# Distill (journal)

> Source: https://aiwiki.ai/wiki/distill
> Updated: 2026-11-18
> Fact-checked: 2026-11-18
> Categories: AI Research, Interpretability, Machine Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Distill (journal)." aiwiki.ai, 18 Nov 2026. https://aiwiki.ai/wiki/distill
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Distill** was an open-access, peer-reviewed online journal of [machine learning](https://aiwiki.ai/wiki/machine_learning) research that published web-native articles, many of them built around interactive diagrams, between 2016 and 2021.[1][2] It was founded by [Chris Olah](https://aiwiki.ai/wiki/chris_olah) and Shan Carter, then both at [Google Brain](https://aiwiki.ai/wiki/google_brain), and was formally launched on March 20, 2017 in collaboration with [OpenAI](https://aiwiki.ai/wiki/openai), [DeepMind](https://aiwiki.ai/wiki/google_deepmind), YC Research and others.[2][3] The journal was meant to give academic standing to explanations, visualizations and other work that did not fit in a PDF, and it published a series of [interpretability](https://aiwiki.ai/wiki/interpretability) papers, including "Feature Visualization", "The Building Blocks of Interpretability", "Activation Atlas" and the multi-part Circuits thread.[1][4][5] On July 2, 2021 its editors announced a one-year hiatus that might be extended indefinitely, citing volunteer burnout and doubts about whether a journal was the right structure; as of November 2026 the site still describes Distill as being on an indefinite hiatus and no longer accepting submissions.[1][6][7]

## History

### Origins and launch (2016-2017)

Distill's first articles appeared in the second half of 2016, before the journal existed in a formal sense. "Attention and Augmented Recurrent Neural Networks" by Olah and Carter is dated September 8, 2016, and was followed by "How to Use t-SNE Effectively" (Martin Wattenberg, Fernanda Viégas and Ian Johnson), "Deconvolution and Checkerboard Artifacts" and "Experiments in Handwriting with a Neural Network" before the end of the year.[8] The editors later wrote that Distill "operated informally for several months before launching itself as a journal", and the journal's review policy states that articles published before March 20, 2017 were reviewed informally, without an official policy.[5][7]

The formal launch was announced on March 20, 2017 in coordinated posts. Writing on the Google Research blog, Carter and Olah described Distill as "a new open science journal and ecosystem supporting human understanding of machine learning", built in collaboration with OpenAI, DeepMind, YC Research and others, and said it was "an independent organization".[2] Their argument was that machine learning had many conferences and journals but no venue dedicated to explanatory and interactive work, partly because traditional venues could not, by virtue of their medium, carry interactive visualizations; without such a venue, they wrote, many contributions did not count as "real academic contributions".[2] The launch package had three parts: a research journal, prizes for outstanding work, and tools for building interactive articles.[2]

Michael Nielsen, then a research fellow at YC Research, announced the launch on the [Y Combinator](https://aiwiki.ai/wiki/y_combinator) blog, calling Distill "an interactive, visual journal for machine learning research" and naming Olah and Carter as founding editors.[3] He listed the original steering committee as [Yoshua Bengio](https://aiwiki.ai/wiki/yoshua_bengio) (University of Montreal), Mike Bostock (creator of D3), Amanda Cox (The New York Times), [Ian Goodfellow](https://aiwiki.ai/wiki/ian_goodfellow) (Google Brain), [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy) (OpenAI), Shakir Mohamed (DeepMind), Nielsen himself and Fernanda Viégas (Google Brain).[3] DeepMind's own post, by Shakir Mohamed and Helen King, said DeepMind was a contributing sponsor of the Distill Prize.[9]

