AI is the new Printing Press (another trite take)

Jul 03, 2026

It has been a while since I last sat down to do personal writing.

I was taking a shower today and had a neat idea sneak into my brain. With a semi-quiet day and an hour or two before my next obligation, I figured I would sit down and jot it out.

I should say up front that this is what this forum is for: these kinds of thoughts. For the more interesting blogs, research, and essays related to my work with Zerg, I will direct the reader to zergai.com, where we will soon be publishing a veritable flood of material we have been working on. I have made the mistake of mixing this channel with that one. No longer. We have a blog for those things.

So this will be less formed. It will still be run through my usual "Fake Idan" virtual editor, so I can promise the reader some degree of AI-enabled sheen, but my process is still basically:

Write out a borderline train-of-thought set of ideas

Go back and shape it into a coherent mass

Run it through my "Fake Idan" AI digital twin and argue with it until we get to some kind of first draft

Manually review and edit, then rinse and repeat with step 3 until it feels good to publish

Dear reader, I will first take a moment to share what has waylaid me from my usual shower-thought-driven musings and woolgathering.

Things (work, life, kids) have been crazy.

I am not going to be the eleventieth-million-and-first person to throw another shovelful onto the ever-growing pile of hype and praise about how amazing AI has been for me day to day. How much I can accomplish. How much more it feels like I have to do. But since I have already said those things, I will admit they are true. Zerg has been, as the kids like to say, "on a tear."

Every founder I know will always tell you that today is the best day ever, so I have learned to identify a tear by how much I am slipping up. When I am not slipping up, we are not on a tear. A tear is specifically defined by having so much going on that you have something of value to lose, and inevitably you will make some bad calls, lose sleep, chew glass, and so on just to attempt recovery (all the while things are actually going well probably).

Anyways. Enough preamble.

I want to talk about AI. Artificial intelligence, not to be mistaken for its other meaning: Actually Idan. More specifically, I want to talk about how I have come to view AI as the next step change in humanity's dominion over the natural sciences, technology, our planet, and potentially beyond.

I will do so by comparing it to the printing press.

Not original, you say? What is anymore? Either way, do not worry. I have a twist.

Many have compared AI to the printing press, but the comparison is usually socioeconomic. Before the printing press, books were limited in number and not available to the average person. One reason was the means of production: books required manual labor and armies of scribes who replicated manuscripts by hand.

The printing press removed many of those constraints. To be precise, printing did not come from nowhere. Woodblock printing already existed broadly in Europe, and movable type had a much earlier history in Asia. But Gutenberg's printing press changed the economics by orders of magnitude in the environment where it clicked. Demand, technology, auxiliary conditions, and the Latin alphabet all combined to make book production and distribution possible at a scale that had not existed before. The economic-history version of this argument is laid out well in Jeremiah Dittmar's work on the printing press and European city growth.

Why China and Asia did not see the same kind of resulting techno-economic boom is also a fascinating (and controversial) subject, alas for another day.

This is usually the argument people make when they say LLMs are like the printing press. Software used to be a manual endeavor. You needed large groups of specialized people building systems one character at a time. Now AI makes that cheaper and faster. You can apply the same argument to accounting, writing, music, art, and almost any other domain.

I am not going to argue that point here. It is mostly true. But I think there is a more sober version of this argument, dare I say a more responsible way to think about it, one that does not sound like "AI is coming for us" any more than the internal combustion engine, electricity, or any other technological step change did.

AI and LLMs, really deep learning, allow us to do for intelligence and creativity what aerodynamics did for flight. Planes are not birds. But the important part is not birdness. It is the physics of lift. As NASA's basic aerodynamics guide puts it, lift is an aerodynamic force produced by motion through air, with wings and airfoils doing most of the work by shaping and turning that flow. At the limit, if you force an airfoil through air fast enough, at the right geometry and angle, you produce lift.

In the same way, the empirical lesson from modern language models is not that they are little humans in a box. It is that loss, capability, and task behavior scale with data, model size, and compute. The scaling-laws literature showed power-law improvement in language-model loss with scale, and GPT-3 showed that sufficiently large autoregressive language models could perform many tasks from instructions and examples alone. Force a ton of data through a model with sufficient capacity, and you get a probability distribution that approaches creativity and intelligence.

This is not meant to cast doubt on these technologies. It is just a more responsible way to think about them. At the limit, AI can approximate human thought and intelligence. That is not to say it is human thought and intelligence. But in a world largely measured and valued by the ability to process and transform information, the approximation can become functionally equivalent.

