Centaur AI: Becoming Human
By learning to predict our choices across diverse scenarios, the model's internal structure began to resemble our neural patterns.
With every article and podcast episode, we provide comprehensive study materials: References, Executive Summary, Briefing Document, Quiz, Essay Questions, Glossary, Timeline, Cast, FAQ, Table of Contents, Index, 3k Image, and Fact Check.
Please support my work by becoming a free subscriber. (Your subscription helps me continue providing evidence-based analysis in a media landscape increasingly hostile to inconvenient truths.)
We're living through a moment that feels like science fiction, but it's happening in sterile research labs with the clinical precision of peer review. A new AI model called Centaur can predict human behavior with startling accuracy—not just what we'll choose, but how long we'll take to decide, and even what our brains will look like while we're thinking about it.
This isn't your typical AI story about chatbots or image generators. This is about something far more intimate: a machine that has learned to think like us by watching millions of our decisions.
The Unified Theory We've Been Waiting For
For decades, cognitive scientists have been building models that work like specialists—brilliant at one thing, useless at everything else. AlphaGo can demolish world champions at Go, but it can't help you decide what to have for breakfast. Prospect theory explains why we make irrational financial decisions, but it's silent on how we learn languages or navigate social situations.
The human mind, meanwhile, is the ultimate generalist. We seamlessly switch between choosing cereals, solving math problems, reading emotions, and planning vacations. We're cognitive Swiss Army knives, and until now, no computational model has captured that versatility.
Enter Centaur, built on Meta's LLaMA language model but trained on something unprecedented: Psych 101, a dataset containing over 10 million human decisions across 160 different psychological experiments. Think of it as a massive catalog of human behavioral patterns, from gambling choices to memory tasks to social predictions.
The results are unsettling in their accuracy. Centaur doesn't just predict what people will choose—it predicts how long they'll take to decide, captures the full spectrum of human decision-making strategies, and even mirrors the patterns of brain activity measured in fMRI scans. It does all this despite never being trained on brain data.
The Uncanny Valley of Behavioral Prediction
What makes Centaur particularly eerie is how it captures our blind spots. In one test, it predicted human behavior in a social game with 64% accuracy—pretty good. But when asked to predict an artificial agent's behavior in the same game, its accuracy plummeted to 35%. This perfectly mirrors how humans perform: we're good at predicting other humans, but struggle with simple artificial rules.
The model doesn't just get our decisions right; it gets our mistakes right too. It's developed our cognitive biases, our inconsistencies, our very human way of being predictably irrational. This isn't a bug—it's a feature that reveals something profound about how our minds work.
When Centaur was tested on completely new scenarios—tasks it had never seen before—it still captured human behavior remarkably well. More striking still, after learning about decision-making and memory, it became better at predicting human performance on logical reasoning tasks it had never encountered. The model had developed something like general cognitive abilities.
The Implications Are Staggering
If this technology continues to advance, we're looking at a future where AI doesn't just assist us—it anticipates us. Imagine systems that can predict not just what you'll click next, but how you'll feel about it, how long you'll hesitate, and what cognitive strategies you'll use to make the decision.
This has obvious applications for personalized learning, where understanding individual cognitive styles could revolutionize education. But it also raises profound questions about privacy, autonomy, and the nature of free will itself. If a machine can predict your behavior with 87% accuracy for response times and remarkable precision for choices, what does that say about human agency?
The researchers behind Centaur are positioning it as a tool for scientific discovery—a way to help cognitive scientists build better theories about how minds work. They've already demonstrated this by using the model to refine explanations of human decision-making, essentially using AI as a benchmark for human behavior.
The Democratization of Mind Reading
Perhaps most concerning is how accessible this technology is becoming. The entire Centaur model was trained in just five days on a single powerful GPU. The researchers only had to adjust 0.15% of the original language model's parameters. This isn't some massive, resource-intensive project requiring the computational power of tech giants—it's something that could be replicated by well-funded research labs or even ambitious individuals.
We're approaching a world where understanding human behavior at scale isn't the exclusive domain of psychology departments and tech companies. The tools for behavioral prediction are becoming increasingly democratic, and that has implications we're only beginning to understand.
The Mirror of Our Own Minds
What's most fascinating—and unsettling—about Centaur is what it reveals about us. By learning to predict our choices across diverse scenarios, the model's internal structure began to resemble our neural patterns. It's as if the act of deeply understanding human behavior inevitably leads to brain-like processing.
This suggests something profound about the nature of intelligence itself. Maybe there's a fundamental architecture underlying all cognitive systems, and any sufficiently advanced model of human behavior will converge on brain-like solutions. We're not just building better AI—we're discovering universal principles of how minds work.
The Questions We Should Be Asking
As we stand on the threshold of AI systems that can predict human behavior with unprecedented accuracy, we need to ask harder questions. Who controls these models? How do we ensure they're used ethically? What happens to human agency in a world where our choices are increasingly predictable?
