Distill: a modern machine learning journal

Distill was an online machine learning journal, founded in 2017 and published by a small group of researchers including Chris Olah, Shan Carter, and Ludwig Schubert, with Google Brain and later OpenAI connections among its contributors. It was hosted at distill.pub and ran until it announced a pause in 2021. That's the short answer. The longer answer is the one people actually need, because Distill's defining feature — every article shipped with custom JavaScript, live figures, and interactive diagrams — is also the reason its archive doesn't age well.

Here's the problem you're probably hitting. You found a Distill article, it's exactly the explanation you needed for a concept, and half the page is dead. Sliders that don't slide. Figures that render as blank boxes. A notebook link that 404s. This piece covers what Distill was, why its interactive format was both its signature and its structural weakness, and how to extract the durable ideas from a broken page — including a specific worked example you can follow with a named article and a named notebook.

What was Distill, exactly?

Distill published peer-reviewed articles on machine learning, but it broke from the academic journal template in ways that mattered. Articles were written in web-native form: scroll-driven animations, inline interactive diagrams, and figures you could manipulate rather than just look at. The editorial standard was clarity over formalism — an article succeeded if a reader understood the mechanism, not if it survived a reviewer checklist.

The journal's best-known pieces became reference material for a generation of people learning deep learning. If you've seen an animated diagram of a neural network's weights shifting as you drag a slider, there's a decent chance it originated on Distill. The format wasn't a gimmick layered on top of the writing. The interactivity was the explanation — the reader was meant to build intuition by manipulating the thing, not by reading a description of the thing.

That design choice is what made Distill distinctive, and it's also what makes the archive fragile. A static PDF from 2017 still opens. A JavaScript-dependent interactive figure from 2017 depends on a browser environment, a dependency tree, and a hosting arrangement that all have to keep working.

Why did the interactive format turn out to be so hard to sustain?

Three pressures compound, and understanding them tells you what to expect from any interactive research format — not just Distill's.

Maintenance cost scales with interactivity. A text article is done when it's published. An interactive article is done when it's published and then needs ongoing attention every time a browser API changes, a library deprecates a function, or a CDN goes away. The more custom code an article carries, the more surface area there is to break. A journal with a handful of editors cannot maintain a growing archive of bespoke JavaScript indefinitely.

Authors pay an unusual tax. Writing a normal paper means writing. Writing a Distill article meant writing and building a small software project — custom visualizations, often a companion notebook, sometimes a whole interactive demo. That's a real barrier. It selects for authors who happen to have both research depth and front-end engineering skill, which is a narrow intersection.

Static publishing is a solved problem; interactive publishing isn't. Academic infrastructure — DOIs, PDF repositories, citation managers — assumes static documents. An interactive article doesn't fit cleanly into any of it. You can cite it, but you can't archive it the way you archive a PDF, because the PDF is a degraded copy of the thing that made it valuable.

None of this means the format was a mistake. It means the format had a cost that the surrounding incentive structure didn't pay for. That's a general lesson about any medium where the delivery mechanism is doing explanatory work.

A worked example: extracting value from a broken Distill page

Abstract advice is useless here, so let's do this concretely. Take one of Distill's most cited articles, "Feature Visualization" (Olah, Mordvintsev, Schubert, 2017), which explains how to generate images that show what individual neurons in a neural network respond to. The article is built around interactive figures — you're meant to manipulate optimization parameters and watch the generated image change.

Now suppose the interactive figures on that page no longer render for you. Here's the recovery path, step by step.

• Find the companion notebook. Distill articles frequently shipped with a linked Colab or Jupyter notebook containing the actual code behind the figures. The notebook is often more durable than the embedded figure, because it runs in a maintained environment rather than depending on the article's own JavaScript. For "Feature Visualization," the companion code is the thing you want — it lets you reproduce the visualization yourself instead of watching someone else's.
• Read the figure captions as standalone text. Distill's captions were written to be self-sufficient, because the editors knew readers might print the article. The caption for a feature-visualization figure typically states what parameter was varied and what changed in the output. That's the finding, stated in prose, independent of whether the animation plays.
• Check the article's bibliography for the static version. Many Distill pieces had a companion paper or a version of the same result in a conventional venue. The static version won't have the interaction, but it will have the claim.
• Look for the archived copy. Web archives preserve the HTML and often the rendered figures at a point in time. It's not a live page, but a snapshot of a working one is frequently enough to read the diagrams.

