At some point, once the architecture from the previous post was working, I asked myself a different question: what if each user had their own model?
Intuitively, it sounds smart. More personalization. More adaptation. More memory, tuned specifically to one person rather than shared across everyone. And honestly, when you think about human interactions, it almost feels natural — the more you know someone, the more your understanding of them evolves, shaped specifically by the history you have with that one individual.
So I started exploring that direction seriously.
Where it breaks down
The deeper I looked into it, the more I realized something important: most large-scale AI systems don’t personalize the model. They personalize the context.
The reason becomes obvious once you look at what training or maintaining one model per user actually requires. It quickly becomes:
- Expensive — every user represents a full model’s worth of compute and storage, multiplied across however many users the system serves.
- Difficult to scale — the infrastructure problem grows linearly, at best, with every new user added.
- Hard to maintain consistently — keeping hundreds or thousands of individually-tuned models coherent, updated, and behaving predictably is its own engineering burden, separate from the product problem you were actually trying to solve.
And in many cases, on top of all of that, it’s unnecessary.
What actually varies between users
What actually changes from one user to another is rarely the reasoning engine itself — the part of the system doing the thinking, weighing tradeoffs, generating language. What varies is:
- The memory.
- The preferences.
- The history.
- The goals.
- The relevant context.
All five of those are things the architecture from the previous post already handles — short-term memory, structured state, targeted retrieval. None of them require a different model underneath. They require different inputs to the same model.
The real question
That realization changed my perspective again. The question I’d started with — “how do we train a model for every user?” — turned out to be the wrong question entirely. The one that actually matters is: how do we design systems that adapt intelligently without retraining everything?
One shared model. Different context. Different experience. That’s the shape personalization actually takes at scale, and it’s a data engineering problem before it’s ever a modeling one — which is exactly where this series keeps landing, one wall at a time.