Comment by bbarnett
21 days ago
"Hey, I know! We should spend billions replacing code and data that provide the precise same output every time (or random from data we choose), with completely random, uncurated data that changes with every new model, because why not! It's awesome!", says every company now.
AI is not useful if you want curated fact, if you want consistent output, if you want repeated quality.
How about training an AI on 1990s style encyclopedias, with their low error rate.
Even wikipedia has random yahoos coming in and changing pages about the moon landing, to say it was filmed in a studio.
AI is being trained on random, it outputs random.
Yeah but at the same time if I ask you any question and ask an LLM the same one, in any subject, the LLM is going to beat you 99% of the time.
Not true, but even if so, that's not the point.
Using curated data means Widgets Inc outputs help to a customer the same always. 99% means AI tells the customer "you stink!" or "our products suck" randomly.
It's useless.
If I ask you general trivia spanning the world's knowledge and history, I guarantee an LLM will beat you. You can test it yourself.
1 reply →
Surely even if you train it on an encyclopedia, if you ask a question that isnt in the encyclopedia it'll just make something up still