> This is written of a future. Not the future, but a future – one of many possibilities.

reminded me of the opening line from Tom Scott's The Artificial Intelligence That Deleted A Century https://www.youtube.com/watch?v=-JlxuQ7tPgQ

> This is a YouTube video from a future. Not the future, just a future.

Not sure if intentional or if you're both referencing something else, fun parallel nonetheless.

A second fallacy is that AI art is public. Not true, most of it is seen exactly one time by human eyes. It is a private affair, and doesn't need to be perfect if it captures the intent of the user. Art made for one.

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

Go see live music downtown. AI is only a threat to corporate art (which never was art, ha ha).

I’ve not been buying the fear-mongering. Maybe someone can convince me I’m mistaken, but this piece did not.

Maybe long division wasn't a particularly necessary skill (though I think learning arithmetic algorithms is beneficial). But calculators could only complete a small subset of school assignments, while the set of traditional school assignments that gen AI cannot complete is getting smaller and smaller.

On the other hand, the game of "turn in assignments until you have enough grade letters to get a piece of paper" game seems to be over, at least in its recent form.

This was right as generative AI was coming out and back when it would struggle with this type of analysis.

The first analysis was done the old-fashioned way – you had to understand what you were looking at, you had to know what to do with the data, and had to be willing to work through uncertainty. The assignment: take this raw data and make recommendations to the channel owner on what to do (left open for interpretation).

Results? - 30% of submissions were terrible. - 60% were good, though not really complete. - 10% were amazing, producing creative recommendations, and better than I expected.

The semester after that I did something I never did before: I assigned the same thing. The same data and instructions, but this time with the requirement to use AI in the analysis.

Results? - Much of the worst work decreased. Only 5% of submissions were terrible. - 90%+ were good. - But at most 5% were amazing. AI usage had mostly eliminated the worst results (30% in the first attempt), but it had also hurt the top output (the 10%), with results condensed toward the center. In some cases this might be the outcome you want, but in other situations you might want to take some terrible results so that you can also benefit from the excellent ones.

Or be disciplined enough not to let AI completely think for you.

This doesn’t quite land because free is nothing special. The Internet has lots of free content that nobody cares about. There needs to be something other than “it’s cheaper” that attracts people’s interest.

One reason AI art is becoming popular is that there are tools that make it easy and fun to make your own. And I expect that the tools will get better.

I suppose it’s true that being able to take selfies quickly and easily competes with professional photography. But it doesn’t feel like direct competition when you wouldn’t have considered hiring a photographer or a portrait artist.

oof