The researchers behind Centaur are admirably transparent about the need for more diverse, globally representative data. They acknowledge that current psychological research is biased toward WEIRD populations—Western, educated, industrialized, rich, and democratic. But even with more inclusive data, the fundamental questions remain.
We're building machines that understand us better than we understand ourselves. That's either the beginning of a new era of human flourishing or the end of human unpredictability as we know it. Probably both.
The future is arriving faster than we can process its implications. Centaur isn't just a research breakthrough—it's a preview of coming attractions. And we're all the starring act in this particular show, whether we signed up for it or not.
The question isn't whether AI will learn to predict human behavior. It already has. The question is what we do with that knowledge, and whether we're prepared for a world where the mystery of the human mind is no longer quite so mysterious.
The mind machine is here. The only question is whether we're ready for what it reveals.
References:
A foundation model to predict and capture human cognition
Find us:
YouTube
Substack
Podcast Providers
Spotify
Apple Podcasts
Patreon
FaceBook Group
STUDY MATERIALS
1. Briefing Document
1. The Pursuit of a Unified Theory of Cognition A long-standing goal in psychology is to develop a "unified theory of cognition" and a computational model that can accurately predict human behavior across various scenarios. Current computational models are "domain-specific," excelling at single tasks (e.g., AlphaGo for Go) but lacking the versatility of the human mind. The Centaur project aims to bridge this gap by creating a "foundation model of human cognition."
2. Centaur's Architecture and Training Centaur is built upon Llama 3.1 70B, a "state-of-the-art language model pretrained by Meta AI." It was developed in a "data-driven manner by fine-tuning" Llama on a massive dataset called Psych-101.
Psych-101 Dataset: This unprecedented dataset contains "trial-by-trial data from more than 60,000 participants performing in excess of 10,000,000 choices in 160 experiments." These experiments were transcribed into natural language, covering diverse domains such as "multi-armed bandits, decision-making, memory, supervised learning, Markov decision processes and more." The dataset contains "253,597,411 text tokens" and has been checked for contamination, showing "no evidence of contamination."
Fine-tuning Process: Centaur was fine-tuned using a "parameter-efficient fine-tuning technique known as quantized low-rank adaptation (QLoRA)." This involved adding "low-rank adapters" to the base Llama model, accounting for only "0.15% of the base model’s parameters," and training for "one epoch" on Psych-101. The training took approximately "five days on an A100 80GB GPU."
3. Centaur's Superiority in Predicting Human Behavior Centaur was rigorously tested and demonstrated remarkable ability to capture human behavior:
Held-out Participants: Centaur predicted the behavior of unseen participants "better than existing cognitive models in almost every single experiment." It consistently outperformed both the base Llama model and "domain-specific cognitive models" (e.g., generalized context model, prospect theory, reinforcement learning models).
Generating Human-like Behavior (Open-Loop Simulations): Centaur could "generate human-like behaviour when simulated in an open-loop fashion." For example, in the "horizon-task paradigm," Centaur's performance was "comparable to human participants." It also replicated human learning patterns in the "two-step task," showing "purely model-free, purely model-based and mixtures thereof."
Failing to Predict Non-Human Behavior: Significantly, Centaur "accurately predicted human responses (64% accuracy) but struggled to predict artificial responses (35% accuracy)," mirroring original human study results and confirming its human-like characteristics.
4. Robust Generalization Abilities A key achievement of Centaur is its ability to generalize to novel situations:
Modified Cover Stories: Centaur successfully captured human behavior in the "two-step task" even with an entirely new "magic-carpet cover story," which was not part of its training data.
Structural Task Modifications: It proved robust to changes in task structure, such as the "Maggie’s farm" paradigm (a three-armed bandit task), outperforming domain-specific models that "did not generalize well to this setting."
Entirely New Domains: Centaur generalized to "entirely new domains," including "logical reasoning," despite Psych-101 not containing such studies. It also robustly captured behavior in other out-of-distribution paradigms, including "moral decision-making, economic games, naturalistic category and reward learning, behavioural propensities and a deep sequential decision task."
Predicting Response Times: Beyond choices, Centaur also effectively predicted "human response times," capturing a larger proportion of variance than Llama or other cognitive models.
5. Alignment with Human Neural Activity Despite being trained solely on behavioral data, Centaur's internal representations showed increased alignment with human neural activity:
fMRI Prediction: Centaur's representations consistently "outperformed Llama’s representations in predicting human neural activity" in fMRI measurements of people performing the "two-step task." This suggests that "fine-tuning a model on large-scale behavioural data aligned its internal representations to human neural activity."
General Neural Alignment: This alignment extended to "unrelated settings," such as a "sentence-reading task," indicating that the cognitive fine-tuning did not degrade its general neural alignment.
