The democratisation of AI stopped at the wrong layer

Democratisation is the AI industry's favourite word, and one, you could argue, it has earned the right to use. Less than a decade ago, building anything with machine learning meant a research team and a budget few organisations could afford. Today, the barrier to using advanced AI has lowered significantly, and what used to be a quiet technology in the background has been put in the hands of millions.

But access is the surface layer. Underneath it sits another layer which, arguably, matters more than access to AI itself, and that layer is training. 

Unlike model access, model training has not been democratised at all and this is the quiet asymmetry at the centre of the industry. We have made it easy to consume models and almost impossible to make one. 

Access may be cheap, but training is not.

It helps to be precise about what "democratisation" has actually delivered.

What opened up was consumption. A person with no technical background can now directly interact with an AI model or build an app in an afternoon. And with open-weight releases, anyone can even run one on their own hardware. These are significant advances. 

What did not open up, however, is the ability to train a model. Producing a frontier model requires capital measured in billions, scarce specialist talent, vast datasets, and retraining cycles that have to be repeated every six to 12 months. 

That is not a barrier a startup, or a hospital, or a mid-sized manufacturer can clear. It is a barrier that only a few well-funded organisations in the world can clear, and they have every commercial reason to keep it that way.

So the curve runs in two directions at once. The cost of using a model has fallen toward zero. The cost of making one has climbed toward the limits of what private capital can sustain.

Owning the weights is not owning a model.

The usual objection at this point is open source. Open weights, the argument goes, solve the ownership problem. You can download the model, run it, fine-tune it, and deploy it. Isn't that democratised training?

It is indeed closer but owning the weights is not the same as owning the model itself.

An open-weight model remains the frozen output of a training process run by the organisation that produced it. What users receive is the result of that process. 

They can somewhat adjust the model but doing that well still demands expertise, compute, and clean data. And as the world moves on, the model ages and retraining is needed which is not something users really control. 

In short, open weights don’t offer users the ability to produce and continuously improve a model. That ability is still held by the labs that produced it.

What democratising training and model ownership would actually change.

It is worth asking at this point what the world would look like if democratisation reaches the training layer and allows users the genuine ability to create, train, and own a model. Several structural changes follow.

Ownership survives the vendor. A model trained by its user on data they own, whose weights are theirs without licensing restrictions, does not evaporate if the company that gave them the tools to build it disappears. The relationship stops being a subscription, and that single change removes the dependency the current market is built on.

Models that keep learning. If training is something that can be done continuously rather than a cyclical and centralised event, the model does not have to freeze at deployment. It can keep learning from new data in production without an expensive retraining cycle and without a team of specialists to run it. The workarounds become unnecessary because the underlying limitation is gone.

Accountability that can be traced. When control of what a model learns and how occurs at user level, the chain of responsibility is legible. That is the kind of traceability regulators are beginning to demand and that black-box frontier models cannot currently provide.

Distributed benefits. This is what makes "democratisation" the right word at last. If the ability to train a model is widely held, then the value, the control, and the responsibility are widely held too. 

This is the thesis Boltzbit is built on, and the reason our General Learning Intelligence treats learning as part of the architecture rather than a cost only the largest labs can bear. But the argument stands on its own, independent of any company's answer to it. Access to models is already cheap and getting cheaper. Access to training is not and that is where the next phase of competition will be decided.

‍