INNOVATIONOngoing practice82% confidence

The Jagged Frontier Capability Model

AI advances where money points — not toward general intelligence, but toward premium industry verticals

Problem it solves

Debunks AGI imminence narratives by explaining why capability advances are selective rather than general

Best for

Evaluating AI capability claims and identifying where capability development is structurally directed versus where it is absent

Not ideal for

Short-term capability comparisons between products — this is a structural directional analysis, not a benchmark framework

Overview

Why this framework exists

Hao observes that AI capabilities do not advance uniformly across domains — they advance specifically in the directions companies choose to invest based on revenue potential from high-value industries. Finance, law, medicine, and commerce receive capability investment because those industries can pay premium prices for AI services. Capabilities in domains with lower ability to pay are comparatively neglected. This creates a 'jagged frontier' — peaks of capability in commercially valuable domains, persistent gaps elsewhere.

The framework has two key implications. First, the 'general intelligence' framing is strategically misleading — these are purpose-built tools for specific high-value markets, shaped by market incentives rather than developing toward genuine generality. Second, transfer learning across domains is shallower than marketed — a self-driving car trained in San Francisco cannot transfer to Mumbai without full retraining, because the jagged capability frontier does not generalize across sufficiently different distributions.

For investors and analysts, the jagged frontier model is a counter to AGI imminence arguments: genuine general intelligence would advance across all domains including commercially unimportant ones. Selective advancement in proportion to commercial value is evidence of market-incentive-driven capability development, not emergent general intelligence.

Core principles

5 total
  1. AI capability investment follows commercial value, not the path toward general intelligence.
  2. The jagged frontier means capability peaks in high-value commercial domains and persistent gaps in others — not uniform advancement across all tasks.
  3. Transfer learning across sufficiently different distributions requires near-full retraining, limiting the practical generalizability of capability gains.
  4. General intelligence framing obscures the commercially-determined shape of actual capability development.
  5. Capability announcements should be evaluated against the commercial incentive to develop that capability — if the incentive is weak, the claim deserves higher skepticism.

Steps

4 steps
  1. Map the commercial incentive landscape for capability claims
    Before evaluating any AI capability claim, identify which industries stand to pay for that capability and at what price point. High commercial value plus high automation surface area equals high investment incentive. Low commercial value plus low automation surface area equals low investment — and claimed capabilities in this quadrant warrant higher skepticism.
    Pro tipHao's explicit list of priority commercial domains: finance, law, medicine, healthcare, commerce. Capability claims outside these domains are less likely to be well-resourced.
  2. Test transfer assumptions explicitly
    When a capability claimed in one domain is extrapolated to another, test the transfer assumption by checking how different the training distribution is from the target domain. The San Francisco-to-Mumbai autonomous driving case is the benchmark: if similar geographic or distributional differences exist, the capability does not transfer without retraining.
    Pro tipAsk: was the model trained on data from this specific deployment context? If not, what is the distribution gap, and has the company measured performance on the target distribution directly?
    WarningCompanies often report capability on benchmark datasets with distributions close to training data — real-world deployment on different distributions frequently shows significant degradation.
  3. Identify the capability gap map
    For a complete picture of an AI system's actual capability frontier, deliberately seek performance on domains outside the high-commercial-value priority list. Domains with low ability to pay (indigenous language preservation, subsistence agriculture optimization, community mental health) will have persistent gaps. The shape of the gap map reveals the commercial incentive structure.
  4. Apply the jagged frontier test to AGI claims
    If a company claims their system has achieved or is approaching general intelligence, test whether capabilities are advancing in domains with low commercial value. Genuine general intelligence would advance across all domains including commercially unimportant ones — selective advancement in proportion to commercial value is evidence against the AGI claim.
    Pro tipThe jagged frontier test and the AGI shapeshifting narrative framework are complementary — the jagged frontier provides the empirical test for the narrative claim.

Checklist

Saved in your browser

Examples

2 cases
Self-driving San Francisco to Mumbai transfer failure

Autonomous vehicle systems trained extensively on San Francisco driving conditions cannot be deployed in Mumbai without full retraining. The road infrastructure, traffic patterns, pedestrian behavior, signage conventions, and vehicle types are sufficiently different that the capability does not transfer. This is despite both being urban driving tasks — demonstrating how shallow transfer is even within a single domain when the distribution shifts.

OutcomeCompanies building autonomous vehicles for global markets must invest in market-specific training data and retraining for each major deployment context — the jagged frontier repeats at every distributional shift.
Commercial priority shaping capability investment

Hao documents that OpenAI's capability roadmap prioritizes finance, law, medicine, and commerce based on willingness to pay for AI services, not based on scientific priority or societal benefit. This means AI safety research, climate modeling, materials science for energy transition, and other high-societal-value but lower-revenue domains are systematically underfunded relative to their importance.

OutcomeThe jagged frontier has a commercial shape — capability peaks where money points. This is the empirical consequence of market-incentive-driven capability development rather than mission-driven development.

Common mistakes

3 traps
Treating benchmark performance as real-world performance
AI benchmarks are typically designed on distributions similar to training data. Real-world deployment involves distributions that may be significantly different — the jagged frontier means that even small distributional differences can collapse performance in ways that benchmark results do not predict.
Inferring general capability from demonstrated specific capability
Strong performance on legal document analysis does not imply strong performance on structural engineering analysis, even if both are 'professional expert domains.' The commercial incentive to invest in legal AI is much higher than in structural engineering AI, so the frontier is jagged between them despite surface similarity.
Accepting frontier lab research priority claims as capability roadmaps
When a frontier lab announces research priorities, those priorities reflect commercial return projections, not the most technically important frontiers. Claims that a company is pursuing beneficial AI in low-commercial-value domains should be evaluated against their actual resource allocation, not their stated intentions.

Origin story

How this framework came to be

The jagged frontier observation is implicit throughout Hao's reporting on how AI companies select research priorities. The San Francisco-to-Mumbai transfer failure is cited as a concrete example from the self-driving domain. The commercial prioritization mechanism — 'how they pick capabilities to advance — it's based on which industries would be able to pay them the most money' — is a direct statement from her analysis of OpenAI's capability roadmap.

Source

Traced to primary
Source · PODCAST
AI Whistleblower: We Are Being Gaslit By AI Companies, They're Hiding The Truth!
Karen Hao · 2025
Open source →

Related frameworks

Browse all Innovation →