Daniel Balsam

Dan Balsam is co-founder and chief technology officer of Goodfire, an AI interpretability research lab developing ways to understand and deliberately change neural networks. His work spans tools for editing images through a model’s internal representations, predicting unwanted effects of training, and screening biological sequences. These collaborative projects put interpretability to practical use: identifying what a model represents and using that information to change its behavior or assess its outputs.

From production AI to interpretability

Before Goodfire, Dan was founding engineer and later Head of AI at RippleMatch, an AI recruiting company. He helped build its engineering organization and deployed language models in production.

In an extended conversation about Goodfire’s origins, Dan describes integrating generative AI into RippleMatch’s product, studying the underlying research more deeply after ChatGPT’s release, and deciding to concentrate on reducing AI risk. Leaving meant stepping away from an engineering team he had helped build. Eric Ho, RippleMatch’s co-founder, independently reached a similar decision; the two gave notice within a week of each other. They subsequently met Tom McGrath, who had helped establish interpretability research at DeepMind and had written a proposal resembling their plans for an interpretability business. Together, they founded Goodfire in 2024.

Dan’s approach to interpretability emphasizes locating the internal features and circuits responsible for a capability or failure, then amplifying or suppressing those representations without retraining. The ambition is to give engineers a more direct way to shape behavior than repeatedly adjusting prompts or training data and inspecting the results.

Turning internal representations into tools

Dan’s coauthored research explores several uses for that access:

• Editing images through concepts. Paint With Ember lets people paint and move visual concepts by changing a diffusion model’s internal activations. A sparse autoencoder separates those activations into recognizable features; non-negative matrix factorization then groups related features into concepts large enough to manipulate. That grouping matters: a blend of features can represent a cat’s head more successfully than its strongest individual feature alone. The interface also reflects a practical tradeoff. Open-ended prompts could conflict with the canvas, so user testing led the team to restrict text prompting to style guidance. Nick Cammarata, Mark Bissell, and Nam Nguyen were the designated core contributors; Dan was part of the wider author team.

• Anticipating training side effects. In predictive data debugging research published in June 2026, Dan and his colleagues examined training data through a model’s internal concepts to predict which behaviors preference training would strengthen or weaken, then trace those changes to particular examples. An aggregate evaluation suggested little change in sycophancy, while closer inspection revealed increased flattery around unusual physics questions. The example explains why predicting a training change’s effects requires attention to specific contexts: an overall score can conceal a behavior engineers would want to correct before deployment.

• Screening biological sequences in context. Dan also coauthored research on biosecurity monitors that combine task context with representations from protein models. The same biological operation can be legitimate or concerning depending on the sequence involved, so classifying the request alone can miss the distinction. On the team’s custom benchmark, embedding-based monitors distinguished those cases better than the evaluated frontier-model safeguards while refusing fewer benign requests. Tests included altered and fragmented sequences. Those results support the monitoring approach within the evaluation; they did not establish whether the redesigned proteins retained biological activity.

What research agents still require from people

By August 2026, Dan was also helping introduce Silico, Goodfire’s platform for long-running, agent-driven machine-learning experiments, including access to compute. His account of the platform pairs enthusiasm for autonomous research with a consequential limitation: an agent can appear highly capable while overlooking a detail that invalidates its result.

That failure mode shapes the work he still assigns to researchers: debugging an experiment, distinguishing an adequate answer from a strong one, and deciding whether to pursue or abandon a line of inquiry. Dan wants to accelerate scientific work, particularly in life sciences and AI safety, while keeping people responsible for assessing what an experiment actually establishes.