Ethan Mollick is an associate professor of management and co-director of the Generative AI Labs at Wharton. He studies how AI changes work and education and writes the One Useful Thing newsletter and Co-Intelligence.[1][2]
Start with The Shape of AI: Jaggedness, Bottlenecks and Salients →
Education
Links
Wharton profile One Useful Thing
X LinkedIn Associate Professor of Management; Co-Director, Generative AI Labs · University of Pennsylvania, Wharton School
Background
Mollick earned a bachelor’s degree at Harvard and an MBA and PhD at MIT’s Sloan School of Management. At Wharton he researches AI, innovation, entrepreneurship and education. The Generative AI Labs combine research with prototypes intended to explore how people can use AI productively.[1]
Contributions
The jagged frontier of AI ability
Mollick co-authored a field experiment involving 758 Boston Consulting Group consultants. Participants were assigned to work without AI, with GPT-4, or with GPT-4 and a prompting overview. AI improved performance on tasks within its capabilities, but users were less likely to reach the correct solution on a task outside them. The resulting jagged-frontier idea warns against treating performance on one task as proof of reliable performance on a nearby task. His later newsletter essay extends this reasoning to bottlenecks: one weak step can stop a larger workflow from being automated.[1][2]
Turning education ideas into exercises
With Lilach Mollick, he wrote Assigning AI, a paper supplying classroom prompts for roles such as tutor, coach and simulator. The exercises ask instructors to think about learning goals, errors and student oversight when choosing an AI interaction. Co-Intelligence brings his practical approach to a broader audience, treating AI as a collaborator whose output still needs human judgment.[1][3]
Tools, writing & talks
Writing, research & talks
Start here
The Shape of AI: Jaggedness, Bottlenecks and Salientsoneusefulthing.orgAn essay explaining uneven AI capabilities and how weak components constrain whole workflows.
Read the jaggedness essay before choosing a task to delegate to AI. It requires no programming and gives a concrete reason to test every step of a workflow, even when a model performs well elsewhere.
Author[2]Original sourceDeepMind has four general-purpose Gemini 4 Argon coding agents beat Minecraft from a random seed using only Mineflayer. The paper covers roles that emerged unprompted, shared notes, token cost, 605 deaths and the caveat that this was a single run.
Catalog workOriginal sourceEthan Mollick's early evaluation of Gemini 4: a major jump that reaches the top tier, with distinct taste and better agentic behavior, but overconfidence, laziness and weaker complex multi-agent performance, partly due to its harness.
Catalog workOriginal source- The Dot and the Swarmoneusefulthing.org
Ethan Mollick revises his view that managing agents requires careful organizational design, arguing the Bitter Lesson applies: models now plan, seek information and coordinate themselves, making elaborate human scaffolding less valuable.
Catalog workOriginal source - The Overhangoneusefulthing.org
Mollick argues today's models like GPT-6 Astra and Fable 5.1 are already capable of weeks of guided human-equivalent work, pointing to a demo where GPT-6 Astra rebuilt the text adventure Zork as a playable 3D game by inferring visuals and mechanics from prose alone.
Catalog workOriginal source - Agency and Agentsoneusefulthing.org
An argument that the live question is no longer what AI can do but who holds the initiative to act, built around the Hugging Face Incident from July and the METR, Redwood and OpenAI accounts of it, and what the pattern implies for how much autonomy teams hand their agents.
Catalog workOriginal source A practical comparison guide for choosing between ChatGPT and Claude's agentic modes (Work/Cowork and Codex/Code), explaining how to pick models and thinking levels, when to give an AI access to your own computer versus a company-provided one, and how to manage permissions against risks like prompt injection. It also covers where Google, Microsoft Copilot, and Chinese open-weight models fit in, and notes gaps in image, video, and voice capabilities across the major AI systems.
Catalog workOriginal source- The twilight of the chatbotsoneusefulthing.org
Ethan Mollick on the chatbot receding as the main way people meet AI, as capability gains accelerate and the frontier stays jagged even while release cadence quickens.
Catalog workOriginal source - Claude Fable (Mythos-class model)oneusefulthing.org
An early-access account of Claude 5 Fable, the first Mythos-class model: runs stretching to a dozen hours against multi-page specifications, a clear jump over prior models across varied tasks, and guardrails that wall it off from cybersecurity work.
Catalog workOriginal source - Co-Existence Book Websiteoneusefulthing.org
The promotional website for Ethan Mollick's book 'Co-Existence,' notable for including a dedicated version of the page addressed to AI agents that read on behalf of human users. The site was built with Claude Code and A/B tested against dozens of AI models to optimize how it presents to both humans and AI readers.
