i gotta admit that i do find it sort of interesting that chatGPT or whatever can be, like, 40% accurate at summarizing a news article. it is interesting! it would also be interesting if a horse would do arithmetic correctly 40% of the time, but that doesn't mean I'd give the horse a bookkeeping job
the divide between "huh, this is pretty cool!" and "this is reliable enough to be trusted with important business things" is vast and in many cases uncrossable
Microsoft began using "AI" in its content filter for spam mails, and dozens of work emails I should have received started getting tossed while random nonsense spam and phishing emails flooded in. Total disaster, the IT dept just...shut off the filter entirely to keep things running
It's also distressing that it's not going to necessarily be easy to parse which part is in the the 60% and which part is in the 40% either, and you can't tell without having to research every last fact yourself.
tbh I feel like the environmental argument is super weird and only really works out the bad way if we invent sample-efficient training so that we can exhaust the corpus more slowly
it's weird that people are so worried about externalities of LLMs when it's not clear that LLMs make financial sense in the first place
cars are bad they are worth the cost to the drivers but not the cost to society
it's unclear people will pay what ChatGPT costs *to them* for ChatGPT.
This one might make more sense with a different state, fwiw.
You could run like *4* of those horses off of Oregon. Maybe 6 if you moved CloudFlare somewhere else.
I'm gonna be honest, I don't think it's achievable either as GenAI doesn't have any concept of self or anything else, but even trying to build an AI you can torture feels like maybe they shouldn't be allowed near animals or small children.
Also, negative reinforcement is terrible for children, but also not surprising that they'd think it's normal.
It's likely why they are the way they are.
in blindsight the method they try to figure out how the scramblers, an intelligent but not conscious, ‘beings’ work is by torturing them. more to see reactions than inquiry
yeah - and every so often there is a task where 40% accuracy at extremely high volumes is probably valuable. "Summarize the reviews of this app" or something.
also "it's dramatically harder to factor a very large number than it is to multiply two immensely large prime numbers together" is a way of describing why modern cryptography works, so the prime factoring button would be very useful for breaking encryption