6. Model-Guided Scientific Discovery Centaur and Psych-101 serve as valuable tools for scientific discovery:
Blueprint for Research: The process demonstrated in a case study for a multi-attribute decision-making experiment can be a "blueprint for model-guided scientific discovery in other experimental paradigms."
In Silico Prototyping: Centaur can be used for "in silico prototyping of experimental studies," helping to design studies, estimate effect sizes, and reduce the number of required participants.
Hypothesis Generation: Future research can "probe Centaur’s internal representations to understand how it represents knowledge and processes information," which could generate "hypotheses about knowledge representation and information processing in humans."
7. Future Directions and Limitations The authors acknowledge ongoing development and future plans:
Psych-101 Expansion: Psych-101 is considered an "ongoing process" and will be expanded to include more domains (e.g., psycholinguistics, social psychology, economic games) and data on "individual differences" (age, personality, socioeconomic status), and address the "WEIRD" (Western, Educated, Industrialized, Rich, Democratic) population bias.
Multimodal Data: The long-term objective is to move towards a "multimodal data format" beyond natural language to overcome selection bias.
Cognitive Decathlon: The success of Centaur across numerous experiments is likened to winning "16 cognitive decathlons," indicating that "data-driven discovery of domain-general models of cognition is a promising research direction." The ultimate goal is to translate this "domain-general computational model into a unified theory of human cognition."
Conclusion: Centaur represents a significant step towards a unified theory of cognition. By fine-tuning a large language model on a vast dataset of human behavioral data, the researchers have created a model that not only predicts human behavior with unprecedented accuracy across diverse settings but also demonstrates internal representations aligned with human neural activity. This approach offers a powerful new paradigm for understanding the human mind and guiding future cognitive science research.
2. Quiz & Answer Key
Quiz
Answer each question in 2-3 sentences.
What is the primary objective of creating Centaur, and how does it aim to address a recognized gap in existing computational models of cognition?
Describe the Psych-101 dataset. What kind of data does it contain, and what makes its scale unprecedented?
How was Centaur developed from a "state-of-the-art language model"? What specific technique was used for fine-tuning, and why?
Explain one way Centaur demonstrated its ability to generalize to "held-out experiments." Provide a specific example from the text.
What does it mean for Centaur to perform "open-loop simulations," and why is this considered a stronger test of the model's capabilities?
Beyond predicting human choices, what other type of human behavioral data did Centaur successfully predict, and what law is referenced in relation to this?
How did Centaur's internal representations change after fine-tuning in relation to human neural activity? What specific brain imaging technique was used to assess this?
Briefly describe one of the out-of-distribution evaluations mentioned in the text that Centaur generalized to, despite not being trained on that specific scenario.
What is the concept of "model-guided scientific discovery" as illustrated in the paper? How can Psych-101 and Centaur contribute to this?
What is the "cognitive decathlon" concept, and how does Centaur's performance relate to it?
Answer Key
The primary objective of creating Centaur is to establish a computational model that can predict and simulate human behavior across a wide range of settings, moving towards a unified theory of cognition. It addresses the limitation of most contemporary models being domain-specific, unable to generalize beyond particular problems.
Psych-101 is a large-scale dataset of human behavior, comprising trial-by-trial data from over 60,000 participants and 10,000,000 choices across 160 psychological experiments. Its unprecedented scale and natural language transcription provide a common format for diverse experimental paradigms.
Centaur was developed by fine-tuning a state-of-the-art large language model (Llama 3.1 70B) on the Psych-101 dataset. The specific technique used for fine-tuning was Quantized Low-Rank Adaptation (QLoRA), which efficiently adds trainable parameters without altering the base model.
Centaur generalized to "held-out experiments" by accurately predicting human behavior in tasks with modified cover stories, such as the two-step task with a "magic-carpet" narrative. Despite only being trained on the canonical "spaceship" cover story, Centaur effectively adapted to the new description.
Performing "open-loop simulations" means feeding Centaur's own responses back into the model to see if it can generate human-like behavior without external intervention. This is a stronger test because it assesses the model's ability to produce realistic behavioral distributions rather than just predicting single choices.
Beyond predicting human choices, Centaur successfully predicted human response times. The article references Hick's law, which relates individual response times to response entropies, and found Centaur's entropies captured more variance than other models.
Centaur's internal representations became more aligned with human neural activity after fine-tuning, even though it was not explicitly trained for this. This was assessed using functional magnetic resonance imaging (fMRI) measurements, showing Centaur outperformed the base model in predicting brain activity.
One out-of-distribution evaluation involved Centaur's performance on Maggie's farm, a three-armed bandit experiment. Although Psych-101 contained two-armed bandit experiments, it did not include Maggie's farm, yet Centaur robustly captured human behavior in this structurally modified task.
Model-guided scientific discovery refers to using models like Centaur and datasets like Psych-101 to iteratively improve the understanding and development of cognitive theories. This involves using the model's predictions and errors to refine interpretable, domain-specific cognitive models.
