Item — HN Reader

All the AI algorithms at the moment need to be trained with human input, e.g. Copilot works on largely human written code, GPT-3 is 60% Common Crawl[0], DALL-E uses 400 million images scraped from the internet[1], etc. I suspect weird things like feedback loops will start to happen if you train generative AI algorithms on the output of generative AI algorithms rather than human input.

Worth noting that people have had these sorts of worries for over 200 years - see e.g. the Luddites[2] and Luddite fallacy[3]. While new forms of technology do make certain forms of employment obsolete, they also introduce many new forms, e.g. who would have thought 20 years ago that there would be people making a living from being a "social media influencer", or even writing and maintaining software like Copilot.

[0] https://en.wikipedia.org/wiki/GPT-3

[1] https://en.wikipedia.org/wiki/DALL-E

[2] https://en.wikipedia.org/wiki/Luddite

[3] https://en.wikipedia.org/wiki/Technological_unemployment#The...

3 comments

It strikes me that most of this human input was taken without real consent of the original people who worked to make it. It's not exactly loom smashing to see that as kind of theft.

I was thinking a lot of John Henry today; you make a good analogy.

Is pirating movies and music also 'kind of theft' or do we have different standards for when it's our type of work vs. someone else's.

DRM and traffic shaping to stop file sharing are better examples of loom smashing.

> who would have thought 20 years ago that there would be people making a living from being a "social media influencer"

a very useful, highly sought-after and an important profession!

On the same token, 20 years ago you had a reasonably good chance of becoming a programmer if you knew HTML. Not JS, not Perl, but basic-ass HTML.

The average influencer in 2022 isn't making any more money than the guys that stand outside stores flipping a giant sign to grab the attention of passers-by.

Definitely deserving a seat reservation on Ark Fleet Ship B

Alpha go learns from self play. A google AI spontaneously invented a communications protocol. GANs work fantastically. The need for human input for training is vastly overestimated.