the empircal evidence got put i to a paper, and it was partially funded by SpaaceX. A pdf is out there. From an informational theory point of video, language space is vast. But image space is vast too. If you just randomly type letters (27 to kedp it simple). You will fome up with an English work or two rather easily. But if you even take a 20x20 image, and dont even have colors just black or white - you will never get anything resembling a character or smiley. In fact, that 20x20 image is larger than we have molecules in the universe. So if you go for thr whole time kf the universe, you might be able to run through all permutations. There is a great video by Steve Brunton (eigensteve on TH-cam) about this topic. Language and images are highly compressible. Since only the a tiny part of those possibilites make sense. And thatd why language models work. Language models simply assing a probability to how likely a sequence of inputs it. That's what is exploited by transformers and products like ChatGPT on the highest end. And those work on tokens, not characters.
I thoroughly enjoyed this Mitch. The thought that occurs to me is that among the works of the infinitely-prolonged-typing-monkey (besides the complete works of Shakespeare) will also be a version which is very nearly correct but has a typo on page 23 of The Tempest.
@@hackmodular but seriously. If any of this is serious.... I think as a programmer, so my own solution to the 'linguistic quantiser' would be to do it in software. It would be very simple to make a microcontroller receive data as serial at whatever baud, modify it according to some rules and spit it out as serial. If you want to talk about this, get in touch.
@@shieladixon cheers Shiela. I think it will come to that eventually, but will be quite fun to see if the Heath Robinson machines can get anywhere close!
Did you ever notice the Tank Museum in Bovington is right next to Monkey World - that’s an accident waiting to happen!
Gary Larson would caption that "Trouble brewing..."
Haha 😄
the empircal evidence got put i to a paper, and it was partially funded by SpaaceX. A pdf is out there.
From an informational theory point of video, language space is vast. But image space is vast too. If you just randomly type letters (27 to kedp it simple). You will fome up with an English work or two rather easily.
But if you even take a 20x20 image, and dont even have colors just black or white - you will never get anything resembling a character or smiley. In fact, that 20x20 image is larger than we have molecules in the universe. So if you go for thr whole time kf the universe, you might be able to run through all permutations.
There is a great video by Steve Brunton (eigensteve on TH-cam) about this topic.
Language and images are highly compressible. Since only the a tiny part of those possibilites make sense. And thatd why language models work. Language models simply assing a probability to how likely a sequence of inputs it. That's what is exploited by transformers and products like ChatGPT on the highest end. And those work on tokens, not characters.
Very interesting thanks
Equivalent of a scrying mirror 🪞
Music is reaching its limits?
just getting started
I thoroughly enjoyed this Mitch. The thought that occurs to me is that among the works of the infinitely-prolonged-typing-monkey (besides the complete works of Shakespeare) will also be a version which is very nearly correct but has a typo on page 23 of The Tempest.
😂😂😂 Of course! Only 33 billion years to get that close! Amazing that’s made me chuckle
@@hackmodular but seriously. If any of this is serious.... I think as a programmer, so my own solution to the 'linguistic quantiser' would be to do it in software. It would be very simple to make a microcontroller receive data as serial at whatever baud, modify it according to some rules and spit it out as serial. If you want to talk about this, get in touch.
@@shieladixon cheers Shiela. I think it will come to that eventually, but will be quite fun to see if the Heath Robinson machines can get anywhere close!
@@hackmodular that would definitely more in keeping with your stuff