The "cognitive decathlon" was proposed as a rigorous evaluation framework for unified cognitive models, testing them across multiple experiments. Centaur's consistent outperformance of established models across what is equivalent to 16 such decathlons suggests the promise of data-driven, domain-general models.
3. Essay Questions
Discuss the significance of Centaur's ability to generalize across different levels (held-out participants, modified cover stories, structural task modifications, entirely new domains). How does this address the limitations of previous computational models in cognitive science and what does it imply for the pursuit of a unified theory of cognition?
Analyze the role of the Psych-101 dataset in the development and evaluation of Centaur. How did its scale and natural language format contribute to Centaur's capabilities, particularly its generalization and alignment with human neural activity?
Compare and contrast Centaur's performance in predicting human choices versus predicting human neural activity. What do the results from each type of evaluation (behavioral goodness-of-fit, open-loop simulations, fMRI alignment) collectively suggest about Centaur's ability to capture human cognition?
The paper suggests that Centaur can be a valuable tool for "model-guided scientific discovery" and "automated cognitive science." Elaborate on these potential applications, providing specific examples of how Centaur could be used to advance cognitive research in the future.
Reflect on the broader implications of Centaur's success for the field of psychology and artificial intelligence. What are the ethical considerations or potential challenges that might arise from the development of highly predictive models of human behavior, and how might future research address these?
4. Glossary of Key Terms
Unified Theory of Cognition: A comprehensive framework or model that explains a wide range of human cognitive abilities and behaviors across different domains, rather than focusing on specific, isolated aspects.
Domain-Specific Models: Computational or theoretical models designed to explain and predict behavior within a very narrow, specific area of cognition (e.g., decision-making in financial contexts, visual perception).
Foundation Model: A large-scale, pre-trained model that can be adapted (fine-tuned) to a wide range of downstream tasks and domains, serving as a versatile base for various applications.
Centaur: The specific computational model introduced in the paper, designed as a foundation model of human cognition by fine-tuning a large language model on human behavioral data.
Psych-101: A large-scale, curated dataset comprising trial-by-trial human behavioral data from 160 psychological experiments, transcribed into natural language, used to fine-tune Centaur.
Large Language Model (LLM): A type of artificial intelligence model trained on vast amounts of text data, capable of understanding, generating, and processing human language. Centaur uses Llama 3.1 70B as its backbone.
Fine-tuning: The process of taking a pre-trained model (like an LLM) and further training it on a smaller, more specific dataset to adapt its capabilities to a new task or domain.
QLoRA (Quantized Low-Rank Adaptation): A parameter-efficient fine-tuning technique that adds small, trainable "low-rank adapters" to a frozen, quantized base model, reducing computational cost.
Held-out Participants/Experiments: Data or participants that were not included in the training set and are used to evaluate the model's ability to generalize to new, unseen data within similar conditions or entirely new experimental contexts.
Negative Log-Likelihood: A common evaluation metric in machine learning and statistics, used to quantify how well a model's predicted probability distribution matches the actual observed data; lower values indicate a better fit.
Open-Loop Simulation (Model Falsification): A rigorous testing method where a computational model generates its own responses, and these responses are fed back into the model, allowing researchers to evaluate if the model can produce realistic, human-like behavioral trajectories without external human input.
Horizon Task: A two-armed bandit paradigm used in behavioral experiments to study different types of exploration strategies in decision-making, such as uncertainty-guided directed exploration.
Two-Step Task: A well-known experimental paradigm used in cognitive science to differentiate between model-free and model-based reinforcement learning strategies in humans.
Cover Story: The narrative or scenario presented to participants in a psychological experiment, which can be modified to test a model's robustness to superficial changes.
Structural Task Modifications: Changes to the underlying design or rules of an experiment, such as adding more options to a choice task (e.g., moving from a two-armed to a three-armed bandit).
Neural Alignment: The degree to which the internal representations or patterns of activity within a computational model correspond to patterns of activity observed in the human brain (e.g., via fMRI).
fMRI (functional Magnetic Resonance Imaging): A neuroimaging technique that measures brain activity by detecting changes associated with blood flow, often used to study neural correlates of cognitive processes.
Response Times (RTs): The duration between a stimulus and a participant's response, often analyzed in cognitive experiments to infer cognitive processes.
Hick's Law: A psychological law stating that the time it takes to make a decision increases logarithmically with the number of choices available.
CogBench: A benchmark designed to test the cognitive abilities of large language models, including various behavioral metrics derived from cognitive psychology experiments.
Model-Guided Scientific Discovery: A research approach where computational models are used as tools to generate hypotheses, analyze data, and refine theories about human cognition, often in an iterative process.
In Silico Prototyping: The use of computational models to simulate and test experimental designs or interventions in a virtual environment before conducting real-world studies, potentially reducing costs and optimizing designs.
