The democratization of AI stopped at the wrong layer
Democratization may be the AI industry’s favorite word, and one that, in fairness, it has earned the right to use.
At the beginning of the decade, building anything with machine learning meant a large research team and a significant budget, and that put it out of reach for many.
Today, however, the barrier to using advanced AI has fallen almost entirely, and what used to be a quiet technology in the background has been thrust to the forefront of everyday work and everyday conversation.
Latest Videos FromTechRadar
But access is just the surface layer. Underneath it sits another layer which, arguably, matters more than access to AI itself, and that layer is training.
CEO and co-founder of Boltzbit.
Unlike access, model training has not been democratized at all. The industry has made it easy to consume a model and almost impossible to make one.
This needs to change.
A gatekeeper to innovation
If we look closely at what “democratization” has really delivered thus far, we see that what has opened up is only consumption. Anybody can now interact directly with an AI model, even without a technical background, and can build an app in an afternoon with minimal effort. Open-weight releases mean anyone can even run a model on their own hardware, free of an API. These are genuine advances that have helped define the AI-powered world we now live in.
What has not open up is the ability to train a model. Currently, producing a frontier model still requires capital measured in billions, scarce specialist talent that only a handful of labs can attract and retain, vast datasets, and retraining cycles that have to be repeated every six to 12 months just to stay current.
It is a cycle that is stifling progress, and a barrier that no startup, cash strapped healthcare organization, or mid-sized manufacturer can clear. It functions as a gatekeeper to innovation, and it is one that well-funded organizations have every commercial incentive to keep locked, because the scarcity of training capability is precisely what protects their market position.
There is a dichotomy. 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.
Can open source help?
Some might argue that the answer is to turn to open-weight models. Open weights, the argument goes, solve the ownership problem. You can download the model, run it, fine-tune it, and deploy it without asking anyone’s permission. This is not democratized training though. It may be 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 organization that produced it. What users receive is the result of that process. It is true that they can somewhat adjust the model. But doing that effectively demands expertise, compute, and clean data. And as the world moves on, the model ages and retraining is needed.
So despite their advantages on the surface, open weights do not offer users the ability to produce and continuously improve a model. That ability is still held by the labs that produced it.
The change that could revolutionize innovation
In my view, if democratization reaches the training layer and allows users the genuine ability to create, train, and own a model, several important structural changes follow that could meaningfully accelerate innovation across industries.
Ownership survives the vendor. If a model is trained by its user on the data they own, the weights are theirs without licensing restrictions, it will not evaporate if the company that gave them the tools to build it disappears. The relationship stops being a subscription. That single change alone removes the dependency the current market is built upon.
Models must learn continuously. If training is something that can be done continuously rather than a cyclical and centralized event, the model does not have to freeze at deployment. It can keep learning from new data in production without expensive retraining cycles and without a team of specialists being needed to run it. The workarounds become unnecessary because the underlying limitation is gone.
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 expecting and that black-box frontier models simply cannot provide.
These are the distributed benefits that make “democratization” the right word at last. If the ability to train a model is widely held, then the value, control, and responsibility are widely held too. General Learning Intelligence treats learning as part of the architecture rather than a cost only the largest labs can bear.
Access to models is already cheap and getting cheaper. However, access to training is not and that is where the next phase of competition will be decided.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit