Telescopic Content

I find the best non-fiction writing to be fractal, or telescopic. It scales nicely with the amount of attention that you decide to bestow on it, and still makes sense and is tightly packed at every scale.

Christopher Alexander's "A city is not a tree" is a great example of that. It works at a slogan level. Just the title itself in 6 words outlines the key idea. And yet - at the essay level it works too and is not just a lot of extra words without much extra signal. I can also imagine that if Christopher Alexander wrote a book-length exposition - it would still be powerful in its own right. Without feeling like a "a book that could have been a blogpost".

Writing telescopically is difficult. It pushes you to get to the core of your argument in a similar way that Picasso's famous bull exploration did. What are the key lines that make this bull really what it is?

One of the benefits of writing telescopically is clarity. But the other is - stickiness. Because the author goes through the work of compressing and decompressing the thought, the reader has it easier. They don't have to invent their own shorthands and heuristics if they don't want to. The author has already provided them.

Reading can also be telescopic. For example, when you skim over a page and pick out the headlines - before diving in. Or when you ask an LLM to summarize the content for you. That is the shrinking part of it. But you can also expand: when you read commentaries to the original text. Or if you search for other references from footnotes, or read the author's biography to understand the context.

Telescopic by default... or by design

In the age of LLMs, all content can and will be telescopic by default. LLMs can shrink and expand any text as you wish - either according to their default ideas of what's important, or according to your specific instructions. The upside is obvious: the intentionality of attention that you are willing to give as a reader, and the assistance that you get as a writer in distilling and expanding the argument. But there are downsides as well. Namely, because the attention that you are using to shrink or expand is not exactly yours - there is a real danger of enshittification and tiktokification of content becoming a new norm. Washed out, generalized TL;DRs and AI-generated summaries that kill the spirit of the original text. And if this kind of text is what you grow up reading, it shapes not just your knowledge, but your attention and most importantly - your taste. Once our tools shape our taste - there is no going back.

Still, the genie is out of the box. Someone somewhere must have made a browser extension that makes any content you land on shrink or expand to taste. All content is telescopic by default - whether you as a writer want that or not. So I feel like the only option we have as writers - is to embrace this. And build telescopic by design.

That's what I'm trying to do here. This essay, as well as all other essays on this website - are from now on telescopic. You can decide how much attention you are going to give them. And they scale accordingly - based on the prompt and examples that I provide as the writer, giving me at least some degree of control on how my content will be fractalized.

From the technology standpoint, building it was not hard, but there were a few interesting choices. For example, how do I make sure that the generated alternative still sounds as much like me as possible? For now it's just careful prompting and supplying all the examples to a large context window model along the task and hoping for the best. There is probably fine-tuning coming later, but I'm not yet convinced that it's better, given that all the content fits into the context window.

Another interesting challenge was UX - how do we present the different "zoom" levels. And how do we show to the user the "natural", default one - the way it was written originally. Also as a writer - I want to have a way of "approving" a summarized or expanded version - then it becomes "official" (as indiciated by a green dot). For now the way this works is very simple - when a new version is autogenerated - I receive an email with this new version. I can then edit it, add it to the repo and redeploy. And now this version will have a green dot, signifying that this version has been author-approved.

Should you be interested, you can see the implementation on github. It's all very basic and mostly written by Claude for me with some guidance. But it works as a proof of concept.

The magic of free lunch

The most joyful moments of all of this for me as a writer were when I got an expanded version of something - and it had new and powerful ideas or parallels that I haven't seen myself. That was absolutely magical and felt like free lunch. Because it sort of was, I guess. Is it inherently less valuable as a result? Does the amount of work you have to put in somehow make the output truly better? Does the process count? I am not sure about this yet. Most of the time I end up rewriting pretty much everything that LLMs write for me. But there are words, sentences, sometimes even full paragraphs that are starting to stick.

Now, if you take this to its logical conclusion - you don't even need an essay to start with, do you? All you need is a title. So as a continuation of this experiment - I've created a system where 404 pages on this website are a little different. Try it out. You can put any title after the slash. It will generate a new essay (5 min version) based on the title and my other writings. And it will save it and display it for everyone to see. It will also send me an email so that I can look at it, edit and add to the official corpus if I choose so. In the future there will need to be a better way of doing this - where I can log in as the author and approve / edit / delete the generated content. But that's for later.

So all content is from now on infinite. You can zoom in and zoom out at will. You can imagine what a non-existant essay on a topic would look like. How much "reference data" is enough for this to actually work? My 20-ish essays provide a surprisingly decent (at times) training corpus. When I read some of the autogenerated stuff - it does actually sound very much like me if I was desperately trying to make it work a certain title. Check this one out for example. Written completely from the title given in the URL of a 404 page. No edits. It's not always this good. It's a bit of a hit and miss. But it starts to work.

Which begs a question:

What would a blogging platform look like if it was reimagined and built from the ground up for the world of infinite telescopic content, a world in which readers prompt the writer's doppelganger to generate what they really want to read, and the real writer then comes in and edits it to their liking?

And is it time someone builds this?