whether a model’s choices are guided by norms or merely conform to them.
Two companion websites let you explore this research interactively. Normativity in LLMs maps over a thousand normative concepts inside two language models and lets you watch them at work as the models reach snap judgements and deliberate. TaxiGPT lets you ride along with a transformer trained only on sequences of moves through Manhattan, and look inside it as it tracks its position and steers by an internal compass.
This work is deeply informed by the history of philosophy. I have reconstructed how early modern philosophers from Ramus and Keckermann to Leibniz and Kant came to treat the systematicity of thought as a cognitive ideal, and how Hume and Nietzsche, Wittgenstein, and Williams explained the emergence and evolution of our concepts and norms by tracing them to the needs they serve. Out of that tradition, I developed an account of how concepts can be evaluated, engineered, and adapted to our needs, which now guides my work on the vocabulary of mind. History and AI research meet in the wheel above: Llullian concept wheels inspired Leibniz’s calculus of concepts, which in turn inspired McCulloch and Pitts’s artificial neural networks. I see the emerging field of mechanistic interpretability as continuous with the early modern project of mechanistically reverse-engineering cognition, and I seek to develop insights from the philosophical tradition through interpretability work of my own. I am a member of the Digital Society Initiative and a founder of the Philosophy with Mechanistic Interpretability Network.
I have published three books with Oxford University Press, and my articles have appeared in Mind, Philosophy and Phenomenological Research, Philosophical Studies, Philosophers' Imprint, The Philosophical Quarterly, European Journal of Philosophy, The Monist, Archiv für Geschichte der Philosophie, British Journal for the History of Philosophy, Synthese, Analysis, The Journal of Political Philosophy, Minds and Machines, and Philosophy & Technology. In 2022, I received the Amerbach Prize of the University of Basel and the Lauener Prize for Up-and-Coming Philosophers.
I have organized my articles by whether they would be most of interest to someone coming from the theoretical, practical, or historical side of philosophy. To facilitate navigation by themes, I also built a system of tags and a search interface. The full site is available in English, French, or German. My last name is pronounced with a silent ‘z’ [ ˈkəloː ]. To contact me, you can send me an email. To be notified of new publications, follow me on Google Scholar.