The rub is what you consider the “natural” output.
If you give it a picture of a blackboard is the natural output pictures of other blackboards, is the natural output similar pictures of rooms with similar color schemes ,is it the life story of the poet who wrote the quote written on the blackboard, internet posts which include the quote ,is it the details of the famous historical event where a politician quoted a line from the same poem or is it the details of the car reg in a photo on the wall?
If you upload a photo of a screwdriver should it give you info on how/where to buy that kind, lists of different types of screwdrivers or pictures of the type of screw it’s designed to fit into?
A major problem you run into with this kind of thing is that you get so very very very many potential links. take a normal photo and there’s thousands, possibly millions of things that link to it reasonable only 1 node away and you need to not just filter but to also prioritize.
Except where it’s pretty safe to assume like with math problems you have to give some kind of hint about what kind of thing you’re looking for.
Yes. At first, you have only several relations known for such a blackboard. But the GLT updates automatically, via NN for example. Just as it was an indexing machine. Which it is.
Many paths lead from such a blackboard picture. Perhaps as much as one million or more. Perhaps a window is near this blackboard and the Saint Peter basilica in Rome is clearly visible through. Thus, a whole new line of relations is opened here. You can filter them in and out.
Did I mentioned that this table is giant? It would dwarf Google. In fact, every Google query can be only added to it. Another possible relation in the GLT. Along with the IP, date, time, OS—where and when the query has been made.
Input bit blob (string), output bit blob (string). Those kind of tuples, along with some “meta-data”, are the GLT’s (retrievable of course) members.
Spurious connections would likely be a massive headache like patterns on the wall matching patterns in the shadows of some random photos taken 2000 miles away while the handwriting style gets matched to a pair of Ukrainian schoolchildren who have never been within 3000 miles while the sentence writing style itself gets matched to an internet post by someone completely unrelated who’s never been within 5000 miles talking about potato dishes.
I get that the table is giant but it sounds almost like an expert system which you don’t ask questions but rather throw info at and hope it comes back with what you want.
Even bounded these things can be a headache. I’ve written code that tried to identify duplicate image regions between 2 images and you’d be surprised how many little sections of images that brute force searches can find matches for in others. little areas of sand, particularly generic trees, shapes in clouds, actual duplicated areas which do match but which are rotated through 27 degrees so that you can’t do a straight pixel by pixel match or which are slightly more compressed slightly less compressed.
If, for example, your system stores info about links between every image that has the same pattern of stars in it then you’d likely need more storage space than you could get by turning the earth into computronium. Exponentials are a bugger.
You hit the same problem one 1000000x worse if you’re trying to match on everything everywhere everywhen.
The rub is what you consider the “natural” output.
If you give it a picture of a blackboard is the natural output pictures of other blackboards, is the natural output similar pictures of rooms with similar color schemes ,is it the life story of the poet who wrote the quote written on the blackboard, internet posts which include the quote ,is it the details of the famous historical event where a politician quoted a line from the same poem or is it the details of the car reg in a photo on the wall?
If you upload a photo of a screwdriver should it give you info on how/where to buy that kind, lists of different types of screwdrivers or pictures of the type of screw it’s designed to fit into?
A major problem you run into with this kind of thing is that you get so very very very many potential links. take a normal photo and there’s thousands, possibly millions of things that link to it reasonable only 1 node away and you need to not just filter but to also prioritize.
Except where it’s pretty safe to assume like with math problems you have to give some kind of hint about what kind of thing you’re looking for.
Yes. At first, you have only several relations known for such a blackboard. But the GLT updates automatically, via NN for example. Just as it was an indexing machine. Which it is.
Many paths lead from such a blackboard picture. Perhaps as much as one million or more. Perhaps a window is near this blackboard and the Saint Peter basilica in Rome is clearly visible through. Thus, a whole new line of relations is opened here. You can filter them in and out.
Did I mentioned that this table is giant? It would dwarf Google. In fact, every Google query can be only added to it. Another possible relation in the GLT. Along with the IP, date, time, OS—where and when the query has been made.
Input bit blob (string), output bit blob (string). Those kind of tuples, along with some “meta-data”, are the GLT’s (retrievable of course) members.
Spurious connections would likely be a massive headache like patterns on the wall matching patterns in the shadows of some random photos taken 2000 miles away while the handwriting style gets matched to a pair of Ukrainian schoolchildren who have never been within 3000 miles while the sentence writing style itself gets matched to an internet post by someone completely unrelated who’s never been within 5000 miles talking about potato dishes.
I get that the table is giant but it sounds almost like an expert system which you don’t ask questions but rather throw info at and hope it comes back with what you want.
Even bounded these things can be a headache. I’ve written code that tried to identify duplicate image regions between 2 images and you’d be surprised how many little sections of images that brute force searches can find matches for in others. little areas of sand, particularly generic trees, shapes in clouds, actual duplicated areas which do match but which are rotated through 27 degrees so that you can’t do a straight pixel by pixel match or which are slightly more compressed slightly less compressed.
If, for example, your system stores info about links between every image that has the same pattern of stars in it then you’d likely need more storage space than you could get by turning the earth into computronium. Exponentials are a bugger.
You hit the same problem one 1000000x worse if you’re trying to match on everything everywhere everywhen.
Crazy ideas thread, isn’t it?
Still, it isn’t more crazy that Google would look like in say 1990.
Likely so, but manageable, one way or another.
That would justify 6 out of 20 zeros, wouldn’t it?
Oh I like the idea, some kind of massive expert system would be awesome.
I’m just running through some of the problems since I’ve played with some things in related domains.