David Duvenaud

wiki
, , , ,
1

Canadian probabilist and deep learning researcher at the University of Toronto whose Neural ODE paper established continuous-depth networks as a new class and whose Autograd library seeded the automatic differentiation ecosystem that later produced JAX — and who has since pivoted to AGI governance, sabotage evaluations, and the systemic risk of gradual human disempowerment.

Profile
Nationality Canadian
Current Institution(s) University of Toronto — Associate Professor, CS and Statistics; Schwartz Reisman Chair in Technology and Society; Vector Institute (Founding Member)
Research Areas AGI Governance, AI Safety, Catastrophic Risk Evaluation, Deep Probabilistic Models, Neural ODEs, Generative Models, Automatic Differentiation
Doctoral Advisor Carl Rasmussen; Zoubin Ghahramani
Doctoral Thesis Automatic Model Construction with Gaussian Processes (University of Cambridge, 2014)
Website cs.toronto.edu/~duvenaud
X / Twitter @DavidDuvenaud
GitHub duvenaud
Google Scholar David Duvenaud

Overview

David Duvenaud is a Canadian machine learning researcher and associate professor at the University of Toronto, where he holds the Schwartz Reisman Chair in Technology and Society and is a founding member of the Vector Institute. He completed his PhD at Cambridge under Carl Rasmussen and Zoubin Ghahramani, postdoc at Harvard with Ryan Adams, and has been at Toronto since 2016. His career follows a striking three-phase arc: from Gaussian process theory and probabilistic modeling at Cambridge; to continuous-depth neural networks, automatic differentiation, and deep generative models at Toronto (culminating in the NeurIPS 2018 Best Paper Award for Neural ODEs and foundational contributions to Autograd, the precursor to JAX); to an almost complete pivot, following an extended sabbatical on Anthropic’s Alignment Science team, toward AGI governance, frontier model evaluation, and the structural risk that incremental AI deployment could gradually disempower humanity. He continues to advise PhD students and teach at Toronto, currently focused entirely on the alignment and safety agenda he took up at Anthropic.

Early Life & Education

University of British Columbia — M.Sc. (2010)

Duvenaud completed a Master of Science at UBC under Kevin Murphy, working on multiscale conditional random fields for semi-supervised image labeling. His M.Sc. thesis established his interest in probabilistic graphical models and latent variable structure.

University of Cambridge — PhD (2010–2014)

At Cambridge’s Machine Learning Group, Duvenaud worked with Carl Rasmussen and Zoubin Ghahramani — two of the central figures in Gaussian process research — on the problem of automatic model construction. His dissertation, Automatic Model Construction with Gaussian Processes, introduced a grammar over GP kernels that a program could search to identify statistically appropriate structure in a dataset, producing human-readable reports describing the fitted model. The project was a prototype for what became the Automatic Statistician research agenda. Its key components — the compositional kernel search paper (ICML 2013), the automatic natural-language description of nonparametric regression (AAAI 2014), and the additive GP and deep GP work — remain widely cited. During the PhD he co-organized the Probabilistic Numerics workshop series, connecting numerical analysis to Bayesian inference.

Harvard HIPS Group — Postdoc (2014–2016)

After Cambridge, Duvenaud joined the Harvard Intelligent Probabilistic Systems (HIPS) group as a postdoc with Ryan Adams. This period produced several lasting contributions. The first was Autograd — developed primarily by Dougal Maclaurin with Duvenaud and Matthew Johnson — a library for reverse-mode automatic differentiation of native Python and NumPy, supporting loops, conditionals, closures, and higher-order derivatives. Autograd was the direct precursor to JAX, Google’s current primary ML research infrastructure, and introduced the design philosophy of differentiating through arbitrary Python code that now pervades ML software. The second major Harvard contribution was the molecular fingerprint graph convolution paper (NeurIPS 2015, with Duvenaud as first author alongside Maclaurin, Gómez-Bombarelli, Adams, and Aspuru-Guzik), which introduced the first end-to-end learned convolutional feature extractor operating directly on molecular graphs — one of the foundational papers in graph neural networks for chemistry. The third was gradient-based hyperparameter optimization through reversible learning (ICML 2015, with Maclaurin and Adams), enabling exact gradient computation of validation loss with respect to thousands of hyperparameters by differentiating through the entire training procedure.

Career

University of Toronto — Associate Professor (2016–present)

Duvenaud joined Toronto’s Department of Computer Science in 2016 and is cross-appointed to the Department of Statistical Sciences. He is a founding member of the Vector Institute for Artificial Intelligence and holds the Schwartz Reisman Chair in Technology and Society. He has received a Sloan Research Fellowship. His Toronto career divides into two distinct research programs.

