A speaker explains why most AI personalization fails and what it actually requires. Machine-level interpretability lets you see what the model learned. Human-level interpretability lets you explain why it made a choice. Both matter for AI health coaching to work.
Personalization in AI is a function of the fidelity of interpretability of the system. Let me explain what I mean by that. So you have machine level interpretability, which is, as a developer or technical product manager that is in charge of building this AI agent or LLM application. Do I understand why the LLM made that response to the user query? Do I understand the user intent, that type of thing? That's machine level interpretability. Do I as a human understand why the machine said this in response to this? Then there's human level interpretability. And that's where we talk about this a lot with Epistemic Me, the open source project that I work on, which is can we understand and model the self? What does it mean that this person is a self? Their representation of their belief systems and how those evolve over time and interactions with your software system or your chatbot. So good example of your building AI health coach. You would want to know what does this user believe about their health, their sleep, their diet, their exercise. You baseline that just as a coach would. Then you give recommendations for them to reach their target self or their target goals. So personalization in a nutshell is a function of the fidelity of machine level interpretability available to your system and the human level interpretability available to your system.
