Mark Bissell is a member of the technical staff at Goodfire, where he develops mechanistic interpretability techniques for understanding, controlling, and learning from advanced AI models. His work applies insights from neural-network internals to production reliability, research agents, creative tools, Alzheimer’s biomarkers, and materials discovery.
Bissell studied computer science at Williams College and Oxford, alongside economics and history, and graduated as valedictorian of Williams’s class of 2023. Before Goodfire, he spent three years working on public- and private-sector healthcare projects at Palantir as a forward-deployed software engineer. That experience informed his focus on reliability, accuracy, and privacy in consequential applications. His professional background and subsequent research trace a shift from deploying healthcare software to investigating how models actually produce their answers.
- Activation-based model steering. Bissell treats internal model features as engineering controls. Instead of relying exclusively on prompt changes, expensive model-as-judge pipelines, or fine-tuning that can introduce unintended behaviors, developers can identify relevant activations and adjust them directly. In a practical interpretability demonstration, he strengthens a feature associated with sensitive information to discourage a language model from disclosing confidential data. He also describes Rakuten’s use of Goodfire’s Ember platform for multilingual detection of personally identifiable information.
- Painting With Concepts. Bissell co-authored research on diffusion-model latents that transforms learned visual concepts into editable interface elements. Users can position a pyramid or lion on a canvas, alter its intensity, move it spatially, or adjust component features such as an open mouth. The project makes interpretability a mechanism for direct creative control.
- Interpreting frontier reasoning models. As a co-author of research examining DeepSeek R1, Bissell helped train and release sparse autoencoders for the 671-billion-parameter reasoning model. These interpreter models expose internal features that researchers can investigate, while leaving room for newer interpretability methods to supersede them.
- Research agents built for experimentation. Bissell was the primary author of a guide to agent-assisted interpretability research distinguishing experimental agents from coding agents. His team released a Jupyter notebook/server MCP reference implementation and interpretability tasks spanning circuit discovery and linear-probe analysis. He describes unreviewed experimental outputs as shallow validation debt: they require scrutiny without locking researchers into an opaque, agent-generated software system.
- Scientific discovery from model internals. As a primary author of Alzheimer’s biomarker research, Bissell helped identify DNA fragment-length patterns associated with disease detection in an epigenetics foundation model. He also co-authored materials-discovery research using internal activations to guide diffusion-model sampling and work on interpretable genomic variant prediction. Together, these projects pursue interpretability-driven scientific discovery: converting patterns learned by powerful models into hypotheses researchers can inspect and test.
