Ethan Mollick argues in One Useful Thing that AI advances through bottlenecks

Ethan Mollick argues in One Useful Thing that AI advances through bottlenecks

Ethan Mollick’s article was published on December 20, 2025. He argues that AI’s “Jagged Frontier” — uneven abilities that can make AI excel at some tasks and fail at others — shapes how quickly it can take on work. Small weaknesses can become bottlenecks: a system may handle most of a task yet still need people for the parts it cannot do, or for processes such as clinical trials and regulatory review that depend on human time and institutions. Mollick describes a study in which GPT-4.1, with prompting and support, reproduced and updated Cochrane reviews in two days, while human expertise remained necessary to handle edge cases. He uses “reverse salients” for the technical or social problems holding a system back, and argues that solving one can produce a sudden advance. His example is Nano Banana Pro: paired with Gemini in NotebookLM, image generation lets the system make slides as images rather than code, opening new visual styles for material that AI could already summarize. The images remain imperfect and uneditable, he notes. Mollick’s conclusion is that AI progress may arrive in lurches as bottlenecks break, but each advance leaves tasks where people are needed. He sees both potential job disruption and opportunities for human work, and urges readers to watch bottlenecks rather than benchmarks. Which weakness will become the next bottleneck to break is unknown.

Publication: d18d09fc-0c99-8706-2cc2-b3e64f62b252

attribution — pass: The reading attributes the article to Ethan Mollick in One Useful Thing, which matches the byline and publisher. Arguments are framed as his throughout. The Cochrane study, Pueyo and Hughes are not misattributed.

The Shape of AI: Jaggedness, Bottlenecks and Salients And why Nano Banana Pro is such a big deal Ethan Mollick

summary — pass: The reading accurately conveys the argument chain: jaggedness, then bottlenecks, then reverse salients, then lurches. It also gives the conclusion to watch bottlenecks rather than benchmarks. Jagged Frontier, bottleneck and reverse salient are explained in plain language, and benchmark is defined beside the reading. Cochrane reviews are not explained, and the omission of the roughly 12 work-years scale weakens the example's force, but neither prevents understanding. The question follows from the source's own closing questions.

If you want to understand where AI is headed, don’t watch the benchmarks. Watch the bottlenecks. When one breaks, everything behind it comes flooding through.

date — pass: December 20, 2025 matches the dateline and the UTC day of originalPublished. modifiedAt equals the publication time, so there is no later edit. The reading includes no hindsight framing or later events, and the forecasts remain attributed to Mollick. This is a publisher page, so no archive disclosure is needed.

Ethan Mollick Dec 20, 2025

support — pass: Each claim is supported by the source. The statement on human edge cases in the Cochrane study is backed by the text, though not by the claim's excerpt. Job disruption is supported by the source saying areas of work that used to be only human become something AI can do. The note that the images are imperfect and uneditable is source-stated; the source also says editing will apparently be fixed soon.

Twelve work-years become two days, but only if a human with expertise in how science is actually done handles the edge cases.

media — pass: No media is selected, and no verified candidates were supplied, so there is nothing to misidentify.