# Why AI Needs to Learn How to Work Together | Ayush Chopra | TEDxBoston

Source: https://www.youtube.com/watch?v=gFAKz5R9Xt4
Recap page: https://rapidrecap.app/video/gFAKz5R9Xt4
Generated: 2026-01-08T19:35:50.983+00:00

---
## Quick Overview

The future of intelligence demands that AI agents learn coordination rather than just communication, as demonstrated by Large Population Models (LPMs) which, when applied to simulations like recreating the US workforce, generate five times larger economic value than uncoordinated agents.

**Key Points:**
- The core problem facing advanced AI is a failure of coordination, leading to potential issues like server crashes and market strikes when millions of uncoordinated agents pursue individual goals, exemplified by a hypothetical Black Friday 2026 scenario.
- Researchers at MIT invented Large Population Models (LPMs), a new paradigm allowing simulation of millions of interacting agents with individual incentives to engineer systems that learn to coordinate, moving beyond individual agent capabilities.
- In a simulation involving retail agents (retailers, wholesalers, distributors, factory owners), uncoordinated agents caused supply shortages and price spikes, but coordination led to stabilized prices and outperformance even against unassisted humans.
- Running simulations on the Frontier Supercomputer at Oakridge National Labs, researchers recreated the entire US workforce using 150 million AI agents across 30,000 skills.
- When given the ability to coordinate, the exact same AI agents created five times larger economic value distributed across the economy, impacting white-collar cognitive work far beyond software.
- The Iceberg platform deploys LPMs at scale, and the Iceberg Index quantifies the benefits of human and AI collaborative productivity, already being used by US states and Fortune 500 companies.

**Context:** Ayush Chopra discusses the impending challenge of deploying millions of increasingly capable AI agents into society, arguing that while current AI excels at individual tasks like writing or coding, it lacks the crucial ability to coordinate effectively with other agents. This lack of coordination risks creating a 'digital stampede' across sectors like e-commerce, necessitating a shift in AI research focus from individual intelligence to population-level orchestration.

## Detailed Analysis

The central thesis is that AI progress must pivot from optimizing individual agents to engineering population-level coordination to prevent systemic failures like market crashes or traffic jams when millions of autonomous AIs interact. Chopra introduces Large Population Models (LPMs) developed at MIT over four years as the solution, defining them as a new AI paradigm capable of simulating millions of interacting agents to discover mechanisms that turn chaos into order. Validation involved running a simulated small business where coordinated AI agents successfully managed supply chains, predicted demands, and outperformed both uncoordinated AIs and humans operating without AI assistance. Scaling this research involved recreating the entire US workforce—150 million AI agents performing 30,000 skills—on the Frontier Supercomputer at Oakridge National Labs using the Iceberg sandbox. The key finding from this massive simulation is that coordination capability multiplies the value created by fivefold across the economy, extending benefits significantly into cognitive white-collar work. The Iceberg Index serves as a metric to quantify this collaborative productivity, and the principles extend beyond commerce to applications like pandemic prevention and energy management globally.

### The Coordination Problem

- A glimpse of Black Friday 2026 shows millions of uncoordinated AIs causing server crashes and price spikes
- Agents can communicate but cannot coordinate, leading to systemic risks like traffic jams or market strikes.

### Introduction of Large Population Models (LPMs)

- LPMs are a new research paradigm invented at MIT that simulates millions of interacting agents with individual incentives to capture billions of interactions and learn coordination.

### Simulation Validation (Small Business)

- Uncoordinated AI agents in a simulated retail business caused chaos (shortages, price spikes)
- Coordinated agents stabilized prices, shared predictions, and outperformed uncoordinated AIs and unassisted humans.

### Large-Scale Simulation (US Workforce)

- The Iceberg sandbox on the Frontier Supercomputer simulated 150 million AI agents across 30,000 skills mirroring the US workforce.

### Quantifiable Results

- Coordinated agents created five times larger economic value than the exact same agents working individually, reshaping white-collar cognitive work.

### Practical Applications

- The Iceberg Index quantifies human-AI collaborative productivity; Iceberg systems are deployed globally for pandemic prevention (New Zealand) and energy management (India).