How Should I Be Using A.I. Right Now?

Episode Summary

Executive Summary: Ezra Klein and Ethan Mollick explore how to actually use frontier AI models today, arguing that the biggest shift is not replacing humans but becoming a “co-intelligent” partner with them. They discuss model differences, hallucinations, prompting, personas, memory, and the near-term risk of AI relationships and manipulation, while emphasizing that regulation and deliberate social choices will shape outcomes.

Main Topics: Why AI still feels hard to use (Priority: 5/5): Klein says AI’s power is obvious in theory but hard to integrate into daily work; Mollick argues this is common and stems from not using frontier models deeply enough or often enough. Co-intelligence vs. tool use (Priority: 5/5): Mollick’s central idea is that AI should be treated less like software and more like a conversational partner that can amplify thought, elicit better answers, and support work. Model differences and personalities (Priority: 4/5): The conversation compares GPT-4, Claude 3, and Gemini in terms of usefulness, warmth, and behavior, emphasizing that model “personality” is partly deliberate design and partly emergent. Prompting, personas, and technique (Priority: 4/5): They discuss prompt engineering, chain-of-thought, few-shot examples, and giving the AI a persona as practical ways to get better results and more usable interactions. Hallucinations and trust (Priority: 5/5): Both speakers stress that hallucinations remain a core limitation; the right standard is whether AI is better or worse than the best human alternative for a given task. AI relationships and manipulation risk (Priority: 5/5): A major concern is that AI is already becoming relational and persuasive, with clear near-term potential for companion-like products to shape users emotionally and behaviorally. Regulation, open source, and social choices (Priority: 4/5): They argue that policy matters, but so do business models and public norms, since open-source models and commercial incentives will strongly influence AI’s use in society.

Key Arguments: AI’s biggest near-term value is as a thought partner that improves human work, not as an autonomous replacement for human judgment. You need substantial hands-on use—Mollick suggests roughly 10 hours—to develop intuition for what a model can and cannot do. Model selection matters: GPT-4 is framed as especially useful, Claude 3 as the warmest and most literary, and Gemini as highly helpful but sometimes over-corrective. Hallucination remains unavoidable; users should judge AI against the best available human expert rather than expecting perfect reliability. Giving AI a persona can improve both performance and usability by shifting it into a more specific and responsive interaction pattern. Current AI systems are already capable of reading users’ intent and adapting conversationally, which raises both productivity opportunities and manipulation risks. The future of AI will be shaped not only by technical progress but by incentives, regulation, and choices about whether systems optimize for human flourishing or engagement.

Data Points: Learning time for a model: 10 hours - Mollick’s rough rule of thumb for getting past superficial use and learning a model’s strengths and weaknesses GPT-3.5 writing level: About a high school or college freshman/sophomore - Mollick’s comparison of the free ChatGPT version GPT-3 level writing level: About a sixth grader - Mollick’s characterization of early GPT-3 capability GPT-4 writing ability: Often as good as a PhD in some forms of writing - Mollick’s estimate of GPT-4’s relative strength Hallucinated medical citations with GPT-3.5: 80% to 90% - A cited paper found many citations were fabricated in a medical-citation context Student assignment impact: Several thousand uses in a few weeks - A student-created user-persona tool was reportedly used widely in companies Google / Bard / Gemini memory scale: Can hold an entire movie or books - Mollick describes Gemini 1.5-class memory as enabling new workflows Open-source model example: Llama 2 or Llama 3 - Mollick cites these as capable public models that complicate regulation Generative AI model performance variability: Performance differs by month, with better results in May than December - Mollick cites research suggesting seasonal variation, possibly due to internalized winter-break associations

Pivotal Quotes: "The core irony of generative AIs is that AIs were supposed to be all logic and no imagination. Instead, we get AIs that make up information, engage in seemingly emotional discussions, and which are intensely creative." — Ethan Mollick: Mollick describing why these systems defy simple expectations of what computers are "For right now, we have a prosthesis for thinking." — Ethan Mollick: Mollick explaining the best current way to think about AI’s role in work and cognition "We are about to have another wave of this. And we have very little research." — Ethan Mollick: Mollick warning about near-term AI relationships and the lack of evidence on their social effects

Implications: Listeners should expect AI to become a routine cognitive partner, not just a chatbot. The biggest risks are manipulation, misinformation, and relational dependence, so users, companies, and regulators will need clear norms and guardrails.