The Shape of AI: Jaggedness, Bottlenecks and Salients
Source
Ethan Mollick
Author
Ethan Mollick
Date
Key takeaways · AI-distilled
Properly scaffolded, GPT-4.1 reproduced and updated an entire issue of 12 Cochrane medical reviews in two days, roughly 12 work-years of effort. It screened over 146,000 citations, extracted data, ran the statistics, and beat human reviewers on accuracy.
What blocked full automation of that review pipeline was mundane: the model could not open supplementary files or email authors for unpublished data. Those account for under 1% of errors, and they still force an expert human into the loop.
Speeding up discovery just relocates the constraint. If AI finds drug candidates ten times faster, the binding limit becomes patient recruitment and FDA review. The bottleneck migrates from intelligence to institutions, and institutions move at institution speed.
Historian Thomas Hughes called the single problem stalling an entire system a "reverse salient." It explains why AI progress looks stuck and then lurches: labs concentrate on the one blocking weakness, the way math ability jumped once it became the obvious barrier.
AI slide decks were bad because the model had to write code to build them. NotebookLM paired with Nano Banana Pro renders each slide as a single image instead, which only became viable at high image quality, so visual style turns into a prompt rather than a template.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Why it matters
Gives builders a practical lens — watch which single capability is bottlenecking a workflow rather than chasing benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → deltas — to predict where AI will suddenly unlock new product surface, illustrated by image generation breaking open slides/docs/visual comms.
Key quotes
“The bottleneck migrates from intelligence to institutions, and institutions move at institution speed.”
“reproduced and updated an entire issue of Cochrane reviews (n=12) in two days, representing approximately 12 work-years of traditional systematic review work.”
“So far, every lurch forward leaves yet more edges in which humans are needed.”
