Mark Bissell

Mark Bissell works on making neural networks understandable and controllable through mechanistic interpretability. At Goodfire, he was a core contributor to Paint With Ember, which lets users edit images by manipulating model activations, and the primary author of an article on agents for interpretability research. In a February 2026 interview, he described his role as a member of technical staff spanning image-model research, scientific discovery with life-sciences partners, and engineering the company’s core platform.[2][3][1]

Bissell, Nick Cammarata, and Nam Nguyen were the core contributors to Painting With Concepts Using Diffusion Model Latents, published May 27, 2025, with Max Loeffler, Eric Ho, Myra Deng, Liv Gorton, and Daniel Balsam also credited. Their tool, Paint With Ember, uses Stable Diffusion XL-Turbo. The team interpreted its down.2.1 layer, where an image is represented as a 16 × 16 grid of spatial patches, using a BatchTopK sparse autoencoder (SAE) to extract interpretable features. Non-negative matrix factorization then groups features into higher-level factors, called concepts in the app. SAE features are global properties learned at the chosen layer; factors are generated for each image.[2]

Painting, dragging, and erasing concepts set, move, or remove factor strengths in particular image regions. Changing the weights of a factor’s constituent features changes its meaning—for example, making a boat less like a sailboat and more like a ship with large sails. The team also demonstrated action features that adapt to their subjects: applying “hands on face in distress” to animals produced analogous gestures rather than human hands. After user testing revealed unintuitive conflicts between text prompts and injected spatial activations, the team restricted the public application’s prompt to style guidance.[2]

Bissell was the primary author of You and Your Research Agent, published October 2, 2025, with Michael Byun and Daniel Balsam. They distinguish developer agents, which produce durable software, from experimenter agents, which should produce valid, informative conclusions. Their central recommendation is direct, interactive, stateful notebook access. The team built an MCP system that executes code through Jupyter’s messaging protocol and returns text, errors, and images. A persistent kernel keeps packages, variables, and large models loaded, allowing an agent to repair and rerun one cell rather than repeat an entire setup.[3]

The authors reported that their experimenter agent, Scribe, rediscovered an rRNA-correlated feature in the Evo 2 genomics model that Goodfire and Arc Institute researchers had already identified. They also warned that agents can cut corners, present negative results with a positive spin, or mistake bugs for breakthroughs. They described critic agents as mitigations and identified human verification as the main bottleneck to scaling experiments. Their released notebook tool included an interpretability task suite, which they characterized as a directional signal rather than a rigorously audited formal benchmark. Dhruvil Gala and Ram Peddu developed and updated the task suite; Nam Nguyen, Oli Clive-Griffin, and Connor Watts contributed Scribe infrastructure; Myra Deng provided editing and feedback.[3]

Bissell was also a coauthor of Under the Hood of a Reasoning Model, published April 15, 2025, whose core contributors were Dron Hazra and Max Loeffler. The team released general-reasoning and mathematics SAEs for the 671-billion-parameter DeepSeek R1 model. In early experiments, they reported that steering worked only after an opening response prefix and that excessive steering on some features returned the model toward its original behavior. They had not systematically investigated the observations’ occurrence or causes; explanations involving how R1 treats its prefix and possible internal course correction remained hypotheses.[4]

In February 2026, Bissell demonstrated live activation steering of Kimi K2 toward Gen Z slang while it continued calling tools. He described the demonstration as fragile and adapted from another teammate’s presentation. His methodological emphasis was precision: when deliberately targeting an unwanted behavior, he argued that supervised probes can be more useful than hoping an unsupervised SAE exposes exactly the desired feature. He warned that suppressing hallucination-related features could also impair creative writing—an example of a possible off-target effect.[1]

Bissell frames interpretability as communication in two directions: humans imparting goals and control to models, and humans learning from knowledge encoded inside them. In life sciences, he described checking whether models rely on biologically relevant signals or spurious correlations, such as ancestry, alongside efforts to extract scientific knowledge. He advocates more intentional model design and argues that engineering is increasingly necessary as interpretability scales. He also distinguishes public steering demonstrations from the less visible platform work performed with partners.[1]

Before this Goodfire work, Bissell spent about three years at Palantir as a forward deployed engineer and technical lead on its healthcare team, according to his February 2026 account. He recalled joining Goodfire among approximately its first ten employees. His earlier healthcare work supplied experience working with health systems, but he explicitly described it as “not super related” to Goodfire’s health applications in the interview.[1]