Two days after the launch, Distill published "Research Debt" by Olah and Carter, an essay that set out the journal's rationale.[10] It compared the struggle of reaching research-level understanding of a field to climbing a mountain, and argued that this difficulty was a form of debt, made up of poor exposition, undigested ideas, bad abstractions and notation, and noise.[10] The authors used the term "interpretive labor" for the effort that either an explainer or an audience must spend for an idea to be understood, and called the opposite of research debt "research distillation".[10] They argued that the "research distiller" should be a recognized role alongside theoreticians, experimentalists and research engineers, and presented the Distill journal, the $10,000 Distill Prize and Distill's publishing infrastructure as a first attempt at an ecosystem to support that role.[10]

### Early years (2017-2018)

In 2017 the journal published "Why Momentum Really Works" by Gabriel Goh, "Feature Visualization" by Olah, Alexander Mordvintsev and Ludwig Schubert, "Sequence Modeling with CTC" by Awni Hannun, and "Using Artificial Intelligence to Augment Human Intelligence" by Carter and Nielsen.[8] A Nature blog post in April 2018 reported that only 10 papers had appeared so far, all with DOIs, and quoted Olah: "We're not publishing that many papers. We're trying to create a prototype."[11] The same post said nearly 2,500 of the 93,351 readers of a recent Olah article (2.6%) had interacted with its accompanying code notebook.[11]

In an editorial update published on August 14, 2018, the editors reported that Distill had had over a million unique readers and more than 2.9 million views, and that its papers had been cited 23 times on average according to Google Scholar.[5] They noted that this would place Distill in the top 2% of journals indexed by Journal Citation Reports by that metric, while calling the comparison "somewhat unfair" because a small journal can more easily publish mostly impactful papers.[5] The update also acknowledged a low publication volume (12 papers at the time) and announced several changes: separating mentorship from the evaluation of articles, a clarified review process with a new reviewer worksheet, a Slack workspace for the community, a simpler procedure for appointing "acting editors" when an editor has a conflict of interest, a clearer dual-submission policy, and the addition of Arvind Satyanarayan to the editorial team to oversee a portfolio covering human-computer interaction and AI.[5] The journal's policy page dates the removal of the mentorship policy and the new acting-editor procedure to July 24, 2018.[7]

### Later publications (2019-2021)

Articles published between 2019 and 2021 included "AI Safety Needs Social Scientists" by [Geoffrey Irving](https://aiwiki.ai/wiki/geoffrey_irving) and [Amanda Askell](https://aiwiki.ai/wiki/amanda_askell), "Activation Atlas", "A Visual Exploration of Gaussian Processes", "Open Questions about Generative Adversarial Networks", "Growing Neural Cellular Automata", "Visualizing Neural Networks with the Grand Tour", "Exploring Bayesian Optimization" and "Communicating with Interactive Articles".[8][12] In 2020 Distill also introduced "threads", collections of short related articles published over time: one on circuits in vision models and one on differentiable self-organizing systems.[4][13] The last articles were two companion pieces on [graph neural networks](https://aiwiki.ai/wiki/graph_neural_network), "A Gentle Introduction to Graph Neural Networks" and "Understanding Convolutions on Graphs", published on September 2, 2021, after the hiatus announcement.[8][14]

Distill's main publications carry DOIs numbered from 10.23915/distill.00001 (September 2016) to 10.23915/distill.00033 (the graph neural network introduction), with thread installments and discussion responses given sub-numbers such as 10.23915/distill.00024.001.[12]

### Hiatus (2021)

On July 2, 2021, the editorial team published "Distill Hiatus", drafted by Chris Olah, Nick Cammarata, Sam Greydanus and Janelle Tam.[6] It announced a one-year hiatus "which may be extended indefinitely"; papers already under review were not affected, published threads could keep adding articles, and limited commentary might still appear.[6] Authors were encouraged to keep using the open-source Distill template and to self-publish or submit to venues such as VISxAI.[6]

The essay said the founding theory, that a journal publishing explanations and interactive articles would let their authors benefit from the academic incentive system, no longer seemed correct.[6] The editors wrote that publication in Distill did not appear to change how seriously most institutions took non-traditional work, and that the main bottleneck was the effort and unusual mix of scientific and design skill such articles require, not the lack of a venue.[6] They described specific tensions in the journal format:

