Jagged Frontier (3D)
About this app
Two 3D stars — human and AI: each spike is a task domain, its length the capability level. The AI grows extremely unevenly (code and proofs explode, physical common sense lags); slid together the stars interpenetrate — there is no single overtaking moment (after Karpathy).
The jagged edge of AI
The term “jagged frontier” describes how unevenly the abilities of today’s AI systems are distributed. On some tasks they clearly outperform people, on other seemingly simple ones they fail. A common observation is that progress comes faster in domains with easily checked results, such as coding or mathematics, than in everyday understanding or long-term planning. It follows that there is no single moment when AI overtakes “the human”. It overtakes domain by domain. Andrej Karpathy has repeatedly described this unevenness as jagged intelligence.
What the app shows
Two stars stand in a rotatable 3D scene: a human star and an AI star. Each spike is a task domain and its length the capability level; the same direction means the same domain on both stars. Eight domains are named: arithmetic, coding, math proofs, translation, factual breadth, physical common sense, long-horizon planning and reliability. The human star is slightly jagged too. The readouts show in how many domains the AI is at least level, and name the strongest and weakest AI domain. The lengths are schematic, not measured data.
Things to try
Move the AI progress slider from 0 to 100 % or let the AI grow on its own, and watch which spikes race ahead and which lag behind. Set the distance to 0 so that both stars overlap: AI tips pierce out of the human star and vice versa, and neither contains the other. The spikes slider sets between 8 and 44 spikes, reroll redistributes the extra spikes, and domain labels can be switched on.
Frequently asked questions
- What does jagged frontier mean in AI?
- The jagged frontier is the uneven boundary of what AI systems can do. Inside it they handle tasks well, outside it poorly, and the boundary does not follow what people find hard or easy. The term became widely known through a 2023 study by Harvard Business School and Boston Consulting Group on AI use in consulting work.
- Why is AI strong at maths but weak at everyday tasks?
- In areas like coding or mathematics it is possible to check automatically whether a solution is correct. That makes it easier to train with many verified examples and feedback. For physical common sense, long plans or reliability such clear checks hardly exist, so abilities there grow more slowly.
- When will AI overtake humans?
- The question assumes a single overtaking moment, which the jagged frontier picture says does not exist. AI already outperforms people in some domains and lags well behind in others. It is more useful to ask, task by task, how well AI handles it.
- What can I try in this app?
- You control AI progress, slide the two stars into each other and change the number of spikes. The readouts show in how many domains the AI matches the human and which domain is strongest and weakest.
Subject: Neural networks
More from Neural networks
- Agent loop
- Attention Visualiser
- Backpropagation — blame in reverse
- Backpropagation: learning from one mistake
- Basin of Attraction
- Your job under AI
- The temperature dial
- Diffusion model — images from noise (3D)
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps