Nick Cammarata

Nick Cammarata is a machine learning researcher who works on mechanistic interpretability, the study of how the internal components of neural networks compute. At OpenAI he was a co-author of the 2020-2021 Distill "Circuits" thread with Chris Olah and colleagues, first author of its articles on curve detectors, and a core contributor to OpenAI's 2023 paper "Language models can explain neurons in language models," whose contributions statement credits him with the idea that started the project.^[2][4][15] He later joined the interpretability company Goodfire, where in October 2026 he published the essay Neural Networks Are Already Clean, which argues that most neurons in modern networks are individually interpretable.^[20][21]

Circuits research at OpenAI (2020-2021)

Cammarata's early interpretability work centered on InceptionV1 (GoogLeNet), an image classifier that the Circuits collaboration tried to understand neuron by neuron. The collaboration was published as an invited Distill thread, "Thread: Circuits," which lists Cammarata among its authors.^[12] The opening essay, "Zoom In: An Introduction to Circuits" (10 March 2020), was written by Olah, Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov and Shan Carter.^[2] Its author-contributions note says Olah primarily wrote the text and that Cammarata "was deeply involved in developing the framing and revising the final text." On the research side, it says Cammarata "took up this line of research to characterize much larger and deeper circuits, greatly expand the number of neurons we understood mechanistically, and performed detailed, rigorous characterizations of curve detectors," and that he introduced the essay's connection to systems biology.^[2]

Cammarata was lead author of two of the thread's articles. "Curve Detectors" (17 June 2020) studied a family of InceptionV1 neurons that respond to curves. Its contributions section credits him with the radial tuning curve, with proposing to render attribution vectors of hand-labeled dataset examples, and with linking the neurons to classical curve detection; it also notes that he hand-labeled more than 800 images for one experiment.^[4] "Curve Circuits" (30 January 2021) reverse-engineered how those neurons are built from earlier line and edge detectors and used the result to hand-write an "artificial artificial neural network" that reimplements the algorithm. Cammarata and Olah wrote its text.^[7]

He was also co-author of "Multimodal Neurons in Artificial Neural Networks" (4 March 2021), a study of OpenAI's CLIP model that reported "the existence of multimodal neurons in artificial neural networks, similar to those found in the human brain." The article marks him as an equal contributor. According to its contributions section, he drew the connection to multimodal neurons in the human brain, which became the article's overall framing, labeled more than 1,500 images for its conditional-probability plots, found that negative pre-ReLU activations are often interpretable, and wrote the sections on identity and emotion neurons.^[9]

The Circuits papers he co-authored are:

"Multimodal Neurons" was published as a standalone Distill article, not as part of the Circuits thread.^[9][12] "Zoom In" set out three speculative claims about features, circuits and universality, and gave examples of a polysemantic neuron and of a circuit it titled "Cars in Superposition," in which a feature appeared to be spread across many neurons (see superposition).^[2] In his 2026 essay Cammarata wrote that he and Olah found the "cars in superposition" circuit shown in "Zoom In" in 2019.^[20]

Distill

Cammarata is listed as staff on the Distill journal's "About" page.^[14] He was one of four people, with Olah, Sam Greydanus and Janelle Tam, who drafted the editorial "Distill Hiatus" (2 July 2021), which announced that the volunteer-run journal would stop publishing for a year, possibly indefinitely, citing editor burnout and doubts about the journal model.^[13]

Automated interpretability at OpenAI (2023-2024)

On 9 May 2023 OpenAI published "Language models can explain neurons in language models" by Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu and William Saunders. Cammarata is marked as a core research contributor.^[15] The method showed an explainer model, GPT-4, pairs of tokens and activations for each MLP neuron in GPT-2 XL and had it write a natural-language explanation. A simulator model then predicted the neuron's activations from that explanation, and the explanation was scored by how well the simulated activations matched the real ones.^[15] The contributions statement says "Nick effectively started the project by having the initial idea to have GPT-4 explain neurons, and showing a simple explanation methodology worked." It also credits him, with William Saunders, with the initial codebase and the neuron viewer, and with an early proof of concept for explanation revisions.^[15]

