[Epistemic status: half-baked musing that I’m writing down to clarify for myself]

For the past 15 years there’s been an economic niche, where a single programer develops a useful tool, utility, or application, and sells it over the internet to a few thousand people for a small amount of money each, and make a decent (sometimes passive or mostly-passive) living on that one-person business.

In practice, these small consumer software businesses are on the far end of a continuum that includes venture-backed startups, and they can sometimes be the seed of an exponentially scaling operation. But you only need to reach product-market fit with a few thousand users for a business like this to sustainable. And at the point, it might be mostly on autopilot, and the entrepreneur has income, but can shift most of their attention to other projects, after only two or three years.

Intend (formally complice), is an example of this kind of business from someone in my circles.

I wonder if these businesses will be over soon, because of AI.

Not just that AI will be able to do the software engineering, but that AI swarms will be able to automate the whole entrepreneurial process from generating (good) ideas, developing early versions, shipping them, getting user-feedback, and iterating.

The discourse already imagines a “one person-unicorn”, where a human CEO coordinates a company of AIs to provide a product or service. With half a step more automation, you might see meta-entrepreneurs overseeing dozens or hundreds of separate AI swarms, each ideating, prototyping, and developing a business. Some will fail (just like every business), but some will grow and succeed and (just like with every other business venture) you can invest more resources into the ones that are working.

Some questions:

• How expensive will inference be, in running these AI entrepreneurs? Will the inference costs be high enough that you need venture funding to run an AI entrepreneur-systems?
  • Estimating this breaks down into roughly “how many tokens does it take to run a business (per day?)?” and “How much will an inference token cost in 2028?”
• What are the moats and barriers to entry here? What kind of person would capture the gains to this kind of setup.
• Will this eat the niche of human-ideated software businesses? Will there be no room left to launch businesses like this and have them succeed, because the space of niche software products will be saturated? Or is the space of software ideas so dense, that there will still be room for differentiation, even if there are 1000x as many products of this type, of comparable quality, available?

. . .

In general, the leverage of code is going to drop over the next 5 years.

Currently, one well-placed engineer will write a line of code that might be used by millions of users. That because there’s 0-marginal cost to replicating software and so a line of code written once might as well be copied to a million computers. But it’s also representative of the relative expense of programming labor. Not many people can write (good) code and so their labor is expensive. It’s definitely not worth paying $100 an hour for an engineer to write some software when you can buy existing off the shelf software that does what you need (or almost what you need) for $50 a month.

But, as AI gets good enough that “writing code” becomes an increasingly inexpensive commodity, the cost-benefit of writing custom software is going to shift in the “benefit” direction. When writing new software is cheap, you might not want to pay the $50 a month, and there will be more flexibility to write exactly the right software for your particular usecase instead of a good-enough off the shelf-version (though I might be overestimating the pickiness of most of humanity with regards to their software). So more people and companies will write custom software more of the time, instead of buying existing software. As that happens the number of computers that run a given line of code will drop, in the process.