A few reasons:
-2.5 years is a pretty short time for a new tech development, even if it fails eventually
- usually when a new tech is introduced, the biggest gains happen when the environment is changed to fit it. That takes time: libraries, api's, verification tooling, rl environments, skilling users, etc.
- possibility of orders of magnitude hardware cost reduction - Optical. Analog. Rram. Many others. Something will work. And internal improvements in the model architecture. And there's scaling in reasoning time.
Exactly, I genuinely believe we are far off from future capabilities, as in models will practically "zeroshot" most use cases without mistakes and it's likely we will directly develop in some sort of language made specifically for AI that is then translated into WASM or the likes which will drastically improve accuracy, there is no point using "human languages" when we will not even be checking codebases (not even filetree) anymore.
IDK. i played with Fable. It's coding abilities are really impressive.