Organizing Intelligence
summary
The gist
Agents, whether human or artificial intelligence (AI), sort endogenously into competing firms, and a firm's productivity is contingent upon its agents' capabilities.
In short
The paper models how agents sort into firms based on their capabilities and productivity, distinguishing between competence (productivity shifts) and coordination (scaling with size). It analyzes four AI regimes, showing that AI expertise is reproducible while human expertise is fixed. This framework explains how organizational structure and firm size evolve as AI advances.
Key concepts
- Competence
- This refers to the specific skills or knowledge agents possess, which directly influences a firm's productivity level. Increasing competence means agents can perform tasks better, thereby shifting the overall output capacity of the organization.
- Coordination
- This concept describes how productivity scales as a firm grows in size. It focuses on how well agents work together or how effectively organizational structure manages increasing scale, determining whether larger firms become more or less productive.
- Technology Effect
- This effect measures the change in wages caused by technological advancements that modify the value of human labor within existing firms. It represents how changes in agent capability alter the direct productivity gains for workers already employed.
- Reorganization Effect
- This effect captures wage changes resulting from adjustments to firm size and composition, rather than just technology. This can reverse initial wage effects by showing how shifts in organizational structure influence worker compensation.
Terminology used across episodes
This episode discusses
- Organizing Intelligence · Paper Radio
- Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction
- Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents
- The Ringelmann Effect in Multi-Agent LLM Systems: A Scaling Law for Effective Team Size
- Why Do Multi-Agent LLM Systems Fail?
- Open Problems in Cooperative AI
- A Model of Artificial Jagged Intelligence
- The Turing Valley: How AI Capabilities Shape Labor Income
- Towards a Science of Scaling Agent Systems
- GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks
- Agents of Chaos
- PaperBench: Evaluating AI's Ability to Replicate AI Research
- APEX-Agents
- The AI Productivity Index (APEX)
- RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
- Understanding Agent Scaling in LLM-Based Multi-Agent Systems via Diversity
The paper
Organizing Intelligence · Read on arXiv
Alex Imas, Andrew Koh, Shane Legg
Google DeepMind
When will advances in AI lead to economic transformation? We develop a model of how intelligence, the ability to achieve goals, matters for organizations, the key unit of economic production. Agents (human or AI) with multidimensional capabilities sort endogenously into competing firms. A firm's productivity depends on its agents' capabilities; we distinguish competence, which shifts productivity levels, from coordination, which governs how productivity scales with size. Two differences between AI and humans drive transformation: AI expertise is reproducible while humans' are in fixed supply, and AIs may not face the same human limits on coordination. We analyze how equilibrium market structure, productivity, employment, and wages evolve as AI capabilities improve. When AI is more capable than the average human, organizations of complementary humans can nonetheless outperform AI. Increases in AI competence can change the composition of workers within firms and decrease firm size. As AI coordination improves, however, this can break the Coasean ceiling on firm size and the market structure undergoes a phase transition toward large, concentrated AI-heavy firms. Changes in human employment are gradual when AI displaces workers within mixed human-AI firms, or sudden when AI firms outcompete human firms. We decompose wage changes into a technology effect, which changes the value of humans within existing firms, and a reorganization effect as firms adjust size and composition. Reorganization can reverse the technology effect on wages.
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Organizing Intelligence".
Warren: Here's the gist of the paper. Agents, whether human or artificial intelligence (AI), sort endogenously into competing firms, and a firm's productivity is contingent upon its agents' capabilities.
Ruth: The paper distinguishes between competence—which shifts productivity levels—and coordination—which dictates how productivity scales with organizational size.
Warren: A key transformation driven by the difference between AI and humans is that AI expertise is reproducible, whereas human expertise is in fixed supply, and AIs may not encounter the same human limitations regarding coordination.
Ruth: The model analyzes how market structure, productivity, employment levels, and wages evolve as AI capabilities advance across four regimes: Minimal AGI, Supercompetent AI, Supercooperative AI, and Artificial Superintelligence (ASI).
Paper summary: Iris: So we covered the basics of "Organizing Intelligence," which develops a model about how intelligence, defined as the ability to achieve goals, affects organizations—the key unit of economic production.
Ruth: The main thesis is that agents, whether human or AI, sort themselves into competing firms based on their capabilities and that a firm's productivity depends directly on those agent capabilities.
Warren: They clearly separate competence from coordination; competence shifts the productivity level up or down based on domain-specific skills like robotics.
Iris: And coordination is about how agents work better together as the organization grows, capturing things like reduced communication costs or better understanding of norms.
Ruth: The paper highlights two key ways AI differs from humans: AI expertise can be reproduced at a cost of compute, while human expertise comes with a fixed supply of people.
Warren: That fixed supply for humans creates what they call Ricardian scarcity rents in the form of higher wages for increasing expertise, which AI doesn't face because copying an AI model doesn't have that compute cost.
Iris: They analyze how equilibrium market structure, employment levels, and wages change as these AI capabilities improve across four stages: Minimal AGI, Supercompetent AI, Supercooperative AI, and Artificial Superintelligence.
Ruth: A key finding is that when the average human agent can't keep up with the AI capability level, organizations with complementary human agents can still achieve superior performance compared to pure AI systems.
Warren: They also show that increases in competence for the AI lead to changes in worker composition within those firms and a reduction in firm size.
Iris: Conversely, improvements in AI coordination seem capable of breaking the Coasean ceiling on firm size, which drives a transition toward very large, highly concentrated AI-heavy firms.
Ruth: The model decomposes wage changes into a technology effect that modifies human labor value and a reorganization effect from changes in firm size and composition.
Warren: Crucially, the authors show that this reorganization effect can even reverse the initial technology effect on wages entirely.
Iris: They conclude by showing that structures based on firm support and active equality are efficient, and under constant returns to scale, the mean capability of the technology corresponds to everyone having equal time allocation across all tasks.
Ruth: This whole study is about mapping out how market structure, productivity, employment levels, and wages evolve when AI advances across these four regimes.
Conclusion: Warren: So looking at "Organizing Intelligence," the authors are essentially proposing a framework that links agent capabilities directly to organizational outcomes like size and employment.
Iris: The paper really gets at how the difference between reproducible AI expertise and fixed human supply creates a unique dynamic for economic transformation.
Ruth: It suggests that simply looking at technology as something that makes workers more productive isn't enough; you have to account for how the organization is structured around those capabilities.
Warren: The implication for us is that we can't just predict a single outcome based on AI skill alone; we need to model the interaction between competence and coordination.
Iris: They’re giving us a way to think about why rapid increases in AI capability might not always result in proportional economic transformation if the organization doesn't adapt its composition or size correctly.
Ruth: The title, "Organizing Intelligence," captures that idea—it’s not just about the intelligence itself, but how it gets organized within a firm.
Warren: It means we need to focus our analysis on those two elements, competence and coordination, because they determine the ultimate market structure and what people actually earn.
Iris: It sets up a clear direction for future work by suggesting specific benchmarks for measuring output scaling with AI copies at fixed compute per copy alongside isolated competence.
Ruth: And we need to track how different worker skill profiles sort across firms to isolate that reorganization effect from the initial technology effect on wages.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19