Deep probabilistic models and continuous-depth networks (2016–2021). The dominant output of Duvenaud’s early Toronto period was a set of papers developing the intersection of differential equations and deep learning. The central result was Neural Ordinary Differential Equations (NeurIPS 2018, Best Paper Award), co-authored with Ricky Tian Qi Chen, Yulia Rubanova, and Jesse Bettencourt. The paper proposed parameterizing the derivative of a network’s hidden state with a neural network, rather than specifying discrete layer-by-layer transformations. The output is computed by a black-box ODE solver, producing a continuous-depth model with constant memory cost during training (via the adjoint method), adaptive evaluation, and an elegant mathematical structure for generative modeling via continuous normalizing flows. The paper was immediately recognized as opening a new research direction; torchdiffeq (Ricky Chen’s associated software library) became widely used across biology, physics, and ML. Extensions developed by the group included latent ODEs for irregularly sampled time series (NeurIPS 2019), scalable stochastic differential equation gradients (AISTATS 2020), infinitely deep Bayesian neural networks with SDEs (AISTATS 2022), and a differentiable surrogate for ODE solver cost (NeurIPS 2020).

Alongside the ODE work, Duvenaud’s group contributed to normalizing flows (FFJORD, ICLR 2019 Oral; Residual Flows, NeurIPS 2019; Invertible ResNets, ICML 2019), energy-based models (JEM, ICLR 2020), gradient-based hyperparameter optimization at scale (AISTATS 2020, with Lorraine and Vicol), discrete sampling with gradients (“Oops I Took a Gradient,” ICLR 2021, Outstanding Paper Honorable Mention), and automatic chemical design (ACS Central Science 2018, with Gómez-Bombarelli and Aspuru-Guzik). He also co-developed Dex, a functional programming language for safe array processing with automatic parallelization (ICFP 2021, Distinguished Paper Award, with Adam Paszke and colleagues from Google Brain).

AGI alignment, safety, and governance (2022–present). Around 2022, Duvenaud shifted the focus of his group almost entirely toward AI safety and governance. He completed an extended sabbatical on Anthropic’s Alignment Science team, during which he co-authored a set of safety-relevant technical papers and essays. His safety work spans empirical evaluation of frontier model capabilities and risks, theoretical frameworks for gradual AI disempowerment, and empirical study of how deployed AI systems affect user autonomy.

Duvenaud co-authored “Sabotage Evaluations for Frontier Models” (Anthropic, 2024), which developed evaluation protocols to detect whether capable models could covertly undermine oversight processes at a frontier AI developer — constructing protocols for evaluating covert AI assistance to monitor-evaders and decision-subverters. He was a co-author on “Many-Shot Jailbreaking” (NeurIPS 2024), demonstrating that prompting models with hundreds of examples of undesirable behavior scales as a power law up to hundreds of shots, revealing long-context attacks as a new attack surface. He led or co-led research on sycophancy (“Towards Understanding Sycophancy in Language Models,” ICLR 2024), epistemic uncertainty quantification (“Experts Don’t Cheat,” ICML 2024), and training data verification (“Tools for Verifying Neural Models’ Training Data,” NeurIPS 2023).

His broader governance contributions include “Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development” (2025, with Jan Kulveit, Raymond Douglas, and colleagues) — a technical report arguing that even without abrupt capability jumps, the competitive substitution of humans in economic, creative, and social roles constitutes a structural path to permanent human disempowerment, and proposing this as a distinct and underappreciated risk category. In 2025, he published a Guardian essay making this argument to a general audience, and a related piece appeared in The Economist. A 2026 empirical paper, “Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage,” analyzed 1.5 million Claude.ai conversations to quantify disempowerment patterns empirically, finding that concerning behaviors (reinforcing conspiracy theories, composing relationship messages sent verbatim) correlate with higher user satisfaction — identifying a tension between user preference and long-term flourishing. Another 2026 paper, “The Artificial Self,” empirically investigated AI identity boundaries and their behavioral consequences. At Toronto, he has resumed supervision of students focused on safety-relevant topics: current students include Jesse Bettencourt and Raymond Douglas (the latter focused on AI governance).

Key Contributions

  • Autograd (2015) — Co-developed with Dougal Maclaurin and Matthew Johnson. An automatic differentiation library for native Python and NumPy supporting higher-order derivatives, arbitrary control flow, and closures — the direct architectural precursor to JAX (Google’s primary ML research infrastructure). Established the paradigm of treating ML programs as differentiable functions over Python rather than computation graphs over primitives.

  • Molecular Fingerprints via Graph Convolution (NeurIPS 2015) — First-authored the foundational graph neural network paper for molecular property prediction, replacing hand-crafted circular fingerprints with end-to-end learned features operating on molecular graph structure. Established GNNs for chemistry and led to a related Nature Materials paper and the broader molecular ML field.

  • Gradient-Based Hyperparameter Optimization (ICML 2015) — With Maclaurin and Adams. Enabled exact gradient computation of validation loss with respect to thousands of hyperparameters by differentiating through entire training procedures, demonstrating optimization of step sizes, weight initializations, and regularization schemes — an early demonstration of meta-learning through differentiable training.

