I'm Mei, an ML research engineer at Renaissance Philanthropy. I study what language models know about their own answers before they say anything about them.
The recent result: a linear probe on a model's hidden state predicts whether its answer is right better than the model's own stated confidence does. The model knows more than it reports. Sometimes it uses what it knows. Sometimes it does not.
Before this I built brain-computer interfaces. MASc in Systems Design Engineering (Waterloo), BHSc in Health Sciences (McMaster), three papers on EEG. On the side I run Free The Agents, an agentic AI studio, and build small instruments for intuition and sound.
- Making beautiful things
- Thinking about how we think
- Producing music, finding my unknown unknowns
Research
- Knowing Before Saying
Internal correctness signals in mathematical reasoning. MATH-AI workshop, NeurIPS 2026.
- GeoGenBench
Can a language model draw a geometry diagram? A compiler checks. Talk at AIME-Con 2026.
- Internal state, behaviour, self-report
An open lab book on where the three agree and where they come apart.
Building
- The Intuition App
Intuition is a muscle. Scored drills, honest feedback, one line a day on iOS.
- Digital Singing Bowl
Trace the rim, tap to strike. Audio synthesis in the browser.
- Free The Agents
Agentic systems for pharma, finance, and clinical research.
Writing
- Digital Witchcraft
We used to call them witches. Now we call them engineers.
- How to Spawn Your OpenClaw
A persistent, always-on AI assistant: server, Signal, memory, routines.
Transmissions
Intermittent signals on what I'm making, thinking, and finding. No schedule, just the interesting parts.