A Research Program

The Relationship
Singularity

The hypothetical point at which computational models of human personality reach sufficient resolution that relationship outcomes become predictable before two people ever meet. Not a dating app. A research frontier.

The Problem

We Are Pre-CLIP for Human Personality

We are living in a social dark age. Married people report happiness equivalent to a $50,000 salary increase. Finding a one-in-a-million match instead of one-in-a-hundred doesn't make someone a little happier — it satisfies something at the level of the soul. The goodness of relationships follows power laws, and we have almost no infrastructure for navigating them.

The Big Five personality model is a 5-pixel image of a person.

Before CLIP, computer vision meant classifying images with a handful of hand-labeled features. After CLIP, every image on earth could be embedded in a single space at whatever resolution was needed. The same transition is coming for human personality — and it changes everything about how we find the people we're meant to know.

This is not speculation. Alan Cowen at Hume AI already proved the paradigm for human emotion: unsupervised high-dimensional representations recover structure — awe vs. surprise, nostalgia vs. melancholy — invisible to prior label-based frameworks. We are extending this to personality, and from personality to relationship outcome prediction.

The Research Program

Six Open Research Directions

Each represents a tractable first experiment with existing data. Together they form a path from personality measurement to relationship prediction.

01

Personality as Embedding Space

Unsupervised, high-dimensional, multimodal personality vectors learned from text, voice, video, behavior, and digital footprints. If I can predict every token you'd generate, I have a model of your personality. The conditional embedding is the personality.

02

The Cultural Trope Hypothesis

All important personality trait combinations eventually become tropes in a culture's stories. TV Tropes is a bipartite graph of characters and tropes — unsupervised personality clustering by thousands of human editors. Node2Vec on this graph produces a character embedding space where similar characters cluster. The training data already exists.

03

Relationship Outcome Manifolds

Compatibility is not a number. It's a structured manifold — a probability distribution over the kinds of relationship dynamics two people will discover. A compatibility score is to a relationship outcome manifold as a temperature reading is to a full weather model.

04

Emotion Embeddings as Proof of Concept

Alan Cowen's work at Hume AI proved this paradigm for emotions. Self-reported labels are a lossy pre-CLIP taxonomy. Unsupervised representations recover structure invisible to prior frameworks. If it works for emotion, it works for personality. The style-GAN moment is imminent.

05

Moneyball Mutual Attraction

Generate the manifold of all possible faces, walk users through latent space to map aesthetic preferences, and identify regions of high mutual attraction that neither person would have found by swiping. High-variance appearance gets more messages than high-average. The alpha is in the tails.

06

LLM Relationship Simulation

Embed two personalities as conditional prompts for language models. Monte Carlo simulate their conversations, conflicts, and resolutions. Run 10,000 dates before the first one. Generative agents in a town was a proof of concept — the next step is generative agents in a relationship.

Love Symposium · November 2024

The Full Research Roadmap

Two hours covering personality embedding spaces, the Cultural Trope Hypothesis, Cowen's emotion manifolds, Moneyball mutual attraction, LLM relationship simulation, geographic optimization, and the path forward.

What’s Happening

Get Involved

This research program lives across several venues. Here’s where the work is happening and how to participate.

Press & Resources New York Times “Can You Optimize Love?” — January 2026 Paper A Theoretical Roadmap to the Relationship Singularity YouTube Full 2-hour research talk — Love Symposium 2024 Conference Love Symposium — San Francisco
References

Key Papers & Resources

The research program draws on work across representation learning, computational social science, personality psychology, and affective computing.

