Location has GPS.
Proximity has Bluetooth.
Identity has OAuth.
Cognition has nothing.
The Human Modeling Landscape
Every company teaching machines to understand humans — and the layer none of them are building.
A working thesis by Abhas Singh, founder of Mahakram.
Everyone trying to understand humans — Meta, Palantir, Google, and every startup on this map — observes from the outside: video, speech, clicks, purchases, self-reports. So do we. There is no other way to observe a person. The difference is what gets fitted to the observations.
The field fits unstructured latents — embeddings, chat memories, population priors — and guesses at the person underneath. Mahakram fits an explicit, structured, falsifiable model: named dimensions of cognitive architecture, each with a testable opposite, each graded separately. This page maps everyone fitting unstructured latents, and the empty square where the explicit layer should be.
The top-right quadrant is empty. That's the bet.
Yes — this is a founder's 2×2 with our name in the good corner. The axes aren't arbitrary: they are the two properties the rest of this page argues are economically decisive. Every placement carries its source.
| Camp | Players | Money | A person is… | Graded on | What's missing |
|---|---|---|---|---|---|
| Memory / context layers | Mem0, Zep, Letta, Cognee, Personal AI, Syke, Plastic Labs (Honcho) | Seed–A | What they SAID (chat exhaust) | Benchmark recall | Per-app, emergent, no typology, never graded on behavior |
| Simulation / synthetic audiences | Simile ($100M, Karpathy + Fei-Fei Li), Prior Computers, People Make Things, Artificial Societies, Synthetic Users, Electric Twin, Panoplai, aiphrodite | $100M+ | What a CROWD like them does | Survey/interview parity | Aggregate insight; hours of interviews per twin; no actuation yet. The one camp pointed at the empty square — treat as a trajectory, not a point |
| Psychometric traits | Humantic, Crystal, Kosinski research lineage | Modest | Scores on universal dimensions | Sales anecdotes (but: +40% field study) | Static reports for humans to read; no engine, no outcomes loop |
| Outcome decisioning | Persado, OfferFit → Braze | Acquired | NO explicit model — an RL policy that is an implicit model of each customer | Real behavioral lift | Black-box, non-portable, cold-starts from zero every client |
| Emotion / expression | Hume, Nuance Labs | ~$60–80M each | How they FEEL right now | "Feels human" / naturalness | State, not structure — the model resets every conversation |
| Cognition foundation models | humans& ($480M seed), Centaur (Nature), Be.FM, OdysSim, Large Behavioral Models | $480M+ | An implicit latent of humanity-in-general | Psych benchmarks | Population priors, not one person; no actuation; Be.FM and Centaur are open — priors are becoming free |
| Individual twins & companions | Delphi ($16M, Sequoia), Tolan/Portola, Replika, Character.AI; Dot (dead) | $36M+ among explicit twins (companions like Replika and Character.AI have raised orders of magnitude more — listed for model shape, not money) | One specific person — explicit (Delphi) or implicit (companions) | Creator revenue / engagement | Models to EXPRESS or RETAIN a person, not to MOVE one |
Memory / context layers
Simulation / synthetic audiences
Psychometric traits
Outcome decisioning
Emotion / expression
Cognition foundation models
Individual twins & companions
All seven camps fit unstructured latents — or, like Persado and OfferFit, no explicit model at all. Nobody holds an explicit, structured model of one individual graded on what that individual does.
The category has no name yet. Every lab coins its own language — 'human foundation model', 'digital minds', 'emotional layer', 'user understanding'. Keyword search fails; only scene channels surface these companies. Pre-consensus vocabulary means the window is still open — though it is starting to close: 'foundation model of human behavior' now appears in both humans& and Simile materials.
| Explicit model of ONE person | Real-time actuation | Individual behavioral ground truth | Portable across contexts | |
|---|---|---|---|---|
| Memory layers | ||||
| Simulation twins | ||||
| Psychometrics | ||||
| Persado/OfferFit | ||||
| Hume/Nuance | ||||
| Foundation models | ||||
| Delphi/companions | ||||
| Mahakram |
The square is empty for an economic reason, not a technical one: behavioral ground truth can't be scraped, licensed, or synthesized. It has to be earned — one real person, one real decision at a time.
Mahakram's row is deliberately not four filled cells. Three are properties of the architecture; the fourth is a claim that has to be earned. The Test below is how it gets earned — or killed.
KNOW
Memory layers, psychometrics, Delphi. They understand the person — but never act, and are never graded.
PREDICT
Simile, Centaur, Be.FM, synthetic audiences. Graded on survey parity and benchmarks — what people SAY.
ACT
Persado, OfferFit. Graded on behavior — but hold no explicit, portable model of the person.
Nobody holds an explicit model of an individual AND behavioral ground truth about that individual at the same time. The leading synthetic-audience players advertise ~95% survey parity. The whole field grades itself on self-report or vibes.
