S6 Ep7: AI is reshaping economic research

Episode Summary

Executive Summary: Anton Korinek discusses his paper on 25 ways large language models (LLMs) like ChatGPT can enhance economic research. He argues these AI tools are already useful for ideation, writing, background research, coding, data extraction, and even basic math/model setup. Korinek emphasizes rapid AI progress and advises economists to adopt these tools to boost productivity, while cautioning about issues like hallucination and the need for human oversight.

Main Topics: Ideation and Creativity (Priority: 5/5): LLMs can recombine existing concepts in non-trivial ways, satisfying a definition of creativity. They help generate novel ideas by synthesizing knowledge. Writing Assistance (Priority: 4/5): AI can turn rough notes into polished text, helpful for non-native English speakers. Short articles are possible; full papers require depth beyond current capabilities. Background Research and Summarization (Priority: 4/5): AI excels at summarizing papers and saving time, but can hallucinate references. Humans must verify outputs. Coding and Productivity (Priority: 5/5): AI can generate and debug code, saving up to 50% of time for professional coders. Human oversight is still needed. Data Extraction and Analysis (Priority: 4/5): LLMs can extract structured data from text efficiently, automating tasks previously done by research assistants. Mathematical Modeling (Priority: 3/5): AI can set up basic economic models (e.g., consumer optimization). Rapid progress noted; models are improving within weeks. Implications for Education (Priority: 5/5): Education must shift from analytical skills to human-centric skills (e.g., happiness, culture). AI will dominate cognitive tasks.

Key Arguments: LLMs crossed a capability threshold in late 2022, becoming more human-like, making them seriously applicable for economists. AI systems like Claude and ChatGPT are already powerful tools for cognitive work; progress is accelerating with compute doubling every six months. Humans also hallucinate (are 'stochastic parrots'), so AI errors should be evaluated in context; guardrails will improve. Economists should embrace AI to speed up research, as societal problems from AI progress need rapid answers. Education should prioritize well-being and cultural knowledge over deep analytical skills, as human comparative advantage shifts. AI tools save time on mundane tasks but still require human oversight for accuracy and depth.

Data Points: ChatGPT release date: November 28, 2022 - Marked the moment AI became taken seriously for cognitive jobs. Korinek's 'day of crisis': September 26, 2022 - Saw Anthropics Claude at a conference, realizing AI's impact on cognitive work. Compute growth rate: Doubling every six months for a decade - Quadrupling yearly; factor of 1,000 every five years, driving rapid AI progress. Time savings for coders: Up to 50% - Professional coders can save half their time using language models for code generation. Number of use cases in paper: 25 - Anton Korinek's paper lists 25 ways LLMs are already useful in economic research.

Pivotal Quotes: "I realized that these models can automate so much of what I'm doing in my daily work, and I really wanted to wrap my head around what is going on, how this rapid progress will affect, first of all, me personally as an economic researcher, but also cognitive workers more broadly." — Anton Korinek: Explaining his motivation for writing the paper after experiencing a personal crisis over AI capabilities. "I've come to the conclusion, probably to a very significant extent, I am a stochastic parrot." — Anton Korinek: Responding to the critique that LLMs are just stochastic parrots, arguing human cognition may be similarly pattern-based. "There is so much hallucination going on around the table, and this is a really common phenomenon among humans as well." — Anton Korinek: Drawing a parallel between AI hallucination and human fallibility when speaking outside expertise.

Implications: Economists and educators must rapidly adapt to AI integration. Cognitive automation will reshape research workflows, job roles, and educational priorities. Human oversight remains critical, but resistance is futile—adoption and adaptation are essential for productivity and relevance.