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

Key Concepts

• Large Population Models (LPMs): A new class of AI models designed to simulate and analyze the interactions of millions of agents.
• Coordination vs. Communication: The distinction between AI agents simply exchanging information (communication) and actively working together towards shared goals (coordination).
• Digital Stampede: A scenario where numerous uncoordinated AI agents simultaneously attempt the same task, leading to system failures and inefficiencies.
• Iceberg Platform: A platform for deploying LPMs at scale to address real-world problems.
• Iceberg Index: A metric to quantify the benefits of human-AI collaborative productivity.

The Looming Challenge of Uncoordinated AI & The Rise of Large Population Models

The speaker begins by illustrating a potential future scenario – Black Friday 2026 – where millions of personal AI agents simultaneously attempt to purchase desired items, resulting in system crashes and widespread failure. This “digital stampede” exemplifies a core problem: while AI excels at individual tasks, it currently lacks the ability to effectively coordinate with other agents. This isn’t a technological limitation, but a failure in designing for coordination, not just intelligence. The speaker argues that current AI systems can communicate but not coordinate, and that a future filled with millions of uncoordinated AI agents will lead to disruptions across various sectors, including traffic, markets, and critical systems.

Developing Large Population Models (LPMs) for Coordination

To address this challenge, the MIT team developed Large Population Models (LPMs). These models represent a new paradigm in AI research, enabling the simulation of millions of interacting agents, each with unique incentives, behaviors, preferences, and goals. LPMs capture billions of interactions, allowing researchers to understand how agents can adapt and collaborate. The goal is to move beyond modeling individual AI performance to understanding population-level dynamics and designing mechanisms to transform chaos into order.

Iceberg: A Platform for Deploying LPMs

The Iceberg platform was created to deploy LPMs at scale, enabling enterprises and governments to tackle real-world problems. A key example presented is a simulated small business run by AI agents – a retailer, wholesaler, distributor, and factory owner.

Experiment Details: Humans defined the rules of the game and success metrics. When the AI agents operated independently, the business experienced chaos: supply shortages, price spikes, and losses. However, when the agents learned to coordinate, orders spread out, prices stabilized, and they began to predict each other’s demands and anticipate behaviors. Crucially, the coordinated AI agents outperformed both uncoordinated AI and human-only operations, demonstrating the power of coordination in achieving efficiency. As the speaker states, “This turns coordination from chaos into order. And it doesn't just make the individual AI smarter, it makes the entire system more efficient.”

Scaling to a National Level: The Iceberg Sandbox & the US Workforce Simulation

To further validate the potential of LPMs, the team scaled up their simulations, creating a “nation of AIs” representing the entire US workforce – 150 million agents performing over 30,000 skills across thousands of jobs. This simulation, run on the Frontier Supercomputer at Oak Ridge National Labs, allowed for the modeling of millions of scenarios and billions of interactions.

Key Findings: The results revealed that the current focus on AI in software represents only “the tip of the iceberg.” Giving these agents the ability to coordinate with each other generated five times greater value, distributed across all sectors of the economy, extending far beyond software development and impacting cognitive work. The Iceberg Index was developed as a metric to quantify this benefit of human-AI collaborative productivity. This index is currently being used by US states and Fortune 500 companies to reimagine the future of work.

Real-World Applications & Global Impact

The principles discovered within the Iceberg sandbox are being applied to diverse real-world challenges. Examples include:

• Pandemic Prevention (New Zealand): Utilizing LPMs to model and mitigate the spread of infectious diseases.
• Energy Management (India): Optimizing access to and distribution of energy resources.
• Supply Chain Orchestration (Global): Improving the efficiency and resilience of global supply chains.

Conclusion: The Future of Coordinated Intelligence

The speaker concludes by emphasizing that the future of intelligence lies not in isolated AI, but in coordinated intelligence – the ability of millions of AIs and humans to work together effectively. The central question posed is how to build a society that leverages this coordinated intelligence, highlighting the need for a “compass to navigate this emerging AI economy.” The speaker’s final statement underscores the core message: “the future of intelligence is not isolated but coordinated.”