Comments on “David Chapman, artificial intelligence, and AI risk”

For what it's worth,

Ian Stewart 2023-02-18

I still liked “Unsafe At Any FLOPs” :-)

Unsafe At Any FLOPS

David Chapman 2023-02-18

I did too :)

“Better without AI” is bland but it has no specific defects. All the other candidate titles had something actually wrong with them.

“Unsafe at any FLOPS” has two things wrong with it. Or maybe three! Many people wouldn’t get the reference; others would find it cheezy; and most people don’t recognize the word FLOPS.

Wrong link?

Kenny 2023-02-21

You wrote:

I concluded that, once again, backprop had fooled researchers into thinking it was doing something it wasn’t. I summarize that analysis in Are language models Scary.

Should that link be to something else? Or do you intend to, eventually, summarize your 2014 analysis on that page?

I think

Ian Stewart 2023-02-24

the only reason I get the Ralph Nader reference is because my grandfather was a car buff who lived through the era. I don’t remember Nader bringing it up too much during his Green Party presidential campaigns, as the Corvair was ancient history by then.

More to the topic at hand, maybe a “Corvan Greenbrier maximizer” wouldn’t be so bad? We’d just have to remember to avoid going around corners very fast.

https://en.wikipedia.org/wiki/Chevrolet_Greenbrier

How backprop fooled vision researchers

David Chapman 2023-02-25

Hi, Kenny, I’ve posted that bit, and fixed the link.

To save you going back to the page and following the link, it is to:

https://betterwithout.ai/understanding-AI-systems#AI-surface-features