Is AI a speculative bubble? I honestly don’t know. There’s indeed tremendous value to the underlying technology, but that doesn’t mean that it’s not currently a bubble.

The numbers are huge. A few US tech companies are spending (combined) close to a trillion dollars per year in AI (compare to 3-4 trillion of total capex in the US). Their own revenues are a few trillion per year, so they can so far afford it. The yearly revenue for AI is difficult to find; it is growing, but it hasn’t hit yet a hundred billion yearly.

AI feels very much like an arms race where the big players just want a seat at the table, with survival at stake. Unless AI revenues scale dramatically, the arms race is a better explanation than return on investment, even with a five year horizon. My gut tells me that a 10x jump in AI revenues would translate to massive layoffs, which would produce a massive demand gap.

I honestly have no idea how it will turn out and I’m not betting on either possibility. In this article I want to explore some tech consequences of AI being a bubble, in the same spirit of Paul Graham’s article, where he analyzes what the web got right, even if it the early web was a massive bubble.

Two quotes from Paul’s essay really stand out in our context:
“The fact is, despite all the nonsense we heard during the Bubble about the “new economy,” there was a core of truth. You need that to get a really big bubble: you need to have something solid at the center, so that even smart people are sucked in. (Isaac Newton and Jonathan Swift both lost money in the South Sea Bubble of 1720.)”

“The prospect of technological leverage will of course raise the specter of unemployment. I’m surprised people still worry about this. After centuries of supposedly job-killing innovations, the number of jobs is within ten percent of the number of people who want them. This can’t be a coincidence. There must be some kind of balancing mechanism.”

Bubbles seem to be moments where majorities forget about fundamentals and pursue upside. Downturns are moments of pondering on fundamentals. So let’s have some fun and ponder on the fundamentals during a moment where not many are thinking about fundamentals.

If AI turns out to be a bubble, what did the bubble still got right?

1. Markdown

Text is powerful. Markdown is a great vessel for text: enough formatting to make it more readable, but still extremely portable and easy to edit. The fact that most LLMs are trained on markdown is a vindication of the power of text and open, lightweight formats, as the vehicle and repository of most knowledge.

2. Generative AI

Definitely useful for a lot of things. Even if we hit a snag and LLMs don’t get any more useful than they are now, they have completely revolutionized software development, many forms of writing, translation and image/video generation.

3. Allow non-coders to create software

This one is huge. Personal software has only existed for those programmers who were motivated enough to create their own scripts. This ability is now democratized: I predict that before the end of the decade, most people of working age will have little pieces of personal software that they use on a regular basis to solve problems for themselves. And this is a great thing, because those humans will be empowered by software that fits their need like a glove.

4. Software built faster

Software development has always been far slower than what us impatient humans expect. AI has dramatically lowered the amount of time required to ship software. Whether that software is any good is still the topic of fierce discussion. But what’s clear is that even when used carefully and conservatively, AI increases the speed of software development. This will likely not go away.

5. More channels to break up information siloes

MCP and their ilk, and the software industry rushing to create integrations for them, open up new channels to your data. They allow agents to go into siloed information and extract it and combine it in novel, freer ways. This shifts power from those owning the apps to those using them. This is in essence what happened with APIs. As long as companies keep on supporting them, users will be relatively more empowered. Perhaps this only happens at times of high uncertainty, where companies need to (paradoxically) lower the castle walls to survive, then raise them when times are more conservative.

6. Lower the market price of slop

Before AI, organizations created a ton of slop. Most slideshows, internal memos, emails and public announcements were repetitive, content-free exercises that wasted the attention of everyone who was subjected to them. But organizations being ritualistic, these slop artifacts were (and still are) considered essential. Before AI, slop required proof of work: a human still had to handcraft those artifacts and show a certain level of competence at it. Now, the vast majority of that slop can be generated, at a higher level of quality, by a badly thought out and misspelled prompt. Because these artifacts were so unoriginal and following preestablished form, AI is able to nail them with just a trace of intent from the prompter.

My anarchistic, optimistic mind hopes that with the price of slop becoming zero, organizations will begin to part ways with it.