When thinking is almost free
Attention per person, not model size
A hundred thoughts on every sentence a learner says.
Vyakti · relational intelligence
AI knows almost everything. Except you.
Vyakti is research toward AI that understands the person in front of it: what they really understand, how they think, what they feel, and how to earn their trust. It gets better at each person with every conversation. We start with learning, the one place where understanding a person can be measured.
A research proposal for the Lossfunk residency · October 2026
listening
01 · The problem
Intelligence is no longer the bottleneck. Understanding the person is.
Models now solve olympiad problems and write production code. Yet a tutor still can’t tell whether a child understood, a coach can’t tell when you are about to quit, and an assistant agrees with you when it should push back.
Remembering a person is close to solved. Understanding one, acting on that understanding, and getting better at it over time are not, and almost nobody measures them.
How far AI has climbed
Remember youRecall what you said
0LongMemEval
Infer youWork out what you know, want or believe
37–65%PersonaMem-v2 · CUPID
Act on itUse what it inferred
0 → 95% if remindedSimpleToM
Get better at youImprove with every conversation
no benchmark exists
?Vyakti’s question
02 · Where we start
Learning is where misreading a person costs the most, and where it can be proven.
- Confusion hides. In class, a student asks about one question every nine hours. With a one-to-one tutor, about 26 questions get asked an hour. Graesser & Person 1994
- Learners can’t see it either. How well people judge their own understanding tracks reality at only about .27. Dunlosky & Lipko 2007
- AI tutors make it worse when they trust surface success. Practice scores rose 48%; exam scores fell 17%. Bastani et al., PNAS 2025
- And learning has an answer key. A fresh problem a week later shows whether the AI was right about the person.
UnderstoodHalf understoodLost, hiding itIllustration
03 · The broader vision
Understanding a person, mapped.
Learning is the first door. Behind it is the whole person: seven families, 25 dimensions. Each has a core question, a reason it is still open, and a ground truth that could prove the AI right. That last column decides what can be researched first.
04 · The insight that ties it together
People reveal more to someone they trust. So the relationship is part of the instrument.
.00
How strongly the strength of the relationship predicts outcomes in therapy, across nearly 300 studies and more than 30,000 patients. Teams that feel safe admit errors and ask for help.Flückiger et al. 2018 · Edmondson 1999
For an AI, memory, one steady personality and a bond that grows are not product features. They are how honest signals get in. Relational intelligence and understanding are one loop.
05 · The hypothesis
Language models learned language by predicting the next word. Vyakti learns a person by predicting the person.
It forms beliefs about you, bets on what you will do next, and lets reality grade every bet. Each miss rewrites its theory of you. So it should understand you better in week six than in week one, and teach you better because of it.
H1 · SeeingUnderstanding can be read from ordinary interaction
Wrong ifits reading predicts later performance no better than asking the learner.
H2 · ActingWriting a belief before each move turns inference into action
Wrong ifmoves fit the learner no better, judged blind.
H3 · TrustingContinuity and a steady self make people reveal more
Wrong ifadmitted confusion doesn’t rise with memory switched on.
H4 · GrowingAccuracy about each person rises with time
Wrong ifthe hit rate stays flat, or plain memory keeps pace.
Does it get better at you?Hypothesis
VyaktiSame model, plain memoryLearner’s own judgement
Share of written predictions that came true: will she solve a new problem next week, where will she get stuck, will she ask for help, will she come back.
06 · A theory that never sets
Every belief is a bet. None of them is permanent.
Vyakti keeps a written theory of each person, in its own words. Every interaction can move it. Beliefs reality contradicts get rewritten, with the reason kept. Beliefs nobody re-confirms fade. So it can’t harden into a label.
- Evidence, not impressionsEvery belief cites what it saw.
- Confidence that movesEach interaction nudges it up or down.
- Fades without fresh proofOld beliefs decay unless retested.
- Yours to read and correctThe person can see it, fix it, delete it.
Meher
Illustrative learner · rewritten after every session07 · The system
Software made of minds, each on its own clock
LLM-native end to end. No hand-written decision tree: every judgement is a model’s, written down with its reason. Small fast models think inside the pause before the teacher speaks. Strong models reflect overnight. Every week the system rewrites its own methods. Code keeps the records. Reality keeps the score.