Representation Learning

  1. Learning Transferable Visual Models From Natural Language Supervision. Radford et al. ICML 2021. The CLIP paper — the governing analogy for the entire research program.
    arxiv.org/abs/2103.00020
  2. node2vec: Scalable Feature Learning for Networks. Grover & Leskovec. KDD 2016. Graph embedding method applied to the TV Tropes bipartite character-trope graph.
    arxiv.org/abs/1607.00653
  3. A Style-Based Generator Architecture for Generative Adversarial Networks. Karras, Laine & Aila. CVPR 2019. StyleGAN — the “style-GAN moment for personality” analogy.
    arxiv.org/abs/1812.04948
  4. Efficient Estimation of Word Representations in Vector Space. Mikolov et al. 2013. Word2Vec — early proof that embedding spaces have geometric structure (king − man + woman = queen).
    arxiv.org/abs/1301.3781
  5. LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models. Schuhmann et al. NeurIPS 2022. Five billion image-text pairs that made CLIP's embedding space comprehensive.
    arxiv.org/abs/2210.08402

Emotion & Affect

  1. Self-report captures 27 distinct categories of emotion bridged by continuous gradients. Cowen & Keltner. PNAS 2017. Foundational evidence that discrete emotion labels are lossy — unsupervised high-dimensional representations recover richer structure.
    doi.org/10.1073/pnas.1702247114
  2. What music makes us feel: At least 13 dimensions organize subjective experiences associated with music. Cowen et al. PNAS 2020. Extension of the emotion embedding approach to music — demonstrating cross-modal consistency.
    doi.org/10.1073/pnas.1910704117
  3. Hume AI. Alan Cowen’s company building emotion AI infrastructure. Proof of concept that the embedding paradigm works for psychological constructs.
    hume.ai

Personality & Social Prediction

  1. Computer-based personality judgments are more accurate than those made by humans. Youyou, Kosinski & Stillwell. PNAS 2015. With 275 Facebook likes, a model predicts Big Five traits more accurately than a spouse.
    doi.org/10.1073/pnas.1418680112
  2. The Big Five personality dimensions and job performance: A meta-analysis. Barrick & Mount. Personnel Psychology 1991. Foundational Big Five work — the “5-pixel image” we aim to surpass.
    doi.org/10.1111/j.1744-6570.1991.tb00688.x
  3. The Mathematics of Marriage: Dynamic Nonlinear Models. Gottman et al. MIT Press 2005. Mathematical modeling of marital interaction — early work treating relationships as dynamical systems.
    mitpress.mit.edu

Social Simulation & Agents

  1. Generative Agents: Interactive Simulacra of Human Behavior. Park et al. UIST 2023. Stanford’s generative agents in a town — the proof of concept for LLM relationship simulation.
    arxiv.org/abs/2304.03442
  2. Simulate Before You Act: LLM Agents Simulate Turn-Taking in Relationships. RELATE-Sim. LLM agents initialized with user personas simulate “turning point” conversations, predicting relationship outcomes.
    arxiv.org/abs/2502.05058

Data Sources

  1. TV Tropes. A wiki of tens of thousands of character tropes linked to tens of thousands of fictional characters. The bipartite graph that encodes unsupervised personality clustering by collective cultural observation.
    tvtropes.org

This Research Program

  1. A Theoretical Roadmap to the Relationship Singularity. Fisher. 2025. The full theoretical framework for personality embeddings and relationship outcome prediction.
    relationshipsingularity.org/ai4hc/roadmap.pdf
  2. Relationship Outcome Prediction Tech. Fisher. Love Symposium, November 2024. Two-hour research talk covering the full roadmap.
    youtu.be/2pQrl_LsjKU
  3. “Can You Optimize Love?” The New York Times, January 2026. Coverage of the Love Symposium and this research program.
    nytimes.com
  4. Love Symposium. A conference in San Francisco exploring technology’s role in human connection.
    symposium.love
About

Matthew Fisher is an AI researcher working on personality embeddings and relationship outcome prediction. He co-founded the Love Symposium in San Francisco, a conference exploring technology and human connection covered by The New York Times.

His research draws on representation learning, computational social science, and the hypothesis that human personality can be embedded at arbitrarily high resolution — and that relationship outcomes between any two embedded personalities can be predicted as structured distributions, not scalar compatibility scores.

staff@relationshipsingularity.org