Hacker News Reader

Yep, and even this comic is now out of date, "identify a bird in the photo? easy!"

Because she got her research team and five years, and shared the results with the world.

Yup. That comic is from 2014!

Funny, because Alexnet came out in 2012 and Inception in 2014. So the problem was already more or less solved when the comic was drawn, but the DL craze hadn't reached the masses yet.

I think we all forget how nonexistent on-device inference was back then. If you had a 500mb CNN model, there was absolutely no way you were including in a mobile app.

You could maybe host it on a server behind an API, but that would be significant costs.

We are spoiled these days. I can drag an ONNX file into a handful of services and do inference for fractions of a cent.

That was no problem, because back then you only really needed a GPU for training. People who only learned about DL through transformers and LLMs might find this hard to believe, but you could comfortably run inference on AlexNet with 250mb RAM on a CPU and Inception only used like 50mb. No fancy hardware or accelerators necessary.

Reply on news.ycombinator.com