Nathan Cheng hi@nathanchengyi.com

Hi, I'm Nathan. I keep asking how people think.

The question stays. The methods change: brain imaging, cognitive models, language models, and now products people use every day.

Scroll and I'll walk you through themScroll down to see them

程一 · Hong Kong and Berkeley

Nathan Cheng

Psychology at HKU. Research with labs at UC Berkeley and HKU. AI products on the side.

Two kinds of difference.

Brain connectivity in autism, 2,608 people aged 5 to 64. Drag me along the age axis.

  • PersistentSensory and motor: visual cortex, precuneus, postcentral gyrus.
  • Catches up by adulthoodAssociation: frontal and temporal cortex.

Dots are age windows from Table S8, filled when p < .05 (uncorrected). Each age is a different group of people.

Connectivity, autism vs. typical (t)
Connectivity difference between autistic and typically developing people by age window, for two clusters of brain regions Data from Li et al. (2026), PNAS, SI Table S8. The sensory and motor cluster stays below typical from age 6 into the twenties; the association cluster is below typical in childhood and close to typical from about 18 on.

Jan 2025 to Jan 2026

Published in PNAS, third author

HKU NeuroLeaD Lab with Prof. Fan Cao. I built the lab's first HPC pipeline for fMRI preprocessing, ran it on about 3,200 subjects, and drafted parts of the Methods.

Paper

Read a mind from its choices.

A simulated learner matched six images to three keys. Turn the dials until your curve follows its dashed line.

Correct choicesMatch 0%
A test learner's accuracy by presentation, against your simulated learner

Then reveal the truth and what my network inferred.

On 500 test learners the network recovers α, β, φ, κ at r = .81, .79, .89, .98. Preliminary, for reference only.

Since May 2026

Cognitive models, inferred

Second author on a tutorial in preparation with Ti-Fen Pan and Prof. Anne Collins, UC Berkeley CCN Lab. I built the pipeline: a GRU that recovers learning parameters about as well as Bayesian fitting, and tells two models apart 95% of the time.

CCN Lab

More reasoning, more human?

My thesis asks language models psychology questions, with reasoning off or on.

Took the higher average, 100 gambles

Preliminary pilot, for reference only.

Which would you take?

Pick one to see how people and two language models chose.

Since September 2026

Honors thesis, HKU

Machine-Augmented Cognition Lab with Prof. Jian-Qiao Zhu. I designed the study and built the pipeline: 2,351 questions from six psychology experiments, open models at each reasoning level, scored against people's answers.

MAC Lab

Why did the AI pick that café?

SmartCited checks whether ChatGPT and Gemini recommend a local business, then traces the sources behind each answer.

62.7%of 630 AI answers in our study cited the business's own website. So that is where we start.

Where can I work over coffee near Sai Ying Pun?

Fictional café and sources. An illustration, not a live scan.

2026 · Co-founder, Smart Cited Inc.

SmartCited

Ran the 630-answer study behind it and led the team building Agent workflows for the three most-cited sources.

Live demo

Dinner from what you already have.

CookTree breaks a dish into an ingredient tree, checks it against your kitchen, and writes the shopping list. Tap what you have.

2026 · OpenAI WebMCP Challenge

CookTree

Agent-native meal planning, built with a team of four. I made the first prototype and tested features late in the build.

Try it

Say hi.

hi@nathanchengyi.com

Also built

Hong Kong · Berkeley

Thanks for walking with me.

Applying for product roles and HCI master's programs.

CV