The systems-centric era
Most AI products today use the same models. So why don’t they create the same value?
Behind the frontier, extracting value from AI is less about choosing a model and more about building the system around it. Memory, tools, retrieval, workflows, skills, trust, interfaces, evaluation, and how all of those parts work together.
This is where products become much better or much worse, and is the defining characteristic of what I’m calling the systems-centric era.
We got here by building on what came before. First was the model-centric era. Researchers treated datasets as static and focused mostly on code; tweaking architectures and hyperparameters to beat benchmarks. It drove incredible progress.
But benchmarks saturated, and we needed a new path for the complex problems in the real world.
So the focus had to change, and we looked to the data. We learned that massive quantity mattered, but quality mattered just as much. Better data produced better performance, even when the model stayed the same. This was the data-centric era.
There’s still a lot of progress to be made here, and we haven’t left either era behind. Models and data continue to improve.
But we don’t need to wait for the next breakthrough to extract more value. There’s already a lot of unused capability in the models we have. The question is how much of it we can get out. And that depends on the system around the model.
We saw this with the Claude Code leak. The value didn't come from the model alone. It came from the system around it: the context management, prompts, tool use, and product judgement. It showed that the most popular AI product of our time is differentiated by everything but the model.
The same will be true as AI moves into the physical world. Sensors, actuators, and control loops will become part of the systems around models, expanding what the AI can do.
Models will keep improving, and that matters. Everything is downstream of this. But the defining products of the next decade will be the ones that turn raw capability into outcomes. And the model is just the start.
This is the systems-centric era.
Postscript
What does this mean for local models?
Small models that can run locally on laptops, phones, or other devices will become more capable and good enough for more real work. More capability is squeezed into fewer parameters every month. They'll be embedded in our lives in ways that are hard to imagine now.
And a good system can turn a weaker model into something much more useful than its raw capability suggests. A weaker local model, inside a strong system, may outperform a stronger frontier model used poorly.
Work that needs a frontier model today may only need a local model tomorrow, if the system around it is good enough.
The systems-centric era doesn't favour the biggest model. It favours the best system. And increasingly, that system will run in your pocket.