In order for AI to be a tool that empowers as many people as possible to express who they are, it must understand the breadth of the human experience beyond information and knowledge to wisdom and meaning. Meaning in a human sense cannot exist without both logic and emotion. Which means we must understand how LLMs respond to not only logical reasoning but also ‘functional emotions’ as inputs.
The study of these emotional inputs fascinates me. So I am exploring how human-model interactions impact model behavior, one layer of abstraction at a time — rowbyroh.com/research.
At the surface, I am curious how my word choices or the shape of a conversation can produce measurable changes in a model’s output. Below the surface, the field of interpretability in its early phases has already produced fascinating work on understanding the impact of functional emotions in LLMs. Are there ways an LLM can interpret audio beyond transcripts into tone, rhythm — the intended emotion behind a piece of music? And will that ability help us better understand, influence and anticipate how to best work with these models?
The common thread I have noticed in my work is that I connect the dots between two unrelated worlds — taking inspiration from modular furniture to scale freight pricing systems or applying supply/demand dynamics to reframe recruiting CRM strategy.
As an Asian-American immigrant — a liberal-arts grad who wanted to be a union carpenter, now building enterprise software — never fully belonging to any one world, I learned to notice the simple, overlooked thing that is valuable in another context.
More: rowbyroh.com/about
- sound-lab — pre-registered hearing tests on an open-weights audio model (Emotional Resonance series)
- 944-parts-watchlist — which Porsche 944 parts can still be bought, with evidence for every call
- claude-harness — hooks, scheduled jobs and session tracking behind a personal Claude Code setup (Field Notes)
- obsidian-mcp — MCP server connecting Claude to a ~670-note Obsidian vault