Why Joe

Joe is, in part, a product — but mostly it’s a suite of related experiments. Here’s the thinking behind them.

At this point I’m convinced that LLMs will – in ten or twenty years – have been as transformative of labor and society as the internet was.

The bad news for AI investors and the good news for the rest of us is that we don’t seem to be seeing a hyperbolic increase in model capabilities, and models are already trained on an appreciable fraction of all the available data.

Far from seeing the serving of large and super capable models becoming a moat which entrenches the market position of major players and creates a natural monopoly, we’re seeing more and more capable small models. We’re also seeing advances in running large LLMs on constrained and near-consumer hardware.

we are witnessing a company who thought they invented the ham sandwich realizing in real time that they sell ham, bread, and mustard at the grocery store

[image or embed]

— ceej (@ceej.online) July 12, 2026 at 7:10 PM

Given the history of the computing industry to date, it seems absurd to argue that LLMs will remain as compute intensive as they have been to date. The entire history of the industry is one of expanding the capabilities of available hardware, and of finding ways to stretch existing hardware farther.

I haven’t dug up the talk, but I remember a video of Alan Kay saying that part of “inventing the future” is “computing in the future” – paying the hardware premium today in order to have first-generation access to resources which will eventually be available at consumer prices and aren’t yet.

That’s how I feel about LLMs. They’re currently expensive, and I don’t think they’re going to develop into super-intelligence like some boosters think, but bots are capable of doing meaningful direction following, reasoning and computing with language.

Which brings me back to Joe.