Anyone who has spent real time using AI has had the same experience.
One moment, it performs what feels like a magic trick. It writes working code in seconds. It summarizes long reports instantly. It solves problems that used to require serious technical expertise.
Then five minutes later, it produces an answer so obviously wrong you wonder if the machine just forgot how to think.
This strange duality has become one of the defining characteristics of modern AI.
A recent New York Times opinion piece captured this perfectly, describing how AI systems built by Google and OpenAI managed to correctly answer five of six problems from the International Math Olympiad, one of the most difficult math competitions in the world for elite high school students. Yet those same systems struggled with a simple reasoning question about whether someone should walk or drive a car only 50 meters to a repair shop.
The contrast is almost comical. Superhuman at one task. Clueless at another.
Researchers have a name for this phenomenon. They call it “jagged intelligence.”
The term was popularized by Andrej Karpathy, one of the founding researchers at OpenAI and former head of AI at Tesla. His point was that AI does not improve in a smooth, linear way, the way human intelligence does. Instead, it develops sharp spikes of capability surrounded by deep valleys of weakness.
Humans tend to get better across many cognitive abilities at the same time as we grow and learn. Our reasoning, intuition and problem-solving skills develop together.
AI doesn’t work that way.
It can be astonishingly good at some things while being bafflingly bad at others that appear far simpler.
That’s why AI can now generate complex software code, solve difficult math problems and analyze enormous datasets, yet still struggles with common sense reasoning or simple real-world planning.
The experience of using AI today often feels like driving through a mountain range. You hit these towering peaks of capability where the technology looks almost magical. Then suddenly, you drop into a valley where it behaves in ways that make you question how intelligent it really is.
Understanding why this happens starts with how these systems are trained.
Large language models learn primarily by analyzing massive amounts of digital data. They study patterns in text, code, images and other information gathered from across the internet. When trained at enormous scale, they become incredibly good at recognizing and reproducing those patterns.
That makes them extraordinarily effective at tasks involving language, mathematics, programming and structured problem solving.
But the internet captures only a fraction of human knowledge.
Most of what humans know comes from interacting with the physical world. We learn through experience, observation, trial and error, and social context. Much of that knowledge never gets written down in a way that AI systems can easily absorb.
As a result, AI models develop very deep expertise in some domains while remaining weak in others that require intuition, planning or real-world judgment.
This jagged structure of intelligence has important implications for the debate over AI and jobs.
Every few months, we hear claims that AI is about to wipe out entire professions. The narrative usually focuses on one particular capability where AI has made dramatic progress, such as writing code or generating text.
But jobs are rarely defined by a single skill.
Most professions are made up of dozens of tasks. Some of those tasks may be easy to automate. Others may be stubbornly resistant.
Take software development as an example. AI can now generate code snippets incredibly fast. Tools like Codex and Claude can produce working code that once required hours of effort.
But writing code is only one part of software engineering.
Designing systems, understanding business requirements, debugging complex interactions and making architectural decisions still require human judgment. AI can assist with those tasks, but it doesn’t yet replace the people responsible for them.
History gives us plenty of examples of technologies that dramatically improved productivity without eliminating entire professions.
The pocket calculator could perform arithmetic faster than any human accountant. That did not eliminate the accounting profession. Instead, it changed the nature of the work. Accountants moved up the value chain, focusing more on analysis, planning and strategy.
Spreadsheets had a similar effect on financial analysts. They made certain tasks far easier but ultimately expanded the scope of what analysts could do.
AI is likely to follow a similar pattern, though on a larger scale.
The technology will automate certain tasks while amplifying others. Workers who learn how to collaborate with these systems will become more productive. Those who ignore them may struggle to keep up.
At the same time, it would be a mistake to assume the jagged nature of AI will remain forever.
The technology is improving at an astonishing pace. Techniques like reinforcement learning allow models to learn from feedback and gradually close some of their capability gaps. Each new generation of models tends to fill in a few more of the valleys.
Tasks that seemed impossible for AI only a few years ago are now routine.
But the jagged structure has not disappeared.
Even the most advanced models still produce strange failures that remind us how different this form of intelligence is from our own. That unpredictability is part of what makes AI both exciting and unsettling.
It also explains why the debate around AI so often swings between extremes.
On one side are those who believe artificial general intelligence is right around the corner and that machines will soon replace most knowledge workers. On the other are those who dismiss AI as little more than a sophisticated autocomplete engine.
Both views miss the bigger picture.
AI is not simply becoming a digital version of human intelligence.
It is evolving into something entirely different.
A form of intelligence that can outperform humans in certain narrow domains while still struggling with tasks that people solve instinctively.
The more we understand these jagged peaks and valleys, the better we will be able to predict how AI will shape the future of work and technology.
For now, the mountains are impressive. The valleys are still deep. And navigating that terrain is going to be one of the defining challenges of the AI era.
