[BSOhealth]

As a parent, I think people overestimate the average quality of pre-AI children’s content.

Also interesting to look beyond children’s work at the incredible amount of “book mill” content that has dominated publishing for hundreds of years.

We rightly celebrate the good ones, but most content before AI was not good. So not surprising that the AI trained on that corpus is of similar quality.

That being said, I think this is just an opportunity to improve AI content, which is a human/computer interface design challenge. This technology is here to stay. Our focus should be on detection and improving the LLMs.

I’ve had luck formalizing this into some post-LLM rules to clean crappy default AI content before I work with it: https://slopwash.com

[gus_massa]

I remember ~2020, when youtube have a lot of "10 white car fall down a ramp one by one, get paint in different color an a lot of stars appear", with the same annoying music and sound effects. Replace cars with balls, with plastic ducks, with bananas, with panes, ... Also "Baby something, baby something where are you?". My daughter used to love them. [And the infamous babyshark [1], but babyshark has a higher level.]

[1] turututuue