Cognitive Decathlon: A concept proposed by early pioneers of unified theories of cognition as a rigorous evaluation framework, where competing models are tested across a diverse set of ten cognitive experiments and judged on cumulative performance.
5. Timeline of Main Events
Early 1980s (Specifically 1983):
"The Architecture of Cognition" by J. Anderson is published. This work likely contributed to the foundational understanding of cognition, preceding the push for unified theories.
Early 1990s (Specifically 1990):
The concept of "unified theories of cognition" is explicitly stated as crucial for intellectually controlling the "wonderful, increasing fund of knowledge" in the field. This highlights a long-standing goal within psychology, emphasizing the need for integrated computational models.
2013:
"Prospect theory" is revisited in a handbook of financial decision making. While an influential account of human choice, it's noted as domain-specific, motivating the need for broader models.
2014:
Research on human exploration strategies in the "horizon-task paradigm" is published. This task later becomes a benchmark for evaluating Centaur's ability to capture human-like exploration.
2015:
Research on "human-level concept learning through probabilistic program induction" is published. This work by Lake, Salakhutdinov, and Tenenbaum highlights advancements in specific cognitive learning capabilities.
2016:
Research exploring when "model-based control" is beneficial is published. This informs the understanding of model-free vs. model-based learning, which Centaur later demonstrates.
2017:
AlphaGo, a computer system by Google DeepMind, masters the game of Go. This is cited as an example of a highly successful but domain-specific computational model, contrasting with the goal of a general model like Centaur.
Research on "Building machines that learn and think like people" by Lake, Ullman, Tenenbaum, & Gershman is published. This further underscores the aspirations of the field to create more human-like AI.
Research highlighting the "importance of falsification in computational cognitive modeling" is published. This principle guides the rigorous testing of Centaur.
2018:
The "Moral Machine experiment" is conducted. This study on moral decision-making is later used as one of the "entirely new domains" to test Centaur's generalization abilities.
2019:
"fMRIPrep," a robust preprocessing pipeline for functional MRI, is introduced. This tool is used in Centaur's neural alignment analysis.
2020:
Research on the "two-step task" is published, suggesting humans primarily use model-based inference. This paradigm becomes crucial for testing Centaur's ability to capture model-based and model-free learning.
Research on "Play, curiosity, and cognition" by Chu & Schulz is published. This contributes to the understanding of human versatility in learning and action.
2021:
QLoRA (Quantized Low-Rank Adaptation) for efficient finetuning of quantized LLMs is introduced. This technique is central to Centaur's training process.
Research on "Integrating explanation and prediction in computational social science" is published. This informs the broader scientific context of Centaur's development.
Research on "A rational model of the Dunning–Kruger effect" is published. This study relates to cognitive biases, an area that general models of cognition might need to address.
Research on "Novelty is not surprise: human exploratory and adaptive behavior in sequential decision-making" is published. This study is among those used to test Centaur's generalization to new domains.
2022:
Research on "Latent motives guide structure learning during adaptive social choice" is published. This work on social prediction games is used to test Centaur's ability to predict human vs. artificial agent behavior.
Research on "Value-free random exploration is linked to impulsivity" is published. This contributes to the understanding of exploration strategies, a behavior Centaur aims to capture.
"The globalizability of temporal discounting" research is published. This meta-study is among those that Psych-101 intends to eventually include to reduce bias.
2023:
Research asking "Can AI language models replace human participants?" is published. This highlights a contemporary debate relevant to the implications of models like Centaur.
"metabench" is introduced as a sparse benchmark for reasoning and knowledge in large language models. This benchmark is used to verify that Centaur's performance on pretraining tasks remains stable.
Research on "Human-like intuitive behavior and reasoning biases emerged in large language models but… disappeared in ChatGPT" is published. This indicates the challenge of retaining human-like characteristics in LLMs, which Centaur aims to overcome.
Late 2024 (Specifically October 26, 2024):
The research paper detailing Centaur is submitted.
Early 2025 (Specifically May 29, 2025):
The research paper detailing Centaur is accepted for publication.
Mid 2025 (Specifically July 2, 2025):
The research paper "A foundation model to predict and capture human cognition" introducing Centaur is published in Nature. This marks the official release and public disclosure of the Centaur model and the Psych-101 dataset.
Ongoing/Future:
Development of Psych-101 continues to include more diverse domains (psycholinguistics, social psychology, economic games), individual differences data (age, personality, socioeconomic status), and a shift towards multimodal data formats.
Future research aims to probe Centaur's internal representations using tools like sparse auto-encoders and attention map visualization to generate hypotheses about human knowledge representation.
Investigation into different model architectures (attention-based vs. vector-based memory, incorporating neuroscience theories) using the Psych-101 dataset is planned.
Centaur is expected to find more applications in automated cognitive science, such as in silico prototyping of experimental studies.
