The Economics of AI: Prediction Machines and Their Impact with Ajay Agrawal

Episode Summary

Executive Summary: Ajay Agrawal argues that AI is best understood as a drop in the cost of prediction: as prediction gets cheaper, society uses more of it and the value of complements like data, judgment, and decision-making actions rises. He says this reframes many tasks—from translation and vision to driving and generative text—as prediction problems, while short-run inference costs are real but likely to fall as compute supply expands.

Main Topics: AI as cheaper prediction (Priority: 5/5): Agrawal’s central framework is that AI advances primarily by reducing the cost of quality-adjusted prediction, rather than introducing a mystical new capability. This explains why AI spreads across many domains. Recasting tasks as prediction problems (Priority: 5/5): He explains how translation, vision, driving, fraud detection, and language generation can all be modeled as prediction problems, which made them amenable to machine learning and deep learning. Substitutes and complements in economics (Priority: 5/5): Lower prediction costs reduce the value of substitutes like human prediction while increasing the value of complements such as data, human judgment, and downstream actions taken on predictions. Surprise at generative AI performance (Priority: 4/5): Agrawal says even experts were surprised by how quickly scaling large models improved performance, especially in language, where ChatGPT became near-indistinguishable from humans. Inference costs and compute constraints (Priority: 4/5): He agrees that current model usage has real marginal inference costs, but argues these are temporary bottlenecks tied to scarce compute and GPU supply, not a permanent limit. Labor-market implications (Priority: 4/5): The discussion turns to how AI may reshape work by substituting for some forms of prediction while raising the value of judgment, relationships, and other complements, with no clear consensus among experts. Long-run bottlenecks: electricity and infrastructure (Priority: 3/5): Agrawal suggests that as compute becomes commoditized, electricity may become the key constraint, implying future AI economics will depend on energy markets and infrastructure.

Key Arguments: AI should be viewed as a fall in the cost of prediction, which helps explain both existing and novel applications. When prediction gets cheaper, people and firms use more of it, consistent with downward-sloping demand. Tasks once not seen as prediction problems—like driving or translation—can be reframed and solved with AI. The value of human prediction falls relative to machine prediction, but complements such as data, judgment, and action-taking increase in value. ChatGPT’s success was surprising even to NLP experts because scaling up models produced unexpectedly large gains. Current inference costs are real but reflect a temporary supply-demand imbalance in GPUs/compute; they should decline as capacity expands. In the long run, the binding constraint on AI may shift from compute to electricity. There is no settled expert consensus on AI’s labor-market effects, but the framework suggests both displacement and complementarity effects.

Data Points: Podcast date: Friday, May 24th - Host introduction ChatGPT release date: November 30th, 2022 - Agrawal references the launch as a surprise to NLP researchers Model version: 3.5 - He notes ChatGPT version 3.5 was in many cases indistinguishable from human Scaling magnitude: a couple of orders of magnitude - He says OpenAI scaled models significantly, which drove large performance gains Book publication year: 2018 - He references the original book where these ideas were imagined Timeline reference: five years ago - Used to illustrate that driving was not commonly seen as prediction then

Pivotal Quotes: "AI getting better and better, as a drop in the cost of prediction, meaning the cost of quality-adjusted prediction is falling as AI advances." — Ajay Agrawal: Defines his core economic framework for understanding AI "Most people five years ago would not have characterized driving as a prediction problem. But that's how we're solving it today." — Ajay Agrawal: Example of reframing a non-obvious task as prediction "In the long run, the compute's no longer the bottleneck. Now it's electricity." — Ajay Agrawal: Describes the likely future constraint on AI scaling

Implications: Listeners should expect AI to expand where prediction is valuable, while raising the value of data, judgment, and action. Near-term costs may slow deployment, but longer-term limits may shift toward energy and infrastructure.