The general rule this illustrates: separate the article's claim from its delivery mechanism, and recover the claim first. The claim in "Feature Visualization" — that you can optimize an input image to maximize a neuron's activation, and that the resulting images reveal what the neuron is tuned to — survives completely without the interactive figures. The interactive figures made the claim easier to feel. They were never the claim itself.

This is the part I'd push back on if someone told me it as a general principle without evidence. It sounds like a tidy aphorism — "prefer snapshots to dashboards" — and aphorisms are usually hiding a trade-off. The trade-off here is real: recovering the claim from a broken interactive article costs you the intuition the interaction was designed to build. You get the fact. You lose the understanding-by-manipulation. That's a genuine loss, not a rounding error. The recovery path above gets you the finding; it does not get you the feeling of having discovered it by dragging a slider.

What Distill's shutdown actually tells you about research publishing

The journal announced it was pausing in 2021, and the reasons were structural rather than dramatic. Interactive publishing is expensive in a way that doesn't show up on a citation count. The editors who built the format were doing work that academic incentives don't reward — maintaining infrastructure, reviewing code-heavy submissions, and keeping an archive alive.

If you're choosing where to publish or where to read, the practical takeaway is about durability, not prestige. A conventional paper in a repository with a DOI will still be readable in twenty years. An interactive article will be readable for as long as someone keeps paying the maintenance cost. That's not an argument against interactive formats — it's an argument for treating them as perishable and archiving the durable parts yourself.

For readers, that means: when you find an interactive explanation that works, save the claim, not just the bookmark. Copy the key figure caption. Note the parameter and the result. Grab the notebook if there is one. The page may not be there when you come back.

For anyone producing explanatory content today, the same tension applies. The tools have shifted — a lot of the prompt-writing and formatting overhead that used to sit between an idea and a published explainer can now be handled by a zero-prompt generator like AI-Mind, where you describe what you want and pick a content type rather than engineering the prompt yourself. That lowers the cost of producing the text. It does nothing about the cost of maintaining an interactive figure for a decade, which is the cost that actually killed the format.

Where this advice breaks down

Two honest limits.

First, the recovery path assumes a companion notebook or a static companion paper exists. Not every Distill article had one. For pieces where the interactivity was the entire argument and no notebook was published, there may be no clean recovery — you're left with captions and whatever the web archive captured. That's a real dead end, not a solvable problem.

Second, if your actual goal is to use a technique rather than understand it, recovering the claim from a broken article is the wrong move entirely. Go find a current implementation. A 2017 interactive explanation of feature visualization is not the tool you'd use to visualize features today; it's the explanation you'd read to understand what the tool is doing. Confusing those two goals wastes time.

Key Takeaways

• Distill was an online ML journal founded in 2017 by researchers including Chris Olah and Shan Carter, hosted at distill.pub.
• Its interactive figures were the explanation, not decoration — which is exactly why the archive breaks as browsers and libraries change.
• To recover value from a dead interactive page, find the companion notebook first, then read figure captions as standalone text.
• Separate the article's claim from its delivery mechanism: the claim usually survives, but the intuition-building interaction does not.
• Interactive research formats are perishable; archive the durable claim yourself rather than trusting the bookmark.

Sources

• Olah, Mordvintsev, Schubert, "Feature Visualization," Distill, 2017. Explains how to generate images showing what individual neurons respond to, with companion code.
• Distill, "Distill Update," 2021. The journal's announcement that it was pausing publication.
• AI Tool Database, internally verified snapshot, 2026. Pricing and capability records for 360 AI tools, most recently verified 2026-09-24. Note: this database does not cover Distill or research-publishing tools.