Karpathy: Software 3.0 arrives as agents flop

Karpathy's talk, delivered at Y Combinator's AI Startup School, is built around a historical arc rather than a product announcement. He argues that software's method of creation has flipped twice in a short period: from humans writing code, to data training neural networks whose weights act as code, to natural language serving as the programming interface for an LLM. He points to Tesla's self-driving work as firsthand evidence, describing how handwritten C++ code was progressively replaced by neural networks.

A second thread concerns where this technology spreads. Karpathy notes that LLMs reached ordinary users before militaries and large companies, an inversion of the usual adoption path, and uses the metaphor that if an LLM goes down, the planet gets dumber — treating it as infrastructure like electricity or water.

The stakes, on his account, sit at the boundary between demo and production. He recalls a Waymo demo in 2013 that looked perfect yet remains unfinished 12 years later, and describes his own experience letting an agent run free until a 1,000-line diff became unreadable. His prescription — build Iron Man's suit, not his robot — suggests the near-term gains may go to teams that keep a human in the loop rather than chase full automation.