Skip to content
View jeffreywilliamportfolio's full-sized avatar
  • Business Intelligence Group LLC (New)
  • Salt Lake City, UT

Block or report jeffreywilliamportfolio

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse

Jeffrey W. Shorthill

Independent AI interpretability researcher, new to the field in 2026 and learning it in public. I also ship production software. Portfolio: jwshorthill.vercel.app

Research

Five preprints on mixture-of-experts routing and safety behavior (June to August 2026), plus the Expert 114 compendium and a philosophy paper. One claim per paper, rules fixed before capture, nulls reported. The five share one method question: which parts of a routing readout are the behavior under test, and which are the position, length, regime, or context the text sits in. Each revision records what I learned about measurement since the last.

  • In-Context Availability Reorganizes Expert Routing: A Measurement Confound for Multi-Turn Evaluations of MoE Language Models. Replicates Qwen3.5 to 3.6. doi:10.5281/zenodo.21892999 · v1.1 record
  • Bundled Dialect Contrasts Measure a Mixture. A feature-isolating matched-pair audit (DMRA) of safety-critical support. doi:10.5281/zenodo.20449546 · v4.3 record
  • Suppressing the Leading Refusal-Associated Expert Index at Every Layer Redistributes Routing and Leaves Refusal Intact. Qwen3.5-35B-A3B. doi:10.5281/zenodo.20786891 · v1.2 record · repo
  • Specialization Is Not a Myth: Subject-Specific Expert Sets Recur During MoE Generation. v1.2 retitled from "Generation-Time Routing Reveals Expert Specialization That Prefill Measurements Miss". doi:10.5281/zenodo.20779604 · v1.2 record · repo
  • Read Routing Entropy at a Fixed Position: A Cross-Model Study of the Prefix Effect in MoE Routing. v2.0 retitled from "Sequence Position Explains an Apparent Complexity Gradient in MoE Routing Entropy". doi:10.5281/zenodo.20779602 · v2.0 record · repo
  • Banana in, Bostrom out. One token-direction swap flips a fixed prompt to paperclip maximization. LessWrong post · repro repo (browser, under a minute)
  • Survival of the Fitted. A Jacobian lens fitted on one checkpoint reads the next one, unmodified.
  • J Volume. Native macOS and web viewer that renders lens captures as an explorable 3D field.

Products

  • Gift Connect Parenting Suite. Live on the App Store and Google Play. Sole engineer, design to store submission.
  • Six deployed web tools for the same nonprofit: realtime voice agent, song and story generators, SMS assistant, analytics and cohort dashboards.

Background

Prompt Engineer at a Fortune 500. Full-stack indie dev for the last three years, focusing on cross-platform mobile apps. Prior education and career in recording engineering.

Contact

jws299792@icloud.com

Pinned Loading

  1. dmra-paper dmra-paper Public

    Preprint (v4.3): Bundled Dialect Contrasts Measure a Mixture — a feature-isolating matched-pair audit (DMRA) of safety-critical support in language models. Concept DOI 10.5281/zenodo.20449546

    TeX

  2. paper-expert-specialization-generation paper-expert-specialization-generation Public

    Preprint + data (v1.1): Generation-Time Routing Reveals Expert Specialization That Prefill Measurements Miss. Concept DOI 10.5281/zenodo.20779604

    Python

  3. paper-token-position-confound paper-token-position-confound Public

    Preprint + reproducibility materials (v1.1): Sequence Position Explains an Apparent Complexity Gradient in Mixture-of-Experts Routing Entropy. Concept DOI 10.5281/zenodo.20779602

    TeX