Sean MacLean: that's also the funniest thing about these public facing models: they're already tainting themselves. A strong model is only as good as its training data, and deciding "I dunno, as much of the internet all the time!" and then companies trying to turn that into "content" is just rapidly enshitifying

Another way to think about this is that every day ChatGPT uses a typical US family’s electrical consumption for 65 _years_. That’s the better part of a lifetime and they’re doing it to make computers less reliable.

I knew this, but holy shit anyway.

this is a big thing I think about when I see people going "but AI will get better!" Like, at what cost?

Also, no it won't; its pool of training data is now too tainted by the output of earlier LLMs.

that's also the funniest thing about these public facing models: they're already tainting themselves. A strong model is only as good as its training data, and deciding "I dunno, as much of the internet all the time!" and then companies trying to turn that into "content" is just rapidly enshitifying

Post