But here is where I want to step back. When have human beings not been doing this?

Go back to the origin of language. Language unlocked epistemic transfer between generations without the need for genetic material. Genetic material is itself a form of this transfer, only slower and governed by evolution. Increasingly, the modern picture is richer than random mutation filtered by selection: the extended evolutionary synthesis, niche construction, developmental bias, and stress-induced mutagenesis literatures all point to ways organisms can shape the search space of evolution, steering which variants are likely to appear and which environments their descendants inherit (lookin' nice!).

Regardless, the origin of life requires some ability to transfer oneself to the incoming generation. Or, at minimum, to distill from the trajectory of one life information that is valuable to future generations.

We experience this every day. We try to be good examples to our children and communities. We write books and long ambling Substack posts so that current or future generations may regard us highly, or at least benefit from the neuronal firings that we got to experience during our brief moment in this plane of existence.

Viewed from a more utilitarian lens, a scratch on the bark of a tree, hieroglyphics etched into sandstone, stories of heroism and defeat passed mouth to mouth, and books written by hand are all versions of the same thing: the forward propagation of information for the advancement of the human condition. Or, more bluntly, the advancement of human dominion over nature, science, technology, and ultimately our own existence.

So now, 20 minutes in, I will make the core argument.

AI is like the printing press because the printing press did not create either the demand or the supply for the underlying good it produced. The printing press did not invent books. It did not invent the nature of books as token generators. In the same way, AI did not invent the token generator. It made the means of generation economically de minimis.

Let me take a step back. Again.

Putting aside the means of production, the technology we recognize as "a book" is much more complex than bound paper. Imagine you had a magic box that, when asked a question, could search through every word and passage ever written, then answer not by novel generation but by quoting the right passage from the right book.

There is an old movie called Explorers, one of my favorite movies of all time. Fun fact: it is also the movie where both River Phoenix and Ethan Hawke made their silver screen debuts. In the movie, kid aliens make contact with kids on Earth. They initially communicate using clips from TV that they had been receiving ambiently through radio transmissions from the planet. Having binged a lifetime equivalent of MASH, they are keen to meet kids from Earth, so they incept the designs for a zero-inertia locomotion technology in the dreams of some random children, who then figure out how to build a spaceship with it and end up going to outer space to meet their new friends.

If you have not watched it, I highly recommend it.

In other words, at the limit, such a black box would be no different from ChatGPT or an unbounded human intelligence to a bounded observer. The box did not write the original books, but you would never have the capacity to read all of those books and hold the box accountable by saying, "Hey, you are just quoting this book on page X."

I am being critical for the sake of argument. Such a box would obviously still be infinitely valuable. The point is simply that in terms of token generation, nothing new has been invented in the context of our magic quote-a-book box.

And this is the point.

Before AI, and before the printing press, books were already token generators. In picking up a book, you were tuning into the token generation of people living long before you. Even before books, human and proto-human language were token generators. And even before language, the propagation of genetic material for the forward propagation of information and experience was also a form of token generation. As such, the printing press simply made the distribution of tokens radically more efficient, in the same way that AI is making generation and recombination radically more efficient today.

I would argue that nearly every technological step change can be viewed through this lens. It is simplistic, I admit. I understand the math and theory underpinning modern LLMs, and the complexity of serving them to ever-growing demand: continuous batching, tensor parallelism, KV cache, Flash/Paged Attention, quantization, monitoring, and all the other production mechanics that sit between a model and a usable product. There is so much reality where the rubber meets the road that I would be the last person to say we can boil all of this down to token transfer.

With that said, I wanted to take a different angle on AI as it continues to drive revolution after revolution in domains we thought were impervious to technological disruption.

AI is an inevitable technological step change that furthers our ability, as human beings, to forward-propagate our experiences and intelligence.

No ambling deep dive is complete without a few predictions, but I will really make only one.

Language ultimately made a deep biological impact on the human animal. The careful way to say this is that humans appear to have a specialized, functionally differentiable language network. Fedorenko, Ivanova, and Regev call this a "natural kind" in the brain: a left-lateralized, language-selective network that is separable from neighboring perceptual, motor, and cognitive systems. That is close to what Anderson and Lightfoot meant by a "language organ": not a literal new anatomical organ, but a functional cognitive faculty shaped by human biology. It is built through evolutionary modification of older primate systems, but it is still one of the clearer differentiators between us and other mammals. As such, I anticipate something similar will occur as a result of AI.

[truncated for AI cost control]