The problem has always been that every token is suspect.

It's the whole answer being correct that's the important thing, and if you compare GPT 3 vs where we are today only 5 years later the progress in accuracy, knowledge and intelligence is jaw dropping.