About
Things that get me out of bed in the morning: Solving problems with math, science, and coding; Collaborating towards meaningful projects; Coffee.
My current focus is on probabilistic agentic AI systems and safety applications.
In particular, experimenting with Active Inference and Variational Free Energy-directed learning structures—mathematical formulations
entwining concepts from physics first-principles with theories of intelligence.
I am also interested in probabilistic programming, Bayesian methods, and generative world modeling.
My personal concern with AI-threat models bends towards nearer-term harms (misinformation, malicious actors, socioeconomic disparities, etc.) vs. catastrophic x-risk scenarios
(though I don’t rate such events as safely ignorable).
I’m a recent alum from the SPAR research program, and am an active participant and volunteer in the AI Governance & Safety Canada (AIGS) group, looking to ensure
positive outcomes for AI in Canada.
In past work, I was as Senior Backend Developer at QED, developing genericized platform tools and architecture
for custom instantiation in business and data management contexts. Since 2022
QED has been creating custom software for New Zealand’s
“largest transport infrastructure project ever”.
In other work placements, I joined the Hyper-K
neutrino detecting efforts—the successor to the 2015 Nobel Prize winner
Super-K—at Canada’s premier accelerator lab,
TRIUMF; researched galactic cloud astrophysics via simulation at McMaster University with James Wadsley’s group;
and worked on an anomaly detection ML model at Canada’s Communications Security Establishment. I am also an extended volunteer for Climate Change AI.
Outside academic and professional pursuits, I love hosting and making food for friends and family; taking my Olympus OM-1 outdoors for some
film shooting (though my bank account would prefer I don’t); Volunteering at The Local Community Food Centre;
and masochistically running Linux on my PC.
What I like to do: ML, Simulations, HPC, NLP, Graph Networks, Shelling, DevOps, API/Backend development, \(\LaTeX\), Git
How I like to do it: pandas, scipy, gymnasium, scikit, tf/pytorch, xarray, jax, seaborn, huggingface, pymc, langchain
Works & Publications
- AI-safety paper reviews and other writings on my blog, http://quarryblog.substack.com.
- Volunteering with AIGS: Contributions to the weekly Toronto AI Safety Meetup, and various website improvement efforts.
- RiskRay. Using ML to detect hips at risk of fracture. Patent pending.
- Traffic prediction Graph Neural Net. View the proof-of-concept toy model.
- My MSc. thesis: Machine learning topological defects of liquid crystals in two dimensions. Available online on UWSpace (2019).
- M. Walters, Q. Wei, J. Z. Y. Chen. Machine learning topological defects of confined liquid crystals in two dimensions. (Phys. Rev. E 99, 062701 / Read PDF) (2019).
- M. B. Bennett et al. Detailed study of the decay \(^{31}Cl(\beta\gamma)^{31}S\). Physical Review C 97 (6), 065803 (2018).
- E. Aboud et al. Toward complete spectroscopy using \(\beta\) decay: The example of \(^{32}Cl(\beta\gamma)^{32}S\). Physical Review C 98 (2), 024309 (2018).
- [Editor’s suggestion] M. B. Bennett et al. Isospin mixing reveals \(^{30}P(p,\gamma)^{31}S\) resonance influencing nova nucleosynthesis. Phys. Rev. Lett. 116, 102502, (2016).
- M. B. Bennett et al. Isobaric multiplet mass equation in the \(A = 31, T = 3/2\) quartets. Physical Review C 93 (6), 064310 (2016).
- C. Wrede et al. \(\beta\) Decay as a Probe of Explosive Nucleosynthesis in Classical Novae. Physics Procedia 66, 532-536 (2015).
Almae Matres
University of Waterloo
MSc. Physics
2017 - 2019
Thesis-based Master’s under Prof. Jeff Chen researching the use and efficacy of neural networks in
learning phases of matter and topological defects, using a simulated liquid crystal system
as a testbed. Findings included the novel use of Recurrent Neural Networks to learn topological features in an off-lattice setting.
Additional thesis research was done using Principle Component Analysis on simulated localized defects to then identify their presence in
large-scale liquid crystal arrangements.
