Future Timeline

https://seekingalpha.com/article/426967 ... armageddon

If only these numbers could've been achieved by NOW, we might've averted a climate catastrophe. Better late than never.

From 2014. This is the start of contemporary AI's explosion

Yuli Ban wrote: ↑Thu Jun 17, 2021 11:47 pm

What a confusing graph.

It's not the best designed by far, but I like these sorts of S-graphs because they show how new paradigms often start lesser than older, more refined paradigms.

A good example can be found in synthetic media AI.
GANs and CNNs have been creating images for years and current GANs are so good as to almost be indistinguishable from real photographs or artwork.

Transformers have only just recently begun being used for image synthesis, such as with DALL-E and CLIP. And while DALL-E specifically is really good for what it is, it's still only about as good as GANs and CNNs were circa 2017. CLIP even less so. It's closer to GANs circa 2015 during the heyday of DeepDream.

So older image synthesis methods rose sooner and are technically superior, but they're reaching their upper asymptotic limit of what they can do. Multimodal models like transformers are currently less qualitatively impressive, but will become far more impressive and capable in the future.