The paper reported that "out of a total of 307,200 neurons, 5,203 (1.7%) have top-and-random scores above 0.7 (explaining roughly half the variance)."^[15] OpenAI released the code and a dataset of explanations for every GPT-2 XL neuron.^[16] Cammarata later described the work as "openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had."^[1]

Cammarata is also one of the eleven listed authors of Transformer Debugger, an OpenAI tool released on GitHub in 2024. Its README says it was developed by OpenAI's Superalignment team and that it combines automated interpretability with sparse autoencoders to investigate behaviors of small language models.^[17]

Goodfire

When Goodfire announced its $7 million seed round on 15 August 2024, its press release quoted Cammarata, describing him as "a leading interpretability researcher formerly at OpenAI." He said: "There is a critical gap right now between frontier research and practical usage of interpretability methods. The Goodfire team is the best team to bridge that gap."^[18]

He is first-listed author and a core contributor on Goodfire's 27 May 2025 post "Painting with concepts using diffusion model latents." It launched Paint With Ember, a tool that lets users generate and edit images by painting sparse-autoencoder features onto a canvas, and Goodfire open-sourced the underlying SAE, which was trained on one layer of SDXL-Turbo, a 3.5-billion-parameter latent diffusion model.^[19]

Neural Networks Are Already Clean (2026)

On 15 October 2026 Goodfire published Cammarata's essay "Neural Networks Are Already Clean," with contributions from Curt Tigges (language neuron research) and Tarun Gogineni (editing).^[20][21] The essay argues that the canonical evidence for polysemanticity and superposition in early vision models was misread. On his account, much of the apparent tangling came from weak tools, older architectures and a pattern he calls "recursive generalization," and most neurons can be studied directly without first decomposing the network.^[20] It says that with ordinary architectural improvements, and no change to model size or training data, "99-100% of neurons in that last layer are easily interpretable" in an InceptionV1-sized vision model, and that in "preliminary audits of Qwen3-32B" the team estimates "roughly 90% of neuron poles have one main job." The essay describes using AI agents at scale to map neurons and circuits and to test explanations, and says a technical companion paper will follow.^[20]

In his post announcing the essay, Cammarata wrote that he had come away from the early OpenAI work "pessimistic, because neural networks appeared to fold into arbitrary tangled knots," and that "with modern agentic tools and a much closer look, the picture looks totally different."^[1] In an interview with MIT Technology Review published 19 October 2026, he said that people had taken the claims too far: "It's not infinitely good news; it's not everything solved." The same article quoted skeptical responses from Stanford's Chris Potts, Jacob Hilton of the Alignment Research Center and Neel Nanda of Google DeepMind.^[22] The essay's arguments and reception are covered in detail in Neural Networks Are Already Clean.

Selected works

See also

• Chris Olah
• Mechanistic interpretability
• Goodfire AI
• Neural Networks Are Already Clean
• Polysemanticity
• Superposition
• Circuit discovery
• OpenAI Microscope