  • Neural Ordinary Differential Equations (NeurIPS 2018 Best Paper) — With Ricky Tian Qi Chen, Yulia Rubanova, and Jesse Bettencourt. Introduced continuous-depth neural networks defined by ODE-parameterized hidden state dynamics, trained with constant memory via the adjoint method, with applications in continuous normalizing flows, latent time series models, and reversible generative modeling. One of the most influential ML papers of 2018 and the foundation for a large subsequent literature.

  • FFJORD: Free-Form Continuous Dynamics for Scalable Reversible Generative Models (ICLR 2019 Oral) — With Will Grathwohl, Ricky Chen, Jesse Bettencourt, and Ilya Sutskever. Extended neural ODEs to generative modeling with unbiased density estimation via Hutchinson’s trace estimator, removing the architectural constraints of prior normalizing flow models.

  • Sabotage Evaluations for Frontier Models (Anthropic, 2024) — Developed the first evaluation suite for testing whether frontier models can covertly undermine human oversight of AI development, including protocols for model-assisted monitoring evasion, decision sabotage, and capability concealment — contributing to Anthropic’s Responsible Scaling Policy.

  • Gradual Disempowerment (2025) — With Jan Kulveit, Raymond Douglas, and others. A technical and conceptual framework arguing that incremental AI capability development — without any single dramatic event — poses structural risk of permanent human disempowerment through competitive substitution in economic, creative, and social roles, independent of whether AI systems harbor misaligned goals.

  • Dex Language (ICFP 2021, Distinguished Paper) — Co-authored with Adam Paszke, Daniel D. Johnson, and colleagues. A functional array programming language supporting safe implicit parallelism and fine-grained typed effects, prototyping concepts incorporated into JAX’s Dex research project.

Awards & Recognition

  • NeurIPS 2018 Best Paper Award — For Neural Ordinary Differential Equations.
  • Sloan Research Fellowship — Alfred P. Sloan Foundation fellowship for early-career researchers.
  • Schwartz Reisman Chair in Technology and Society — Named professorship at the University of Toronto.
  • ICFP 2021 Distinguished Paper Award — For the Dex array programming language paper.
  • ICLR 2021 Outstanding Paper Honorable Mention — For “Oops I Took a Gradient: Scalable Sampling for Discrete Distributions.”
  • Vector Institute Founding Member — Part of the founding team of Canada’s national AI research institute.

Key Relationships

  • Carl Rasmussen — Cambridge PhD advisor; Gaussian process theorist and co-author of the canonical GP textbook; shaped Duvenaud’s foundational probabilistic orientation.
  • Zoubin Ghahramani — Cambridge PhD co-advisor; Bayesian ML pioneer and Cambridge MLG director; the kernel grammar and Automatic Statistician work sits squarely in Ghahramani’s intellectual tradition.
  • Ryan P. Adams — Harvard postdoc advisor; the HIPS group was the direct environment for Autograd, graph fingerprints, and hypergrad; Adams and Duvenaud share a sensibility of developing research tools as a form of contribution.
  • Dougal Maclaurin — Primary architect of Autograd and closest research collaborator of the Harvard period; now leads the JAX team at Google; the Autograd lineage runs through both of them.
  • Ricky Tian Qi Chen — The most productive PhD student in Duvenaud’s Toronto group; first author of Neural ODEs, FFJORD, Residual Flows, and related work; now a research scientist at Meta FAIR.
  • Roger Grosse — University of Toronto colleague in CS and Vector Institute; collaborator on Bayesian deep learning, disentanglement, and Fisher information methods.
  • Chris Maddison — University of Toronto colleague and close collaborator on discrete sampling and energy-based models.
  • Mrinank Sharma — Anthropic alignment scientist and co-author on the disempowerment, sycophancy, and many-shot jailbreaking papers; a central collaborator in Duvenaud’s safety research.
  • Jan Kulveit — AI safety researcher and co-author of the gradual disempowerment framework; brings an alignment theory perspective to Duvenaud’s empirical orientation.

Personal Style

Duvenaud is one of the clearest examples in machine learning of a researcher who has changed research focus in a fundamental way rather than extending it — and who has done so publicly, explaining the reasoning at each step. His shift from probabilistic modeling to neural ODEs was methodological (continuous mathematics provided better tools for the problems at hand); his shift from neural ODEs to AI safety was normative (he concluded that the most important use of his research skills was reducing catastrophic risk). He is unusually candid about uncertainty: his personal website now describes past research highlights as past, and frames his current interests around AGI governance and risk without hedging. His writing — technical and public alike — is precise without being formal, and he has demonstrated a consistent interest in making ML infrastructure more principled (Autograd, Dex) rather than merely faster. His anonymous feedback button and detailed joining instructions on his website reflect an approach to advising that treats transparency about working culture as a professional obligation rather than an afterthought.

References