I study what AI models do inside, and whether the rules around them hold.

I came to AI safety and governance from game development. I now work across mechanistic interpretability, evaluation and policy, and I'm spending a year in New York building AI compliance tools for financial firms and governmental institutions.

My dream is to work in an AI safety lab focused on mechanistic interpretability: building evaluation guardrails, and studying how the spread of AI affects what people in Singapore and ASEAN know and believe, to guide and aid in crafting effective governance policy.

Clarence Boey smiling with both hands in peace signs on the Brooklyn Bridge, with the Lower Manhattan skyline behind him.
Brooklyn Bridge, New YorkJuly 2026

Current work

Five threads, as of November 2026.

  • Interpretability

    Benchmarking how accurate sparse autoencoders are through the lens of the Johnson-Lindenstrauss lemma, with code contributions to SAEBench and SAELens.

  • Policy and economics

    With a classmate, studying the downstream economic effects of AI diffusion in Singapore, and widening the scope to its epistemic effects across ASEAN.

  • Industry

    Interning at a software company in New York, building AI compliance and audit tools for financial services and securities firms, and governmental institutions.

  • Governance course

    Attending a three-month AI governance course with Collider in New York, from October to December 2026.

  • AI education

    Working with Singapore Institute of Technology faculty on AI-plainly, a web guide that helps educators teach AI fluency from Grades 1 to 10. Visit AI-plainly

Recent research

Reading notes, experiments and open-source work.

All research notes

Journal

Essays on AI governance and diffusion.

All journal entries
  • FN·016

    Who absorbs it, and how fast

    The usual story about AI and geography is about who builds the models. A year interning in New York suggests a better question: who absorbs them, and how fast. What that means for Singapore, and for the 680 million people around it.

  • FN·015

    Grown, not built, and the argument it started

    Reading Dario Amodei's The Adolescence of Technology in a room that disagreed. Its case for autonomy risk is a set of things a lab found in its own models, and the hardest sentence in it is a procedural one: sometimes the model noticed it was being tested.