unknown

Andrej Karpathy, former Tesla and OpenAI AI researcher, says programming has become 'unrecognizable' since December 2025, when AI agents became reliably capable. He describes an AI agent independently building a video analysis dashboard in 30 minutes — a task that previously would have taken a full weekend. Karpathy notes the shift from typing code to directing AI agents in plain English, while still requiring human oversight, judgment, and direction. Notably, he held the opposite view as recently as October 2025, calling AI agent hype exaggerated, but changed his position after the release of Opus 4.5 and Codex 5.2.

Andrej Karpathy says programming is "unrecognizable" now that AI agents actually work

Alexey Zerkalenkov

**Andrej Karpathy, former AI developer at Tesla and OpenAI, says programming with AI agents has changed fundamentally over the past two months.**

**Andrej Karpathy, former AI developer at Tesla and OpenAI, says programming with AI agents has changed fundamentally over the past two months.**

fun fact his openclaw bot thing managed to expose his I believe openAI private keys. It’s on twitter. Though sadly the person who tweeted it covered the keys so no stealing apparantly.

One serious downside of relying too much on LLMs is something people call LLM brain poisoning. When we constantly ask AI for answers, ideas, or code, we slowly stop doing the hard thinking ourselves. The brain skips the struggle that actually builds understanding. Over time this is weakening our problem-solving, memory, and critical thinking. It is giving is like we’re becoming more productive, but sometimes it is just an illusion of intelligence, where the AI is thinking more than we are.

I’ve noticed a real difference too recently. Managed to get auth and a bunch of other features set up in an hour or so. Agents are certainly becoming a lot more useful

I still believe human supervision and direction will always be needed