Why AI is best understood as 'jagged intelligence'

If a colleague could solve world-class Olympiad math puzzles in their head but consistently failed at everyday common-sense reasoning, you would think their mind was fascinatingly uneven.

That is essentially where modern artificial intelligence stands today. On April 15, researchers and economists increasingly rallied around a concept that captures this exact paradox: "jagged intelligence."

The phrase was coined by AI researcher Andrej Karpathy to describe models that perform miracles in complex domains while tumbling over basic practical tasks.

Elite math skills, zero street smarts

Recent benchmarks show just how sharp the contrast has become.

AI models managed to crack five out of six notoriously grueling problems at the 2025 International Math Olympiad. It was a performance that rivaled the sharpest young mathematical minds on Earth.

Yet bring those same underlying systems into everyday reality, and they still stumble over basic, common-sense reasoning prompts that an average human wouldn't think twice about.

This is what Karpathy means by "jagged." Progress is not expanding as a smooth, tidy circle. It stretches out into aggressive spikes of superhuman capability right alongside deep valleys of bizarre failure.

Because systems advance at such uneven rates, knowing where an AI tool will succeed and where it will randomly break down remains one of the hardest guessing games in tech.

Automation takes tasks, not titles

This jagged frontier is quietly reshaping how economists look at the future of work.

For the past few years, the prevailing anxiety has centered on entire job titles getting wiped out overnight, from coders to paralegals.

Economists now argue that view is fundamentally flawed. Automation is not poised to swallow professions whole; it is targeting specific, narrowly defined tasks.

An AI model might easily draft complex boilerplate code, prove a theorem, or synthesize reams of technical data in seconds. But it will still fail to piece together the messy, context-heavy operational steps required to deliver an actual project.

The real disruption won't be a mass pink slip for an entire industry. It will be the internal restructuring of how our daily workday gets split up.

Why it matters

We think adopting the "jagged intelligence" mental model is the most practical reality check the AI ecosystem has received in months.

For individual professionals, it offers a grounded roadmap. Stop worrying about whether your job will vanish next quarter, and start mapping out which parts of your daily workflow land on the AI's jagged peaks. Delegating deep computation or synthesis while holding onto real-world reasoning and context is the winning move right now.

For product builders, it exposes the biggest flaw in the current push toward fully autonomous AI agents. Shipping tools that pretend the jagged edge does not exist is a recipe for embarrassing edge-case blowups. The smartest teams will build software designed for graceful handoffs between AI strengths and human common sense, rather than pretending the gaps have already been bridged.