Interpretability researcher / Utah
I’m Jeffrey Shorthill, an independent interpretability researcher investigating how computational structures and representations develop, persist, and change across related language models.
Interpretability research · Applied AI · iOS development
Scroll to explore01 / Selected work
I build experiments around open model weights, then share the measurements and code so others can follow the evidence. All publications
Routing & causal interventions
An investigation of expert 114 in Qwen3.5-35B-A3B, where controlled routing experiments and causal interventions distinguish a readout associated with phenomenological language from the upstream computations producing that representation.
Context & expert routing
A controlled study of how text already available in a conversation reorganizes expert routing, with the final prompt and scored response tokens held fixed. It examines how contextual availability affects measurements of model internals.
A survey of repeated decoder blocks across 31 Qwen checkpoints, two Qwen-derived models, and SOLAR as a calibration control, with separate studies of weight correspondence and neuron responses.
The release includes model revisions, tensor hashes, measured tables, and tools for reproducing the results.
How alike are the responses of corresponding neurons 24 layers apart in Qwen3-32B?
I examined 288 neuron pairs on 10,000 fixed text windows. Corresponding partners often ranked above random alternatives, while absolute response resemblance was usually modest.
Read the response study
Weight correspondence
Separate Qwen2.5 and Qwen3 studies examine how weights correspond across 14B and 32B checkpoints.
Readout transfer
Tests whether a Jacobian lens fitted on Qwen3.6-27B remains informative on Qwen3.8-27B. The evaluated semantic readouts survive transfer, with changes in calibration and performance across layers.
Generation & expert routing
Generation-time expert sets track the local subject of model outputs, revealing specialization that pooled prefill measurements can obscure. The study includes cross-checkpoint replication and controls for prompt design, position, and register.
Evaluation & safety-critical support
A matched-pair evaluation protocol that isolates individual linguistic features contributing to differences in safety-critical assistance, including responses to AAVE-marked prompts. The release includes the paper, revision history, and links to sanitized public artifacts.
Causal interventions & refusal
An exploratory Qwen3.5-35B-A3B case study: suppressing the leading refusal-associated expert index across all 40 routers redistributes routing while refusal persists on the tested prompts.
Open methods, retained evidence, and code you can inspect.
Browse the codeApplied work
Alongside research, my work includes applied AI and building tools people can use.
Previous AI work
I previously worked on AI at Progressive Insurance, a Fortune 500 insurance company.
Company website
Tools that
support families.
Gift Connect / iOS development
I built the iOS app for Gift Connect’s Parenting Suite, bringing parenting resources and AI-powered voice, song, and story tools into one app.
02 / The question I’m following
As a model grows, what do its neurons keep, change, and become?
I’m tracing computational structures across related models to investigate how representations emerge, persist, specialize, and reorganize—including those associated with self-description and situational understanding.
Talk research03 / About me
I’m Jeffrey Shorthill, an independent mechanistic interpretability researcher investigating how computational structures and representations develop and change across language models.
Before moving into full-time research, I helped architect disaster-relief chatbots for Progressive Insurance in 2023 and, over the past two years, worked on AI platforms for underserved families at Gift Connect, a 501(c)(3) nonprofit. That same commitment to using AI to benefit people continues to guide my research.
Questions about AI welfare and moral consideration also motivate my research. I approach them through mechanistic interpretability, studying how the internal organization supporting self-description and experiential language develops and changes.
I use structural analysis, activation measurements, and causal experiments to understand how models organize information, particularly around self-description and situational understanding.
I publish my methods, data, and findings openly to support independent verification, criticism, and collaboration.
04 / Get in touch
Questions, critiques, or a research idea?
I’d like to hear from you.