Are you jaded about the trend where as hardware gets better, software gets less efficient and more bloated? Buckle up because it's about to get a lot weirder.
As AI inference trends towards faster and cheaper, and as the models' accuracy improves, there may come a time where a computer is no longer running on code; it's running on the AI telling it what to do in real time. Now you might think, isn't that the same as AI writing code? No it's not; the AI won't need to write code anymore; it will sit between the user and low-level operations in real time, and decide immediately what pixels to show the user and what I/O operations to do based on the context and the user input.
How did I come up with this [dumbass|genius] idea you might ask. One of the holy grails of software engineering is to eliminate bugs. Obviously bugs can't be completely eliminated outright, but if you'll notice, most bugs have a clear intended behavior. The next time you encounter a software bug, ask yourself whether the bug would've happened if, instead of software controlling everything, it were a human with very good memory that could think at 1000x speed.
Won't this introduce a ton of other bugs and unexpected behavior, you might ask. Yes, that's why I agree my idea sounds ridiculous, and would only work if AI models became really good and really fast.
Can a giant neural net really replace a huge amount of code like an entire operating system? It seems like a crazy idea, but crazier things have happened in the past (like the fact that text and images can be generated coherently today would've been considered pretty crazy 10 years ago, which for some reason people now dismiss saying "of course computers can do that" just because it's not as good as a human).
We can see an extremely nascent version of a similar idea in DeepMind's Genie 3, which generates pixels directly rather than using game state as an intermediary (the game state is actually stored in the neural net itself with no manually coded logic)
No comments:
Post a Comment