- **Mentorship versus judgment.** Editors had sometimes given more than 50 hours of help to a single early article, yet also had to accept or reject the same authors, and deep editing could amount to co-authorship, creating conflicts of interest.[6]
- **Editor articles.** Because the community interested in this style of work was small, a non-trivial share of articles came from editors, which the essay acknowledged had led to criticism and to Distill declining to publish some editor articles.[6]
- **Burnout.** The volunteer team reported significant burnout, which it linked to conflicting goals and very high standards, and said its review process had become slower and more like a typical journal's.[6]

The editors argued that self-publication, either on one-off websites or on a hypothetical "Distill Arxiv", was probably the future for most articles, pointing to [David Ha](https://aiwiki.ai/wiki/david_ha) and collaborators' use of the Distill template on GitHub Pages (for example, the [World Models](https://aiwiki.ai/wiki/world_model) article) as a model for neutral, multi-author self-publishing.[6] They chose a hiatus rather than handing the journal to new editors because they wanted to preserve what they called Distill's "weird and idiosyncratic vision".[6]

## Editorial model

### Scope and article types

Distill described itself as an academic journal in machine learning whose distinguishing trait was "outstanding communication and a dedication to human understanding"; articles often, but not always, used interactive media.[7] It accepted four kinds of work: research articles with novel results; exposition (reviews, tutorials, primers and perspective pieces, with particular interest in explorable explanations); commentary, such as essays on public policy or on science itself; and non-traditional resources and artifacts such as datasets and tools.[7] There were no word limits. The journal warned that a suitable article often took more than 100 hours to write and illustrate.[7]

Reviewers assessed three things: whether an article significantly advanced the research community's dialogue, whether its communication met a very high standard, and its scientific integrity, including whether high-level descriptions matched the results and weaknesses were disclosed.[7] Distill was a primary publication, so it would not publish work substantially identical to something published elsewhere. It did accept papers previously posted to [arXiv](https://aiwiki.ai/wiki/arxiv), on the understanding that Distill was the formal publication, and expanded "journal versions" of workshop or conference papers that substantially advanced on them.[7]

### Review on GitHub

Each submission lived in a GitHub repository, and reviews and editing were handled through GitHub issues.[7] When an article was published, its repository was made public and transferred to the distillpub organization, so the peer reviews of every published article are public.[7] Review had three stages: an editorial assessment of fit, peer review by external reviewers chosen by the handling editor, and open-ended post-publication review in which readers could raise issues; significant unresolved issues could be shown in the article margin.[7] Editors could not handle papers they had written or could not judge objectively, and in those cases an outside "acting editor" from the relevant research community was appointed and named in the public review log.[7]

The 2018 editorial update described this setup as a form of continuous integration for scientific papers: drafts were built automatically and served at password-protected URLs, and after publication readers sent pull requests to fix typos or suggest edits.[5]

### Indexing, licensing and tools

Distill registered an ISSN (2476-0757) and a DOI prefix with Crossref, and its articles were indexed in Google Scholar.[7][3] Articles had to be released under the Creative Commons Attribution license; the code for interactive diagrams was released under the MIT license unless noted otherwise.[7][15] The FAQ said static diagrams were usually drawn in vector tools such as Adobe Illustrator, Sketch or Inkscape and dynamic diagrams were generally built with D3.js.[15]

Articles were written in HTML using Distill's own framework, a CSS framework plus custom web components for bylines, citations, footnotes and appendices. In its simplest form, a Distill post was a single HTML file with one script tag that applied the house styling in the browser.[16] The hiatus essay listed this template, which it said many non-Distill authors use, among the main ways Distill had produced value.[6]

