Myra Deng

Myra Deng was on Goodfire’s founding team and coauthored research applying mechanistic interpretability—understanding neural networks through their internal computations—to practical model control. In February 2026, she was Goodfire’s Head of Product, working across research and engineering to turn frontier research into repeatable platforms and products. Her contributions include shared primary authorship of a production PII-detection project with Rakuten and coauthorship of interfaces for steering language and image models.[2][1][5][3][4]

In her February 6, 2026 Latent Space interview, Deng described product work as identifying important real-world problems, translating research into usable tools, and explaining what interpretability can accomplish. Customer and researcher conversations informed the research agenda; she noted that chief scientist Tom also spent substantial time talking to customers and taking their problems back to the research team.[1]

Deng said Goodfire’s first steering API fell short of prompting or fine-tuning, leading the team to revisit its approach. She credited Ekdeep and Atticus with research toward more powerful, robust steering. Her ambition extended beyond changing a model’s style: adapting a model into an expert legal reasoner, for example, would require more sophisticated interventions and breakthroughs in learning algorithms. She presented that as a research direction, while acknowledging that the eventual product interface remained unclear.[1]

The immediate focus was post-training. Deng pointed to sycophancy, reward hacking, and enterprise customization that teaches a model noise or fails to teach the intended task. Her question was how an understanding of what a model knows and does could guide its learning. She also thought the techniques could apply to pre-training, without presenting a demonstrated pre-training result.[1]

Deng was a primary author of the October 28, 2025 Rakuten PII report, alongside Goodfire’s Nam Nguyen, Dhruvil Gala, and Michael Byun and Rakuten’s Kenta Naruse and Felix Giovanni Virgo, with additional collaborators. The project detected personally identifiable information in English and Japanese user messages so it could be filtered before further processing. High recall and lightweight inference were priorities, and privacy requirements meant training on synthetic examples rather than real users’ PII.[5]

The system used Llama 3.1 8B as a sidecar model, running alongside the model producing user-facing responses. Small classifiers predicted PII classes for each token, using either the sidecar’s raw hidden activations or features extracted by a sparse autoencoder (SAE). The selected SAE probe used a random forest classifier; the best raw-activation probe used XGBoost. Token-level predictions enabled precise filtering, while evaluation measured whole PII entities, such as a full address.[5]

The joint report found SAE probes more robust than its other probe approaches when transferring from synthetic training data to real production data. It reported 96% F1 for SAE probing versus 51% for using the same Llama 3.1 8B model as a black-box judge, and costs 10–500 times lower than the compared GPT-5 Mini and Claude Opus 4.1 judging configurations. These were PII-task results. A fine-tuned DeBERTa baseline matched the SAE probe on exact entity boundaries and exceeded it on partial coverage; probes offered the separate advantage of adding detection tasks with negligible marginal inference cost. Rakuten deployed the SAE probes.[5]

Deng’s broader assessment of SAEs was qualified. In the February 2026 interview, she reported that raw-activation probes often performed better on downstream detection tasks, suggesting that SAE concept spaces were not always as clean or accurate as expected. Rakuten was an important exception: she attributed its useful generalization partly to the noisier data setting. She treated the comparison as dependent on the task and data rather than a universal ranking of methods.[1][5]

Deng framed alignment and safety around models that follow human intent and remain open to oversight. As she put it in the interview, “we want to make sure that we can build models that do what we want them to do and that we have scalable oversight into what these models are doing.” She connected that view to concrete technical problems such as reward hacking and keeping humans involved in the deployment of increasingly capable agents.[1]

She also described ongoing work to detect hallucinations through internal signals of uncertainty or user-pleasing behavior, and to investigate mitigation within the model. This was a project under development in February 2026, without a reported performance figure or claim that hallucinations had been eliminated.[1]

Deng recalled healthcare institutions approaching Goodfire when the company was about three months old, asking for help understanding scientific models they had trained. The small team had not used those models before and began investigating them. She distinguished explanations needed for patient-facing predictions from the possibility of extracting previously unknown scientific information. For image and video models, she argued that directly visible features could shorten the interpretability feedback cycle compared with language features that require automated interpretation.[1]

Those February 2026 discussions followed Deng’s earlier work on modeling infrastructure and Goodfire’s experimental interfaces for exposing and manipulating model features.[2][3][4][1]

Before Goodfire, Deng built large-scale simulation modeling infrastructure at Two Sigma as a quant engineer and product lead. She holds an MBA and an MS in Computer Science from Stanford and a BS in Computer Science from Columbia, graduating summa cum laude.[2][6]

In September 2024, Deng coauthored Understanding and Steering Llama 3 with Sparse Autoencoders with Thomas McGrath, Daniel Balsam, and Eric Ho. McGrath conceived and led the research; the team contributed implementation and writing. The preview combined Llama-3-8B-Instruct with an SAE so users could inspect features and intervene on them. Its interface and API were deprecated in February 2026.[3]

The team chose a conservative ReLU SAE, prioritizing simplicity and development speed over investigating concerns about high-frequency features in newer architectures. Training on activations from LMSYS-Chat-1M produced features better suited to chat interventions than the tested non-chat data. Steering changed selected feature strengths, decoded them back into activations, and retained the reconstruction error—the portion the SAE had not captured—before continuing inference through the language model.[3]

Deng also coauthored the May 27, 2025 Paint With Ember release. Nick Cammarata, Mark Bissell, and Nam Nguyen were marked as core contributors; the other coauthors were Max Loeffler, Eric Ho, Liv Gorton, and Daniel Balsam. The tool let users generate and edit images by manipulating SDXL-Turbo’s internal activations through a canvas.[4]

A BatchTopK SAE exposed granular visual features, and non-negative matrix factorization grouped them into per-image concepts, such as a dog or grassy field. Painting, dragging, and erasing changed those concepts’ strengths and positions across image patches. After user testing revealed unintuitive conflicts between open-ended prompts and spatial edits, the team restricted prompts to style guidance. These were shared project mechanisms and interface decisions.[4]

Alongside this research work, Deng helped organize Apart Research’s Women in AI Safety Hackathon in March 2025, an event encouraging women and underrepresented groups to contribute to AI safety.[6]