Andrej Karpathy: Software 3.0, Vibe Coding, and the Future of Engineering

In a recent insightful interview at AI Ascent, legendary AI researcher Andrej Karpathy discussed the rapid evolution of AI programming tools and a profound paradigm shift in how we build software. Even Karpathy admitted to feeling “behind” as a programmer until he fully embraced the new wave of agentic tools.

You can watch the full conversation here: YouTube Video.

The transition he describes isn’t just about faster autocomplete; it’s a fundamental move towards what he calls “Software 3.0,” where Large Language Models (LLMs) act as the primary interpreter.

The Shift to Software 3.0

Computing paradigms are evolving rapidly:

Software 1.0: Traditional programming relying on explicit, hand-written rules.
Software 2.0: The deep learning era, where neural networks learn from data via weights.
Software 3.0: The agentic era, where prompting an LLM acts as the core programming mechanism.

Programming is increasingly shifting from writing raw code to writing natural language text that an agent executes. Tasks that previously required complex, brittle scripts can now be achieved by feeding context directly to an agent. As Karpathy noted, December marked a clear turning point where these tools saw a massive leap in reliability, allowing programmers to trust AI to write functional chunks of code without constant manual correction.

Vibe Coding vs. Agentic Engineering

Karpathy draws a crucial distinction between two emerging styles of AI-assisted development:

Vibe Coding: This approach democratizes software creation. It’s about raising the floor, allowing non-programmers to build functional software by focusing on high-level goals—essentially creating by “vibes” while the AI handles the implementation.
Agentic Engineering: This preserves the quality ceiling. Professional engineers are becoming managers who write detailed specs, establish testing protocols, and direct AI agents safely to ensure secure and robust code.

“Vibe coding is about raising the floor… agentic engineering is about preserving the quality bar,” Karpathy explains.

Verifiability and “Jagged Intelligence”

Why do AI models automate certain domains like math and coding so much faster than others? The answer lies in verifiability.

AI improves rapidly in verifiable domains because Reinforcement Learning (RL) models can easily check if an answer is mathematically correct or if code compiles and passes tests. This training method creates what Karpathy calls jagged entities—models that are superhuman in narrow, verifiable domains (like refactoring a 100,000-line codebase) but highly flawed in basic common sense or logic outside their training distribution.

If you are an AI founder, Karpathy advises targeting these verifiable problems to ride the wave of AI improvements, while treating current AI agents as intern-level entities that require clear direction.

Predicting 2026: The Neural Computer

Extrapolating current trends to 2026, Karpathy envisions a world of “neural computers.” In this future, the neural network acts as the core host process, directly receiving raw video and audio inputs. User interfaces will be rendered dynamically via diffusion models, reducing the traditional CPU to a mere coprocessor dedicated to specific, deterministic tasks.

Outsourcing Thinking, Not Understanding

With AI executing more of the raw work, what happens to human value?

Karpathy emphasizes that human motivation, aesthetics, and true creativity remain firmly outside the easy grasp of current RL-driven models. Current models actually struggle with simplifying code because simplicity isn’t heavily prioritized in their RL training, leading to a tendency for AI to produce bloated, overly complex solutions.

Human intervention is constantly required to trim this bloat and enforce elegant design. While you can rely on AI to generate the code, you must fundamentally understand the underlying concepts to properly debug and direct the AI.

As Karpathy powerfully summarizes: “You can outsource your thinking, but you can’t outsource your understanding.”