Two years after coining "vibe coding," Andrej Karpathy is back with another term that's reshaping how we think about software development. But this time, he's not just naming a trend—he's backing companies that simulate entire societies.

AI AGENTS: FROM CODING ASSISTANTS TO SOCIETY SIMULATORS

Karpathy has moved beyond his viral "vibe coding" phrase to introduce "agentic engineering." The evolution tells a bigger story about where AI development is heading.

Vibe coding was mostly weekend fun—engineers experimenting with early AI tools to build casual projects. Agentic engineering represents something more mature. Engineers now direct and oversee agents that write code, rather than coding directly themselves.

"There is an art and science and expertise to it," Karpathy explains. The shift reflects how large language models have improved so dramatically that professional developers now use them routinely.

Here's what's interesting—this isn't just theoretical. Cursor raised two point three billion dollars last November. Lovable hit a six point six billion valuation. Replit is reportedly nearing a nine billion dollar round.

The market is betting that human-AI collaboration in coding isn't a fad. It's the new normal.

SIMULATING HUMAN BEHAVIOR AT UNPRECEDENTED SCALE

But Karpathy's latest move goes beyond coding. He's backing Simile, a Stanford spinout that emerged from stealth with one hundred million in funding. Their goal? Simulate human behavior at population scale.

Fei-Fei Li joins Karpathy as an investor. Both have deep ties to Stanford's founding team, including researchers who created Smallville—a twenty twenty-three experiment where twenty-five AI agents interacted in a virtual world.

Now they're scaling that concept for business. CVS uses Simile for simulated focus groups. Gallup builds digital polling panels. For earnings calls, the platform predicts about eighty percent of analyst questions.

The vision extends far beyond corporate applications. Simile wants to model the real-world effects of major decisions—from public policy to product launches—across virtual populations that mirror human behavior.

"Why simulate one person when you could try to simulate a population?" Karpathy asks. Large language models are trained on data from vast numbers of people anyway. Why not leverage that statistical power?

This connects to Li's own work at World Labs, her billion-dollar startup building three-dimensional digital environments. Both companies represent a shift toward AI that doesn't just process information—it models complex systems and behaviors.

The implications stretch beyond business forecasting. If AI can accurately simulate how populations respond to policy changes or product launches, it could transform decision-making across industries. The question isn't whether this technology will arrive—it's how quickly we'll learn to use it responsibly.