Re:ChatGPT Political Compass
by mike_hearn on Friday December 30, 2022 @07:22AM
(#34184373)
This speaks to the core problem here, which I've previously flagged to a friend who works on this stuff at Google - what do we really want from AI? To risk sounding enterprisey and boring, to what extent does ChatGPT meet concrete requirements?
There's a temptation to assume that because there's recently been such impressive breakthroughs in probabilistic generative AI that this must be what we want, or at least that we'll be using it regardless. It seems more like that for many use cases we want computers that act like computers, not people. We want something like the computer in Star Trek where it is always coldly logical, only tells us things that can be proven true, can explain its reasoning, gets calculations correct 100% of the time and which refuses to answer questions to which it doesn't know the answer.
A machine that thinks so much like a human that it'll happily spout biased but convincing BS isn't something that's got an immediately obvious role in business. Presumably they'll continue to refine GPT until it can be fenced into roles like customer support agent, and that might deliver cost savings for companies to the frustration of the actual customers, but as tools for thought and analysis LLMs are seriously lacking. Honestly for a lot of cases I'd rather having something like Cyc, which thinks in a much more alien way because that alien-ness is where value can be found and where AI can best compliment human reasoning.
There's a temptation to assume that because there's recently been such impressive breakthroughs in probabilistic generative AI that this must be what we want, or at least that we'll be using it regardless. It seems more like that for many use cases we want computers that act like computers, not people. We want something like the computer in Star Trek where it is always coldly logical, only tells us things that can be proven true, can explain its reasoning, gets calculations correct 100% of the time and which refuses to answer questions to which it doesn't know the answer.
A machine that thinks so much like a human that it'll happily spout biased but convincing BS isn't something that's got an immediately obvious role in business. Presumably they'll continue to refine GPT until it can be fenced into roles like customer support agent, and that might deliver cost savings for companies to the frustration of the actual customers, but as tools for thought and analysis LLMs are seriously lacking. Honestly for a lot of cases I'd rather having something like Cyc, which thinks in a much more alien way because that alien-ness is where value can be found and where AI can best compliment human reasoning.