I stumbled across Distill.pub back in 2017 while trying to understand how neural networks actually see images. You know that moment when a concept finally clicks? That's what happened. But it wasn't the code that made it click. It was the interactive diagrams — the ones where you could drag a slider and watch the activation layers shift in real time. I remember thinking: why doesn't every ML paper do this?
Five years later, I'm still asking that question.
Most machine learning papers are PDFs. Static. Dense. Full of equations that assume you already know the notation. They're written for reviewers, not for people who want to understand. Distill took a different approach entirely. It treated machine learning explanation as a design problem, not just a writing problem. And while the journal stopped publishing new articles in 2021, the approach it pioneered is more relevant now than ever — especially with AI tools making interactive content creation actually feasible for regular practitioners.
Let me explain why Distill mattered, what made it different, and how you can apply its principles today without needing a team of web developers.
What Was Distill, Actually?
Distill launched in 2016 as an online journal focused on clarity in machine learning. Not new algorithms. Not benchmark scores. Clarity. The founders — Chris Olah, Shan Carter, and a few others — had backgrounds at Google Brain and had already built a reputation for creating stunning visual explanations of neural networks. They wanted a venue where the quality of explanation mattered as much as the quality of research.
Here's what set it apart: every article was a web-native experience. You'd get interactive diagrams where you could hover over neurons and see what activated them. You'd get side-by-side comparisons where sliders controlled parameters. You'd get prose that actually walked you through the intuition before hitting you with the math. It was like the difference between reading a cookbook and watching someone cook while they explain why they're doing each step.
The journal published roughly 25 articles over five years. That's not many. But each one took months to produce. The team had dedicated designers, developers, and editors working alongside researchers. According to a retrospective by the editors in 2021, the production cost per article was simply too high to sustain without institutional backing. They tried. It didn't work out.
But the format? The format was never the problem.
Why Static PDFs Are Holding Machine Learning Back
I've read hundreds of ML papers. Maybe thousands. Most of them I've forgotten within a week. The ones I remember? They had something visual. Something I could play with.
The standard academic paper format was designed for print journals in the 1900s. It assumes your reader is sitting with a physical copy, underlining passages with a pen. But machine learning concepts are dynamic. Gradient descent is a process. Attention mechanisms are relationships between tokens. Convolutional filters are pattern detectors that activate differently depending on input. None of this is well-served by a static figure with a caption.
There's a deeper problem too. PDFs hide complexity behind notation. When an equation fills half a page, only a tiny fraction of readers will work through it. Most will skip it and hope the text summary is enough. Interactive explanations don't eliminate the math — but they give you a ladder to climb up to it. You can see what the equation does before you parse every symbol.
This isn't just my opinion. Research on science communication consistently shows that interactive visualizations improve comprehension of complex systems. A 2018 study in the Journal of Science Communication found that readers who used interactive graphics retained significantly more conceptual understanding than those who viewed static equivalents. The effect was strongest for topics involving dynamic processes — which is basically all of machine learning.
Distill proved the concept. The problem was always the production cost. But that's changing.
What Distill Got Right About Explanation Design
If you go back and read Distill articles now, you'll notice patterns. The team had a philosophy, not just a format. I've reverse-engineered what I think were their core principles:
1. Progressive disclosure. They never dumped everything on you at once. An article would start with a simple interactive element — maybe a single neuron responding to handwritten digits. Then they'd layer on complexity. By the time you reached the technical depth, you'd already built intuition through exploration.
2. Multiple representations of the same concept. A mathematical formula would appear alongside a visual diagram, a textual description, and often an interactive widget. Different people learn differently. Distill didn't make you choose.
3. Frictionless experimentation. The best Distill articles let you change inputs and immediately see outputs. No setup. No installing libraries. No cloning repositories. You moved a slider and the network responded. That immediacy is what builds understanding faster than any amount of reading.
4. Narrative structure. These weren't reference documents. They told a story. There was a question, an exploration, moments of surprise, and a resolution. You felt like you were discovering something alongside the author, not being lectured at.
