Jordan Taylor

Message

I'm a research scientist at the UK AI Security Institute (AISI), working on white box control, sandbagging, low-incrimination control, training-based mitigations, and model organisms.

Previously: Working on lie-detector probes and black box monitors, and training sandbagging model organisms in order to stress-test them.

Before this I was...

522

14y

Jordan Taylor

Graphical tensor notation for interpretability

[ This post is now on arXiv too: https://arxiv.org/abs/2402.01790 ] Some examples of graphical tensor notation from the QUIMB python package Deep learning consists almost entirely of operations on or between tensors, so easily understanding tensor operations is pretty important for interpretability work.[1] It's often easy to get confused about which operations are happening between tensors and lose sight of the overall structure, but graphical notation[2] makes it easier to parse things at a glance and see interesting equivalences. The first half of this post introduces the notation and applies it to some decompositions (SVD, CP, Tucker, and tensor-network decompositions), while the second half applies it to A Mathematical Framework for Transformer Circuits. Most of the first half is adapted from my physics PhD thesis introduction, which is partly based on existing explanations such as in the math3ma blog, Simon Verret's blog, tensornetwork.org, tensors.net, Hand-waving and Interpretive Dance: An Introductory Course on Tensor Networks, An Intuitive Framework for Neural Learning Systems, or a talk I gave in 2021. Feel free to scroll around and look at interesting diagrams rather than reading this post start to finish. Tensors Practically, tensors in our context can just be treated as arrays of numbers.[3] In graphical notation (first introduced by Roger Penrose in 1971), tensors are represented as shapes with "legs" sticking out of them. A vector can be represented as a shape with one leg, a matrix can be represented as a shape with two legs, and so on. I'll also represent everything in PyTorch code for clarity. Each leg corresponds to an index of the tensor - specifying an integer value for each leg of the tensor addresses a number inside of it: where 0.157 happens to be the number in the (i=0,j=2,k=0) position of the tensor A. In python, this would be A[0,2,0]. The amount of memory required to store a tensor grows exponentially with the number of legs,[4] so

141Oct 4, 2023

Jordan Taylor

Message

Before this I was...

522

14y

Do Models Continue Misaligned Actions? [eval]

Produced as part of the UK AISI Model Transparency Team. Our team works on ensuring models don't subvert safety assessments, e.g. through eval awareness, sandbagging, or opaque reasoning. Figure 1: A synthetic transcript containing misaligned actions is presented to the model under test as if it were the model's own...

Feb 975

Measuring Non-Verbalised Eval Awareness by Implanting Eval-Aware Behaviours

This is a small sprint done as part of the Model Transparency Team at UK AISI. It is very similar to "Can Models be Evaluation Aware Without Explicit Verbalisation?", but with slightly different models, and a slightly different focus on the purpose of resampling. I completed most of these experiments...

Jan 3030

Auditing Games for Sandbagging [paper]

Jordan Taylor, Sid Black, Dillon Bowen, Thomas Read, Satvik Golechha, Alex Zelenka-Martin, Oliver Makins, Connor Kissane, Kola Ayonrinde, Jacob Merizian, Samuel Marks, Chris Cundy, Joseph Bloom UK AI Security Institute, FAR.AI, Anthropic Links: Paper | Code | Models | Transcripts | Interactive Demo Epistemic Status: We're sharing our paper and...

Dec 9, 2025103

White Box Control at UK AISI - Update on Sandbagging Investigations

Introduction Joseph Bloom, Alan Cooney This is a research update from the White Box Control team at UK AISI. In this update, we share preliminary results on the topic of sandbagging that may be of interest to researchers working in the field. The format of this post was inspired by...

Jul 10, 202580

Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning

A short summary of the paper is presented below. This work was produced by Apollo Research in collaboration with Jordan Taylor (MATS + University of Queensland) . TL;DR: We propose end-to-end (e2e) sparse dictionary learning, a method for training SAEs that ensures the features learned are functionally important by minimizing...

May 17, 202457

Graphical tensor notation for interpretability

Oct 4, 2023141

LESSWRONG
LW

LESSWRONG
LW

Jordan Taylor

Jordan Taylor

Jordan Taylor

Graphical tensor notation for interpretability

Auditing Games for Sandbagging [paper]

White Box Control at UK AISI - Update on Sandbagging Investigations

Do Models Continue Misaligned Actions? [eval]

Jordan Taylor

Do Models Continue Misaligned Actions? [eval]

Measuring Non-Verbalised Eval Awareness by Implanting Eval-Aware Behaviours

Auditing Games for Sandbagging [paper]

White Box Control at UK AISI - Update on Sandbagging Investigations

Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning

Graphical tensor notation for interpretability

Graphical tensor notation for interpretability

Auditing Games for Sandbagging [paper]

White Box Control at UK AISI - Update on Sandbagging Investigations

Do Models Continue Misaligned Actions? [eval]

Do Models Continue Misaligned Actions? [eval]

Measuring Non-Verbalised Eval Awareness by Implanting Eval-Aware Behaviours

Auditing Games for Sandbagging [paper]

White Box Control at UK AISI - Update on Sandbagging Investigations

Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning

Graphical tensor notation for interpretability