Why BUILD isn't good at machine learning

2024-01-19

One thing I like about software development is that it has, or at least seems to have, a relatively low barrier to entry. In theory, you can set up a business if you are armed with nothing more than a laptop and an internet connection. If you give yourself some luxuries like some programming know-how and a few hundred dollars for cloud services and APIs, you can change the world.

It was certainly possible to change the world with these meager resources in this way in the nineties and early aughts. Though I agree that it's gotten much harder, I'd argue it's still possible now. Software and writing are the two things I know that allow me to create matter like this.

Machine learning inverts how software is developed. Traditional programs are rules that produce data when run. Machine learning models are rules derived from data. I can write rules, but I cannot manufacture data. Success in the machine learning space depends on data, which is capital, and capital is a barrier to entry.

All BUILD has right now is talent and labor. We try, unsuccessfully, to get into machine learning, and only now do I understand why we keep failing. Capital has forcefully re-asserted that it is a critical factor of production after twenty years of us being able to pretend otherwise. Maybe it's a failing on our part that we can't work around our lack of capital, but it is doubtlessly much harder to get started. We just haven't been able to defeat that inertia.