Andrej Karpathy on Jagged Intelligence: Why LLMs Excel at Code but Fail at Everyday Logic
Andrej Karpathy never fails to deliver fresh, thought-provoking insights, and his recent appearance at Sequoia Ascent 2026 was no exception. In a wide-ranging conversation, the former OpenAI co-founder and Tesla AI leader unpacked the current state of large language models and the rapidly evolving AI landscape. Among the many gems he shared, four key ideas stood out as especially relevant for anyone building with or relying on AI today: the inherent “jaggedness” of LLM capabilities, the emergence of a new unified application paradigm, the distinction between vibe coding and agentic engineering, and the critical reminder that while you can delegate thinking, you cannot delegate true understanding.
These points aren’t just academic observations—they reflect the practical reality developers, product builders, and business leaders are encountering right now in 2026. Let’s dive deeper into each one.
The Jagged Nature of Modern LLMs
One of the most striking concepts Karpathy highlighted is what he calls the “jaggedness” (or jagged intelligence) of today’s LLMs. You’ve probably experienced this yourself. A model can refactor a massive 100,000-line codebase flawlessly, solve complex mathematical problems at an Olympiad level, or generate production-ready code in seconds. Yet ask it something seemingly trivial—like whether you should drive or walk to a car wash 50 meters from your house—and it might earnestly suggest walking, completely missing the obvious real-world context.
Why does this happen? It comes down to how these models are trained. LLMs don’t grow up playing outside, learning through trial and error in the physical world the way humans do. Instead, they’re shaped by massive datasets rich in technical documentation, textbooks, code repositories, and verifiable problems. High-quality data around everyday common-sense reasoning is far scarcer. Labs pour reinforcement learning rewards into domains where outputs can be automatically checked—math, coding, chess—because verification loops accelerate improvement there. In messy, ambiguous real-life scenarios, those clear reward signals are missing, so performance remains rough and unreliable.
This isn’t a bug in the architecture; it’s an expected outcome of the training process. Karpathy describes LLMs as “jagged entities” that spike dramatically in verifiable domains while staying flat or erratic elsewhere. Understanding this jagged profile helps us use AI more effectively: lean on it for what it’s superhuman at, and stay vigilant in areas where common sense should rule.
From Fragmented Apps to One Powerful Chat Window
Another major shift Karpathy described is the move away from specialized, standalone applications toward a single, dynamic chat interface. In the old world, you had a different app for everything—car rental here, spreadsheets there, photo editing somewhere else. Each had its own rigid interface and limited scope.
Now, everything is collapsing into one conversational window. The LLM can generate a custom interface on the fly, perfectly tailored to whatever task you’re handling in the moment. Need to analyze data? It spins up the right tools and visualizations. Want to edit images or book travel? The interface adapts instantly. This new paradigm feels more fluid and human, reducing the friction of switching contexts and learning multiple UIs. It represents a fundamental rethinking of how we interact with software.
Vibe Coding vs. Agentic Engineering: Raising the Floor and the Ceiling
Karpathy revisited and expanded on the “vibe coding” concept he introduced previously, contrasting it with the more serious discipline emerging now: agentic engineering.
Vibe coding lowers the barrier to entry. Thanks to natural language becoming the dominant programming interface, far more people can build simple applications. English—or even Russian—lets non-engineers prototype ideas quickly without wrestling with syntax. This democratizes creation and raises the floor for what the average person can achieve.
Agentic engineering, on the other hand, raises the ceiling dramatically. Experienced builders who deeply understand system architecture, combined with powerful AI agents, can now accomplish what once required entire teams. They orchestrate agents to handle research, coding, debugging, and iteration at unprecedented speed and scale.
The result? The gap between junior and mid-level developers is narrowing because AI handles so much routine work. But the divide between strong senior developers and true senior architects is widening. Those who don’t evolve into architects—capable of designing large, reliable, maintainable systems—may find themselves competing with boosted mid-level talent. The message is clear: adapt or risk being outpaced.
You Can Outsource Thinking, But Not Understanding
Perhaps the most profound takeaway is this: agents can do the heavy lifting—searching information, analyzing data, writing code, and executing tasks—but they cannot replace human judgment about what truly matters.
You can outsource the mechanics of thinking, but you cannot outsource understanding. Someone still needs to decide priorities, evaluate trade-offs, define success, and maintain the bigger picture. The human role is elevating from doing the work to directing it with intention and taste. This shift demands stronger strategic thinking, clearer vision, and better evaluation skills than ever before.
In Karpathy’s view, the people who thrive will be those who master this higher-level orchestration—using AI as a powerful collaborator while retaining ownership of goals and values.
What This Means for Developers and Builders in 2026 and Beyond
Karpathy’s insights paint an exciting yet demanding picture. AI is not a smooth, uniform intelligence but a jagged, specialized tool that rewards thoughtful use. The tools are becoming more powerful and accessible, but the premium on human insight, architecture, and judgment is only increasing.
For developers, this is both an opportunity and a call to level up. Embrace vibe coding to prototype faster, master agentic workflows to multiply your output, and invest deeply in the architectural skills that AI can’t yet replicate. For product leaders and founders, the unified chat paradigm opens new possibilities for delightful user experiences, but success still hinges on understanding real human needs.
The AI revolution isn’t replacing humans—it’s elevating those who learn to guide it wisely. As Karpathy continues to demonstrate through his own work and thinking, staying curious, building taste, and focusing on what matters most will separate the winners from the rest.
