Matteo He

I want to help make increasingly capable AI systems safe, and I’m especially interested in understanding and preventing misalignment that emerges and generalises when models act autonomously in novel situations. I’m open to projects across AI safety, from interpretability to new alignment methods to evaluation and oversight.

I completed an MPhil in Advanced Computer Science at Cambridge with Distinction. My dissertation studied how language models learn to turn internal information into predictions. The resulting first-author paper, Learning to Read Out, was accepted to NeurIPS 2026.

I’m seeking AI safety research fellowships and research scientist or research engineer roles. Get in touch →

Recent updates

  1. Learning to Read Out released on arXiv.

  2. Learning to Read Out accepted at NeurIPS 2026.

  3. Sparse Readout Prism preprint released on arXiv.

  4. Completed the MPhil in Advanced Computer Science at Cambridge with Distinction.

Selected research

Learning to Read Out

Unembedding Dynamics in Language Model Pretraining

Matteo He, William F. Shen, Alex Iacob, Andrej Jovanovic, Xinchi Qiu, Nicholas D. Lane

First authorNeurIPS 2026

A model’s outputs can lag behind what it already represents.

Evaluations usually judge a model by what it says. But during training, everything it says passes through a readout that is still being learned, and that is the confound we identify: it’s like reading a ruler while the markings are still being drawn. Probing hidden states directly, we found task information there before the model’s answers showed it, and on harder tasks the gap never fully closed. If you’re watching outputs for early warning signs, you may find out late, or not at all.

Sparse Readout Prism

Explaining Logit-Lens Scores in Features Instead of Tokens

Matteo He, William F. Shen, Xinchi Qiu, Nicholas D. Lane

First authorPreprint · Under review

A lens reading tells you about the lens as well as the model.

Lenses let you read a model’s hidden states partway through, as if it had to name the next token right there, and people use them to ask things like whether a model works in English. The confound we identify is that a lens has opinions of its own: two lenses fitted on different text can read the same hidden state as different tokens. Sparse Readout Prism is our attempt at a fairer reading. It breaks any token’s score into features of the model’s own output weights, and is upfront about how much it can’t explain.

Education & distinctions

University of Cambridge

MPhil, Advanced Computer Science

Distinction · AI Alignment Fellowship

University of St Andrews

BSc (Hons), Computer Science & Mathematics

First Class · 18/20 average · Dean’s List in all three years

Top Student Medal for the highest academic achievement in the Direct Entry Computer Science cohort.

GRE: 340/340 · taken in September 2021, before I started university

Experience

Amazon Alexa AI

Software Development Engineer Intern

On the Alexa NLU team I built ML data and experiment infrastructure for model evaluation, used by 20+ applied scientists to iterate on datasets faster.

University of St Andrews

Research Intern · STARIS

I automated data collection for research on energy-aware algorithm selection, gathering hundreds of hours of solver measurements, then analysed the data and identified the features that most improved prediction accuracy.

University of St Andrews

Research Intern · Vertically Integrated Projects

I built ML classifiers to detect dolphins in acoustic recordings, combining signal processing with neural networks, and refactored the project’s Python research codebase so it ran across platforms.

Interactive explainers

An Atlas of Loss Surfaces

I wanted to properly understand the optimisation literature, so I learnt it by building this. It covers forty results, from saddle points and the edge of stability to mode connectivity and the safety basin. Each one is a 3D landscape you can spin round while a dozen optimisers roll across it. Zoom in and the particles become stick men, who scratch their heads at saddles and sit round a campfire once they settle. I hope the pictures help someone else build the same intuition.

Dusk at a rest point: campers sit by the fire, and one lies back to point at the stars.

Projects

vigil-gpu

Open source · Python

I built a terminal monitor for rented cloud GPUs: every instance’s training logs and metrics on one screen, with alerts when a run stalls, hits NaNs or plateaus.

Four rented GPU instances in one terminal, each with a live loss sparkline and training log.

Japanese Grammar Analyzer

I built an app for learning Japanese from real text. Paste in anything you’re reading and it marks each word’s grammar and level, with furigana over the kanji. Translation runs on a language model inside your browser, so it costs nothing to serve.

Pasting in the opening of Sōseki’s I Am a Cat, switching highlights and clicking 猫 (cat).

3D Game of Life

I took Conway’s Game of Life into three and four dimensions. A 3D cell has 26 neighbours rather than 8, so the hard part is finding rules that work: you can tune them by hand, or let a rule finder try hundreds and keep the ones that stay alive.

A fractal crystal growing from a single cell, with older cells turning amber, then rust.

Soodokoo

I built Sudoku in up to six dimensions. Every number appears once along each row, column and line through the layers, and you play one 2D slice at a time, sliding through the rest.

A number entered on the flat board, then the same cell in Cube view, where its four lines run through five cubes, one per world.

Art

Matrix Digital Rain

I turned the falling code from The Matrix into a corridor you fly through, steering with the mouse or a finger. The space bar warps it into a tunnel.

Flying down the corridor grown from the seed “matteo”, with a warp into the tunnel.

Matrix Cave

I lit a generated cave with nothing but the falling code: each string is water running over the rock, so you see the walls only where it traces them. Every sound, down to the muffled drips, is synthesised in the page.

A pulse outlines the tunnel ahead, then a walk down it past rock lit only by the falling code.

Contact

For enquiries about research roles, fellowships, or collaborations, email me.

matteohe.research@gmail.com

I’m also on LinkedIn.