Phonology at Home: Reflections on Hosting OCP23 at Caius

In mid-January 2026, Gonville & Caius College — my home for the past five and a half years — hosted the 23rd Old-World Conference in Phonology. For three days, the Old Court that has seen everything from plague to the Reformation heard talks on tonal systems in Bantu languages, the phonetics of Icelandic aspirated stops, and the information-theoretic structure of phoneme inventories. It was, by any measure, the nerdiest and most satisfying thing to happen in that courtyard in recent memory.

I joined the Organising Committee alongside the lead organiser Yury Makarov and programme chair Professor Bert Vaux. My role was primarily logistical — room bookings, catering, the usual machinery of academic conferences — but the real privilege was being close to the science.

Phonology meets computation

Fermín Moscoso del Prado Martín and I presented two oral papers. One examined what we are calling convergent equilibria in cross-lingual phoneme surprisal distributions: the observation that unrelated languages tend to settle at similar information-theoretic points in their phoneme inventories, even when their surface phonologies look quite different. The second paper zoomed out further, asking about the roles of chance, functional cost, and cross-tier integration in shaping phoneme systems globally.

These projects grew out of an extended collaboration that began when Fermín arrived at Cambridge and immediately started asking the kind of questions that make you rethink everything you thought you knew about language. The Mental Lexicon keynote in Montréal last June — which preceded OCP23 — was the first public airing of this work, and the response there sharpened the arguments considerably.

What I find genuinely exciting about this line of research is that it sits at the intersection of three traditions that rarely speak to each other: classical phonological typology, information theory in linguistics (the Zipfian tradition from Piantadosi to Coupé), and the emerging NLP literature on byte-level models and tokenisation. ByteSpan — the tokenisation work I presented at ICML in Vancouver — was partly inspired by thinking about how prediction error spikes at lexical boundaries, which is itself a story about surprisal. The threads connect in unexpected ways.

What phonology has to teach AI

Sitting in talks about allophonic variation, moraic theory, and vowel harmony, I kept thinking about a tension I've been living with in my NLP work: how much of what we think of as learned by a language model is genuinely structural, and how much is an artefact of the training distribution?

Phonologists have been worrying about a version of this question for decades. The Optimality Theory vs. rule-based debate, the nativist vs. emergentist split, the question of whether phonotactics is knowledge or just frequency — these are not so different from the mechanistic interpretability question of whether a transformer has learned a grammar or a very compressed lookup table. One of my NeurIPS CogInterp papers this year argued that theoretical linguistics should constrain the causal abstractions we use in interpretability research. OCP23 only reinforced that conviction.

On being a host

There is something strange and moving about hosting a conference in a place you know intimately. I know which rooms are cold in January (most of them), which staircases confuse visitors (all of them), and which College staff will quietly save the day when the AV cable fails (they know who they are). But I also know the Hall where I sat my first-year exams, the bar where I celebrated finishing my MEng, the garden where I sometimes go to think when a proof won't come together.

Seeing 150 phonologists navigate all of it — mostly cheerfully, occasionally with mild bewilderment — was a particular kind of pleasure. Cambridge at its best is a place where serious intellectual work and the ordinary pleasures of college life coexist without apology. OCP23 felt like that.

More phonology soon. Stay tuned for the papers.