Persado and OfferFit prove behavioral grading works — and their black boxes are the strongest argument for the explicit version. An implicit policy is non-portable and cold-starts from zero at every client. An explicit model compounds: classify once, act everywhere, audit every dimension, carry it across contexts. That is the difference between a service and a protocol.
Built on the Objective Personality System — a decade of operator-based development. 9 independent binary dimensions of cognitive architecture, crossed with 4 social types: 29 × 4 = 2,048 distinct types. Observer-based, not self-reported: independent typologists type separately from observed behavior — video, speech, decisions — then compare dimension-by-dimension. Every dimension has a falsifiable opposite.
1. Fingerprint
A multimodal pipeline compresses video, audio, text, and behavioral signal into a single numerical fingerprint.
2. Triangulate
Embeddings are cross-retrieved against a hand-typed corpus through multiple independent similarity layers, validated where they agree.
3. Resolve
A proprietary elimination algorithm collapses 2,048 candidates to one, with confidence scored per dimension.
POST api.mahakram.in/v1/classifyOne API call returns the cognitive architecture of a consenting user: 9 binary dimensions plus social type, per-dimension confidence, an evidence chain. Classify once — one expensive call, unlimited cheap briefs.
Bet one: the layer should exist. An explicit, per-person model graded on individual behavior is missing from the map for economic reasons, and worth building. Trait-based targeting already moves behavior — a 2026 field study reports +40% — which grades the category's potential, not any typology. Bet one survives even if bet two dies.
Bet two: OPS is the right first model of the layer. Discrete cognitive types may not be real. Academic psychometrics moved from types to continuous traits for good reasons, and the system's claimed inter-rater agreement has never been independently replicated. This is a falsifiable bet, not a belief. Phase 1 is that replication — run double-blind, published either way.
Double-blind inter-rater reliability across 500+ participants.
Pass bar: Cohen's kappa > 0.6 per dimension
Predictive behavioral experiments — construct validity.
Longitudinal stability — directly testing the ~50% retest failure that breaks MBTI.
Neurological and biological correlates.
If independent raters can't agree, the type model is wrong — and I'll say so publicly. Most people building on personality frameworks are trying to prove them. I'm trying to kill mine; whatever survives is real.
Dot (New Computer)
"a living mirror of yourself"
beautifully built, shut down Sept 2025. The founders' stated reason: their north stars diverged. The reported reality: roughly 25,000 lifetime iOS downloads against claims of "hundreds of thousands" of users.
Deep user modeling with no paid outcome attached = death. Revenue funds the science; the science makes the product defensible.
Identity is dynamic. The bet here is that the architecture generating it is not. State shifts by the hour; structure — if it exists — doesn't. Whether it exists is exactly what The Test measures.
The next trillion users on the internet won't be people — they'll be AI agents. They need to understand the humans they serve, negotiate with, and act for. Today's models guess at mood and intent from context; no typology-labeled, behavior-grounded dataset of individual cognitive architecture exists to learn from. That dataset is what we're building — and it can't be bought with compute: the labels come from real people making real decisions under observation, one at a time.
Multimodal AI finally makes observation scalable. Typing a person from video and speech took trained operators hours; a pipeline now compresses behavioral signal into features at scale — a boutique practice becomes infrastructure.
Psychology can't fix this from the inside. No lab can recruit 1,000 typed individuals per type. The internet can.
"Everyone is building machines that understand people; they grade themselves on what people say. We grade ourselves on what one specific person does next."
Consent-first by design: classification runs only on people who opt in. EU AI Act biometric-categorization provisions treated as a hard constraint, not an afterthought.
▶Sources
humans& — humansand.ai ($480M seed, $4.48B val — Crunchbase News, Jan 2026)
Simile $100M Series A — simile.ai/blog/the-simulation-company, Bloomberg Feb 2026, indexventures.com
Nuance Labs — nuancelabs.ai, geekwire.com, lsvp.com
Hume $50M Series B — hume.ai/blog/series-b-evi-announcement
Delphi $16M Series A — delphi.ai/blog
Plastic Labs — plasticlabs.ai
Centaur — Nature s41586-025-09215-4
Be.FM — arXiv 2505.23058
Large Behavioral Models — research.unboxai.com
OdysSim — arXiv 2606.14199
Matz/Kosinski PNAS 2017; SSE 2026 field study (+40% / -20%)
OfferFit → Braze — braze.com; Persado — persado.com
Tolan/Portola — geekwire.com
Dot shutdown — techcrunch.com Sept 2025; shutdown note — new.computer; downloads figure — Appfigures via TechCrunch
Electric Twin, Synthetic Users — the leading synthetic-audience players advertise ~95% survey parity (electrictwin.com, EY/Evidenza study)
Prior Computers, People Make Things — newcomer.co
© 2026 Mahakram · mahakram.in · thesis last updated July 2026