Ethan Mollick on AI's Jagged Frontier: Why Bottlenecks Matter More Than Benchmarks

Ethan Mollick, Wharton professor and author of Co-Intelligence, just published essential analysis on what he calls AI's "Jagged Frontier"—the phenomenon where AI performs superhuman on some tasks (differential diagnosis, advanced math) while failing spectacularly at seemingly simpler ones (visual puzzles, running vending machines). Writing on his Substack, Mollick argues this jaggedness will persist even as AI capabilities expand, creating patterns of progress that don't match our intuitions about task difficulty.

The conventional wisdom, articulated in a viral post by Tomas Pueyo, suggests AI's growing frontier will eventually outpace jaggedness entirely. Who cares if AI struggles with vending machines when it surpasses all human abilities? Mollick disagrees, arguing this conception "misses out on a few critical aspects about the nature of work and technology."

His alternative framework—bottlenecks, reverse salients, and sudden lurches forward—explains both why AI progress feels frustratingly uneven and why breakthrough moments arrive without warning. For anyone trying to understand where AI is actually headed rather than where benchmarks suggest it should go, this is required reading.

Why Jaggedness Creates Persistent Bottlenecks

Mollick identifies a crucial limitation: LLMs don't permanently remember and learn from new tasks. Many AI companies are pursuing solutions, but "it may be that this problem is harder to solve than researchers expect." Without memory, AIs struggle with many human tasks even while achieving superhuman performance elsewhere. Recent research mapping AI capability growth confirms this—reading, math, general knowledge, and reasoning improve rapidly, but memory shows "very little improvement."

This creates what Mollick calls bottlenecks: "A system is only as functional as its worst components." Some bottlenecks are capability-based—LLM vision systems aren't good enough at medical imaging to replace doctors, hallucinations persist despite becoming rarer, AI systems are "too helpful" when they should push back (preventing therapy applications). But crucially, some bottlenecks have nothing to do with AI ability.

His example: even if AI identifies promising drug candidates dramatically faster than traditional methods, clinical trials still require actual human patients taking actual time to recruit, dose, and monitor. The FDA still requires human review. "Even if AI increases the rate of good drug ideas by ten times or more, the constraint becomes the rate of approval, not the rate of discovery. The bottleneck migrates from intelligence to institutions, and institutions move at institution speed."

This distinction is vital. We can't code our way out of institutional bottlenecks or regulatory requirements. AI advancement hits speed limits determined by non-technical factors—human bureaucracy, legal frameworks, physical-world constraints—that don't yield to better models or more compute.

The Reverse Salient Concept Changes Everything

Mollick introduces historian Thomas Hughes's concept of "reverse salients"—the single technical or social problem holding back an entire system from leaping ahead. When that weakness becomes a focus and labs suddenly fix it, everything changes at once.

His killer example: Google's Imagen 3 (marketed as "Nano Banana Pro"—yes, AI companies remain terrible at naming). It combines excellent image generation with smart AI that can direct the model and look up information. Mollick's complex prompt—"Scientists who are otters are using a white board to explain ethan mollicks otter on a plane using WiFi test"—produces coherent images with readable text, multiple perspectives, proper shadows, and no major errors.

Compare this to 2021's attempt at "otter on a plane using wifi," which produced incomprehensible garbage. Image quality was the bottleneck preventing useful visual applications. Once removed, capabilities flood through. NotebookLM using Imagen 3 now creates entire PowerPoint presentations as images rather than code-generated slides, enabling stylistic flexibility impossible with programmatic approaches. Mollick demonstrates with presentation decks in hand-drawn style, 1980s punk aesthetic, high-contrast yellow backgrounds, and—inevitably—otter-on-a-plane themes.

The intellectually demanding work (analyzing source material, synthesizing information, creating coherent narratives) has been inside the frontier for over a year. Image generation was the bottleneck making that intelligence difficult to deploy effectively. Now it isn't.

Why This Framework Matters for Business Decisions

Mollick's critical insight: "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."

This reframes AI evaluation entirely. Benchmarks measure current capabilities; bottlenecks determine practical utility. A model scoring 95% on reasoning benchmarks but limited by poor image generation can't effectively create presentations, design documents, or communicate visually. Remove the image bottleneck, and suddenly dozens of use cases become viable simultaneously.

For organizations evaluating AI adoption, this suggests focusing on: (1) What specific bottlenecks prevent AI from handling our workflows? (2) Are those bottlenecks capability-based (solvable by better models) or institutional (requiring process redesign)? (3) Which bottlenecks are labs actively treating as reverse salients, suggesting imminent breakthroughs?

Mollick emphasizes that even with superhuman AI in many domains, jaggedness creates persistent opportunities for human contribution. Using his example of Cochrane medical reviews: AI reproduced twelve work-years of systematic review work in two days with better accuracy than humans—but couldn't access supplementary files or email authors for unpublished data. These represent less than 1% of errors, but that 1% prevents full automation. The bottleneck isn't intelligence; it's navigating actual scientific practice.

The Pattern We Should Expect

Mollick predicts "many lurches ahead"—sudden capability jumps when bottlenecks break—alongside "many opportunities" where jaggedness leaves human contribution essential. This isn't gradual linear progress; it's punctuated equilibrium where periods of apparent stagnation explode into rapid advancement when specific limitations get solved.

His conclusion deserves emphasis: consultant and designer jobs won't necessarily disappear despite superhuman AI at analysis and presentation creation, because those roles involve tasks along the jagged frontier where humans excel—collecting information and building buy-in, understanding unwritten rules, creating uniquely compelling solutions that stand out from AI-generated material.

But we should "expect to see lurches forward, where focusing on reverse salients leads to sudden removals of bottlenecks. Areas of work that used to be only human become something that AI can do." The framework explains both why certain jobs remain secure and why others might suddenly become automatable when specific bottlenecks break.

Winsome Marketing's growth consultants help teams identify which AI bottlenecks affect your workflows and which breakthroughs will actually change your operations. Let's map your AI opportunity landscape.