Before generative AI, there were two gaming GPUs

Before the Web, there was a black cube. Before generative AI, there were two gaming GPUs.

In 1990, Steve Jobs' NeXT Computer was considered a commercial flop. It had sold fewer than 50,000 units. Yet in a lab in Geneva, Tim Berners-Lee sat down at one of those black cubes and wrote the software for the very first World Wide Web server.

An expensive commercial failure quietly became the backbone of the modern internet - and later, the foundation for macOS and iOS.

2012, and two cards meant for video games

Fast forward to 2012. Neural networks had been dismissed as an academic dead end for decades, because standard CPUs couldn't handle the math.

Then Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton wired together two off-the-shelf NVIDIA GTX 580 gaming cards - hardware designed for rendering video games - and used them to win the ImageNet competition by a staggering margin.

That single experiment proved that deep learning plus GPUs was the path forward. It didn't just solve computer vision. It laid the direct groundwork for the Transformer architecture, for ChatGPT, and for the entire generative AI landscape we are navigating today.

The pattern

Technological revolutions rarely start with a massive, grand product launch. More often they happen in the shadows, when someone takes existing hardware, pairs it with an ambitious idea, and builds something the original creators never anticipated.

What do you think is today's unsung technology that will power the next decade?

References: ImageNet Classification with Deep Convolutional Neural Networks (Krizhevsky, Sutskever, Hinton, 2012); Attention Is All You Need (Vaswani et al., 2017); Welch Labs, "The moment we stopped understanding AI".

← All writing