📖 Business
The Jagged Frontier
AI capabilities are not a smooth gradient from "can't do" to "can do." They form a jagged, unpredictable frontier — a landscape of peaks and valleys where AI excels at surprisingly complex tasks while failing at seemingly simple ones. Mollick uses the term to replace the naive binary model ("AI can do X but not Y") with a nuanced map of capability that must be explored empirically rather than assumed. An LLM can write a convincing essay on 18th-century French politics (a task most humans would find very hard) but might fail to count the letters in "strawberry" (a task any child can do). The frontier is jagged because LLMs work through pattern matching and statistical prediction, not through understanding.
2
Minutes
2
Concepts
+45
XP
1
How It Works
Why the frontier is jagged:
- LLMs learn statistical patterns from training data, not causal reasoning
- Tasks that are common in training data (essays, code, summaries) tend to be peaks
- Tasks that require precise counting, spatial reasoning, or novel logic tend to be valleys
- The relationship between human difficulty and AI difficulty is weak — hard human tasks can be easy for AI and vice versa
Properties of the frontier:
- Unpredictable — you can't reliably predict whether AI will be good at a task without testing it
- Constantly moving — each model generation shifts the peaks and fills in some valleys
- Task-specific, not domain-specific — AI might be great at one kind of legal analysis and terrible at another
- Context-dependent — the same capability can vary based on how you prompt, what context you provide, and which model you use
Practical implications:
- Don't pre-filter tasks — try things with AI before assuming it can't help
- Test empirically — run your actual task through the AI rather than reasoning from analogies
- Expect surprises in both directions — AI will both impress and disappoint you in ways you didn't predict
- Benchmark regularly — what failed 6 months ago may work now as the frontier moves
- Build workflows that accommodate failure — since the frontier is jagged, your process must handle cases where AI output is unreliable