Daniel D. Johnson


Hi! My current focus is on building tools and techniques for understanding and shaping model behavior and character, and sharing the findings and tools as part of a broader public science of model behavior. I'm especially interested in understanding the self-concepts and psychological states of language models, and using this understanding to build models that are robustly good.

Before this, I worked at Transluce, where I developed techniques for surfacing unwanted model behaviors, led the development of the fxtr library for building and iterating on bespoke model evaluations and model-powered analyses, and contributed to Transluce's Mental Health Behavior Report.

During my PhD at the University of Toronto, I studied how to get powerful models to accurately report their uncertainty, so that we know when to trust them even if we can't directly verify their reasoning. You can read more about this in my blog post on "uncertain simulators" and my paper "Experts Don't Cheat". I've also explored ways to summarize the uncertainty in generative models of code, described in my paper about the R-U-SURE system.

While at Google DeepMind, I released a JAX neural network library, Penzai, and an interactive pretty-printer, Treescope, that together make it easy to inspect, modify, and visualize parts of pretrained models. I designed Penzai and Treescope with the hope that they could lower the barrier of entry for research into understanding neural networks and steering them toward safe behaviors.

In the past, I have worked on generative models of discrete data structures (such as trees, sets, and graphs), theoretical analyses of self-supervised learning, a strongly-typed language (Dex) for building unconventional machine learning models, generative models for music, and many others. See my research page for more information.

I worked at Google Brain and Google DeepMind from 2019 to 2024, first as an AI Resident and later as a Research Scientist. Before that, I worked on applied machine learning at Cruise from 2018 to 2019. I received my Bachelor's degree from Harvey Mudd College in 2018 as a CS/Math joint major.

In my free time, I enjoy playing indie video games! Current recommendations include Outer Wilds, Baba is You, Monomoon, A Monster's Expedition, and Tunic.