U.S. healthcare is in the midst of a massive shift – moving complex, long-term care out of expensive institutions and into the home. Successfully shifting care into the home depends almost entirely on one group: informal caregivers – family and friends doing the work of care. But the informal caregiver system is fragile. When it comes to supporting caregivers, we know what works: professionals providing training, education, respite, navigation, and other services. Professional-driven models work, but they don’t scale.
And that’s Healthcare’s Big Bet – that we’ll figure out how to support informal caregivers at scale before the whole system collapses.
And that’s why I’m building Ren – a hybrid AI Partner in Care that takes what works in professional-driven models and delivers it to caregivers everywhere at scale. Hybrid AI is an approach to artificial intelligence that deliberately synthesizes multiple forms of AI technologies and traditions. The goal is to build on the strengths of each tradition while compensating for the limitations of any single paradigm.
One thing that makes professional-driven models work is how professionals engage with caregivers to learn what’s going on and how they can help. When they first meet a caregiver, they don’t immediately demand to know the birthday, address, dietary preferences, and daily schedule for the person the caregiver is supporting. They don’t force a caregiver to fill out a 75-part intake questionnaire. They ask what’s relevant to the context and the relationship in that moment. Questions like “How are you?” or “How is your mom doing today?” or “How did it go with the PT this morning?”
In the last few days, I’ve been working on two ideas that sit at the core of how Ren learns: computational epistemology (“curiosity”) and epistemic exploration (“inquiry”). In simple terms, this is about helping an AI know what it knows, recognize what it doesn’t, and then act on that ignorance in an intelligent, inspectable way.
This matters because it allows Ren to operate more like a human professional – only asking the questions that matter. And it’s part of what allows Ren to take what works in professional-driven models and deliver it to caregivers everywhere at scale.
Computational Epistemology (“curiosity”)
You can think about computational epistemology in two parts.
The first part is what the system knows. In Ren, knowledge is stored as facts in a graph (which is essentially a specific database structure), propositions (which are statements about what is stored in the graph), and hunches (which are representations of incomplete information). I go into a lot of detail about hunches, propositions, and the graph in my blog post on Knowledge Representation & The Epistemic Substrate – so I won’t belabor it here.
The second part is what the system knows about what it knows. This includes things like confidence (the system understands how certain it is about its knowledge), provenance (the system understands where knowledge came from like user conversations, documents, inferences, and so on), and gap identification (when there is missing data that the system wants to have for reasoning or other purposes).
Ren uses curiosities to identify gaps. Curiosities are concrete representations of things Ren wants to know and the ways Ren has for checking whether it knows what it wants to know. Curiosities get triggered in a number of ways – like when particular events happen or Ren’s knowledge changes. If, in response to a trigger, Ren concludes “I don’t know” or “I’m not confident”, it creates a desire to know what it doesn’t know or to know with greater certainty. The desire to know is the beginning of epistemic exploration – discussed below.
The AI community has been working on computational epistemology and related topics for over half a century, beginning with early work on knowledge representation and epistemic logic in the 1960s. Most of that work has stayed in academic circles and has never become mainstream. Even today, very few systems explicitly reason about what they know and don’t know.
Epistemic Exploration (“inquiry”)
Epistemic exploration is what the system does when it detects a gap or wants higher confidence – how it seeks, asks, or verifies. Basically, epistemic exploration is how Ren gathers new information – how it learns.
As of the writing of this post, Ren has one way to seek out new information – ask the user. When Ren develops a desire to know something, it checks if it has an inquiry strategy for asking the user for the desired information. Inquiries represent two things: how to formulate a question to ask, and how to interpret the user’s response. Finding an inquiry that matches a desire to know is Ren’s way of saying, “Oh! I know how to ask about this and how to understand what the user says.” If Ren is able to find an inquiry for the desired knowledge, then Ren will generate a desire to speak with the user and ask a question.
Eventually, Ren will read documents and search the internet. But for now, Ren holds conversations with the user when it wants to know something.
The idea of systems actively seeking information isn’t new. Active learning, where systems identify informative examples to query, has been studied since the 1990s. Question-answering systems and dialogue agents have explored how to ask clarification questions when uncertain. But these approaches typically operate within narrow, predefined contexts – asking about specific data points for classification tasks or clarifying ambiguous user commands. Few systems combine the ability to identify arbitrary knowledge gaps with strategies for pursuing that knowledge through natural conversation. The integration of epistemic self-awareness with proactive information-seeking remains largely unexplored territory.
Example: “Are you a caregiver?”
One of the most important things Ren needs to know is whether a user is a caregiver. This knowledge fundamentally shapes how Ren interacts with that person – because Ren is specifically built to support caregivers.
When Ren encounters a new user, a curiosity triggers: “Is this person a caregiver?” Ren has an inquiry strategy that matches this desire to know, so it formulates a question like “Are you providing support for someone in your life?” And if the user responds, “Yeah, I care for my mom who lives with us.” – Ren can notice “The user expresses involvement in caring for their mom who lives with them.” and conclude that the answer is “yes” – this person is a caregiver.
Once Ren asks and receives an answer, it can update its knowledge graph and adjust how it interacts with that person. It’s a simple example of how curiosity and inquiry combine to supporting targeted knowledge seeking and learning.
Why it Matters
Computational epistemology and epistemic exploration are the foundation for targeted knowledge seeking. Ren doesn’t do passive information processing like almost every other “AI” system. And Ren doesn’t spam the user with dozens of irrelevant questions to get to the one or two questions that really matter. Instead, Ren engages in active learning. That’s part of what makes Ren different.
In part, professional-driven caregiver support models work because professionals ask targeted questions. They don’t robotically demand every possible piece of information. When a care navigator meets a caregiver, they don’t immediately ask for a complete medical history, daily schedules, financial records, and family dynamics. They ask what’s relevant to the moment and the relationship – building understanding progressively as the relationship develops and needs emerge.
This is exactly what Ren does. Curiosities are triggered by specific situations and needs – it asks about caregiver status when it encounters a new user because that knowledge shapes the interaction. It doesn’t ask about things it doesn’t need to know, or things that aren’t appropriate to the moment. This selectivity is what makes the learning process efficient, respectful, and contextually appropriate – just like professional support.
This is how Ren takes what works in professional-driven models and makes it scalable. Instead of requiring a human professional to conduct targeted inquiry with each caregiver, Ren can do it – learning incrementally through natural conversation, filling in exactly what it needs when it needs it, and building understanding over time the way a good professional would.
Leave a Reply