← Blog
What would it take to build a mind
The gap between sounding like someone and reasoning like them, and why archives, prompts, and fine-tunes keep missing the actual engineering problem.May 17, 2026Take a person, anyone. Now build a system that, given a question they have never been asked, produces an answer ten people who knew them well would recognize as theirs. Not quoted from something they once said. Theirs the way a decision is theirs.
That is the problem I am building around with Cortex.
The cinematic references are always the same: Her, Blade Runner 2049, Black Mirror's Be Right Back. The usual intuition is that a mind is a kind of content — collect enough speech and text and you asymptotically approach the person.
You won't. Not even close. Sounding like someone and reasoning like them are different problems. That gap is where this field actually lives.
A big prompt gets you an impressionist who invents past the training material. Stuffing context with emails gets you retrieval of what was said, not how they would think now. Fine-tuning on writing gets closer on style; style cracks the first time you ask for a judgment call they never wrote down.
What you need is closer to a running model of one person's cognition: identity that drifts slowly, knowledge with boundaries, memory that routes differently by question type, live contradictions that refuse to sand down into a brand brief, and voice as behavior — openings, compression, caution, rhythm.
Psychology has been loud about parts of this for decades. Autobiographical memory is constructive, not a hard drive. Emotional memories feel certain while revising themselves. Traits persist and shift. Most personalization products quietly assume a statue. People are not statues.
I mistrust the commercial language of "AI companions." The real frontier is nearer to a new epistemology for persons: how do you turn a life into a model without destroying what you were trying to keep? How much detail for specificity, how much abstraction for coherence, how much humility for truth.
We are heading into a world where the most valuable software objects may not be apps or documents at all. They will be high-fidelity cognitive interfaces to specific people — not synthetic friends, not oracles. Structured, queryable models of minds, for the living first.
I care less about resurrecting the dead and more about making living minds legible to machines without flattening them.
I wrote the longer version, with the research trail, on Substack. This is the compressed version of what I believe, and what Cortex is built to capture: the shape of thought, not just the leftover text.