Robert McNeesJul 18, 2023

If anyone is wondering how google is doing, it is giving incorrect answers to the query “How old is the Universe?”

Instead of serving up scientific consensus (just shy of 14 billion years) it is latching onto recent media coverage of a questionable study (tired light, time-dependent coupling constants) claiming a much larger figure.

Notably, it gives me the right answer from an incognito window. But elevating popularity metrics over scientific consensus is a real problem!

Show threadMarc Etienne

@mcnees I think the problem is that AI models are rewarded solely with approval, which spirals into telling us what we want to hear, not what is correct.

And this leads to the problem that language model apps have no way to test, to verify. They are not connected to sensors, not even in a secondary sense. All they can do is make statements and see which answers get voted up or down.

Jul 18, 2023 at 7:25pmWeb