Douw Marx
Summary
AI safety researcher working on agentic evaluations, risk estimation, and alignment. Contributing to the European Commission Technical Assistance for AI Safety, Redwood Research, AI Futures Project, and METR via Equistamp. PhD from KU Leuven, Belgium, applying machine learning and signal processing to fault detection in rotating machines.
Experience
Equistamp, Research Engineer
-
Contributing to EU AI Office evaluations on CBRN, loss of control, and harmful manipulation
-
Contributing to Redwood Research's LinuxArena, a control setting for highly privileged AI agents
-
Contributor to Redwood Research's ASMR-Bench, a benchmark for detecting AI sabotage in ML research
-
Merged upstream contributions to open-source eval infrastructure: METR's Hawk and Redwood Research's Control Tower
-
Baselining for Andon Labs, Redwood Research, and METR
-
Equistamp communications on Substack, including evaluation selection approaches based on risk modelling and value of information
-
Scenario evaluations for the AI Futures Project
-
Credited for feedback on AI 2040: Plan A, AI Futures Project
-
Grant writing and associated automations
KU Leuven, Marie Skłodowska-Curie PhD Fellow
Research on fault detection of rotating machines using unsupervised learning and differentiable signal processing methods.
-
Proposed and led 4 Master's research projects
-
Research secondments at INSA Lyon, France and SafranTech, Paris, France
-
Runner-up in PHM 2023 Conference Data Challenge, Salt Lake City, Utah, USA
Wolfram MathCore, SystemModeler Intern
Develop Virtual labs using Wolfram System Modeler and Mathematica.
XRAM Technologies, Data Analyst and Mechanical Engineer
Build data-driven measurement models for electrode length prediction in electric-arc furnaces.
-
Furnace electrode length prediction
-
Chemical process data analysis
-
Bayesian structural models and Gaussian process regression
Wolfram Summer School, Participant
-
Cellular Automata
-
Inverse Kinematics
Publications
Scientific articles
Constitutional Sensitivities of Preference Models
Apart Research Hackathons: AI monitoring at low audit budgets, and Automated risk modelling for evaluation value-of-information estimation
Blog posts on Equistamp's Substack: Value of information and the seven deadly evalua-sins, and No more snapshots: forecast AI risk by backtesting
Workshop: Agentic AI safety
Talk: Does your LLM care about the same things you do? (75th Data Science Leuven Meetup, slides)
Patent for a novel parallel kinematic planar mechanism: WO2020208551A1
Education
KU Leuven, Fault detection in rotating machinery
Marie Skłodowska-Curie PhD fellow at KU Leuven.
-
Thesis: Fault Detection in Rotating Machinery Using Domain Knowledge and Reference Data
-
Regularised unsupervised deep learning models
-
Differentiable signal processing methods
University of Pretoria, Honours and Master's in Mechanical Engineering
-
Machine learning applied to vibration monitoring
-
Numerical methods and optimisation
-
Thesis: Towards a hybrid approach for diagnostics and prognostics of planetary gearboxes
University of Pretoria, Mechanical Engineering (With distinction)
Skills and Tools
Evaluation infrastructure
Programming/Mathematical Modelling
Python (PyTorch, Pandas, SciPy, Dash, LangChain, Scikit-learn), Matlab, C++, Mathematica
CLI Tools
HPC with Slurm, Linux, LaTeX, Git, Vim, Emacs
Languages
Fluent in English and Afrikaans, understands Dutch.
Certificates
AI Alignment Course: Bluedot technical alignment course with project: Constitutional sensitivities of preference models.
EA Introductory Program: Effective Altruism Introductory Program
Interests
Effective altruism, music, woodworking, hiking