Why AI Creates a Monopoly, Not a Democracy: Runaway Growth and Pandemic Lessons

I've been thinking a lot lately about AI, a subject I explored back in grad school and postdoc. In this episode, I walk through a hackathon project to democratize AI by distributing neural network training across home computers. I explain why it failed: first the obvious reason, a bad actor can silently poison a distributed model, then a subtler one rooted in the physics of computation. I break down how neural networks work, why a single nonlinearity lets them approximate almost anything, and why these models need the low latency bandwidth only a data center can provide.

From there I connect to something bigger: why AI is destined to concentrate power into a few organizations, and why that oligopoly could tip into an outright monopoly. I use the mathematics of pandemic spread to show how a tiny edge in AI self improvement compounds into a runaway advantage, and I compare it to UK land ownership to explain what it means to be a rent seeker over an entire economy. I close with why every business is now a software business, why AI generated code brings new unintended behavior, and why we need to stay engaged rather than assume it's someone else's problem.

In this episode you will learn:

(00:00) The hackathon idea that started this, distributing AI training across home computers so everyone could have a stake, and why it failed, starting with undetectable poisoning by bad actors.

(03:44) How neural networks work, from linear functions in algebra to the single nonlinearity that lets them approximate virtually any function.

(06:36) The matrix computations and graphics chips these models need, and why low latency communication means the work must happen inside one data center, not distributed machines.

(10:00) Why AI is destined to be an oligopoly controlled by a handful of corporations and nation states, and how AI improving AI can turn a breakout model into a runaway one.

(12:57) The mathematics of pandemic spread, the replication number separating exponential growth from decline, and how a small edge in AI self improvement compounds into an overwhelming advantage.

(15:56) The idea of a rent seeker, illustrated through UK land ownership dating back nearly a thousand years, and how a runaway AI could end up in a similar position over the global economy.

(18:53) Dan Kennedy's insight that whoever can spend the most to acquire a customer wins, and why that means every business is now a software business built on AI.

(21:32) The unintended behaviors in AI generated software, and why engaging with these risks now matters more than treating them as a distant problem.

Let’s connect!