Hi, I'm Jayde, an AI researcher at NTU CCDS. My path runs through three worlds — big tech, then a startup of my own, and now back in research — but one thread holds across all of it: I'm curious about people, and skeptical that machines really understand them.
It started at ByteDance, where I first saw up close how a machine "understands" you — as a bundle of labels, plus a guess drawn from the average of people like you. Clever, useful — but that's the crowd you resemble, not you. Can a machine go further, and understand a specific person?
So I work on evaluation. The targets I care about have no answer key — does a model understand this person, is it actually aligned, should it speak right now? The field crowds the layer that's easy to fake: the group, the aggregate, what people generally do, where models look fine. The real test is where you can't fake it — a specific individual, or the causal question: change one thing, and does this person shift the way a real one would? There, models collapse, and a bigger, pricier model doesn't fix it. That's where I build the ruler.
Human Mirror — the ruler for the level you can't fake: a specific individual, and the causal. Group simulation already has a standardized benchmark (SimBench), and individual survey answers can be reproduced (Stanford's 1,000-agent study) — but the causal layer still has no equivalent, and that's the gap Human Mirror fills.
Early results are stark: a model reproduces a group's behavior well, but reproducing one specific individual comes back close to chance. The sharpest tell: on weak-signal behavior, a stranger's history can predict the person as well as their own — the model was leaning on the crowd, not them. To be clear, the goal isn't a better model of any individual; it's an external check that catches when a simulator only looks like it has one.
So I measure fidelity where it can't be faked — anchored on the person's own behavior, normalized against their own self-consistency, causal part grounded on real randomized interventions — and I report where a simulator is unfaithful rather than certify that it "understood." A paper is in progress.
Two more in the same spirit, earlier-stage:
Flatland (in design) — meaning lives beneath the literal: "you look great" is a compliment or a jab depending on context. Models tend to flatten that into the safest literal reading; Flatland is built to catch where they do — in real context, not a quiz.
Tact (early) — when an AI should speak in a live interaction, when it shouldn't, and when the smallest move beats saying anything.
Before research, I founded MosuMosu — a core full-time team of about ten at its peak, plus a larger group of volunteers who did the emotion-tagged text annotation — and we shipped two versions:
MosuMosu v1 — a vertical fan-fiction search engine for Japanese readers: semantic search over a corpus of Chinese, English, and Japanese fan works, so one query reaches across all three, with a rewritten Japanese summary for each result — a Perplexity for an entertainment niche. The hard part: meaning here is fandom-specific — character names, ship tags, and in-group slang don't survive literal translation across the three languages, so off-the-shelf multilingual search just retrieves the wrong things.
MosuMosu v2 — a Copilot for fans of mid-tier Japanese celebrities: for each, an AI-assembled channel and an assistant giving fans real-time updates and deeper Q&A, pulled from Fandom, Wikipedia, Twitter, and fan circles. The hard part: modeling each fan — continuously inferring what about a celebrity that fan cares about, and refreshing their profile (down to their location) so it can push the events they'd actually want, before they go looking.
Running it took me deep into anime/comics/games/novels (ACGN) fan subcultures across Tokyo, China, and the US — and taught me that a community doesn't only have meanings, it has boundaries: to many fan-fiction authors, being easier to find isn't opportunity, it's the unease of being exposed. And here's the throughline to now: getting that per-fan model right was the hardest, least-verifiable thing we did — so I stopped trying to build the model and started measuring whether anyone can.