Coursework covered quantum mechanics, quantum many-body systems, Monte Carlo methods, and stochastic processes.
2017 Provost Graduate Scholarship recipient
McMaster University
BSc. Hons. Physics
2012 - 2017
Bachelor’s degree covering essential physics topics. With my co-ops and electives I further focused my education and experience on computational/numerical methods, quantum mechanics, quantum computing, and mathematical physics.
Past & Ongoing
ActInf Research
Research lead
May 2024 - Present
Better decision-making with AI risk estimation
In my recent research, I’ve been putting into practice theory from Active Inference and Free-Energy-based learning directives, in collaboration with Rafael Kaufmann and Justice Sefas.
We are demonstrating how ActInf formalisms can provide first-principle risk metrics to stakeholders, aiding to bridge inner and outer alignment between stakeholders and agent objective functions. In one scenario we demonstrate how agent forecasting can signal population collapse in a fisheries context. In another, we investigate how our risk metric can inform a gatekeeper module in autonomous vehicle decision-making.
These experiments utilize probabilistic programming, generative modeling, and the gymnasium
toolkit.
Have a poke around the AV repo.
We aim to preprint our findings over the next few months, but here are some
recent slides and
writing sample.
SPAR
AI Safety Researcher
Feb. - May, 2024
The Supervised Program for Alignment Research (SPAR) program brings project leads and new researchers together to tackle questions at the forefront of AI safety and alignment. Through SPAR I joined three projects:
With Rafael Kaufmann and Roman Leventov, our group is looking at how systems referencing a shared causal world-model can help us reason about, quantify and improve the safety of AI systems in their interactions with real-world (physical and social) environments.
With Kellin Pelrine, our group is looking into applying hybrid neural network architectures, including LLMs and anomaly detection, to combat malicious coordinated group activities on social media.
With Gabriel Mukobi, our group is investigating algorithms and benchmarks for Unadaptable Foundation Models. If tenable, pretrained models that undergo unadaptability processing would be resistant to finetuning in certain directions, while ideally maintaining the capacity for desirable finetuning directions.
QED
Senior Backend Developer
Feb. 2020 - Aug. 2023
The small team size and startup nature of the company required learning and engaging with the numerous technological layers of the product, from direct database tinkering to frontend UX, and all the surrounding programs/utilities. This demands both breadth and depth of development to get the job done.
More technically, through Django and Postgres I had a significant hand in constructing a suite of genericized models, tools, and architecture that could be instantiated and characterized swiftly yet specifically via YAML files for a given product context. The result allows sophisticated customization of not only model properties but interactions across objects on a per customer basis. This also included building out the REST API, automated testing suites for all corners of the codebase with GitHub Action workflows, and documentation. I also built many other in-house, ad-hoc data management features and reports to grant customers insights into their data.
I worked to both design and code numerous aspects of the platform backend through its iterative growth over the past three years. Other undertakings of note would include: managing a collaboration between a contracted third party to design a locomotive worksite scheduling optimizer using machine learning; Prototyping an Internet-of-Things plug-in to our software; Formulating a multivariate mathematical model for worksite Task Success/Health prediction.
CSEC
Software Developer
Sep, 2016 - Jan, 2017
At the Communications Security Establishment of Canada I programmed a hybrid neural net for anomaly detection. The language was primarily Python, using TensorFlow with a Postgres database.
McMaster University
Astrophysics Researcher
May - Sep, 2016
I simulated galactic disk portions using parallel SPH tree-code Gasoline to investigate behavioural dependencies on certain factors, primarily resolution. I prepared various initial conditions for these runs, with subsequent analysis in Python. The HPC cluster SHARCNET was used for these runs.
TRIUMF
Hyper-Kamiokande R&D
Jan - Sep, 2015
As part of the T2K Canada group, I worked on the successor to the 2015 Nobel Prize-winning neutrino oscillation experiment Super-Kamiokande. My primary work was analyzing the timing resolution of three different flash analog-to-digital signal converter modules. This included acquiring and processing signal data, for which I used the CERN analysis toolkit ROOT. I also simulated neutrino oscillation detection events in the Water-Cherenkov detector, WCSim, using Geant4.