Black Boxes and the One Percent

I never imagined I would see a shift happen right under my nose.

I used to read about how the industrial revolution changed the way things are built. I used to ask myself why handmade things are so much more expensive. I knew the answer, but only from stories. This time I am experiencing it first hand, and I can see how some big companies are reluctant to change.

Yes, I’m talking about AI. But no, I’m not talking about companies being hesitant to use AI to speed up development. They do use it. They are even ecstatic about it. The problem is that the perspective doesn’t change.

Employers, upper management and non-technical people are excited that AI can speed up delivery by, let’s say, 10x. And of course they don’t really care about the quality underneath, like the code and the infrastructure.

At the same time, the tech people are divided, and that division will consume the progress. Early adopters, and some who were skeptical before, adopt AI massively. They vibe code, they deliver fast, and they get addicted to the shortcuts. Meanwhile the old school techies are overwhelmed by the quality of what AI produces. Some of them push back on the idea entirely, citing security and quality.

Big companies don’t seem able to address this effectively and efficiently, which means the drift will become a blocker in the long term. They spend a lot on tokens at the enterprise level and hand them out to every engineer, and each engineer uses their own approach because the company failed to produce a uniform way of working and a shared harness. The amount of money burned this way is large.

At the same time, solo developers who embrace AI and use it as it is can build the same thing at a fraction of the cost, for a fraction of a million.

Let me break it down.

A solo builder uses a non-enterprise plan, which is much cheaper than paying per token. Yes, you will bring up privacy and security. It’s true. But does your entire product contain secrets and very unique logic? No. I would say 80% of any company is not proprietary.

The other thing is that you should not expect AI to consistently deliver the quality of the top 1% of expert, experienced humans. History repeats itself. No matter how advanced the machinery gets, we will always appreciate, admire and value a handmade product, for the same reason: an expert human is special. The same machine, though, can produce up to the quality of the average human.

So how do we approach this differently?

In software development, we can’t evaluate AI output line by line. We won’t be able to keep up, and we will be overwhelmed. I advocate seeing it this way instead.

Treat the output as a bunch of black boxes with very clear contracts. Introduce a very strict harness, something that was nearly impossible in the past and is now much easier to build with the help of AI. Linters, tests, whatever. Throw everything you know at making sure these black boxes behave correctly.

But the interconnections between the black boxes should always stay under the tight supervision of that 1% human. If you don’t have anyone like that on the team, find one. We still need that kind of person to ensure the entire output is secure and tolerable. I repeat, tolerable. Why? Be realistic.

The focus should now shift to architecture and harness, not syntax or super elegant code. Code smells still need to be taken care of, and never lower your guard on them. Contracts and visual representations are what will help those architects most, so lean on them more, and get yourself familiar with every kind of testing and with CI/CD.

This is how, in a few years, we will read more stories about solo developers outcompeting big companies.