Decode the mind behind the words.
Anticipate the behavior that follows.
We’re building a behavioral layer for AI coaching and advisory products: one that turns conversation history into revisable hypotheses about a person’s goals, constraints and likely next actions, and into guidance for the assistant’s next response.
A behavioral layer for your AI.
It surrounds an existing conversational model: evidence in, interpretations and response guidance out, feedback back into the model.
Memory tells an AI what happened. Bionic Mind models what that evidence implies about what the person may do next.
One request in. One record out.
Help an assistant choose a more useful next response.
Evidence, the interpretations that evidence supports, and the response guidance that follows. Read it top to bottom.
The message below is ordinary. What changes its reading is everything that came before it.
“Okay, let’s do it.”
In the history: three plans agreed to in six weeks, none completed. Every one of them had more than two steps.
Competing interpretations, held together and ranked against the evidence. None of them is a label, and none is treated as settled.
Guidance for the assistant, not a script. The partner’s model still writes the reply.
An illustration of the record we intend to produce. Not output from a running system.
What we are building.
Bionic Mind is pre-product. Nothing on this site is a measured result. We are building the behavioral layer and looking for one design partner to build and test it with.
An assistant that remembers everything a person said still treats every statement at face value. Agreement and intention are not the same thing, and the difference is visible in the history.
A layer around the model a partner already runs. Conversation and available history in; ranked interpretations and response guidance out; the person’s actual behavior back in as correction.
Three sources, kept separate: simulated populations, work on real behavioral data, and replications of published behavioral studies.
Held-out choices, scored against an equally informed general AI with memory and summary, and against a variant without structured priors. A layer that cannot beat good memory is not worth integrating.
Four measures agreed before it starts: task completion, user-rated usefulness, how often the user has to correct the assistant, and the cost and latency the layer adds.
One behavioral model. Two applications. Conversational AI first.
Conversational AI
First partners: recurring AI coaching and advisory products, where useful guidance depends on a person’s goals, constraints and previous decisions.
Your coaching assistant recommends plans users agree to and do not complete. That failure is recurring, observable, and expensive: it is the reason people stop showing up.
What we would build together
A bounded behavioral layer over your existing model, reading the interaction history you already keep. No migration, no replacement of the underlying model.
What we would measure
Plan completion against your current baseline, user-rated usefulness, correction frequency, and the cost and latency we add. Agreed before we start, reported whichever way they come out.
A proposal, not a case study. We have not run it yet.
Market simulation
The same behavioral foundation, extended to populations. Composition, representativeness and validation are a substantial problem of their own, which is why it comes second and not first.
Can an AI anticipate what you will actually do?
Not what you say you will do. That is the question the whole company is built around, and the one a pilot is designed to answer — with the misses reported alongside the hits.
Explore the conceptWe are not building another persona system.
We are building a falsifiable behavioral model.
Language is evidence, not the whole person.
Conversation and available history become weighted hypotheses about goals, constraints and tendencies.
People are not personas.
Structured priors are a starting point, never a permanent identity. Individual evidence can overturn them.
Prediction comes before persuasion.
The model commits to a prediction before the outcome, keeps uncertainty visible, and records hits and misses.
Surprise is information.
When behavior contradicts the model, we update the hypothesis instead of explaining the contradiction away.
Simulation must answer to reality.
Forecasts are compared with withheld real observations, including the variance and the failures.
We do not claim to model the human mind. We predict one decision, commit to it before the outcome, and show you every time we are wrong.
Read the full manifestoFrom population priors to the individual.
Archetypal Mind
Structured behavioral priors.
Population-level patterns of motivation, applied as starting hypotheses that individual evidence can overturn.
Adaptive Mind
Longitudinal individual learning.
The model learns one person across interactions, choices and feedback, and carries that forward.
Augmented Mind
Consented contextual signals.
Additional signals enter the model only where they measurably improve prediction.
Both applications begin in Archetypal Mind. Later stages are earned by evaluation, not scheduled.
Research foundations.
The literature tells us the route is real. It does not tell us the layer works. Those are two different jobs.
- Digital behavioral traces carry real signal about stable individual differences.
Youyou, Kosinski & Stillwell, PNAS, 2015
Why we read the history as evidence rather than treating each message on its own. - Models trained on large-scale human choice data generalize to people they have never seen.
Binz et al., Nature, 2025
Why we think behavioral priors can transfer to a person the model has not met. - Agents grounded in long interviews with 1,052 people reproduced those individuals’ own survey answers better than demographic profiles did.
Park et al., arXiv preprint, 2024 · A preprint, and survey answers are not real-world choices. - Language models used as stand-ins for people can flatten identity groups and mislead on simple behavioral tasks.
Wang et al., Nature Machine Intelligence, 2025 · Gao et al., PNAS, 2025
Why the layer has to be evaluated on its own rather than inheriting trust from the model underneath it.
- That the layer moves one agreed, observable outcome against an equally informed baseline.
- That structured priors add something a variant without them does not.
- That predictions stay calibrated on people the model has not seen.
- Where the model systematically fails, and whether that is fixable.
Built at the intersection of behavior and AI.
Founded by YAO, applied behavioral scientist, with research focused on heuristics and decision making.
Build AI that becomes more useful as it learns about you.
We are looking for one design partner: a team with a recurring conversation their assistant gets wrong, willing to define the measure with us before we build.