Cast of Characters
Marcel Binz: Project lead and a primary author of the paper on Centaur. Contributed significantly to data curation, quality control, model training, evaluation, neural analyses, first draft, conception, design, and review/editing. Corresponding author.
Eric Schulz: A primary author of the paper on Centaur. Contributed significantly to the first draft, conception, design, and review/editing.
Peter Dayan: Co-author. Contributed to the conception, design, and review/editing of the paper. Received funding from the Max Planck Society and Humboldt Foundation.
Thomas L. Griffiths: Co-author. Contributed to the conception, design, and review/editing of the paper. Received funding from the NOMIS Foundation.
Marcelo Mattar: Co-author. Contributed to neural analyses, conception, design, and review/editing of the paper.
Fabian J. Theis: Co-author. Contributed to conception, design, and review/editing of the paper. Also has consulting and ownership interests in various biotech and AI companies.
Robert Wilson: Co-author. Contributed to conception, design, and review/editing of the paper.
Miryam Naddaf: Author of the Nature news article "This AI 'thinks' like a human – after training on 160 psychology studies" that reported on the Centaur research.
J. Anderson: Author of "The Architecture of Cognition" (1983), a foundational work in psychology that influenced the development of cognitive theories.
Allen Newell: A pioneer in the field of cognitive science, who, in 1990, emphasized the importance of "unified theories of cognition," a goal Centaur directly addresses.
Daniel Kahneman & Amos Tversky: Researchers known for "prospect theory," cited as an influential but domain-specific account of human choice, highlighting the limitations Centaur aims to overcome.
David Silver et al. (Google DeepMind): Developers of AlphaGo, an example of a highly specialized AI system contrasted with Centaur's generalist ambition.
Sreejan Kumar: Co-author. Supported by a Google PhD Fellowship. Contributed to data curation, neural analyses, and review/editing.
DeepSeek-R1: A language-based reasoning model (specifically, the Distill-Llama-70B variant) used in the paper's case study for model-guided scientific discovery. It was prompted to generate explanations of human decision-making, which were then formalized into computational models.
Llama 3.1 70B (Meta AI): The state-of-the-art open-source language model that serves as the backbone ("base model") for Centaur, pretrained by Meta AI.
Minitaur: A smaller version of Centaur that uses Llama 3.1 8B as its base model. It is noted for being useful for prototyping due to lower hardware requirements.
6. FAQ
1. What is Centaur and what is its primary goal?
Centaur is a computational model designed to predict and simulate human behavior across a wide array of settings. Its primary goal is to contribute to a unified theory of cognition by providing a general-purpose model that can capture human behavior in various domains, moving beyond the limitations of domain-specific cognitive models. It aims to offer a deeper understanding of the human mind by predicting responses, generating human-like behavior, and aligning with human neural activity.
2. How was Centaur developed and what data did it use?
Centaur was developed by fine-tuning a state-of-the-art large language model, Llama 3.1 70B, on a novel and extensive dataset called Psych-101. Psych-101 is an unprecedentedly scaled dataset comprising trial-by-trial data from over 60,000 participants and exceeding 10,000,000 choices across 160 psychological experiments. These experiments were meticulously transcribed into natural language, providing a common format that covers diverse domains like decision-making, memory, and learning. The fine-tuning process involved using a parameter-efficient technique called QLoRA.
3. How does Centaur compare to existing cognitive models and untuned language models?
Centaur significantly outperforms both untuned large language models (like the base Llama model) and existing domain-specific cognitive models in predicting human behavior. It shows improved goodness-of-fit for held-out participants in almost every experiment. This indicates that the fine-tuning process on Psych-101 effectively enhanced the model's ability to capture human-like responses beyond what a general language model or specialized cognitive model could achieve individually.
4. What evidence supports Centaur's generalization abilities to new and modified experimental settings?
Centaur demonstrates robust generalization across several increasingly complex out-of-distribution evaluations. It successfully predicts human behavior even when presented with:
Modified cover stories: For instance, in the two-step task with a "magic-carpet" narrative, despite being trained only on the "spaceship" cover story.
Structural task modifications: Such as generalizing from two-armed bandit tasks (present in training) to a three-armed bandit experiment (Maggie’s farm, not in training).
Entirely new domains: Including logical reasoning tasks, which were not part of the Psych-101 training data, and other diverse paradigms like moral decision-making and economic games.
5. Does Centaur's internal functioning align with human brain activity?
Yes, despite being trained solely on human behavioral data, Centaur's internal representations show increased alignment with human neural activity. This was demonstrated through analyses predicting fMRI measurements in tasks like the two-step task and a sentence-reading task. The model's internal representations consistently outperformed those of the untuned base model in predicting human neural responses, suggesting that fine-tuning on large-scale behavioral data leads to an internal organization that mirrors human brain function.
