The stat that got me: 2012 image nets had 60 million params, now we argue about billions
I was reading an old blog post from Andrej Karpathy on my lunch break last week, the one about the ImageNet days, and he mentioned that AlexNet back in 2012 had about 60 million parameters. I had to read that line like three times. My phone probably has more going on in its camera app now, and we're out here in 2024 acting like a model with 70 billion params is just a normal Tuesday download. I remember when 8 gigs of VRAM felt like a flex, and now half the threads on here are people asking if their 4090 can handle the latest release. It's wild how fast the goalposts moved, and I don't think most of us even noticed it happening. Anyone else have a moment where an old benchmark or paper just completely reset how they think about where this stuff started?