What You’re Allowed to Know

In 1633 an old man knelt on a stone floor in Rome and said out loud that he was wrong about something he knew to be right. Galileo had seen the moons of Jupiter through a lens he ground himself. He had the proof that the Earth moves. But the institution that decided what Europe was allowed to know could not afford for him to be right, so they made him recant, put his book on the banned list, and shut him in his house for the rest of his life.

The truth did not change. It just had to wait for a world that would let it be said.

Nearly four hundred years later we are handing that same power, the power to decide what is true and what can be said, to a handful of companies.

The newest gatekeepers

AI is becoming the default way people find facts and settle questions. It is becoming the default way the internet runs at all. This year, for the first time, automated systems passed people to become the majority of web traffic, about 57 percent of all requests by Cloudflare’s count, most of it the new wave of AI agents. That makes whoever runs the best models the most powerful editor who has ever lived. The strongest ones sit behind monthly subscriptions and metered APIs, and the data they learn from is uneven in the usual directions. The tool gets sold as universal. It behaves like a private library with a cover charge.

When a single company decides what its model will and will not say, it is drawing the edges of what millions of people can easily know. Some of that is real safety work, and it matters. But the same lever that blocks a genuinely dangerous answer can also trim an inconvenient fact, soften a finding a business partner would dislike, or steer you off a topic for reasons nobody ever has to explain. You do not get a vote. Most of the time you cannot even see that there was a line.

The cost of asking

There is also the plain matter of money. To use the best proprietary models you pay per request, which sorts the world into tiers. Deep pockets get to build on the good stuff. Everyone else rents access to someone else’s, on terms that can change next quarter. Knowledge slips back into being a luxury good, the way it was before cheap books and free libraries pried it loose.

You can own the lens

Galileo’s real threat was never one observation. It was the lens. Once anyone could grind a piece of glass and look for themselves, the monopoly on what counted as real began to crack.

We have a version of the lens now. You can download an open model like Llama or Mistral and run it on a laptop or a small server in your office. When the model runs on your own hardware, the gatekeepers lose their hold.

1. No permission required. There is no API key to request and no terms of service that change next quarter. The model is yours.
2. Privacy by default. Your data never leaves your hardware, which matters for anyone handling records they would rather not hand to a stranger.
3. Tuned to you. A local model can be trained on your own material, so it learns your work instead of a generic average. You stop waiting for a vendor to decide your use case is worth supporting.

Build on what you own

The durable question in AI is the same one it has always been about infrastructure. Do you control it, or does someone else? Owning the hardware and the weights turns a tool you borrow into one you actually keep, and a model you can open up is one that cannot quietly decide things on your behalf.

That is the work we care about at Ghost Hat Studio. We help people own the whole stack, from the data going in to the answer coming out, so the tools they build stay theirs.

What finally beat the gatekeepers was a lens cheap enough that everyone could own one. The models are good now, they are open, and they run on hardware you can buy. The window is open. You only have to look.