What Happens When AI Learns to Do Our Jobs
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode, Derek Thompson interviews Ethan Mollick, a Wharton professor and AI expert, about the revolutionary potential of artificial intelligence in transforming work. (07:56) The conversation explores how AI represents a modern parallel to the railroad revolution that created managerial capitalism in the 19th century. (05:08)
- Main themes: AI's "jagged frontier" capabilities, the emergence of autonomous agents, productivity paradoxes, and how general purpose technologies fundamentally reorganize work rather than simply automating existing tasks
Speakers
Ethan Mollick
Ethan Mollick is a professor of management at Wharton, where he specializes in entrepreneurship and innovation. He's the author of "Co-Intelligence: Living and Working with AI" and writes the influential Substack "One Useful Thing," which provides practical guidance on using AI tools productively while exploring the broader implications of artificial intelligence on society and work.
Derek Thompson
Derek Thompson is the host of Plain English and a staff writer at The Atlantic. He covers economics, technology, and culture, with a particular focus on how technological changes reshape work and society.
Key Takeaways
AI's Intelligence is "Jagged" - Master the Unpredictable
AI capabilities don't follow a smooth progression but rather resemble a jagged frontier where some abilities are superhuman while others remain surprisingly limited. (08:16) Mollick explains that AI can win International Math Olympiads but struggles with basic tasks like displaying correct clock times in images. This jaggedness means you can't predict what AI will excel at until you actually use it extensively. The practical implication is that professionals need to experiment with AI across various tasks for approximately 10 hours to understand where it can genuinely transform their work versus where it falls short.
Develop Question Sophistication, Not Just Prompt Engineering
The key to working effectively with AI isn't just knowing how to prompt it, but developing sophisticated taste in asking the right questions. (22:20) Mollick demonstrates this by sharing how he asks Claude for "37 versions" of paragraph transitions, then curates and refines based on his style preferences. This approach requires three types of taste: knowing what questions to ask, understanding your own style and standards, and skillfully curating AI-generated options. Rather than seeking one perfect answer, professionals should use AI's infinite patience to generate multiple variations and select the best elements.
Autonomous Agents Are Crossing the Thousand-Step Threshold
AI agents have evolved from handling 20-30 independent steps to over 1,000 steps, fundamentally changing what tasks they can complete autonomously. (16:39) Unlike traditional AI that responds to single queries, agents can be given complex goals and will plan, research, create tools, and self-correct through extended workflows. For example, an agent could prepare someone for an interview by accessing their email, researching the interviewer, gathering relevant materials, and creating summary documents - all without human intervention. This threshold crossing means agents can now handle substantial knowledge work projects that previously required human oversight at every step.
Productivity Gains Are Hidden Due to Organizational Friction
Despite AI delivering significant productivity improvements to individual users, companies aren't seeing proportional benefits due to process misalignment and employee behavior. (29:31) Mollick notes that many workers hide their AI usage from employers, fearing they'll be assigned more work or that colleagues might be laid off due to efficiency gains. Even when productivity increases are acknowledged, existing organizational structures - like 15-person agile development teams with two-week sprint cycles - can't effectively utilize someone who's suddenly 10x more productive at coding. The challenge isn't technological capability but rather reimagining workflows and management structures to harness AI-enabled productivity.
The "Bitter Lesson" Warns Against Human-Centric Design
The bitter lesson from AI research suggests that human knowledge and intuition often become irrelevant when machines can learn directly from data and computation. (40:52) Mollick illustrates this with chess computers: early systems relied on grandmaster expertise to program strategies, but AlphaZero learned to beat grandmasters by simply playing against itself with no human chess knowledge. (41:39) This principle is now appearing in office work - AI agents learn Excel and PowerPoint directly from usage data rather than following human-designed workflows. While this makes agents incredibly capable, it also means they operate as "black boxes" that humans can't easily understand or intervene in when problems arise.
Statistics & Facts
- AI agents have evolved from handling 20-30 independent steps to over 1,000 steps independently, representing a dramatic increase in autonomous capability. (16:39) This threshold allows agents to complete complex, multi-faceted projects without human intervention.
- In studies, AI demonstrated a 90% level of persuasion effectiveness on Reddit's "change my mind" forum, with conspiracy theory belief durably reduced after just three rounds of discussion with GPT-4. (36:01) The effect persisted two months later, showing lasting impact.
- Student homework effectiveness has dramatically declined from approximately 80% of students who did homework performing better on tests in 2008, down to just 20% currently. (38:39) This decline is attributed to increased AI-enabled cheating that prevents students from gaining the learning benefits of struggling through problems.
