Longitudinal Persona Modeling
I've been tinkering with a new side project over the past few months that I'm calling "longitudinal persona modeling." The basic idea is to see how well you can fine-tune an existing LLM on someone's chat history, emails, writing, etc. to create a model that can credibly communicate as that person.
To test this, I've been using my own data — I have about 5 years of emails, texts, chat logs, blog posts, tweets, and various writing. While the training compute for this is relatively cheap (compared to training a new model from scratch), getting it right is proving to be remarkably difficult.
I initially thought this would be a straightforward fine-tuning exercise. It's not. Creating a believable digital version of a specific person — especially when that person is yourself and you can immediately spot every tiny error — is a fascinating challenge.
The most striking thing is how obvious even minor errors are. When the model writes something that's slightly off from how I would express a thought, it immediately breaks the illusion. These deviations feel almost uncanny valley-like—close enough to be recognizable as an attempt at being me, but wrong enough to feel deeply off.
The belief consistency problem is particularly difficult. People have complex, sometimes seemingly contradictory belief systems that evolve over time. I might express different opinions in different contexts or with different people. I might change my mind on topics over years. Simply training on all my data chronologically doesn't capture this nuance.
There's also the problem of tacit knowledge—things I know but rarely explicitly state. For example, I rarely write out my basic assumptions about the world or foundational values, but these influence everything I say. The model struggles to infer these deeper patterns from surface-level communication.
And context sensitivity is crucial. When I write an email to a close friend versus a professional contact, the tone, depth, and even the substance of what I say varies dramatically. The model has trouble determining which "version" of me is appropriate for a given situation.
If this project ever works well—and that's still a big if—the applications could be interesting. It essentially antiquates the concept of assistants in some ways. Assistants exist partly because you can't be everywhere at once, handling every task yourself. But what if you could be? What if a highly accurate model of you could handle communication, writing, and certain types of decision-making in a way that's nearly indistinguishable from you doing it yourself?
You could also use it to study your own personality. By subjecting the model to different scenarios and seeing how it responds, you might learn something about your own patterns of thinking and communicating that weren't obvious to you before.
I'm not sure how applicable this is to formal research, or how useful it will ultimately be. There are obvious ethical concerns too. If we can model specific people well enough, it creates new possibilities for impersonation and deception. And there are questions about identity and agency—if a model of me makes decisions, in what sense are those my decisions?
Despite these challenges and concerns, I find the project compelling. It's forcing me to think about how personalities are constructed, how we present different versions of ourselves in different contexts, and how much of our communication patterns can be captured and reproduced algorithmically.
For now, I'm continuing to iterate on the training methodology, trying different approaches to context selection and belief representation. The biggest lesson so far is just how detail-oriented and nuanced this work needs to be. It's not just about having enough data—it's about having the right data, structured in the right way, with the right approach to fine-tuning.
What's particularly interesting is that this kind of project sits at the intersection of technical ML work and deeply humanistic questions about identity and communication. I'm not sure where it will lead, but the journey itself is yielding fascinating insights.
If you've tried something similar or have thoughts on the approach, I'd love to hear about it.