We have some escape valves from that future though - one key assumption in the evolution of that future is:
> Because creatives had excluded themselves, their values, and their opinions from model’s knowledge, they found themselves disregarded by history. LLMs are the main way of education and learning and have lost the ability to educate people on human-created art.
Prior to the pushback from creatives (indeed, many prior to the digital age, even), many thousands of books were created which describe and discuss human-created art, including how to make it, how to compare artworks, and what aspects of art we value. The digitisation of these works from the last few hundred years is likely to have been incorporated into the training data for the LLMs.
LLMs know how to manufacture [0] and play [1] a violin, even if they don't include any new concertos. They know how to teach yourself sketching [2], even if no new sketches are added to the training data post 2020. Many other artistic and creative endeavours have been documented over the last few hundred years.
It's true that the state of the art in the early-mid-21st century may be poorly captured in the training data. That people may lose the ability to learn how to make LoFi, Hyperpop, or Seapunk music, because they're "too new" to be in the training corpus, but how to make music won't be excised from the data.
Similarly, the idea that people may never learn how to use old tech like Roland 808s because of AI is also unlikely - there are undoubtedly enough 808 tutorials which will be incorporated into the LLMs prior to the current AI disillusionment amongst creatives preventing the creation of more. So if a future creative finds an ancient 808 in a flea market, they will be able to find out enough about it to repair it and play it. There may not be as much information about newer tech like the Roland TR-8S, but that doesn't mean that the future music production nerds will have nothing to start from.
A weird coincidence: I had a copy of [0], the violin making book. I'd rate it as being sufficient to make a passable violin for an amateur or fiddler, if that person had already seen and played violins. But nothing close to a high quality contemporary instrument.
Also, the violin-making art is evolving over time, and it's speculated that designs will have to change in response to the disappearance of old growth timber. A rumored consequence of global warming is that trees grow faster, changing the grain structure of the wood. There's been great progress in making bows from carbon fiber.
It strikes me as a bit weird that violin playing could be taught from a book, especially Suzuki 1 (both my kids used it, and it's great). Again, theories about how the violin is played have changed since that book was written. For instance, there has been a growing emphasis on ergonomics and physical health, that was glossed over when I learned as a kid.
Of course these are just two books. It's not unthinkable that the knowledge from multiple books could be combined with recordings, videos, and even measurements of known violins if things like 3d imaging become cheap.
Although, the limitation on what is available to use might still exist: a future creative may be able to find out how to repair and play an 808, but they may not be able to acquire the relevant parts, because the manufacturing of such things went out of production...
We have some escape valves from that future though - one key assumption in the evolution of that future is:
> Because creatives had excluded themselves, their values, and their opinions from model’s knowledge, they found themselves disregarded by history. LLMs are the main way of education and learning and have lost the ability to educate people on human-created art.
Prior to the pushback from creatives (indeed, many prior to the digital age, even), many thousands of books were created which describe and discuss human-created art, including how to make it, how to compare artworks, and what aspects of art we value. The digitisation of these works from the last few hundred years is likely to have been incorporated into the training data for the LLMs.
LLMs know how to manufacture [0] and play [1] a violin, even if they don't include any new concertos. They know how to teach yourself sketching [2], even if no new sketches are added to the training data post 2020. Many other artistic and creative endeavours have been documented over the last few hundred years.
It's true that the state of the art in the early-mid-21st century may be poorly captured in the training data. That people may lose the ability to learn how to make LoFi, Hyperpop, or Seapunk music, because they're "too new" to be in the training corpus, but how to make music won't be excised from the data.
Similarly, the idea that people may never learn how to use old tech like Roland 808s because of AI is also unlikely - there are undoubtedly enough 808 tutorials which will be incorporated into the LLMs prior to the current AI disillusionment amongst creatives preventing the creation of more. So if a future creative finds an ancient 808 in a flea market, they will be able to find out enough about it to repair it and play it. There may not be as much information about newer tech like the Roland TR-8S, but that doesn't mean that the future music production nerds will have nothing to start from.
--
[0] Broadhouse, John; Bull, Cie Bornemann, "How to make a violin. And Violin notes by Ole Bull", 1900 https://archive.org/details/howtomakeviolinv00broa
[1] Suzuki, Shinichi, "Suzuki violin school : violin part, volume 1", 1978 https://archive.org/details/suzukiviolinscho00suzu
[2] Crawshaw, Alwyn, "Learn to sketch", 1983 https://archive.org/details/learntosketch0000craw_o0h9