The Shape of AI: Jaggedness, Bottlenecks and Salients

Ethan Mollick has been chewing on this idea since 2023: AI systems ace things that should be brutally hard — differential diagnosis, competition math — while flubbing stuff a toddler could manage, like running a vending machine or reading a simple visual puzzle. He calls it the jagged frontier, and the metaphor going around lately, courtesy of Tomas Pueyo, is that this jaggedness just doesn't matter because the overall AI ceiling is rising faster than human ability ever will. Mollick isn't buying that framing, at least not fully.

His pushback centers on memory. Large language models don't retain what they learn from one task to the next in any durable way, and that single limitation ripples outward, creating whole categories of work AI simply can't take over no matter how smart it gets at reasoning or reading. A recent attempt to actually chart AI capability across skills backed this up — reading, math, and general knowledge are climbing fast, memory basically isn't moving. Colin Fraser's sketches of human-versus-AI ability overlap make the same point visually: some zones are pure AI dominance, some are pure human territory, and plenty barely touch at all.

The more useful concept, Mollick argues, is the bottleneck — a system is only as capable as its weakest linked component, so a single stubborn weakness can stall automation even when everything else is superhuman. Sometimes that's a raw skill gap: LLMs still aren't trusted to read medical scans unsupervised, or push back the way a good therapist should. Sometimes it's not about ability at all — AI can spit out promising drug candidates ten times faster, but clinical trials still need actual human bodies and the FDA still needs actual human reviewers, so the bottleneck just relocates to the institution. A study on AI reproducing Cochrane systematic reviews nails this: GPT-4.1 did roughly twelve work-years of review labor in two days, screening 146,000 citations and beating human accuracy — but it still can't email a researcher for unpublished data, and that gap, under 1% of the work, is exactly why a human still has to sit in the loop.

Mollick borrows a term from historian Thomas Hughes — reverse salient — for the single fixable weakness holding an entire system back. Fix it, and progress lurches forward all at once. His live example is Google's new Nano Banana Pro image model, which suddenly makes AI-generated slides genuinely usable — coherent text, real layout, no code required, unlike Claude or ChatGPT's PowerPoint-via-code approach, which still looks stiff. The intellectual heavy lifting — summarizing a book, drafting a report — was already inside the frontier for over a year. What was missing was simply the ability to render it well visually, and that bottleneck just broke.

Mollick's bigger point is that watching benchmarks tells you less than watching bottlenecks. Consultants and designers aren't obsolete just because AI can now make a decent slide, because their jobs are bundles of jagged tasks — building trust, reading unwritten political rules, inventing something that doesn't look machine-made — that AI still can't touch. But somewhere a lab is staring at memory, or real-time learning, or physical-world action, treating it exactly like Hughes's reverse salient. When that breaks, expect another lurch, and expect it with no warning.