Tristan Hume

Tristan Hume is a computer scientist and engineering team lead whose work spans computer vision, low-latency systems, language-model alignment and interpretability. He researches hands-free input at Waterloo, completes a computer science degree there in 2019, then builds computer-vision software at Cruise and performance tools at Jane Street. At Anthropic, he coauthors studies on helpful and harmless assistants and sparse autoencoders, including research on Claude 3 Sonnet. As of October 2026, he leads Anthropic's performance optimization team; a take-home test he redesigns changes as Claude Opus models match top applicants and helps the company hire dozens of performance engineers.

Now at: Anthropic
Badges: Bridge
Links: Website, X and GitHub

Timeline

2016

Undergraduate research at Waterloo

In Winter 2016, he works as an undergraduate researcher in the UWaterloo HCI Lab. He develops high-accuracy, low-latency audio-recognition algorithms that turn mouth noises such as lip popping into actions such as clicking, and builds a hands-free mouse that combines eye and head tracking.
2018

Computer vision at Cruise

In Winter 2018, he works as a computer vision intern at Cruise. He builds an internal TensorFlow training system and integrates a machine-learning model for detecting lens flares from rain on camera lenses.
April 2019

A Waterloo computer science degree

He completes a bachelor's degree in computer science at the University of Waterloo in April 2019.
2019–2022

Performance tools at Jane Street

From August 2019 to January 2022, he works as a developer at Jane Street, focusing on system latencies and high-performance tracing. In October 2019, he writes about measuring nanosecond-scale latencies at work; in January 2022, he publishes magic-trace, a tool he designs and builds that Jane Street open-sources to capture and visualize program traces.
2022

RLHF and Constitutional AI

In April 2022, he coauthors an Anthropic study that applies preference modeling and reinforcement learning from human feedback to fine-tune language models as helpful and harmless assistants. In December 2022, he coauthors Constitutional AI, which trains a harmless assistant through self-critique, revisions and AI feedback guided by written principles.
October 2023

Interpretable features in transformers

In October 2023, he coauthors Towards Monosemanticity, which uses a sparse autoencoder to extract interpretable features from a one-layer transformer.
May 2024

Scaling sparse autoencoders

In May 2024, he coauthors Scaling Monosemanticity, which scales sparse autoencoders to Claude 3 Sonnet and extracts interpretable features from Anthropic's production model.
As of October 2026

Leading performance optimization

In January 2026, he publishes a design history of the team's take-home after Claude Opus 4 outperforms most applicants under its time limit and Opus 4.5 matches the strongest candidates. The redesigned test helps Anthropic hire dozens of performance engineers. As of October 2026, he leads Anthropic's performance optimization team.