Observing intelligence as it forms — a live look at what I’m training, reading, wondering, and where the work has taken me.
How a second language is introduced during training matters more than model size for human-like reading
Online processing and grammar come apart — staged exposure → reading-time alignment, balanced → grammaticality
Tokenisation is representation learning, not preprocessing
Independently-trained models converge on compatible representations — mergeability is emergent, not a hack
Curriculum benefit scales with typological closeness to the target language
Small, human-scale models can rival 8B+ LLMs on cognitively-meaningful tasks
We can “rewind” a model’s childhood and read off developmental stages — still unconvinced
An MRI scanner for language models
114 bilingual models · 13 first languages · 5 exposure curricula · ~30 checkpoints each
Most people alive speak at least two languages — yet most cognitive science studies monolinguals.
Pico + BabyLM — two Outstanding Paper Awards
CogInterp — two position papers
ByteSpan, Tokenization Workshop
Measuring Grammatical Diversity
BabyBabelLM + MME
Culture × AI
Invited keynote — phoneme distributions
Phonology — organising committee
Departmental colloquium
Poster
Invited seminar
Invited seminar