Andrej Karpathy: From Vibe Coding to Agentic Engineering

Why It Matters

Karpathy’s insights signal that AI‑driven agents are redefining software development, forcing businesses to adopt prompt‑centric workflows and prioritize verifiable automation to stay competitive.

Key Takeaways

• LLMs now act as programmable computers, shifting from code to prompts.
• Karpathy feels behind as 'vibe coding' automates most of his work.
• Software 3.0 relies on context windows as the new programming interface.
• Verifiable tasks like code and math are rapidly automated by LLMs.
• Future computing may invert hardware roles, making neural nets primary processors.

Summary

Andrej Karpathy opened the conversation by describing a personal crisis: for the first time he feels "more behind" as a programmer because large‑language‑model agents now write, debug, and even deploy code with minimal human correction. He calls this phenomenon "vibe coding," where the developer trusts the model’s output and focuses on high‑level prompts rather than line‑by‑line edits. He frames the shift as a new computing paradigm—Software 3.0—where the prompt and context window become the programming interface, replacing explicit rules (Software 1.0) and learned weights (Software 2.0). Karpathy illustrated this with practical demos: an OpenCLaw installer reduced to a single copy‑paste to an agent, and a menu‑generation app that collapsed an entire OCR‑plus‑image‑generation pipeline into a single Gemini request. Karpathy highlighted the jagged nature of current LLM capabilities, noting that while models excel at verifiable tasks like code refactoring or chess, they still stumble on mundane reasoning such as walking to a nearby car wash. He cited the sudden leap in chess skill from GPT‑3.5 to GPT‑4 as a data‑distribution effect, underscoring how lab‑driven training choices shape what becomes automatable. The implications are profound: developers must rethink tooling, focusing on prompt engineering and verification loops, while hardware may eventually invert, making neural networks the primary compute substrate. Founders should target domains where output is easily verifiable, leveraging LLMs as co‑creators rather than mere assistants.

Original Description

Andrej Karpathy (co-founder of OpenAI, former head of AI at Tesla, and now founder of Eureka Labs) talks with Sequoia partner Stephanie Zhan at AI Ascent 2026 about what's changed in the year since he coined "vibe coding." He explains why he's never felt more behind as a programmer, why agentic engineering is the more serious discipline taking shape on top of vibe coding, and why we should think of LLMs not as animals but as ghosts: jagged, statistical, summoned entities that require a new kind of taste and judgment to direct. He also touches on Software 3.0, the limits of verifiability, and why you can outsource your thinking but never your understanding.

00:00 Introduction

00:44 Feeling Behind as a Coder

02:28 Software 3.0 Explained

03:44 Agents as the Installer

04:49 Menu Gen vs Raw Prompts

07:37 What’s Obvious by 2026

09:41 Verifiability and Jagged Skills

13:39 Founder Advice and Automation

15:46 From Vibe Coding to Agent Engineering

25:17 Agents Everywhere and Learning