References

1. ^1 ^2Cammarata, N. Post on X announcing "Neural Networks Are Already Clean," 15 October 2026. x.com/...2110757819427574056. Accessed 27 October 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S. "Zoom In: An Introduction to Circuits." *Distill*, 10 March 2020. doi:10.23915/distill.00024.001. distill.pub/...zoom-in. Accessed 27 October 2026.
3. ^Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S. "An Overview of Early Vision in InceptionV1." *Distill*, 1 April 2020. doi:10.23915/distill.00024.002. distill.pub/...early-vision. Accessed 27 October 2026.
4. ^1 ^2 ^3 ^4Cammarata, N., Goh, G., Carter, S., Schubert, L., Petrov, M. and Olah, C. "Curve Detectors." *Distill*, 17 June 2020. doi:10.23915/distill.00024.003. distill.pub/...curve-detectors. Accessed 27 October 2026.
5. ^Olah, C., Cammarata, N., Voss, C., Schubert, L. and Goh, G. "Naturally Occurring Equivariance in Neural Networks." *Distill*, 8 December 2020. doi:10.23915/distill.00024.004. distill.pub/...equivariance. Accessed 27 October 2026.
6. ^Schubert, L., Voss, C., Cammarata, N., Goh, G. and Olah, C. "High-Low Frequency Detectors." *Distill*, 27 January 2021. doi:10.23915/distill.00024.005. distill.pub/...frequency-edges. Accessed 27 October 2026.
7. ^1 ^2 ^3Cammarata, N., Goh, G., Carter, S., Voss, C., Schubert, L. and Olah, C. "Curve Circuits." *Distill*, 30 January 2021. doi:10.23915/distill.00024.006. distill.pub/...curve-circuits. Accessed 27 October 2026.
8. ^Voss, C., Cammarata, N., Goh, G., Petrov, M., Schubert, L., Egan, B., Lim, S. K. and Olah, C. "Visualizing Weights." *Distill*, 4 February 2021. doi:10.23915/distill.00024.007. distill.pub/...visualizing-weights. Accessed 27 October 2026.
9. ^1 ^2 ^3 ^4Goh, G., Cammarata, N., Voss, C., Carter, S., Petrov, M., Schubert, L., Radford, A. and Olah, C. "Multimodal Neurons in Artificial Neural Networks." *Distill*, 4 March 2021. doi:10.23915/distill.00030. distill.pub/...multimodal-neurons. Accessed 27 October 2026.
10. ^Voss, C., Goh, G., Cammarata, N., Petrov, M., Schubert, L. and Olah, C. "Branch Specialization." *Distill*, 5 April 2021. doi:10.23915/distill.00024.008. distill.pub/...branch-specialization. Accessed 27 October 2026.
11. ^Petrov, M., Voss, C., Schubert, L., Cammarata, N., Goh, G. and Olah, C. "Weight Banding." *Distill*, 8 April 2021. doi:10.23915/distill.00024.009. distill.pub/...weight-banding. Accessed 27 October 2026.
12. ^1 ^2"Thread: Circuits." *Distill*, 2020. doi:10.23915/distill.00024. distill.pub/...circuits. Accessed 27 October 2026.
13. ^Distill Editorial Team. "Distill Hiatus." *Distill*, 2 July 2021. doi:10.23915/distill.00031. distill.pub/...distill-hiatus. Accessed 27 October 2026.
14. ^"About." *Distill*. distill.pub/about. Accessed 27 October 2026.
15. ^1 ^2 ^3 ^4 ^5 ^6Bills, S., Cammarata, N., Mossing, D., Tillman, H., Gao, L., Goh, G., Sutskever, I., Leike, J., Wu, J. and Saunders, W. "Language models can explain neurons in language models." OpenAI, 9 May 2023. openaipublic.blob.core.windows.net/...paper. Accessed 27 October 2026.
16. ^OpenAI. "automated-interpretability" (code and datasets). GitHub. github.com/...automated-interpretability. Accessed 27 October 2026.
17. ^1 ^2Mossing, D., Bills, S., Tillman, H., Dupre la Tour, T., Cammarata, N., Gao, L., Achiam, J., Yeh, C., Leike, J., Wu, J. and Saunders, W. "Transformer Debugger." GitHub, 2024. github.com/...transformer-debugger. Accessed 27 October 2026.
18. ^Goodfire. "Goodfire Raises $7M to Break Open the Black Box of Generative AI Models." Business Wire press release, 15 August 2024 (republished by Silicon UK). silicon.co.uk/...black-box-of-generative-ai-models. Accessed 27 October 2026.
19. ^1 ^2Cammarata, N., Bissell, M., Nguyen, N., Loeffler, M., Ho, E., Deng, M., Gorton, L. and Balsam, D. "Painting with concepts using diffusion model latents." Goodfire, 27 May 2025. goodfire.com/...painting-with-concepts. Accessed 27 October 2026.
20. ^1 ^2 ^3 ^4 ^5 ^6Cammarata, N. "Neural Networks Are Already Clean." Goodfire, 15 October 2026. already-clean.goodfire.com (plain-text version: already-clean.goodfire.com/essay.md). Accessed 27 October 2026.
21. ^1 ^2Goodfire. "Research." goodfire.com/research. Accessed 27 October 2026.
22. ^Heaven, W. D. "'We were looking at it wrong': The explosive claim that's upended the AI world." *MIT Technology Review*, 19 October 2026. technologyreview.com/...thats-upended-the-ai-world. Accessed 27 October 2026.