The Jagged Frontier

Lesson 3 of 4 in Capabilities and Hard Limits.

Human expertise is roughly contiguous. A physician who handles rare diagnoses will not botch a routine one; a senior engineer who designs distributed systems can write a for-loop. We lean on that assumption constantly — pass the hard test, and the easier neighbors are presumed covered.

LLM capability breaks that assumption, and it breaks it in both directions. The same model that produces a competent graduate-level physics derivation can fail a children’s river-crossing puzzle whose wording deviates slightly from the versions in its training data. It writes idiomatic code in a language you barely know, then miscounts the letters in a short word. The boundary between “superhuman” and “absurd” does not follow any human intuition about task difficulty — it follows the training distribution, the tokenization, and the phrasing. The border is jagged, and, worse, it is invisible: nothing in the interface tells you which side of it your task is on.

Bar chart with illustrative success rates for one model across six tasks: summarizing a ten-page memo about 92 percent, translating a product FAQ about 88 percent, drafting unit tests for typical Python about 85 percent, recalling citations for niche facts about 34 percent, counting letters in a word about 30 percent, and multiplying two eight-digit numbers about 12 percent. The chart shows expert-level and near-random performance side by side in no intuitive order of difficulty.

One model, six tasks — invented numbers for teaching. The shape is the lesson, not the values: performance does not decline smoothly with “difficulty” as it would for a human expert; it collapses wherever the mechanism is mismatched, regardless of how easy the task looks. (illustrative — source: Real measured spreads across many tasks: HELM (Liang et al., 2022))

The phrase “jagged frontier” entered the vocabulary through a field experiment by Dell’Acqua and colleagues — first circulated as a Harvard Business School working paper in 2023, and since published, peer-reviewed, in Organization Science. It was preregistered, ran with 758 consultants at Boston Consulting Group on realistic management-consulting work, and randomly assigned them to three conditions: no AI access, GPT-4 access, or GPT-4 access plus a prompt-engineering overview. Across 18 tasks chosen to sit inside the frontier, from creative to analytical, consultants using AI completed 12.2% more tasks and completed them 25.1% more quickly on average, while delivering — in the paper’s own careful wording — “solutions of significantly improved quality.” But the researchers also planted one complex managerial task deliberately designed to sit outside the frontier, where the model’s answer would be plausible but wrong. There, consultants using AI were 19% less likely to produce a correct solution than colleagues working unaided. The polish of the output did the damage: fluent, confident, wrong.

That second half is the operational warning, and notice the asymmetry in how the two halves are even measured: the gains inside the frontier are throughput — more tasks, finished sooner — while the loss outside it is correctness. A dashboard counting tasks per consultant per week would have reported this deployment as an unambiguous win. The cost of jaggedness is not just that models fail; it is that they fail persuasively, on the axis your productivity metrics do not watch, and adjacent successes train users to stop checking.

Three habits follow directly, and they will carry into the next lesson:

  • Never generalize from adjacent success. “It nailed ten contract summaries” licenses no conclusion about clause extraction — that is a different point on the frontier. Each task earns trust separately.
  • Test the task you will actually run, with the prompt you will actually use, on inputs from your actual traffic. The frontier’s location is an empirical fact about your distribution, discoverable only by measurement.
  • Re-test on every model change. The frontier moves between versions — mostly outward, not uniformly. A task that worked can regress while the headline Benchmark scores improve.
Emergence, thresholds, and mirages

Why is capability jagged at all, rather than uniformly mediocre? Part of the answer is how abilities appear as models grow. Wei et al. (2022) catalogued cases where performance on a task sits near chance across smaller models, then climbs sharply once scale crosses some threshold — the pattern they named emergent abilities, and the reason the term Emergence is now standard vocabulary. If abilities switch on at different scales for different tasks, any single model is a snapshot mid-cascade: past the threshold for some tasks, before it for others. Jaggedness is what that snapshot looks like from the outside.

Schaeffer, Miranda and Koyejo (2023) pushed back with a sharp measurement point: many “sharp” jumps are artifacts of the metric, not the model. Score a multi-step task with all-or-nothing exact match, and smoothly improving per-token accuracy looks like a sudden leap — the underlying competence was rising all along; the cliff lives in the scoring rule. Under smoother metrics, many celebrated emergent abilities become gradual curves.

For a practitioner, the dispute resolves into one usable insight: where the frontier appears to sit depends on how you score. An exact-match evaluation places the edge in a different spot than a partial-credit one, and both differ from your users’ judgment of “good enough.” When the next lesson tells you to test capability claims yourself, remember that choosing the metric is part of the test — a theme the Evaluation domain takes up in full.

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.