## Distill Prize

The Distill Prize for Clarity in Machine Learning was announced alongside the journal to recognize work that communicated and refined ideas in machine learning and adjacent fields.[17] Prizes were expected to be $10,000 each, from an initial endowment of $125,000 funded by Chris Olah, [Greg Brockman](https://aiwiki.ai/wiki/greg_brockman), [Jeff Dean](https://aiwiki.ai/wiki/jeff_dean), DeepMind and the [Open Philanthropy](https://aiwiki.ai/wiki/open_philanthropy) Project, which also handled the prize's logistics.[17] Eligible work included tutorials, new framings of known ideas, visualizations and explorable explanations, and filtering or summarizing research; people could nominate others but not themselves.[17] The prize page states that the prize is legally separate from the journal, is not funded or awarded by it, and has no preference for work published in Distill.[17]

Prizes were to be awarded annually from 2018.[17] In August 2018 the editors reported that the 2018 prize had not yet been awarded: they had received 59 submissions, several of them long lecture series, and had no good process for evaluating that kind of material.[5] They said they aimed to award it by Thanksgiving 2018.[5]

## Notable publications

The following table lists a selection of Distill articles, with dates and authors as recorded on the articles and in Crossref.[12]

| Date | Title | Authors | Notes |
|------|-------|---------|-------|
| 2016-09-08 | Attention and Augmented Recurrent Neural Networks | Chris Olah, Shan Carter | First Distill article; covers [attention](https://aiwiki.ai/wiki/attention) and augmented [RNNs](https://aiwiki.ai/wiki/recurrent_neural_network)[12] |
| 2016-10-13 | How to Use t-SNE Effectively | Martin Wattenberg, Fernanda Viégas, Ian Johnson | Shows how [t-SNE](https://aiwiki.ai/wiki/t_sne) plots can mislead[18] |
| 2017-03-22 | Research Debt | Chris Olah, Shan Carter | Essay setting out the case for "research distillation"[10] |
| 2017-04-04 | Why Momentum Really Works | Gabriel Goh | Interactive account of [momentum](https://aiwiki.ai/wiki/momentum) in optimization[8] |
| 2017-11-07 | Feature Visualization | Chris Olah, Alexander Mordvintsev, Ludwig Schubert | How [GoogLeNet](https://aiwiki.ai/wiki/googlenet) builds up its understanding of images[19] |
| 2017-11-27 | Sequence Modeling with CTC | Awni Hannun | [Connectionist temporal classification](https://aiwiki.ai/wiki/connectionist_temporal_classification) for sequence modeling[12] |
| 2018-03-06 | The Building Blocks of Interpretability | Chris Olah and six others | Combines interpretability techniques into interfaces[20] |
| 2019-02-19 | AI Safety Needs Social Scientists | Geoffrey Irving, Amanda Askell | [AI alignment](https://aiwiki.ai/wiki/ai_alignment) and the empirical study of humans[30] |
| 2019-03-06 | Activation Atlas | Shan Carter and four others | Map of features learned by InceptionV1[21] |
| 2019-08-06 | A Discussion of "Adversarial Examples Are Not Bugs, They Are Features" | 12 authors | Commentaries plus the original authors' responses[22] |
| 2020-02-11 | Growing Neural Cellular Automata | Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, Michael Levin | "Differentiable model of morphogenesis"[23] |
| 2020-03-10 | Zoom In: An Introduction to Circuits | Chris Olah and five others | Opening article of the Circuits thread[4] |
| 2020-09-11 | Communicating with Interactive Articles | Fred Hohman, Matthew Conlen, Jeffrey Heer, Duen Horng Chau | Design of interactive articles, drawing on education, journalism and visualization[24] |
| 2021-03-04 | Multimodal Neurons in Artificial Neural Networks | Gabriel Goh and seven others | Neurons in [CLIP](https://aiwiki.ai/wiki/clip)[25] |
| 2021-09-02 | A Gentle Introduction to Graph Neural Networks | Benjamin Sanchez-Lengeling, Emily Reif, Adam Pearce, Alexander Wiltschko | One of two final articles[14] |

