Physics of Intelligence
Open menu
TopPeopleResearchPublicationsOutreachNews

Research

Mechanistic Swarm Interpretability: What Is the Agent Condition? · Part 1

What Shapes Collective Belief Collapse in AI Swarms?

This post opens our series on Mechanistic Swarm Interpretability: understanding how the personas of individual agents and their interactions shape collective personas and behavior.

Blog Author: Hidenori Tanaka · Sep 2026

Read the blog Read the paper

Research program

From steam engines to transistors, the history of industrial revolutions is a history of understanding and harnessing emergence. We build a ladder of conceptual frameworks, from neurons to personas to swarms, to understand and align emerging superintelligence.

From neurons to personas to swarms
  1. A neural circuit with flowing learning trajectories.

    Neural networks

    Neural Mechanics of Learning
  2. One agent with a faint robot face and a visible neural circuit, with an arrow suggesting behavioral steering.

    Individual agents

    Persona Mechanics
  3. A connected group of abstract agents, each containing a small neural circuit.

    Agent collectives

    Mechanistic Swarm Interpretability

More research

All publications

Neural Mechanics of Learning

What laws govern how neural networks learn?

We study how symmetry and symmetry breaking shape learning dynamics, and how concepts and algorithms emerge and compete during training.

Persona Mechanics

What structures underlie model behavior, and how can we steer it?

We seek low-dimensional structures underlying model behavior and study how they change under fine-tuning, in-context prompting, and activation steering.

Mechanistic Swarm Interpretability

How do individual agents and their interactions shape collective beliefs and behavior?

We trace how messages change individual states, how those changes propagate, and which interventions alter collective behavior.

Hidenori has previously worked on the evolutionary dynamics of self-replicating artificial materials and the safety challenges of containing CRISPR gene drives.

Prediction and Forecasting

Ultimately, the success of Physics of Intelligence should be measured by its ability to forecast the future. We aim to make concrete, qualitative predictions about emerging AI safety risks, drawing on a scientific foundation of intelligence to make the unimaginable intelligible.

Jun 2026

Interactive

The Circuit in Your Head

An interactive exhibit on body, alarm, attention, and action. Developed under the supervision of Prof. Mai Uchida and shown at a Boston Museum of Science event.