Karpathy Joins Anthropic: An Industry Shift from Vibe Coding to Agentic Engineering

Breaking News: A Strategic Move by an AI Thought Leader

Andrej Karpathy—a figure regarded as a spiritual leader by countless AI practitioners—has officially announced his joining of Anthropic. As a founding member of OpenAI, the architect of Tesla's pure-vision approach to autonomous driving, and one of the world's most influential AI educators, each of his career moves has almost precisely aligned with critical inflection points in technological evolution.

Karpathy completed his PhD in computer vision under Fei-Fei Li at Stanford University in 2015, with his dissertation focusing on applying convolutional neural networks to image captioning. He then joined OpenAI as a founding member, moved to Tesla in 2017 as AI Director, and spearheaded the controversial decision to completely abandon LiDAR in favor of a camera-only pure-vision autonomous driving system—a choice that sparked enormous debate in the industry but ultimately defined Tesla FSD's technical roadmap. After leaving Tesla in 2023, he founded the AI education company Eureka Labs, and his neural network tutorial series on YouTube has accumulated over ten million views, earning recognition as "one of the best free educational resources in AI."

Karpathy's announcement was remarkably concise, conveying just three core messages: he has joined Anthropic; he believes the next few years of frontier LLM development will be particularly formative and is excited to return to the research frontlines; and he remains passionate about education, planning to resume related work when the time is right. You might have missed that he did not shut down Eureka Labs—he has merely paused his educational endeavors temporarily to go all-in on frontier research.

A Cognitive Reversal: From Criticism to Being Blown Away

To understand the deeper logic behind Karpathy's choice of Anthropic, one must appreciate the fundamental reversal in his perception of AI capabilities over the past six months. As late as the end of 2025, he was publicly criticizing AI agents for insufficient intelligence, weak multimodal capabilities, and lack of continuous learning ability. Yet just a few months later, he openly admitted to being completely blown away by the latest advances in agentic AI for programming.

When underlying foundation models have evolved to the point of autonomously planning project architectures, independently diagnosing complex environment dependencies, and fixing errors on their own, requiring students to spend large amounts of time manually writing low-level backpropagation code starts to feel somewhat misaligned with the cutting edge of technological evolution. Backpropagation is the core algorithm for neural network training—it uses the chain rule to propagate output errors backward through layers, computing each parameter's gradient with respect to the loss function to guide weight updates. Manually implementing backpropagation has long been considered the "rite of passage" for deep learning—only by deriving gradient computations from scratch can one truly understand how neural networks work. Karpathy's LLM101N project followed this tradition, aiming to guide learners from basic neurons all the way to building a complete language model step by step. However, when frontier models can autonomously complete these low-level implementations, the necessity of this educational path genuinely warrants reconsideration—not because fundamentals don't matter, but because both the starting point and endpoint of education need recalibration as technological boundaries shift. The LLM101N repository on GitHub is currently archived, hinting that the old educational paradigm needs to wait for new frontier technologies to stabilize before it can be reconstructed.

The Intellectual Evolution from Vibe Coding to Agentic Engineering

In February 2025, Karpathy coined the term "Vibe Coding," describing a radically new development experience enabled by large language models—where developers need only express requirements in natural language, and AI tools automatically generate code, build websites, and complete deployment.

Vibe Coding is essentially a popularized description of "Prompt-Driven Development." Developers describe requirements in natural language to AI coding tools like Cursor or GitHub Copilot, and models generate corresponding implementations based on code understanding within their context window. This paradigm is extremely efficient for small projects or prototype validation, but suffers from two structural flaws: first, "skill hollowing"—developers gradually lose debugging ability and cannot effectively intervene when AI-generated code goes wrong; second, the "context window bottleneck"—even GPT-4-class models with 128K token contexts struggle with enterprise codebases spanning hundreds of thousands of lines, producing code that is locally consistent but systemically contradictory in the absence of global understanding, creating cascading errors that are difficult to trace.

As his practice deepened, Karpathy keenly identified these two fatal problems. His research agent once automatically ran 700 experiments over two days, during which it attempted to match user IDs across different databases by brute-force comparing email addresses as string formats—profoundly exposing the fragility of early Vibe Coding.

Karpathy then formally introduced the concept of "Agentic Engineering." His "Three Stages of Software Evolution" framework provides crucial context for understanding this transition: Software 1.0 corresponds to traditional imperative programming, where logic is entirely explicitly coded by humans; Software 2.0 corresponds to the deep learning era, where humans design network architectures and loss functions while behavior implicitly emerges from data and gradient descent; Software 3.0 is the agentic programming paradigm built on LLMs as infrastructure. Under this paradigm, the engineer's core work shifts to: designing task decomposition strategies, defining verification criteria, building Tool Use/Function Calling interfaces, and designing orchestration logic for multi-agent collaboration. This demands stronger systems thinking and abstraction capabilities from engineers, rather than faster code typing speed. The engineer's role must evolve from "writer" to "orchestrator."