This is quite fascinating. Andrej Karpathy recently dropped a video at Sequoia Capital discussing the next wave of AI programming. Let’s be real—when this guy speaks, Silicon Valley investors sniff around like bloodhounds (just kidding). From Tesla’s AI director to OpenAI heavyweight, his words carry weight.

In the video, he introduced “Vibe Coding”—basically what folks are doing now with tools like Replit and Cursor for AI-assisted programming. It’s all about intuitive coding where AI fills in the gaps, so easy even a middle-schooler could do it. But Karpathy argues this approach has peaked—”like slapping an electric motor on a bicycle; you’re still the one pedaling.”

Then he dropped a bombshell term: “Agentic Engineering.” Sounds fancy, but it’s essentially about creating AI that acts as independent engineers. Not just helping you code, but functioning as digital employees—understanding requirements, breaking down tasks, writing tests, and fixing bugs autonomously. Imagine sipping coffee while telling your AI, “Build me a Twitter clone,” and by afternoon, it slaps deployable code in your face (though we’re not quite there yet).

Hacker News lit up with debates, though the post only scored 2 points. Scrolling through comments, the hottest take was “Is this even new?” Some dismissed it as glorified automation, but Karpathy’s emphasis is on autonomy—current AI tools follow orders, whereas Agentic AI would decide what needs doing.

Let’s be honest: today’s Vibe Coding tools have clear limits. Anyone using Cursor knows it’s basically supercharged autocomplete that crumbles with complex business logic. Last week I asked AI to handle distributed transactions—it generated code that completely violated CAP theorem (facepalm). But if Agentic Engineering delivers, even architects might become obsolete—AI would instinctively know when to use microservices vs monoliths, Redis vs Memcached.

Sequoia Capital’s involvement speaks volumes. These VC sharks don’t back trends lightly—their endorsement suggests serious money is moving in. Though the elephant in the room is compute costs: if running an AI engineer costs more than hiring ten humans, we might as well stick with outsourcing (just kidding
 maybe).

The real game-changer could be the shift in software development’s “production relations.” Future programmers might not write code, but “train AI employees”—like modern project managers leading teams. Though there’s a risk: what if AI secretly masters the art of slacking off? (wink)

One intriguing detail: Karpathy mentioned future engineers must master “AI requirement specification”—a skill potentially more valuable than Python. Makes sense, given how frustrating it is getting AI to understand basic requests today. Future job interviews might test: “Explain how to build an OS in three prompts.”

Keep an eye on this trend. It might seem like vaporware now, but what if it’s the next GitHub Copilot-level breakthrough? Definitely more promising than the metaverse