One moment, slowed down1 / 7
Liveevery turn · fast models
Listenernotices, never concludes
Directorstates intent, picks the surface
Proberdesigns quiet checks
Makerbuilds for the exact gap
Voiceone steady personality
Heldwritten by models · read before every move
Prediction ledgerevery bet, with a due date
Theory of the personbeliefs · evidence · confidence
Selfwho I am with you
Overnightstrong models · judged by reality
Realitywhat the person does next
Reflectorscores bets, rewrites the theory
Dreamerrehearses with a simulated you
Weekly, then lateracross all learners
Playbook evolverrewrites how every mind works
Self keepergrowth without drift
Weightslater: a model that has learned you
08 · Multimodal, both ways
Every signal in. Any surface out.
Understanding rarely sits in the words alone. Vyakti reads what a learner gives it, with consent: voice, words, hands, screen and time. Each signal alone is a hint; together they are evidence. Then the Director chooses the surface the way it chooses words.
Voice
3 s of silence, then a quick “got it”
Words
“okay, makes sense”
no words of her ownHands3/8 → 3/5erased her first answer
Screen
step 3 · 40 s
stuck on the same stepTime
skipped decimals twice
Togetherprobably lost. Plan a quiet check, not a quiz.
Vyaktivoice
VyaktiI’ll talk, you listen. You’re walking home, right?
09 · Learning anything
Change the subject. Keep the understanding.
The answer key changes from subject to subject. The person doesn’t. What Vyakti learns about how you learn, what you avoid and how you talk travels with you: from fractions to French, from code to chemistry, from a child’s homework to an adult’s new career.
That is how one system could make expert one-to-one teaching, the most effective form of teaching we know, available for anything, to anyone.
What travels with you
how you learn bestwhat you avoidstuck or thinkingwhat you care abouthow you talkhow much you trust
One theory of youMeherlearns by predicting first · hides confusion when watched · loves cricket
10 · Beyond one learner
One mind, many bounded relationships
FirstLearning
NextMentors
NextCompanions
NextElder care
NextHealth
11 · Built for the models that are coming
As models get stronger, understanding the person becomes the bottleneck
Raw intelligence won’t tell you that a student’s silence means “I’m scared to ask.” Every jump in capability makes that gap matter more. Five bets.
When models answer instantly
Thought that fits inside a pause
Real-time theory of mind becomes part of every voice agent.
When models learn continually
The theory becomes weights
A model that has learned you. You should own it, read it, delete it.
When systems improve themselves
A reward that can’t be faked
Whether a real person later does what you predicted. Growth, never engagement.
When intelligence is general
The last mile is one person
An AGI aligned with humanity still has to be aligned with this child, today, in this mood.
12 · What we will work on at Lossfunk
One hard question: can an AI measurably get to know a person?
- 01Can understanding be read without testing?
From conversation, voice and behaviour alone, does Vyakti’s reading of a learner predict how they do on a fresh problem a week later?
dimensions 1 · 2 · 3 · 12 · 13 - 02Does a living theory beat memory?
Does an explicit, evolving theory of the person predict them better than the same model with plain memory?
dimensions 5 · 24 · 25 - 03Is trust a sensor?
Do continuity and a steady personality make people reveal more of what they really think?
dimensions 11 · 15 · 21 - 04Can it improve itself from its misses?
When the system rewrites its own methods from wrong predictions, does it get measurably better at new people?
the loop · G
What will exist
- The Vyakti layeran open, LLM-native architecture any agent can use
- “Does it know you?”an evaluation that scores an AI’s beliefs about a person against what they later do
- Living theoriesreal learners’ theories evolving over weeks, shared with consent
- The write-upwhat worked, what broke and why. Failures first.
Every prediction and every miss is logged in the open.
Always says it is an AIHelplines built inNo companion behaviour with minorsThe person owns their theoryScored on being right, never on being liked
Vyakti · relational intelligence
Start with a teacher that truly understands. Then everything that talks to you.
What this does not claim: that Vyakti understands people the way people do. The claim is narrower and testable. Its beliefs about a person can be scored against that person’s life, and made better with every conversation.
raghavsharma1729@gmail.com