### Interpretability work

A large share of Distill's best-known articles dealt with how [neural networks](https://aiwiki.ai/wiki/neural_network) represent what they learn. "Feature Visualization" (2017) showed how GoogLeNet, trained on ImageNet, builds up its understanding of images over many layers.[19] "The Building Blocks of Interpretability" (2018) argued that interpretability techniques are usually studied in isolation and explored the interfaces that result from combining them.[20] "Activation Atlas" (2019), by Carter, Zan Armstrong, Schubert, Ian Johnson and Olah, used feature inversion to visualize millions of activations from an image classification network and arranged them into an explorable map of the features the InceptionV1 network had learned.[21]

### The Circuits thread

The Circuits thread, launched on March 10, 2020, asked "What can we learn if we invest heavily in reverse engineering a single neural network?"[4] Its introduction noted that InceptionV1 has around 10,000 unique neurons, a number that a group effort could attempt to study one by one along with the connections between them.[4] The editors described the thread as a "living document" of short articles on individual circuits or small families of features, interspersed with critical commentary from experts in adjacent fields, and said its content should be seen as early-stage exploratory research.[4]

The opening article, "Zoom In: An Introduction to Circuits" by Olah, Cammarata, Schubert, Goh, Michael Petrov and Carter, proposed three claims that would justify studying neurons and their connections directly: that meaningful features exist, that meaningful circuits exist between features, and that features and circuits are universal across models.[4] Later installments, mostly by researchers at OpenAI, covered early vision in InceptionV1, curve detectors, naturally occurring equivariance, high-low frequency detectors, curve circuits, visualizing weights, branch specialization and weight banding; the last of these appeared on April 8, 2021.[4][12] "Multimodal Neurons in Artificial Neural Networks" (March 2021), by the same group with [Alec Radford](https://aiwiki.ai/wiki/alec_radford), studied OpenAI's CLIP model; its author contribution statement describes porting CLIP to [OpenAI Microscope](https://aiwiki.ai/wiki/microscope), the discovery of an original "Spider-Man" multimodal neuron, and the investigation of typographic attacks on CLIP.[25]

Olah's later circuits work at [Anthropic](https://aiwiki.ai/wiki/anthropic) was published outside Distill. Anthropic's Transformer Circuits Thread website states that its project was "Inspired by the Distill Circuits Thread" and that "Previously we would have submitted to Distill, but with Distill on Hiatus", the team was publishing on its own site instead, following David Ha's approach of creating standalone websites.[26] Its first major paper, "A Mathematical Framework for Transformer Circuits" (December 22, 2021), described the Distill Circuits thread as a previous project that had attempted to reverse engineer vision models, and set out to begin a comparable effort for transformer language models.[27]

### Discussion articles and the self-organizing systems thread

In August 2019 Distill tried a new format for scientific debate. In response to the paper "Adversarial Examples Are Not Bugs, They Are Features" by Andrew Ilyas and colleagues at MIT, it published six commentary articles by other researchers, including Justin Gilmer, Dan Hendrycks, Gabriel Goh, Reiichiro Nakano, Preetum Nakkiran and Eric Wallace, together with responses from the original authors, describing the project as an experimental "discussion article", a format it had wanted to try for several years.[22] The hiatus essay later cited this kind of discussion article as evidence that reviewers will spend far more effort when they feel a review genuinely matters to the community.[6]

The Differentiable Self-organizing Systems thread, opened on August 27, 2020 by Mordvintsev, Randazzo, Niklasson, Levin and Greydanus, asked how robust, general-purpose self-organizing systems could be built.[13] It gathered "Growing Neural Cellular Automata", which proposed a differentiable cellular automata model of morphogenesis, and follow-up articles on self-classifying MNIST digits, self-organizing textures and the adversarial reprogramming of neural cellular automata, the last published in May 2021.[13][12]