These principles aren't proprietary. Anyone can use them. The barrier has always been execution — turning a research insight into a working interactive widget requires skills most researchers don't have. But the tooling gap is closing fast.
How AI Tools Are Making Distill-Style Content Actually Achievable
Here's where things get interesting. In 2021, building an interactive ML explanation meant knowing JavaScript, D3.js, and probably some WebGL. Today, AI coding assistants can generate that code from natural language descriptions. I've built interactive network visualizations in an afternoon that would have taken me a week three years ago.
The workflow looks something like this: you describe the concept you want to visualize, the AI generates the HTML/CSS/JavaScript, you tweak the parameters, and you embed the result in a blog post or course material. Tools like Claude, ChatGPT, and Cursor can all do this. The quality varies — you'll need to iterate — but the baseline is shockingly high.
There are also purpose-built platforms emerging. Observable notebooks let you create reactive documents where code, text, and visualization live together. They're essentially what Distill would have built if they'd started in 2024 instead of 2016. The learning curve is still there, but it's measured in days, not months.
This is where AI-Mind fits into the picture. If you're creating educational content about machine learning — tutorials, course materials, documentation — you don't need to write prompts from scratch for every interactive element. AI-Mind has content-type templates specifically for technical explanations. You describe what you want to show, and it handles the structure. The first 30 generations are free, which is enough to build several interactive components and see if the approach works for your audience.
The point isn't that one tool solves everything. The point is that the bottleneck has shifted. It used to be technical ability. Now it's clarity of thinking. If you can explain a concept clearly in words, AI tools can help you turn that explanation into something interactive. Distill's vision is suddenly accessible to anyone who teaches machine learning.
The Catch: Why Most ML Explanations Still Suck
I should be honest about something. Having the tools doesn't mean you'll automatically create good explanations. I've seen plenty of AI-generated interactive content that's technically functional but pedagogically useless. Sliders that don't teach anything. Diagrams that look pretty but confuse more than they clarify.
The hard part was never the code. Distill's team spent months on each article not because the JavaScript was complicated, but because they were iterating on the explanation itself. They'd test versions with readers, watch where people got confused, and redesign accordingly. That process doesn't get faster with AI. It might even get harder — because when generation is cheap, the temptation is to publish your first draft instead of your tenth.
There's also a discoverability problem. Distill articles rank well in search because the journal built authority over years. If you publish an interactive ML explanation on your personal blog, nobody will find it unless you're intentional about SEO. The content might be better than a Distill article, but if Google can't parse your interactive elements, you'll lose to static PDFs that have more backlinks.
So yes, the tools are here. But the craft still matters. Maybe more than ever.
Where This Is All Heading
I think we're going to see a resurgence of Distill-style content in the next few years. Not from a single journal — that model proved unsustainable — but from individual creators, research labs, and companies that want their work to actually be understood. The economics have flipped. What cost $50,000 per article in 2019 might cost $500 in 2025, if you know what you're doing.
The formats will evolve too. I'm seeing people experiment with AI-generated video explanations that adapt to viewer questions in real time. Interactive textbooks where every diagram responds to student input. Documentation that doesn't just describe an API — it lets you call the API from inside the documentation and see results immediately.
Distill closed its doors, but the idea it championed — that machine learning deserves better explanations — is more alive than ever. The difference now is that you don't need a team at Google Brain to make it happen. You just need clarity of thought, a willingness to iterate, and the right tools.
If you're teaching ML concepts to anyone — students, colleagues, users of your product — try this: take one concept you explain regularly. Something you've explained a dozen times. Now build an interactive version of that explanation. A slider. A diagram that responds to input. Something the learner can play with. See if it changes how quickly people understand. I'm betting it will.
That's what Distill taught me. Not that interactive articles are nice to have. That they're fundamentally better at doing what explanations are supposed to do: make the complex feel obvious.
Sources: Distill.pub editorial retrospective, "Reflecting on Five Years of Distill" (2021); Journal of Science Communication, "Interactive Visualizations and Conceptual Understanding in Science Education" (2018); Personal experience building interactive ML explanations using AI coding assistants (2023-2025).