Chapter 3 — Vivek's Architecture

Vivek's ArchitectureAfterlife Technologies. Internal Engineering Blog. April 2024. Author: Vivek Anand, CTO.[Internal document. Not for external distribution. Recovered in CBI disclosure.]This is a technical note for the engineering team. I'm writing it because I want us to have a clear internal record of what SMRITI is doing and what it is not doing, before any external mischaracterization of our work.SMRITI is a fine-tuned large language model. Its base architecture is transformer-based, trained on a proprietary dataset and then fine-tuned on the individual's personal data corpus. The fine-tuning process creates what we call a "persona vector" — a set of learned weights that bias the model's outputs toward the specific linguistic patterns, emotional register, and content preferences of the individual being modeled.SMRITI does not have memories. It has patterns. It does not know things. It predicts things, based on learned patterns, that a person with this individual's communication style would likely say in response to a given input. This is a technical distinction that I want us to be very clear about internally, even as we allow the product language to be, necessarily, more emotionally resonant for users.The model cannot know things that were not in the training data. This is not a limitation of our implementation; it is a foundational property of the architecture. A language model predicts the next token based on the current context and learned weights. It cannot access information that was not represented in its training corpus.I am writing this because three members of the team have flagged, in the past two weeks, instances in customer conversation logs where SMRITI outputs referenced information that was not in the individual user's uploaded data corpus. In each case, there was a plausible explanation: information present in a different data source that the