Teaching Machines to Understand Humans

Toward a Foundation Model of Human Behavior

Jonathan Politzki, Jean Technologies. Working paper, 2026. Fourteen pages, with the fields' definitions and objectives collected in the appendices.

Open the PDF

The purpose of this write-up is to outline how humans can teach machines to understand us, how we can practically build it, and survey people who are building similar technology today.

Introduction

All human attempts to understand the human mind have failed, or made modest progress. We can dig deep into history: to the early philosophers Aristotle, Socrates, the psychoanalysts Freud and Jung, up to contemporary cognitive scientists.

At the same time, we have witnessed massive recent success applying deep learning across domains, with these methods winning the Nobel Prizes in Chemistry and Physics in 2024.

In recent years, many have asked whether the success of massive computation and data will also lead to breakthroughs in how we understand and model the human mind. Perhaps this direction holds the answer to questions we may never be able to answer ourselves.

Contributions

In this write-up, we will summarize a recent paper we have published where we:

1. Formalize a definition of a “foundation model for human behavior.”
2. Survey three fields building these models: their methods, results, and trajectories.
3. Prove these fields are converging, and formalize their convergence into a common model of next-behavior prediction.
4. Outline the areas of progress that are most crucial to developing this technology and call for inter-disciplinary collaboration to work on these bottlenecks.

What it Means to “Understand?”

If our goal here is to teach computers to understand humans, we should be precise in how we define what the term means. Then, we will do our best to articulate why it is so important.

It’s not clear what “understand” really means. But our definitions circle around how well we know what causes something or what the explanation of a concept or phenomenon is.

This definition doesn’t fit cleanly onto computers. In machine learning, the common thought experiment is the Chinese Room Experiment, 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, does it even matter? He cared more about the end than the means to achieve it.

We adopt his framing around the problem.

And to formalize this logic further, if we think about how a computer “understands” information, the closest analog is how a computer “represents” information. Bengio defines a good representation roughly as containing the explanatory variations in data, and being general enough such that this representation is useful across a wide and general range of predictions.

We can mathematically define this representation as a learned prior to predict the next token, word a human says, or action they may take.

Without diving deep into math, h represents context. For instance, h can include memory of a user, a stream of actions a user takes, the context, their current environment, their persona, and other context that is useful for “understanding/representing” that person properly.

Why should we care about understanding at all?

Why does understanding, or their proxy representation, even matter? Does it buy us, or computers, anything?

Per the definition above, we actually don’t… But since the downstream predictive performance, how well a computer can predict what a person would do in any situation, relies upon the computer’s ability to represent the person, we do. 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 a computer’s understanding 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, 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.

They learned the complexity themselves, we just gave them a lot of data and let them through this generic objective of predicting the next token. The key ingredients were 1) self-supervised learning, 2) attention, 3) and massive data scale.

The reason we call them foundation models (Bommasani et al., 2021) is that because these models were trained generally, they 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 a model trained on behavioral traces, under a generic self-supervised objective, such that training learns a person-level representation from that person's history alone, and the one resulting artifact serves many tasks.

And one of the most interesting byproducts of this training process was that in order to predict the next word that a human says or writes, you must understand much more than just the world. It actually requires a model to begin to self-learn some of the underlying psychology of the human mind. It learned to understand us.

We can therefore classify existing models as foundation models for human behavior. These models have our psychology built into them. We have already distilled the human mind into them.

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

I cannot claim to have invented this field. Computer scientists have asked these questions for 100+ years. And in recent years, this research direction has picked up steam. We claim that the following three fields are converging on a common foundation model for 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. 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. Each is developing models that become more general every year. Today, each has converged on the exact same formula for prediction.

The Common Formula

The common formula is predicting a_t+1 from h, Eq. 2 above.

The generality is important, but h can be decomposed. It can be decomposed into context engineering, which is familiar terminology, or into context. We can assume infinite and rich context, which is helpful because it makes this clean. But of course, in reality, that is not tractable, or basically not practical. Recommendation keeps the actions and the catalog. HCI keeps a memory of the user and the current screen. Simulation keeps the situation, the environment, and a persona written out in words. Each is Eq. 2 with part of h removed, summarized, or handed in from outside.

More specifically: E could be environment. We might have prior actions, the context of those actions, the environment of those actions. We might have situations. We might have a persona that has been written out. We might have metadata. We might have options, to constrain the output space.

Feature engineering is becoming context engineering. We are seeing this trend across each of these industries. Netflix, on their move to an LLM-backed recommender:

Historically, recommender systems have relied on a wide variety of bespoke architectures: two-tower models, DLRM-style feature interaction networks, custom attention blocks, and large multi-task setups. LLM-centric systems instead standardize on transformer backbones inherited from a foundation model. Innovation moves up a level: from per-task architecture design to questions of data, scaling laws, post-training strategy, and inference optimization.

Classical RecSys stacks can hit diminishing returns due to sparse IDs, heavily engineered objectives, and task-specific architectures. In an LLM-backed setting, the recommendation model shares its backbone with a pre-trained LLM and inherits its data and model scaling behavior. Our studies show clear scaling trends in both data and model size: performance improves monotonically as we add more Phase-2 training data, and larger models consistently outperform smaller ones under fixed training budgets. This brings RecSys closer to the broader LLM paradigm, where scaling laws are treated as a central design guide rather than an after-the-fact observation.

Li, Sehgal, and Rao, GenRec: An LLM-backed Recommendation Ranker at Netflix, 2026

Future Directions

Generality

We should imagine each of the above as narrow, siloed domain experts that each understand a different fragment of a person. Understands what music we like and why, what we look for in relationships, what we aspire to achieve. While narrowly trained, these models can learn a lot from collaborating and sharing information about an individual.

Common lingua

The fault of training these models today has similarly been siloed data. And creating a common language across them. We have to be able to tokenize not just words for a model to process but actions as well, and this requires deep engineering to develop a high fidelity translator of sorts.

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. This is likely the correct direction for the next 20 years.

But simple does not mean easy. The path ahead is massively bottlenecked by data, different languages, and coordination.

Read the paper

The full argument, the survey of the three fields, the reduction of each field's formula to next behavior prediction, the scaling results, the corpora, and a limitations section.

Toward a Foundation Model of Human Behavior (PDF)