The Promise of Distil.pub

Research has a distribution problem. Not in the sense of getting papers onto arXiv or into journals - we solved that decades ago. The problem is making research comprehensible. Making it stick. Making it change how people think rather than just adding to their ever-growing list of papers to skim.

Distil.pub tried to fix this. From 2016 to 2021, it stood as a beacon for what research communication could become: interactive, visual, and designed for human understanding rather than academic credentialing. When I look back at what they accomplished, and what we lost when they went on hiatus¹, I can't help but think we had something extraordinary within reach.

The Old Way of Sharing Ideas

Academic publishing in 2025 looks remarkably similar to academic publishing in 1925. PDFs with dense text, static figures, and the occasional table. The format demands that authors cram complex, dynamic concepts into linear narratives constrained by page limits and print-era conventions.

This creates what Chris Olah and Shan Carter called "research debt"² - the accumulated cost of ideas that are poorly explained, undigested, or wrapped in bad abstractions. When you read a typical machine learning paper, you're often left with questions: What does this algorithm actually do? How sensitive is it to hyperparameters? What happens when you change this assumption?

The traditional format actively fights against understanding. You get a few cherry-picked examples, maybe some ablation studies if you're lucky, and a wall of mathematical notation that may or may not correspond to how the authors actually implemented their method. The paper becomes a kind of elaborate ritual where the goal is to demonstrate technical sophistication rather than transfer genuine understanding.

I spent years reading papers this way. You develop a skill for extracting the key insights from dense academic prose, but it's inefficient and lossy. Most of the actual understanding happens when you implement the method yourself, or when someone creates a good blog post explaining it in plain language.

What Distil.pub Actually Did

Distil.pub wasn't just about making papers prettier. They fundamentally rethought what research communication could be in an interactive medium.

Take their article on feature visualization in neural networks³. Instead of static images of what neurons detect, you could manipulate the optimization process in real-time. Want to see how a "cat detector" responds to different inputs? Drag the sliders and watch the visualization update. The concepts weren't just described - they were explorable.

Or consider their piece on t-SNE⁴. Rather than reading about hyperparameter sensitivity, you could adjust the perplexity setting and immediately see how it changed the embedding. The authors didn't just tell you that "different parameter values produce different results" - they let you experience the parameter space yourself.

This wasn't just pedagogically superior. It was a different kind of thinking tool. When you can manipulate a model's parameters and see the results instantly, you develop intuitions that are impossible to get from static descriptions. You start to understand not just what the algorithm does, but how it feels to use it.

The visual style mattered too. Clean typography, thoughtful use of color, animations that revealed rather than distracted. Everything was designed to reduce cognitive load and let you focus on the ideas themselves. Reading a Distil article felt like having a conversation with someone who deeply understood both the technical content and how to communicate it clearly.

The Economics of Explanation

But here's the problem: creating content at Distil's level of quality was expensive. Not financially expensive - though it did require significant time and design resources - but expensive in opportunity cost for researchers.

A typical Distil article took months to produce. The "Building Blocks of Interpretability" piece involved custom interactive diagrams, novel visualization techniques, and extensive iteration on both the technical content and the presentation. That's months that could have been spent on traditional research, grant applications, or career advancement.

The academic incentive structure doesn't reward this kind of work. Explaining someone else's research brilliantly gets you far less career credit than publishing incremental novelty in a top-tier venue. Even explaining your own research this well doesn't count for much beyond the original paper.

This is the tragedy of research debt. The people best positioned to create excellent explanations - the researchers who deeply understand the work - have the least incentive to do so. Meanwhile, the people with time to focus on explanation often lack the technical depth to do it justice.

Distil tried to solve this by treating explanatory work as first-class research contribution. They offered peer review, DOIs, and academic credibility. But ultimately, they were fighting against an entire ecosystem optimized for different priorities.

The LLM Moment

Which brings me to why this matters right now. We're in the middle of the most rapid research boom in machine learning history. New architectures, training techniques, and applications appear weekly. The rate of genuine insight mixed with incremental publication has never been higher.

Most of this work is communicated in the traditional format: dense papers optimized for getting through peer review rather than building understanding. The research debt is accumulating faster than ever.

But we also have something we didn't have in 2016: large language models that can help with the interpretive labor. LLMs might actually make Distil's vision feasible to scale.

Imagine being able to take any recent paper and automatically generate an interactive explanation in the Distil style. Not just a summary or a reformatting, but a genuine exploration tool where you could:

• Manipulate the model's inputs and see how outputs change
• Explore the hyperparameter space with live feedback
• Compare different approaches side-by-side with shared datasets
• Generate synthetic examples that probe the method's boundaries

The technical building blocks for this exist today. We have models that can understand research papers, generate interactive visualizations, and even implement algorithms from descriptions. What we're missing is the framework - the systematic approach to turning static research into dynamic understanding tools.

The Distribution Problem, Solved

This isn't just about making research more accessible to newcomers, though that would be valuable. It's about fundamentally changing how we think about and validate research itself.

When research is presented interactively, weaknesses become immediately apparent. Claims about robustness crumble when readers can easily probe edge cases. Methodological shortcuts become obvious when the approach doesn't generalize to user-generated examples. Interactive presentation naturally encourages more rigorous research.

It also changes how insights spread through the community. Instead of playing telephone through citation chains, concepts could be transmitted with much higher fidelity. Researchers could build genuine intuition about techniques before attempting to extend them.

The compound effects would be enormous. Better understanding leads to better research questions. More researchers with deeper intuitions about existing methods means more sophisticated extensions and applications. The field's bandwidth for genuinely novel work increases when we're not constantly re-learning poorly communicated basics.

What We're Building Toward

LLMs give us the opportunity to revisit Distil's vision with different economics. What if the interpretive labor could be largely automated? What if transforming research into interactive, explorable form became as easy as running a script?

We're not there yet, but the pieces are falling into place. Models can already analyze research papers and generate explanations. Interactive visualization libraries are more powerful and easier to use than ever. The web platform for rich media continues to improve.

The missing piece is intentional design - understanding that the goal isn't just to make research "more interactive" but to create genuine thinking tools. Tools that help both experts and newcomers develop better intuitions about how these systems actually work.

Distil.pub showed us what this could look like when done with extraordinary care and attention. Now we have the opportunity to make it systematic, scalable, and sustainable. The question is whether we'll choose to build it.

The research debt keeps accumulating. But for the first time since Distil went on hiatus, we have the tools to start paying it down at scale. The promise isn't just better research communication - it's research that compounds more effectively because understanding spreads with higher fidelity.

We just have to decide we want it.