Tesla’s Former AI Director, who feels behind as programmer, warns software engineers again

Programming has undergone a seismic shift, according to former Tesla AI director Andrej Karpathy. He claims coding agents, virtually non-existent before December, now perform complex tasks like setting up a video dashboard in minutes. This marks a departure from traditional coding, with programmers now directing AI agents rather than writing line-by-line code.

Andrej Karpathy, Tesla’s former AI director and OpenAI cofounder, has once again stirred the tech world with a bold claim—programming has fundamentally changed in just two months, and it’s “nowhere near business as usual” anymore. In a detailed post on X, Karpathy wrote that coding agents “basically didn’t work before December and basically work since.” The shift, he said, isn’t the gradual kind the industry is used to. It’s abrupt, disruptive, and already reshaping how software gets built.To make his point, Karpathy shared a weekend project. He asked an AI agent to set up a video analysis dashboard for his home cameras—logging into a server, configuring SSH keys, downloading and benchmarking an AI model, building a web UI, and writing a full report. The agent handled it all in about 30 minutes, troubleshooting errors on its own. “I didn’t touch anything,” he wrote.

Karpathy says programmers now manage AI agents instead of writing code line by line

This is a sharp evolution from the man who coined “vibe coding” in early 2025, where he described casually prompting AI and barely reviewing the output. That was fun for throwaway projects, he’d said at the time. Now, Karpathy is talking about something far more structured—spinning up AI agents, assigning tasks in plain English, and reviewing their work in parallel.“You’re not typing computer code into an editor like the way things were since computers were invented, that era is over,” he wrote.The post comes weeks after Karpathy admitted he’d “never felt this much behind as a programmer,” describing the profession as being “dramatically refactored.” He spoke of needing to master a new layer of abstraction involving agents, subagents, prompts, memory modes, and MCP protocols—all while the underlying AI models keep changing.

AI coding productivity still a mixed bag despite industry hype

Not everyone’s sold on the revolution, though. A METR study from July found AI assistants actually decreased experienced developers’ productivity by 19%. Bain & Company called programming productivity gains “unremarkable.” Yet Google CEO Sundar Pichai has said AI writes over 30% of new code at Google, and Anthropic CEO Dario Amodei claimed Claude was behind 90% of the company’s code.Karpathy, for his part, isn’t calling it magic. “It needs high-level direction, judgement, taste, oversight, iteration and hints and ideas,” he wrote. But the leverage, he believes, is already enormous—and only growing.