Human-written text carries traces of how people understand the world. In describing a situation, a writer gives it a particular structure, making some relationships explicit while leaving others for the reader to recognize. What the writer chooses to explain, connect, or take for granted reflects an interpretation of the situation.

Learning from text may therefore offer indirect access to reasoning and tacit knowledge that are never fully articulated. An explanation can show how evidence supports a conclusion, while a description can reveal which distinctions matter and what background knowledge is assumed. These patterns can convey aspects of human thought even when the writer never describes the thinking behind them. This is what interests me about text as a learning medium, and why I want to understand how representations can preserve the relationships that make these patterns available for learning.

Where understanding the world asks whether machines can interpret the reality around us, understanding human asks whether they can interpret the realities that exist within and between people. Much of what shapes how we think, feel, and communicate is never stated directly: personal memories, routines, preferences, emotional states, cultural references, and the shared context built over time within families, communities, and relationships.

These aspects of human life are not easily captured as a fixed set of semantic labels. They are distributed across behavior, language, history, and relationships, felt rather than fully articulated. Bringing together language, facial expressions, gestures, and patterns of behavior over time may help AI systems recognize these implicit patterns. Such understanding could provide a basis for human–AI symbiosis, helping systems adapt to people’s lives and relationships as those contexts develop and change.

I think of an AI model’s beliefs as working judgments about what is true or likely, given the available information. These judgments guide interpretation and action, and their dependencies matter when evidence changes. Revising an assumption should prompt reconsideration of the conclusions that rely on it, while allowing those with sufficient independent support to remain. Sometimes the appropriate update is a change in confidence rather than a reversal; uncertainty should reflect the strength and limits of the available support.

Language provides a way to express these judgments, but producing a claim is not the same as relying on it. Autoregressive modeling defines a distribution over token sequences, but token probabilities alone do not establish coherent beliefs or calibrated confidence in the claims those sequences express. What interests me is whether a claim expressed in one answer meaningfully constrains subsequent judgments, and whether revisions carry through to the judgments that depend on it.

I see the translation layer between machine representations and human experience as a starting point for system design. For people, the interface should help them understand the system’s behavior, evaluate its responses, and stay in control of the interaction. What a system makes visible and how people can respond to it shape the feedback available to the model. A rejected answer may be factually wrong or simply miss the person’s goal, and the interface can help reveal that difference.

These distinctions should inform learning objectives and algorithms from the outset. In turn, the model’s uncertainty can guide what it asks or reveals next, giving people opportunities to clarify their goals or correct its interpretation. Learning and interaction can be designed together, with each response reflecting the system’s understanding while shaping what it has a chance to learn next.

Expanding a person’s ability to communicate is as important as expanding what they know. A system might help someone convey an experience that is difficult to put into words, make their reasoning visible to others, or explore where their interpretations differ. The question is therefore not only what a person can understand with a machine’s help, but what people can express, discuss, and come to understand together. Exploring these possibilities requires building prototypes and designing and testing these experiences with people.

A design can look convincing while resting on a mistaken understanding of the people it is meant to support. User studies help examine those assumptions, with different methods answering different questions. Observations can reveal routines and workarounds, interviews can explore the reasons behind them, and surveys can compare reported experiences across a broader group.

When evaluating a prototype, the choice of participants, tasks, and measures should follow the question being investigated. Task completion and errors can reveal difficulties, while participants’ explanations can help interpret those outcomes. Unexpected behavior or conflicting feedback calls for a closer look at the assumptions behind the design. Sometimes the most useful result of a study is a reason to rethink what the design was trying to solve.

Design can shape how people notice, feel, remember, and connect with the world around them. This means designing not only individual interactions, but also the relationship between them: how sequential moments build into a narrative, how simultaneous sensory cues reinforce one another, and how an experience can create resonance or evoke a meaningful memory. I am interested in using form, material, interaction, and atmosphere to expand and connect our senses, making ordinary moments feel more immersive and meaningful.