The Journey
Every conversation with an AI starts from zero. This is the story so far of teaching an AI how a specific person thinks, decides and communicates, told as the questions we set out to answer and what each answer made us ask next.
Can an AI remember me?
Every conversation with an AI starts from zero. Memory tools are starting to fix that: they store what you've said and bring it back later. But remembering what you said isn't the same as knowing how you think.
Two people can look at the same facts and reach opposite conclusions. An AI acting for you would need to know how you would read a situation, not just what you've told it before.
What we learned
- ›The missing piece isn't recall. It's interpretation
Which raised the next question: Can an AI learn how a specific person thinks?
Can it learn how I think?
We built a process that reads what someone has written and pulls out short, specific statements about them, each tied to the words it came from. It organises them into three parts: the beliefs they reason from, how they communicate, and how they tend to respond in particular situations. We call the result a behavioral specification: a written portrait of how someone thinks. We tried it on everyday conversations, journals, Benjamin Franklin's autobiography, Frederick Douglass, Mary Wollstonecraft, Theodore Roosevelt, patent filings, and the letters of Warren Buffett and Howard Marks.
Then we tried to break it. A panel of AI reviewers we had added turned out to change nothing. Ten of the fourteen steps made no difference, so the process shrank to four. Some tests failed outright.
In May we published the research as Beyond Recall, revised in August. Across fourteen public-domain autobiographies, a short portrait gave an AI about three quarters of the benefit of reading everything the person wrote, at around a twenty-fifth of the length. It helped most on questions of judgement, and could get in the way on plain questions of fact. Remembering and understanding really are different jobs.
What we learned
- ›The three-part structure is what matters. Most of the machinery around it didn't
- ›Grading AI with AI goes in circles unless the judges are calibrated
Which raised the next question: If the portrait changes how an AI reasons, is it actually right about the person?
Is it really me?
A portrait that changes how an AI reasons isn't necessarily accurate. Telling people apart is not the same as describing them correctly, and only the person can confirm it. That gap shaped everything that came next.
Once someone's writing got long, the process only ever looked at a small slice of it: a limit nobody had chosen was quietly throwing most of the evidence away. Letting an AI decide how much to read didn't fix it; it read far too little. So we rebuilt it to read everything, keep contradictions instead of smoothing them over, and cite its evidence for every claim.
We also found that the easiest way to describe the wrong person is to listen to the wrong voice: text someone pasted in, the AI's own replies, other people in the conversation. Now anything said about a person has to rest on words they actually said or wrote, quoted exactly, and a second AI checks every statement against its quote. When a statement claims more than its quote shows, it is flagged rather than hidden, so the person can see where the portrait is reaching.
What we learned
- ›A limit nobody chose can throw away most of the evidence
- ›People hold contradictions. A good portrait keeps both sides
- ›Every claim should point back to the person's own words, so they can check it and correct it
Which raised the next question: What happens when the portrait has to act?
Can it act for me?
The point of understanding someone is to act well on their behalf. When an AI agent books, sends, approves or spends for you, something has to decide whether to go ahead, stop, or check with you first.
We built an early prototype of that decision step on a public draft standard for agent permissions (AAuth), and started testing how a portrait informs it: which part applies to this situation, what evidence supports it, and when the right move is to bring the decision back to the person.
What we learned
- ›For an agent to follow a rule, it needs to know when the rule applies, why it holds, whether it was checked, and how to fix it
- ›Fixing a mistake is easy. Finding everything the mistake affected is the hard part
- ›Checking with the person is a real choice with a cost, not a failure
Where we are now
The portrait is becoming something an agent checks before it acts for you. Each one is now organised by the situations it applies to, so an AI can find the right part at the right moment, and we are testing a way for it to pull in just those parts during a conversation, with a record of what it used.
What isn't solved yet is the part that matters most: knowing whether the AI actually used what it was given, knowing when it should ask you instead of acting, and making sure you can always see and correct what it believes about you.
We are building portraits for people working in this field, with their permission, and asking them where we got them wrong.
What we learned
- ›We're not looking for validation. We're looking for feedback
Open Questions
What we don’t know yet: the questions that come from where we are now.
The code, the paper, the data and the failures are public. Some research tools aren’t published yet. We’re not looking for validation. We’re looking for feedback.