## Reception and criticism

Early commentators focused on the format. In May 2017 the digital-preservation researcher David Rosenthal examined "How to Use t-SNE Effectively" and concluded that, because its interactive functionality was implemented in JavaScript files loaded with the page, it was in principle as preservable as a PDF; he noted, however, that the Internet Archive's first capture of the article was only partly functional, while a Webrecorder capture worked fully.[28] The 2018 Nature blog post quoted the computational biologist Jennifer Listgarten describing Distill as formalizing something that was already happening informally on blogs, GitHub and notebooks.[11]

The most detailed criticism came from Kent Anderson in The Scholarly Kitchen in June 2018. Anderson described Distill as what appeared to be a Google "house" journal, noting that 21 of its 26 authors to that point were Google employees, that the editors Olah, Carter and Satyanarayan were among its authors, and that its domain was owned by Google; he questioned the launch statement that Distill was an independent organization and said the editors and an advisory board member had not responded to his emails.[29] He also observed that publication had slowed and that no prize had yet been announced.[29] The editors' own 2021 hiatus essay acknowledged that Distill was "sometimes criticized for publishing too many articles with editors as authors", while arguing that editor articles were a strength of the journal.[6]

In assessing its own record, the editorial team concluded in 2021 that Distill's value had come less from its status as a journal than from mentoring authors, providing the Distill template, the articles its own members produced, and the encouragement and community it gave to authors.[6] Anthropic's Transformer Circuits site, for its part, says that interpretability research benefits a lot from interactive articles and that its team would previously have submitted its work to Distill.[26]

## See also

- [Chris Olah](https://aiwiki.ai/wiki/chris_olah)
- [Mechanistic interpretability](https://aiwiki.ai/wiki/mechanistic_interpretability)
- [Interpretability](https://aiwiki.ai/wiki/interpretability)
- [OpenAI Microscope](https://aiwiki.ai/wiki/microscope)
- [TensorFlow Playground](https://aiwiki.ai/wiki/tensorflow_playground)
- [Explainable AI](https://aiwiki.ai/wiki/explainable_ai)
- [Data visualization](https://aiwiki.ai/wiki/data_visualization)

