Jordan Taylor Email & Phone Number

London, Gb

London, Gb

Creating methods and benchmarks for mechanistic anomaly detection in LLMs. These methods should be able to tell if the internal processes used to compute a model's output were "strange", flagging jailbreaks or backdoors at runtime, for example.

Brisbane, Queensland, Australia

Developing new tensor-network theory and algorithms for understanding and simulating entangled quantum systems, and for understanding machine learning systems. Supervised by Ian McCulloch.

Sunnyvale, California, United States

Developing theory and algorithms to better simulate the dynamics of entangled quantum systems, by identifying wavefunction branches in tensor network states. Supervised by Jess Riedel

Berkeley, California, United States

Collaborated with Lee Sharkey and Dan Braun at Apollo Research to develop a new method for training sparse autoencoders (SAEs) for more faithful interpretability of AI systems. See our paper: https://arxiv.org/abs/2405.12241 and our codebase: https://github.com/ApolloResearch/e2e_sae

Berkeley, California, United States

Attended an intense pair-programming bootcamp, coded partial replications of BERT, GPT-2, ResNet50, CLIP, diffusion models, U-Nets, DQN, PPO, in addition to CUDA programming, backpropagation from scratch, and mechanistic interpretability.

Applied machine learning and interpretability techniques to neuroscience research, using data from individual neurons in the brains of rats to predict the timing of audio beeps they were hearing.Testimonial:"Jordan's work in his internship has been exemplary, with a wide variety of techniques and approaches used to deal with a challenging problem, all in a relatively short time and almost entirely on his own. What's more, his work was presented to the team in a deceptively clear and comprehensible manner, showing that he not only has an excellent understanding of the problem and methods used to tackle it, but fantastic skills in boiling down deep technical spaghetti into an easy-to-understand story - a rare and valuable skill." - Maciej Trzaskowski, Head of Research at Max Kelsen

12-month honors project: Developed atomic structure simulations in C++ characterizing errors in the world’s best atomic clocksThesis: bit.ly/32aoQ5z 6-month research project: Created a parallel algorithm to simulate the dynamics of quantum wavefunctions. My C++ implementation is available at https://github.com/jordansauce/iterative_parallel_CN Paper: bit.ly/3jWbp0XSupervised by Dr Michael Bromley

Bachelor Of Advanced Science + First-Class Honours, Physics, Gpa Of 6.6/7 (94%)

10 mathematics courses, 20 physics courses, all graded distinction or high-distinction, with an exchange to Vancouver, Canada.

Doctor Of Philosophy - Phd, Physics

Developing new tensor-network theory and algorithms for understanding and simulating entangled quantum systems, and for understanding.