Fun Times With ChatGPT-Powered Bing Search ➶

Simon Willison:

But… they make things up. And that’s not a current bug that can be easily fixed in the future: it’s fundamental to how a language model works.

The only thing these models know how to do is to complete a sentence in a statistically likely way. They have no concept of “truth”—they just know that “The first man on the moon was… ” should be completed with “Neil Armstrong” while “Twinkle twinkle … ” should be completed with “little star” (example from this excellent paper by Murray Shanahan).

I’d love to hear from expert AI researchers on this. My hunch as an enthusiastic amateur is that a language model on its own is not enough to build a reliable AI-assisted search engine.

# Thursday, 16 February 2023

Prior entry:
Next entry: