SUCHIR SALHAN
I study how machines learn language — and build systems that make AI more understandable, efficient and accessible.
I build the instruments that make model learning observable, then carry what they reveal into products, policy and teaching. Alongside the research I founded Per Capita Media.
Researcher · Builder · Writer
Let's connect
Working on something at the intersection of AI, language, learning, policy or media? Whether you want to collaborate, build, or just compare notes — a message is always welcome.
Research
Understanding how language models learn
Language models are the most powerful learning systems we've built — and we still barely understand how they learn. I build controlled experiments to watch learning unfold: how grammar emerges, how knowledge is represented, and why independently-trained models converge on similar internal structure.
Featured · my instrument
Beetle
Language models as model organisms of human language learning.
Beetle replays learning under controlled conditions. Change one language, one tokenizer, one curriculum — and measure what changes, and what doesn't.
114 bilingual models · 13 first languages · 5 curricula
Open the Beetle page →
Build
Smaller, faster AI you can actually own
Most companies don't need a bigger model — they need one they understand, that adapts quickly, runs cheaply, and they actually own. My research turns into practical tools for exactly that.
Model merging
Combining independently-trained models into one — near-term and practical.
Small LM agents
Efficient, deployable agents built on models you can run yourself.
Multilinguality
AI that works beyond the richest few languages.
Interested in building something? Let's talk →
Ideas & media
AI beyond the lab
AI shouldn't only work for the richest languages, companies or countries. Open, interpretable models make AI something universities, governments and educators can build — not just buy. Research creates ideas; writing tests them; policy puts them to work. That's why I founded Per Capita Media.
Outside research
Research has taken me further than I expected — Seoul, Vancouver, San Diego, Vienna, Morocco. Cambridge still feels like home, but language models have become an unexpected passport.
Read more about me — the full story →
SUCHIR SALHAN
A science of how machines learn language · Build · Govern · Teach · Per Capita Media
University of Cambridge · Gonville & Caius College · sas245@cam.ac.uk