Jesse Hoogland

Executive Director, Timaeus

Berkeley, CA, USA

[email protected] • +1 415 424 0316

Website • GitHub • LessWrong • Twitter • LinkedIn

Summary

I'm an AI-safety researcher and one of the founders of the Singular Learning Theory (SLT) for AI safety research agenda. This led to me co-founding and now directing Timaeus, a nonprofit research organization studying this agenda. Together with my team, we've scaled the organization from 3 to 16 staff and raised $3.5M+ in grants. In addition to co-authoring papers directly, I'm an experienced communicator with 20+ talks at frontier AI labs and academic venues.

Core Skills

  • Research: Singular Learning Theory • Interpretability & Alignment • Statistical Physics
  • Operations: Non-profit leadership • Fundraising • Research management
  • Communication: Technical writing • Public speaking • Data visualization
  • Technical: PyTorch • TPUs • MLOps (HuggingFace, Wandb) • DevOps (AWS, GCP, Docker, K8s)

Current Role

Timaeus — Executive Director

Berkeley, CA | Jul 2023 - Present

I run Timaeus, a nonprofit AI safety research organization working on singular learning theory (SLT) for alignment. SLT establishes a connection between the geometry of the loss landscape and internal structure in models, which we are using to develop scalable, rigorous tools for evaluating, interpreting, and aligning neural networks.

My main responsibilities include:

  • Outreach: My primary responsibility is making sure our research reaches external stakeholders (AI safety researchers, scaling labs, funders, etc.). In practice, this takes the form of (see below) writing distillations and explainers, giving regular talks, and conducting targeted outreach to specific individuals.
  • Research: I work closely alongside our director of research, Daniel Murfet, on developing and implementing our research agenda. For a full list of our papers, see here.
  • Operations: I work closely alongside our director of operations, Stan van Wingderden, on operations, recruitment, and fundraising. Cumulatively, we've raised ~$3.5mm via the SFF, Manifund, LTFF, Open Phil, and AISTOF. We've scaled the team from 3 FTE to 16 FTE as of July 2025.

History of Timaeus

Timaeus was founded in October 2023 after we raised funding from Manifund and later SFF.

Our initial focus was validating basic scientific predictions made by SLT (Do phase transitions really exist in models trained by SGD? Can we really learn new things by studying development? Etc.). This was a basic prerequisite to the rest of the agenda (applying this understanding to safety). As we describe below in "Timaeus in 2024", this effort was successful.

Subsequently, we turned to the next major bottleneck: scaling our techniques to larger models. As of early 2025, we successfully reached the 10B parameter range. We will release our first safety-relevant results from this work in mid-2025.

Past updates:
  • Timaeus in 2024
    February 2025 • Jesse Hoogland, Stan van Wingerden, Alexander Gietelink Oldenziel, Daniel Murfet
  • Timaeus's First Four Months
    February 2024 • Jesse Hoogland, Stan van Wingerden, Alexander Gietelink Oldenziel, Daniel Murfet
  • Announcing Timaeus
    October 2023 • Jesse Hoogland, Daniel Murfet, Stan van Wingerden, Alexander Gietelink Oldenziel

Events

I was introduced to singular learning theory (SLT) and the rest of Timaeus's founding team by my co-founder, Alexander Gietelink Oldenziel. Alexander proposed the idea of an SLT for Alignment conference, which began the first time where the founding team worked together and met in person.

  • ILIAD 2024
    August 2024

    At ILIAD, I was in charge of organizing and coordinating the SLT track, and I also gave two talks.

  • The 2023 Oxford Conference
    November 2023

    I co-organized a one-week summit (November 2023 at Wytham Abbey in Oxford) on developmental interpretability. Videos available here. My responsibilities were similar to the previous conference (fundraising, logistics, and content). This conference’s main impact was to accelerate our initial round of papers and disseminate intermediate progress.

  • The 2023 Berkeley Conference
    June 2023

    I co-organized a two-week summit (June 19th-July 2nd 2023) on singular learning theory and its applications to AI safety together with the rest of the team that would go on to found Timaeus. Together with Alexander Gietelink Oldenziel, we raised on the order of $40k. Together with Daniel Murfet, I prepared a curriculum, giving six talks myself (see below, videos available here). Together with Stan van Wingerden, I helped arrange the logistics of the conference, including the venue, the food, and the schedule. We ultimately brought together around 40 people in person and more than 150 people virtually to learn about singular learning theory and its applications to alignment. This directly culminated in us proposing the “developmental interpretability” research agenda, our initial funding, and the founding of Timaeus.

Prior Experience

SERI MATS 3.0 & 3.1 — Scholar

Amsterdam, NL; Berkeley, USA | Oct 2022 - Jul 2022

I was in Evan Hubinger's Deceptive Alignment track. This is when I first began investigating developmental interpretability and then later SLT, which I've been working on ever since. I co-organized the first SLT and Alignment conference, which led to the subsequent development of the SLT for Alignment research agenda and foundation of Timaeus.

