Software Is Changing Again

A must-watch. This recent talk by Andrej Karpathy really reshaped my understanding of how AI not only reimagines software, but also redefines the very notion of what software is. Here are a few takeaways that struck me.

After decades of hand-coded software (1.0), two rapid shifts redefined the field: first, ML models as weights (software 2.0), then prompts as programmable interfaces (software 3.0)—as Andrej captured it early on, pointing out how English was becoming "the hottest new programming language" in the age of prompts. Just like software 2.0 replaced decades of deterministic software, software 3.0 is likely to do the same across the industry.

LLMs as the new OS—an idea I've long believed in, here powerfully illustrated, hinting at a coming interface revolution and a new era of personal computing. Andrej refutes the simplistic view of LLMs as mere commodities, highlighting instead their nature as sophisticated software systems central to emerging technological ecosystems.

LLMs are flipping the script on technology diffusion—while seeing AI as "the new electricity," pointing out its ascending adoption from consumers to corporations and governments, unlike past disruptive technologies.

There's also an emerging "psychology" specific to stochastic AI systems, with striking cognitive contrasts—often brilliant, fast and broad, but still lossy and unreliable in many unexpected ways.

The latter paradigm shift is the most worrying part to me, especially when we think of how much the digital sphere ultimately redefines real-world interactions. "First we shape our tools, thereafter our tools shape us," as someone once said.

As AI increasingly becomes the OS of our systems, our companies, and even our relationships and minds, there is a growing risk of bypassing the guardrails that help us keep control—the "autonomy slider" Andrej refers to. I take note of his advice to keep the AI on a leash, advocating "partial autonomy" products and human-above-the-loop UIs. This is critical for our software—and even more so for us.