## References

1. Distill. "About". Distill. https://distill.pub/about/
2. Shan Carter and Chris Olah. "Distill: Supporting Clarity in Machine Learning". Google Research Blog, March 20, 2017. https://research.google/blog/distill-supporting-clarity-in-machine-learning/
3. Michael Nielsen. "Distill: An Interactive, Visual Journal for Machine Learning Research". Y Combinator Blog, March 20, 2017. https://www.ycombinator.com/blog/distill-an-interactive-visual-journal-for-machine-learning-research
4. Nick Cammarata, Shan Carter, Gabriel Goh, Chris Olah, Michael Petrov, Ludwig Schubert, Chelsea Voss, Ben Egan, Swee Kiat Lim. "Thread: Circuits". Distill, March 10, 2020. https://distill.pub/2020/circuits/
5. Distill Editors. "Distill Update 2018". Distill, August 14, 2018. https://distill.pub/2018/editorial-update/
6. Editorial Team. "Distill Hiatus". Distill, July 2, 2021. https://distill.pub/2021/distill-hiatus/
7. Distill. "Publishing in the Distill Research Journal". Distill. https://distill.pub/journal/
8. Distill. "Distill" (RSS feed of published articles). Distill. https://distill.pub/rss.xml
9. Shakir Mohamed and Helen King. "Distill: Communicating the science of machine learning". DeepMind Blog, March 20, 2017. https://deepmind.google/blog/distill-communicating-the-science-of-machine-learning/
10. Chris Olah and Shan Carter. "Research Debt". Distill, March 22, 2017. https://distill.pub/2017/research-debt/
11. Anna Nowogrodzki. "Machine learning gets a journal for interactive figures". Nature blogs, April 16, 2018. https://blogs.nature.com/blog/machine-learning-gets-a-journal-for-interactive-figures/
12. Crossref. "Works registered under DOI prefix 10.23915 (Distill)". Crossref REST API. https://api.crossref.org/prefixes/10.23915/works?rows=200
13. Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, Michael Levin, Sam Greydanus. "Thread: Differentiable Self-organizing Systems". Distill, August 27, 2020. https://distill.pub/2020/selforg/
14. Benjamin Sanchez-Lengeling, Emily Reif, Adam Pearce, Alexander B. Wiltschko. "A Gentle Introduction to Graph Neural Networks". Distill, September 2, 2021. https://distill.pub/2021/gnn-intro/
15. Distill. "Distill FAQ". Distill. https://distill.pub/faq/
16. Distill. "How to Create a Distill Article". Distill. https://distill.pub/guide/
17. Distill. "The Distill Prize for Clarity in Machine Learning". Distill. https://distill.pub/prize/
18. Martin Wattenberg, Fernanda Viégas, Ian Johnson. "How to Use t-SNE Effectively". Distill, October 13, 2016. https://distill.pub/2016/misread-tsne/
19. Chris Olah, Alexander Mordvintsev, Ludwig Schubert. "Feature Visualization". Distill, November 7, 2017. https://distill.pub/2017/feature-visualization/
20. Chris Olah, Arvind Satyanarayan, Ian Johnson, Shan Carter, Ludwig Schubert, Katherine Ye, Alexander Mordvintsev. "The Building Blocks of Interpretability". Distill, March 6, 2018. https://distill.pub/2018/building-blocks/
21. Shan Carter, Zan Armstrong, Ludwig Schubert, Ian Johnson, Chris Olah. "Activation Atlas". Distill, March 6, 2019. https://distill.pub/2019/activation-atlas/
22. Logan Engstrom, Justin Gilmer, Gabriel Goh, Dan Hendrycks, Andrew Ilyas, Aleksander Madry, Reiichiro Nakano, Preetum Nakkiran, Shibani Santurkar, Brandon Tran, Dimitris Tsipras, Eric Wallace. "A Discussion of 'Adversarial Examples Are Not Bugs, They Are Features'". Distill, August 6, 2019. https://distill.pub/2019/advex-bugs-discussion/
23. Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, Michael Levin. "Growing Neural Cellular Automata". Distill, February 11, 2020. https://distill.pub/2020/growing-ca/
24. Fred Hohman, Matthew Conlen, Jeffrey Heer, Duen Horng (Polo) Chau. "Communicating with Interactive Articles". Distill, September 11, 2020. https://distill.pub/2020/communicating-with-interactive-articles/
25. Gabriel Goh, Nick Cammarata, Chelsea Voss, Shan Carter, Michael Petrov, Ludwig Schubert, Alec Radford, Chris Olah. "Multimodal Neurons in Artificial Neural Networks". Distill, March 4, 2021. https://distill.pub/2021/multimodal-neurons/
26. Anthropic. "Transformer Circuits Thread". transformer-circuits.pub. https://transformer-circuits.pub/
27. Nelson Elhage, Neel Nanda, Catherine Olsson, et al. "A Mathematical Framework for Transformer Circuits". Transformer Circuits Thread, December 22, 2021. https://transformer-circuits.pub/2021/framework/index.html
28. David Rosenthal. "Distill: Is This What Journals Should Look Like?". DSHR's Blog, May 2, 2017. https://blog.dshr.org/2017/05/distill-is-this-what-journals-should.html
29. Kent Anderson. "Has Google Become a Journal Publisher?". The Scholarly Kitchen, June 11, 2018. https://scholarlykitchen.sspnet.org/2018/06/11/has-google-become-a-journal-publisher/
30. Geoffrey Irving, Amanda Askell. "AI Safety Needs Social Scientists". Distill, February 19, 2019. https://distill.pub/2019/safety-needs-social-scientists/