During this time, I also:

  • Worked as a Research Assistant at David Krueger's group at the University of Cambridge with Lauro Langosco and Xander Davies on studying the links between grokking and (epoch-wise) double descent.
  • Contributed to the "Single-Agent Control" chapter for Dan Hendrycks's textbook at the Center for AI Safety. I wrote most of the chapter and coordinated its completion with other writers.

FTX Future Fund — Grant Recipient

Amsterdam, NL | Oct 2022 - Mar 2023

I received a grant from the FTX future fund to bridge my transition to AI safety research. I worked through several textbooks (e.g., Mathematics for Machine Learning, Pattern Recognition & Machine Learning, Artificial Intelligence, Reinforcement Learning) and courses (ARENA virtual).

Health Curious — CTO

Amsterdam, NL; SF, USA; Brasília, BR | Mar 2021 - Aug 2022

In 2021, as I was finishing my Masters, I started a company with my now wife, Robin Laird, to help small-and-independent healthcare providers build their own virtual care programs. We focused on bariatric surgery, the leading intervention for obesity. We decided to close Health Curious after a year and a half, when I realized I found the problem uninspiring, the work a poor personal fit, and AI too important to ignore.

Bit — Software Developer & Coach

Amsterdam, NL | Mar 2020 - Oct 2022

During my Bachelors and Masters, I worked part-time at Bit, an applied consulting studio. It's a place companies go to quickly prototype new products/solutions with fresh young talent. I worked on two projects with SURF as well as coaching high school students and teaching bootcampers in Bit's educational offshoot, the Bit Academy.

Education

University of Amsterdam — Masters of Physics (Theoretical Physics)

Amsterdam, NL | Sep 2019 - Aug 2021

I focused on dynamical systems theory, and conducted my research under the supervision of Dr. Greg Stephens whose group studies the physics of animal behavior.

GPA: 4.0

Thesis: The Ergodic Theory of Random Neural Networks (8.5/10)

Amsterdam University College — Bachelor of Science (Physics & CS)

Amsterdam, NL | Sep 2016 - Jun 2019

I majored in theoretical physics and computer science and graduated as salutatorian (defined as among the top-10 students of my class). My thesis was my program's exclusive nomination for the VU thesis prize (out of a class of 300+ students).

GPA: 4.0

Thesis: Restricted Boltzmann Machines and the Renormalization Group: Learning Relevant Information in Statistical Physics (9.7/10)

Fox Lane High School — High School Diploma

Bedford, NY | Sep 2012 - Jun 2016

I graduated as salutatorian, with the second-highest GPA out of a class of roughly 400 students.

GPA: 4.0

Research

I work on and help direct the Singular learning theory (SLT) for Alignment research agenda, which was developed by Daniel Murfet. As part of this agenda, I helped establish Developmental Interpretability, an approach to interpretability grounded in singular learning theory and studying changes over the course of the learning process.

Publications

Talks

Other Writing

Projects

Personal Website — JesseHoogland.com

Dec 2020 - Ongoing

I publish much of what I write to my personal website.

Data Visualization Tools

Since reading Bret Victor's What Can a Technologist Do About Climate Change?, I've been obsessed with the idea of developing better tools for embedding dynamic models in writing. I made several attempts to improve on the sorry state of affairs, starting with remark-tangle to add support to Tangle in the remark ecosystem. Second, ddmd to try to subsume Tangle altogether. Later, an unpublished library to try to subsume ddmd (in favor of Solid.js's powerful reactive model). In 2022, obsidian-squiggle to add support for the probabilistic programming language, Squiggle, in the note-taking app Obsidian.

Language-Learning Tools

I love learning languages. A lot. I built my own frontend for Anki in order to speed up flashcard creation (by pulling in content from google searches, wiktionary, Forvo, etc.). At one point, I was ready to start learning Mandarin, so I created a tone trainer for myself. At another point, I was ready to start converting Wiktionary into machine-readable semantic triples, so I could automatically pull more helpful information for my flashcards. That turned out to be pretty audacious (I haven't finished the wikitext parser yet), so it's currently on the backburner.

Note-Taking Tools

Note-taking is an addiction. My personal Obsidian vault is at about 10,000 notes. My personal Anki deck is at about 50,000 flashcards and 500,000 total reviews (encompassing an entire month of my life). I've made a few plugins / modifications (obsidian-squiggle, obsidian-export, an unpublished obsidian-export v2, fork of obsidian-linter, etc.).

Skills

Research

Interpretability, Alignment, Science of Deep Learning, Singular Learning Theory

Machine Learning

PyTorch, JAX, XLA, TPUs, Wandb, HuggingFace

Devops

Docker, K8s, AWS, GRC

Web Development

JS/TS, React, Next.js, Astro, Fast API

Communications

Writing, Lecturing, Lots of Emailing

Operations/Management

Agile/Scrum, Linear

Languages

English

Native

Dutch

Native

French

B2

Spanish

B2

Portuguese

B2

Italian

B1

Japanese

A1

Mandarin

A1

References available upon request.