/sci/ - Science & Math

>>15072158
Machine Learning is basically one big parlor trick. It's good at what it does, which is extending text. All ML-based AIs are text extenders. All modern AI models are ML based.

ML comes up against hard foundational limits any time it has to do thinking of its own instead of just guessing what it thinks the next word should be.

Look at how GPT3 is basically a wikipedia dispenser.

Look at how DALL-E 2 is only good at making approximation drawings.

Want GPT to think for itself, have thoughts of its own, be able to reason and logic on its own, start conversations on its own? That is never going to happen with ML. The only thing ML can continue to do is scale, aka feed it more garbage from the internet so that it becomes a better parrot. A better text extender.

Want DALL-E to stop drawing fucked up faces, deformed appendages and nonsensical details? It will never happen with machine learning. The only thing DALL-E can continue to do is steal from other artists and roughly copy their style. ML DALL-E will never be able to make truly detailed drawings without all of the fucked up imperfections that you see now.

Again, Machine Learning is a neat parlor trick but it's already reaching its limit. There is another AI Winter on the horizon.