6. Beyond predicting choices, what other human metrics can Centaur capture?
In addition to accurately predicting human choices, Centaur also shows an ability to predict human response times. It was found that response entropies derived from Centaur captured a significantly larger proportion of variance in human response times compared to those derived from the untuned Llama model or traditional cognitive models. This highlights Centaur's capability to model aspects of human cognition beyond just the final decision.
7. How can Centaur be used for scientific discovery in cognitive science?
Centaur, alongside the Psych-101 dataset, serves as a valuable tool for scientific discovery. It can guide the development of new, interpretable cognitive models. For example, the study demonstrated how Centaur could be used in a "scientific regret minimization" pipeline: a large language model (DeepSeek-R1) generated an initial explanation of human behavior, and Centaur then highlighted discrepancies, allowing for refinement into a more accurate and interpretable domain-specific cognitive model. This process offers a blueprint for automated cognitive science, including in silico prototyping of experiments to optimize study design (e.g., maximizing effect sizes or reducing participant numbers).
8. What are the future directions and limitations of Centaur and Psych-101?
Future work aims to expand Psych-101 to include more cognitive domains beyond learning and decision-making, such as psycholinguistics and social psychology. There's also a focus on incorporating individual differences (age, personality, socioeconomic status) to enable models to capture variations in human behavior. While the natural-language format is versatile, a long-term objective is to move towards a multimodal data format to overcome biases against experiments not expressible in text. Furthermore, researchers plan to probe Centaur's internal representations to generate hypotheses about human knowledge representation and information processing, and to explore training models with different neural architectures from scratch using the Psych-101 dataset to investigate human cognitive architecture.
7. Table of Contents
0:00 - Introduction and Welcome
Introduction to Heliox podcast and today's exploration of the human mind and computational modeling
1:15 - The Challenge of Unified Cognition
Discussion of human adaptability versus domain-specific computational models like AlphaGo and prospect theory
2:45 - The Need for General Purpose Models
Historical context of unified theories in cognitive science and the limitations of specialized approaches
3:30 - Introducing Centaur: A Foundation Model
Overview of Centaur as a foundation model built on Meta AI's LLaMA 3.1 70B language model
4:15 - The Psych 101 Dataset
Detailed examination of the unprecedented dataset containing 10 million decisions from 160 psychological experiments
5:30 - Training Methodology and Efficiency
Explanation of the fine-tuning process using QLora and the surprisingly efficient 5-day training period
6:45 - Prediction Capabilities
Analysis of Centaur's ability to predict human choices better than existing models across all 160 experiments
8:00 - Behavioral Generation and Model Falsification
Testing Centaur's ability to generate authentic human-like behavior from scratch, including the horizon task
9:30 - Capturing Human Diversity
Discussion of how Centaur generates the full spectrum of human learning strategies, not just averages
10:45 - Social Prediction and Human-Like Mistakes
Analysis of Centaur's performance in social games and its human-like blind spots with artificial agents
12:00 - Generalization to New Situations
Testing Centaur's robustness with out-of-distribution scenarios and structural changes
13:15 - Cover Story Changes and Task Structure
Examples of spaceships to magic carpets and two-armed to three-armed bandit problems
14:30 - Emergence in New Domains
Surprising improvement in logical reasoning tasks despite no training on logic problems
15:45 - Response Time Prediction
Centaur's ability to predict human decision timing using internal uncertainty metrics
17:00 - Neural Alignment Without Brain Training
Remarkable correspondence between Centaur's internal patterns and fMRI brain activity
18:30 - Scientific Discovery Applications
Case study of multi-attribute decision-making and scientific regret minimization
20:00 - In Silico Prototyping
Potential for using Centaur to simulate experiments before running them on humans
21:15 - Future Directions and Expansion
Plans for expanding Psych 101 with more cognitive domains and individual differences
22:30 - Addressing Demographic Bias
Discussion of WEIRD populations and goals for global representation
23:45 - Historical Skepticism and Cognitive Decathlon
Addressing concerns about unified models and the cognitive decathlon concept
25:00 - Centaur's Comprehensive Success
Analysis of performance across 160 experiments as multiple cognitive decathlons
26:15 - Implications for Understanding Ourselves
Philosophical reflection on what Centaur reveals about human cognition and consciousness