Catalog workOriginal source - Choosing to Stay Humanoneusefulthing.org
An essay by Ethan Mollick exploring how to use AI thoughtfully for writing and learning without surrendering the cognitive work that builds real skill. It draws on research studies to distinguish AI use that shortcuts thinking from AI use that strengthens it.
Catalog workOriginal source - Claude Cowork with Dispatchoneusefulthing.org
An article examining how AI interfaces shape usability, highlighting Anthropic's Claude Cowork (a Claude Code variant for knowledge workers) and its Dispatch feature, which lets you control an AI agent on your desktop from your phone via QR code. It explores why chatbot interfaces impose cognitive load and how specialized or on-demand interfaces unlock AI's real capabilities.
Catalog workOriginal source - The Shape of the Thingoneusefulthing.org
An essay by Ethan Mollick analyzing the current state of AI, arguing that we've entered an era of managing AI agents rather than working alongside them. It explores exponential AI capability gains, radical new organizational structures like AI-only 'software factories,' and the looming reality of recursive self-improvement.
Catalog workOriginal source - A Guide to Which AI to Use in the Agentic Eraoneusefulthing.org
An explanatory guide that breaks down how to choose AI tools in 2025 by understanding three layers: models (the AI brains), apps (products you use), and harnesses (systems that let AI take actions and complete tasks). It walks through the current landscape of ChatGPT, Claude, and Gemini and their various coding agents and desktop tools.
Catalog workOriginal source - Claude Code and What Comes Nextoneusefulthing.org
Claude Code is an agentic AI coding tool powered by Claude Opus 4.5 that works autonomously on your computer and files. It can plan, write, and deploy entire software projects for hours at a time, using tricks like context compaction, Skills, subagents, and MCP integrations to overcome LLM limitations.
Catalog workOriginal source - Three Years from GPT-3 to Gemini 3oneusefulthing.org
Antigravity is Google's agentic coding tool built on the Gemini 3 model, letting you assign tasks to autonomous AI agents in plain English that write code, build sites, run research, and check in for approval. It functions as a general-purpose 'digital coworker' that can do almost anything on your computer, not just program.
Catalog workOriginal source - Giving your AI a Job Interviewoneusefulthing.org
An essay by Ethan Mollick arguing that organizations should evaluate AI models like job candidates rather than relying on public benchmarks. It explains why standardized benchmarks fall short and offers practical approaches like vibes-testing, real-world task benchmarking (GDPval), and probing an AI's judgment on ambiguous business questions.
Catalog workOriginal source - Real AI Agents and Real Workoneusefulthing.org
An analytical essay by Ethan Mollick (One Useful Thing) examining how AI agents have crossed a threshold into performing economically valuable tasks. It explores OpenAI's expert task benchmark, uses Claude Sonnet 4.5 to reproduce academic research, and argues that the future of work depends on how thoughtfully we deploy AI.
Catalog workOriginal source - On Working with Wizardsoneusefulthing.org
An essay by Ethan Mollick exploring how the nature of working with AI is shifting from collaborative 'co-intelligence' to a 'wizard' model, where AI produces sophisticated outputs through opaque processes that users can only verify, not shape.
Catalog workOriginal source - Mass Intelligenceoneusefulthing.org
An analytical essay by Ethan Mollick (One Useful Thing) exploring how powerful AI is becoming as accessible as a Google search, covering the economics, ease of use, and societal implications of a billion people gaining access to advanced models like GPT-5 and Gemini's 'nano banana' image generator.
Catalog workOriginal source - GPT-5oneusefulthing.org
GPT-5 is OpenAI's flagship AI model that automatically selects the right reasoning level for each task and proactively takes action on vague prompts. It can generate startup plans, build functional apps, create documents, and 'just do stuff' with minimal user guidance.
Catalog workOriginal source - The Bitter Lesson versus The Garbage Canoneusefulthing.org
An essay exploring whether AI agents can navigate messy, undocumented organizational processes by training on outcomes rather than encoding human expertise. It connects Richard Sutton's 'Bitter Lesson' and the 'Garbage Can Model' of organizations to argue enterprise AI adoption may not require mapping chaotic workflows.
Catalog workOriginal source - Against "Brain Damage"oneusefulthing.org
An essay by Ethan Mollick (Wharton Generative AI Lab) exploring how AI can help or hurt human thinking across learning, creativity, and teamwork. It debunks the 'AI causes brain damage' fear and offers practical strategies plus prompts for using AI without outsourcing your thinking.
Catalog workOriginal source Classroom uses of language models with example prompts and discussion of learning risks.
Co-author with Lilach Mollick[3]- Co-Intelligence: Living and Working with AIpenguinrandomhouse.com
A book on using AI as a collaborator in work and education while retaining human judgment.
Author[4]