The most useful AI of the next few years will not be the most powerful. The headline story is still about scale, and that story has further to run. The shift that may matter most for everyone else is happening somewhere smaller, on models light enough to run on the device already in your pocket. The quieter story is about access, and that is the one set to change who gets to build.
For most of the past few years, progress in AI has meant going bigger, with more data and more compute producing larger models trained on larger budgets. New hardware keeps that trend moving, and as architectures like Blackwell matures and Vera Rubin comes online, the frontier models will keep getting smarter and faster.
For all that capability, those frontier models stay expensive to run and tied to centralised data centres. The budgets behind them sit with a small number of very large players, which keeps that level of capability closer to a luxury than a default for most organisations.
At the same time, smaller and faster models are gaining ground, and they do not need a warehouse of hardware to be useful. A run of new techniques is making this possible. Aggressive quantisation lets a model hold far more of a conversation in the same memory, six to 15 times more with the latest methods, so it stops losing track of what you told it earlier. New architectures, including ternary and binary models, shrink things by an order of magnitude. One open model, Gemma 3, drops from around 13 gigabytes to roughly 300 megabytes when re-trained this way, which is enough to move it from the data centre to the handset.
Apple is taking a related route with its unified memory architecture, a design that lets standard hardware do work that would otherwise demand expensive, specialised chips.
Smaller models will not match the largest ones on every task, and they do not have to. Once a model becomes good enough for a job, it changes what that job costs to do. A capability that used to need a billion dollar budget becomes something a small team can reach for. That is the point where the economics tip.
Running a capable model on your own device opens up more than convenience. A model that lives on your phone keeps your data on your phone, so it never has to travel to the cloud to be useful. It can act as a personal assistant that learns your context over time and gets better the longer you use it.
Changing a model today means retraining it from scratch, which can take months. Research efforts such as MIT SEAL are exploring a different approach, letting a model adjust its own weights as it goes. Run that locally and the model can learn from your experience without anyone else being exposed to what it picks up.
A model like this is at its weakest on the day you first switch it on, before it knows anything about you. From there it only improves, and everything it learns stays on the device
None of this removes the case for the big frontier models, which will keep pushing what is possible. The shift worth watching sits alongside them, in models small enough to run anywhere and cheap enough for far more people to use. When the cost of building with capable AI drops that far, the ideas stop being limited to those who can afford the hardware. Innovation starts coming from the smaller teams and individuals who have always had the ideas and now have the means to act on them.
At ClearPoint, this is the shift we are paying attention to. The opportunity is open to any organisation ready to treat AI as a default and build the foundations now to put it to work wherever it runs from the data centre to the device in your pocket.