This write-up outlines how we might teach machines to understand us, how such systems could be built, and who is building related technology today.
Introduction
Despite centuries of work across philosophy, psychology, and cognitive science, our understanding of the human mind remains limited. In recent decades, we have meanwhile achieved great success applying deep learning across domains, with these methods winning the Nobel Prizes in Chemistry and Physics in 2024.
Many have asked whether this success will also lead to breakthroughs in how we understand and model the human mind (Binz et al., 2025). Perhaps this direction holds the answer to some of humanity’s greatest mysteries.
Contributions
In this write-up, we will summarize a recent paper we have published where we:
- Formalize a definition of a “foundation model of human behavior.”
- Survey three fields building these models: their methods, results, and trajectories.
- Prove these fields are converging, and formalize their convergence into a common model of next-behavior prediction.
- Outline the areas of progress in this line of technology and call for interdisciplinary collaboration to work on these problems.
What It Means to “Understand”
If our goal here is to teach computers to understand humans, we should be precise in how we define the term.
This definition doesn’t fit cleanly onto computers. In machine learning, the commonly cited thought experiment is the Chinese Room, which shows how a machine might be able to perfectly process information without understanding any of it.
Can a machine understand at all? Does it even matter?
Alan Turing argued that these debates often devolved into meaningless arguments and believed that if the outcome is the same as if one understands, the output is what matters.
We adopt his framing around the problem.
If a machine can predict or simulate what a person will do in any given situation, that machine understands the person.
If we think about how a computer “understands” information, the closest analog is how a computer “represents” information. Bengio defines a good representation as containing the explanatory variations in data, and being general enough so that this representation can be used across a wide range of tasks.
We can think of this representation as a compressed, context-dependent state that captures what the model currently understands about a person. The model uses that state to assign probabilities to what the person might say or do next.
p(at+1 | hu,t) = g(zu,t) (2)
Without diving deep into math, h represents context, broadly. Everything the model has been given about the person flows into it. The model compresses all of it into zu,t and predicts from that state. Nothing about the representation is designed; it is whatever the model needs in order to predict well.
Why should we care about understanding at all?
Why does understanding and representation even matter?
It matters because a computer’s ability to predict well relies upon its ability to represent. This is not always in our hands during the training process. A computer self-learns these representations. A good representation is necessary for a good prediction but is obtained indirectly.
And economically, it is very important for companies and people to have software that understands them, so that software can better serve us and provide us more agency. Companies already predict what you like on Netflix, Spotify, and Amazon to give you a better experience.
This trend is still in the very early innings. In the near future, all software will understand you better than you understand yourself.
Foundation models
Now we will ground these claims into practical, achievable steps to build this model. The wonder of modern AI is that we built the most complex, intelligent systems in history, and we did it through a relatively simple process. We simply asked a computer to predict the next word.
These representations were not explicitly programmed. They emerged from training on massive datasets under a generic prediction objective. The key ingredients were self-supervised learning, attention, and scale.
The reason we call them foundation models (Bommasani et al., 2021) is that they were trained generally, and could then be adapted for a wide range of tasks. Today, we use the same models across tasks related to coding, finance, and personal assistants.
Foundation models of human behavior
We define a foundation model of human behavior as follows.
A model trained on behavioral data under a generic self-supervised objective, where the resulting model can be used across a wide range of downstream tasks.
Today’s LLMs provide an early demonstration of the underlying idea. To predict human language reliably, models must learn some representation of human emotions and psychology (Anthropic, 2026).
The rest of this write-up will attempt to take this insight to the limit. We care that a model can represent a human perfectly.
Survey: three fields
We cannot claim to have invented this field. Computer scientists have asked these questions for 100+ years. In recent years, this research direction has picked up steam. We claim that the following three fields are converging on a common foundation model of human behavior:
HCI: developing software and assistants that can better serve us by understanding who we were, who we are, and who we want to become.
Simulation: simulating human nature across a wide range of environments and building the “what-if machine.”
Prediction: building models to predict the next action a user will take (click, purchase, view, etc.) within a platform like Netflix or Meta.
These disciplines occasionally overlap but ultimately focus on different tasks. Importantly, each must understand and represent a user generally, and is developing more general models every year. Today, each has converged on the exact same formula for prediction.
The Common Formula
The common formula is simple.
The probability of the next behavior, at+1, given everything known about the person, h. A behavior can be a click, a purchase, a word said, a survey answer, a step taken in a room.
The generality is important because h can contain many kinds of context. Recommendation systems retain actions and catalog information; HCI systems retain user history and the current interface; simulations retain the situation, environment, and persona.
Even in recommendation systems, feature engineering is becoming context engineering.
The problem shifts “from feature engineering to context engineering.”
“Innovation moves up a level: from per-task architecture design to questions of data, scaling laws, post-training strategy, and inference optimization.”
Li, Sehgal, and Rao, GenRec: An LLM-backed Recommendation Ranker at Netflix, 2026
Future Directions
The convergence described above points to a shared research agenda. Building more capable foundation models of human behavior will require progress along three fronts: generality across domains, a shared language for representing behavior, and access to high-quality behavioral data and computation at scale.
Generality
We should consider each narrow predictive model as a domain expert. A music model, a shopping model, and a health model each learn from a different slice of human behavior. Training across these domain silos could reveal patterns that transfer between them and produce a more general understanding of human behavior than any one model could learn alone.
Lingua Franca
Today, these models are limited by siloed data and the absence of a common language across domains. Models must be able to tokenize not just words, but actions as well. Doing so requires a high-fidelity translation layer across different kinds of behavior.
Scaling
We have observed scaling laws in this space as well. We need to make sure we have the correct data mixture. But ultimately, if you assume we have enough computational resources, this is the path to scaling a machine that understands us better than we understand ourselves.
Recap
How could we teach a computer to understand humans?
We explore three paths and argue that they converge and generalize into a surprisingly simple formula: next behavior prediction.
But simple does not mean easy. The path ahead is bottlenecked by access to high-quality behavioral data in context, a common language across domains, and the coordination required to connect them responsibly.
References
- Bengio, Y., Courville, A., and Vincent, P. (2013). Representation Learning: A Review and New Perspectives. IEEE TPAMI. arXiv:1206.5538
- Bommasani, R., et al. (2021). On the Opportunities and Risks of Foundation Models. arXiv:2108.07258
- Binz, M., et al. (2025). A foundation model to predict and capture human cognition. Nature, 644. doi:10.1038/s41586-025-09215-4
- Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3).
- Turing, A. M. (1950). Computing Machinery and Intelligence. Mind, 59(236).
- Templeton, A., et al. (2024). Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. transformer-circuits.pub
- Anthropic (2026). Emotion concepts and their function in a large language model. transformer-circuits.pub
- Park, J. S., et al. (2024). Generative Agent Simulations of 1,000 People. arXiv:2411.10109
- Shaikh, O., et al. (2025). Creating General User Models from Computer Use. UIST. arXiv:2505.10831
- Zhai, J., et al. (2024). Actions Speak Louder than Words. ICML. arXiv:2402.17152
- Zhou, G., et al. (2025). OpenOneRec Technical Report. arXiv:2512.24762
- Li, Y., Sehgal, S., and Rao, A. (2026). GenRec: An LLM-backed Recommendation Ranker at Netflix. arXiv:2608.10257
- Epic Cosmos and Microsoft Research (2025). Generative Medical Event Models Improve with Scale. arXiv:2508.12104
- Simile (2026). Building the What-If Machine. simile.com