The Jagged Frontier: AI Capabilities, Bottlenecks, and Reverse Salients

Research Date: 2025-12-20
Source URL: https://www.oneusefulthing.org/p/the-shape-of-ai-jaggedness-bottlenecks

Reference URLs

• One Useful Thing - The Shape of AI
• Original Jagged Frontier Research Paper - Dell’Acqua et al., 2023
• Cochrane Review AI Reproduction Study

Summary

Ethan Mollick’s analysis presents a framework for understanding AI capabilities through three interconnected concepts: the “jagged frontier,” bottlenecks, and reverse salients. The jagged frontier describes AI’s inconsistent ability profile—superhuman performance in some domains (differential diagnosis, advanced mathematics) coexists with subhuman performance in seemingly simpler tasks (visual puzzles, physical manipulation). This uneven capability distribution creates bottlenecks that prevent full automation of complex workflows, even when AI excels at the intellectually demanding components. The concept of “reverse salients,” borrowed from historian Thomas Hughes’s work on electrical systems, identifies specific weaknesses that, once resolved, enable rapid system-wide advancement.

The framework challenges simplistic narratives about AI replacing human workers. While AI capabilities continue expanding, the jagged nature of improvement means human-AI complementarity remains essential. Bottlenecks migrate from intelligence to institutions, from technical limitations to process constraints. The practical implication is that monitoring benchmark improvements provides less insight than tracking which bottlenecks break—when they do, previously constrained capabilities flood forward simultaneously.

Main Analysis

The Jagged Frontier Concept

The term “jagged frontier” originated in Mollick’s 2023 research with collaborators, describing how AI ability maps poorly to human intuitions about task difficulty. An AI system may demonstrate superhuman performance on complex medical diagnosis while failing at tasks humans consider elementary.

Colin Fraser’s conceptual diagrams, referenced in the article, illustrate scenarios where AI and human capabilities may never fully overlap. Rather than AI progressively subsuming all human abilities, the two capability sets may remain distinct with different peaks and valleys.

Bottleneck Typology

Mollick identifies three distinct bottleneck categories that constrain AI from automating complete workflows:

The Cochrane review reproduction study exemplifies edge case bottlenecks. GPT-4.1 reproduced twelve work-years of systematic review effort in two days, screening 146,000 citations with accuracy exceeding human reviewers. Yet the AI could not access supplementary files or email authors—comprising less than 1% of errors but preventing full automation.

Reverse Salients and Capability Lurches

Thomas Hughes’s concept of “reverse salients” from the history of electrical systems describes how progress often stalls on single technical or social problems. When AI labs identify and resolve these constraints, capabilities jump forward discontinuously.

Google’s Nano Banana Pro image generation system demonstrates this pattern. High-quality image generation was the reverse salient constraining visual communication applications. NotebookLM can now generate complete slide presentations by creating each slide as an image rather than writing programmatic code—a capability impossible when image quality was a limiting factor.

Implications for AI-Human Complementarity

The jagged frontier framework suggests AI advancement does not follow a simple trajectory toward complete human replacement. Several factors sustain human relevance:

1. Memory persistence: LLMs do not retain learning from new tasks permanently, creating systematic capability gaps
2. Institutional requirements: Regulatory, legal, and organizational processes mandate human involvement
3. Edge case handling: Rare situations requiring judgment, physical access, or social navigation remain human domains
4. Unwritten knowledge: Understanding implicit rules and contextual requirements that determine actual needs

Jobs involving multiple tasks distributed across the jagged frontier offer natural human-AI complementarity. Consulting and design work, for instance, combine analytical tasks (inside AI frontier) with stakeholder management, implicit requirement discovery, and novel solution generation (outside AI frontier).

Visual Analysis

Capability Radar Comparison

The article references a radar chart comparing GPT-4 and GPT-5 capabilities across dimensions including reading, math, general knowledge, reasoning, and memory. The visualization demonstrates uneven improvement—memory shows minimal advancement while other capabilities progress rapidly, consistent with jagged frontier predictions.

Image Generation Evolution

Mollick’s “otter on a plane using WiFi” test illustrates capability advancement in image generation:

• 2021: Incoherent imagery with distorted subjects and nonsensical composition
• 2025 (Nano Banana Pro): Coherent scenes with correct spelling, consistent shadows, multiple angles, and semantic understanding of complex prompts

Key Findings

• AI capability advancement follows an uneven “jagged frontier” pattern where performance varies dramatically across task types, independent of human-perceived difficulty
• Bottlenecks constrain full workflow automation even when AI excels at intellectually demanding components—these bottlenecks may be capability-based, institutional, or edge-case-driven
• “Reverse salients” are specific weaknesses that, when resolved by focused development, unlock multiple previously constrained capabilities simultaneously
• Memory persistence remains a significant weakness with limited improvement, potentially creating permanent gaps in AI-human capability overlap
• The practical strategy for anticipating AI impact is monitoring bottleneck resolution rather than benchmark improvements

Theoretical Implications

Counter-Argument to Linear Replacement Narratives

Tomas Pueyo’s viral visualization suggested AI’s expanding frontier would eventually circumscribe all human capabilities. Mollick’s framework challenges this by noting:

1. The human capability frontier is not fixed—humans adapt and specialize
2. Jaggedness may persist or shift rather than smooth out
3. Non-capability bottlenecks (institutional, edge-case) exist independently of raw AI ability

Complementarity vs. Substitution

The framework supports viewing AI as complementary rather than substitutive. Even “very smart AI cannot easily substitute for humans” due to jaggedness-induced bottlenecks. This has dual implications:

• Positive: Prevents rapid job displacement, creates new collaborative opportunities
• Negative: Limits acceleration potential in areas like scientific research where full automation could dramatically increase progress

References

1. Mollick, E. “The Shape of AI: Jaggedness, Bottlenecks and Salients” - One Useful Thing, December 20, 2025
2. Dell’Acqua, F., McFowland, E., Mollick, E. et al. “Navigating the Jagged Technological Frontier” - Harvard Business School Working Paper, 2023
3. Hughes, T. “Networks of Power: Electrification in Western Society, 1880-1930” - Johns Hopkins University Press, 1983
4. Cochrane AI Review Reproduction Study - MedRxiv preprint, 2024