Pixels are free now
Software is following media's path...others will follow too.
One of the main disruptions of the internet was that it removed the friction of distribution, at least for knowledge. Previously, getting any kind of word out was hard and expensive, and usually had some kind of gatekeeper attached - a TV station, printing press, or movie studio. Lots of businesses had a built-in assumption about this friction, even if they didn’t think they did. An easy one to see is the bundling of newspapers and advertising, which fairly completely fell apart as soon as the frictionless internet made the bundle redundant. There are more subtle ones, places where inefficient distribution of information let one party take advantage of another - car sales, finance, travel, etc.
The catchphrase was something like “information wants to be free” but it was more accurate to say that distribution had become free.
AI is doing something similar, but to pixels. We assume that it’s hard to put a pixel in front of an end user. What kind of pixel? Any kind - images, movies, but also business reports, search results, application UI - any digital good has an assumption baked in that there is cost, friction and a human bottleneck involved. AI is breaking those assumptions.
Don’t be fooled by the weak solutions - the early internet didn’t work at all for things like streamed movies. We got there. Business reports and writing aren’t great? Sure, whatever, those pixels will be free soon enough - the tech will catch up.
It’s easiest to see this in two areas: visual content, and software. Visual content (pictures and video) are almost literally “free pixels” now (not really, but relative to the amount of work it would have taken to produce them a few years ago, they’re rounding to 0).
Software is a bit behind, but the vibe coding energy is the start of the wave. This is very similar to the moment that Blogger showed the world that anyone could post something on the internet. We are about to get an explosion of small, weird, funky, one-off software solutions that the more formal practitioners (e.g. us software engineers) are going to HATE. But it won’t matter - the gatekeeper is being removed, those pixels are getting free, the frothing market is coming.
But this will keep going. Almost anything that ends in a consumed pixel on a screen has baked into it cost assumptions that are breaking, the way the internet broke media. There are going to be all kinds of hard to predict effects here - there might be network effects and influencers, software as fashion or ideology, super marginal use cases that someone puts insane effort into, like YouTube channels meant to educate a small community.
In some cases, the gatekeepers have legal or regulatory protections - those pixels will take longer to become free. But people will start to work around those restrictions - in the internet it was things like Napster or BitTorrent. We are already seeing people use LLMs for legal or medical advice that is not sanctioned or official. This will likely increase, and put pressure on those gatekeepers. Look to the internet for clues as to how this might play out - it’s a good historical analog.
It’s easy to argue against some of this, and sometimes hard to see. But much of what we take for granted now - the abundance of media, connectivity, information, and interaction - barely worked for even elite users at one point. If you look at elite users now, you can start to see where the future is going. And it looks like pixels are going to be mostly free.
Ben Thompson coined the term Aggregation Theory many years ago to describe the effect the digital/Internet revolution had on taking distribution marginal costs to zero, now the AI/LLM revolution will take the production cost of digital goods to zero. It will be an amazing time to be an entrepreneur!
Agree with the sentiment, but disagree with the terminology. Calling AI-generated content an aggregation of 'pixels' is akin to calling a jet plane an aggregation of 'atoms'. There are myriad ways to put atoms together, but there's something special required to put them together into something that is useful to someone. I'd like to think that there's a better term we can find to represent a bit of knowledge that we can compose together to produce a larger bit of useful knowledge. That would seem to be better than adding that semantic onto a term (pixel) with existing and distinct semantics.
In the past you've discussed the idea of replacing a document with a form of knowledge that someone could query or have a conversation with. I'd think that that document would be exactly an aggregation of these 'knowledge bits'. I can imagine extracting various forms of textual or visual representations of that knowledge (e.g. pixels), but the knowledge itself is something different.
Maybe part of the problem is that we currently build huge monolithic models of knowledge and ask the system to extract some of that knowledge into a directly-consumable form given a context/query, whereas I could imagine this being separated into phases (extract/generate a model representing the answer domain (like your document above), and from that generating an answer, but with the ability to query that model further). This opens up questions about how to extract, represent, and compose these knowledge models (what's the api?).
In any case, I think it's worth the effort to use a different term to describe this new elemental form of knowledge. I do think this is a distinct concept from a pixel or a bit, and I also believe that having a new term to describe a new concept facilitates thinking and discussion around that new concept.
Caveat: I'm actually not that familiar with how the model/context/query pieces work today, so these thoughts are based on my partial understanding / current mental model.