27:30 - Personal Reflection and Conclusion
Invitation for listeners to consider their own decision-making processes
28:00 - Closing and Recurring Themes
Summary of boundary dissolution, adaptive complexity, embodied knowledge, and quantum-like uncertainty
8. Index
Index: The Centaur Model Deep Dive
A100 GPU - 5:30
Adaptive complexity - 28:00
AlphaGo - 2:45
Artificial agent prediction - 10:45
Bandit problems - 13:15
Behavioral generation - 8:00
Blind spots - 10:45
Boundary dissolution - 28:00
Brain activity prediction - 17:00
Centaur model - 3:30, 6:45, 8:00, 12:00, 15:45, 17:00, 18:30, 20:00, 25:00
Cognitive decathlon - 23:45, 25:00
Cognitive effort - 15:45
Cognitive science - 2:45, 23:45
Common format - 4:15
Computational model - 1:15, 2:45
Consciousness - 26:15
Cover story changes - 13:15
Decision-making - 4:15, 18:30, 26:15
Demographic bias - 22:30
Directed exploration - 8:00
Domain-specific models - 2:45
Embodied knowledge - 28:00
Entropy - 15:45
Experimental design - 20:00
Fine-tuning - 5:30
fMRI data - 17:00
Foundation model - 3:30
Gambling - 2:45, 4:15
General purpose - 2:45, 3:30
Generalization - 12:00, 14:30
Horizon task - 8:00
Human behavior - 4:15, 6:45, 12:00, 18:30
Human diversity - 9:30
Individual differences - 21:15
In silico prototyping - 20:00
Language learning - 2:45
Language model - 3:30
Learning strategies - 9:30
LLaMA 3.1 70B - 3:30
Logical reasoning - 14:30
LSAT - 14:30
Maggie's Farm - 13:15
Memory tasks - 4:15
Meta AI - 3:30
Model falsification - 8:00
Model-based learning - 9:30
Model-free learning - 9:30
Moral decision-making - 12:00
Multi-attribute decision-making - 18:30
Multimodal data - 21:15
Natural language - 4:15
Neural alignment - 17:00
Neural patterns - 17:00
Out-of-distribution - 12:00
Parameter efficiency - 5:30
Personalized learning - 9:30
Prediction accuracy - 6:45, 10:45, 15:45
Prospect theory - 2:45
Psych 101 dataset - 4:15, 18:30, 21:15
Psychological experiments - 4:15
Psycholinguistics - 21:15
QLora - 5:30
Quantum-like uncertainty - 28:00
Response entropy - 15:45
Response times - 15:45
Scientific discovery - 18:30
Scientific regret minimization - 18:30
Social prediction - 10:45
Social psychology - 21:15
Specialization - 2:45, 13:15
Task structure - 13:15
Two-step task - 9:30, 17:00
Uncertainty-guided exploration - 8:00
Unified theory - 2:45, 25:00
Variance explanation - 15:45
WEIRD populations - 22:30
9. Post-Episode Fact Check
✅ VERIFIED FACTS:
Centaur Model: Centaur is indeed a computational model that can predict and simulate human behavior in any experiment expressible in natural language NaturearXiv, published in Nature.
Psych-101 Dataset: The dataset contains trial-by-trial data from 160 psychological experiments with over 60,000 participants making over 10 million choices Hugging FaceKuakua. The exact figures are 60,092 participants making 10,681,650 choices marcelbinz/Psych-101 · Datasets at Hugging Face.
Base Model: The model was fine-tuned on LLaMA 3.1 Bayram Annakov on X: "Psych 101: How to Predict Human Decisions So, Psych-101. It's a test dataset containing 10 million decisions made by 60,000 participants across 160 psychological experiments. They fine-tuned Llama 3.1 on it, and it looks like this could lead to a new theory of consciousness. But https://t.co/LJH4g9YqXH" / X (though the podcast mentioned LLaMA 3.1 70B specifically).
Publication: The paper was published in Nature on July 2, 2025 Groundbreaking New AI Trained on Psychology Studies Can Predict Human Behavior with Stunning Accuracy - The Debrief, which means this is very recent research.
Research Team: The research was conducted by Eric Schulz and colleagues at the Institute for Human-Centered AI at Helmholtz Munich AI That Thinks Like Us: New Model Predicts Human Decisions With Startling Accuracy.
⚠️ MINOR DISCREPANCIES:
The podcast states "over 60,000 participants" but the exact number is 60,092 participants marcelbinz/Psych-101 · Datasets at Hugging Face
The podcast mentions "more than 10 million decisions" which is accurate as the exact number is 10,681,650 choices marcelbinz/Psych-101 · Datasets at Hugging Face
❓ UNVERIFIED TECHNICAL DETAILS:
The podcast contains many specific technical details (like the 87% variance explanation for response times, the 5-day training period, 0.15% parameter adjustment, etc.) that would require access to the full Nature paper to verify. However, the core claims about the model's capabilities and the basic facts about the dataset are accurate.
⚠️ IMPORTANT NOTE:
Some scientists are skeptical of the claims Researchers claim their AI model simulates the human mind. Others are skeptical | Science | AAAS, as reported in Science magazine, indicating this is an active area of scientific debate rather than universally accepted findings.
OVERALL ASSESSMENT: The podcast appears to be factually accurate regarding the basic facts about Centaur and the Psych-101 dataset. The specific technical details would need verification from the full paper, but the core narrative and claims align with the available information about this recent Nature publication.
10. Image (3000 x 3000 pixels)
