This is interesting and the world models predicting the next frame is huge. I'm also wondering on how this would be applicable to neurological proceesses, thoughts etc.
If thoughts become quantifiable and measurable enough would similar models be able to predict the next one? Wondering what the impact of that would be.
Wow. This is the first take on AI that I find really shocking, and it is totally believable.
But worth noting the difference between a system you interact with and a visualization of such system, injected into your eyeball, is the lack of secondary effects (or first effects?). I suppose those can ALSO be faked and maintained, but at some point the illusion is more work than the system that would normally create it.
It seems likely that the fake system would be much more work than the code needed to generate it for most modern uses, but when we are spinning up bespoke code that will be run once, for a single user? The fake output might be more efficient.
I don't think this means this is the future. But it certainly means it could become a preferred approach for certain use cases, when there is no need for reuse.
I love finding ideas like this, ehich feel like true "AI-native" thinking.
His spreadsheet example is very telling. The difference in energy requirements between an actual spreadsheet program and an AI spreadsheet simulator would be multiple orders of magnitude (at least 6, probably more). So the spreadsheet simulator makes no sense. I bet you have the same problem with a world simulator. The energy requirement is so huge you will have to just write a program. With the help of an LLM of course.
5 comments
[ 3.7 ms ] story [ 26.1 ms ] threadThe future will be weird.
If thoughts become quantifiable and measurable enough would similar models be able to predict the next one? Wondering what the impact of that would be.
But worth noting the difference between a system you interact with and a visualization of such system, injected into your eyeball, is the lack of secondary effects (or first effects?). I suppose those can ALSO be faked and maintained, but at some point the illusion is more work than the system that would normally create it.
It seems likely that the fake system would be much more work than the code needed to generate it for most modern uses, but when we are spinning up bespoke code that will be run once, for a single user? The fake output might be more efficient.
I don't think this means this is the future. But it certainly means it could become a preferred approach for certain use cases, when there is no need for reuse.
I love finding ideas like this, ehich feel like true "AI-native" thinking.
So I'm not buying his thesis.