Human training data still currently holds a place of use, but it’s extremely time consuming to produce and frequently full of errors. Long term, future consumers and producers of information statistically speaking by volume will be AIs, not humans.
My guess PG knows this, but for whatever reasons is saying otherwise — given it’s pretty obvious; even currently, my guess is machine learning algorithms already easily consume and produce more information than humans. Yes, currently one might argue the ML is the result of humans producing it, but only matter of time before the code writes the specs that write the code that collects the data that get converted into information that’s used to repeat the process based on an objective it defines and independently evaluates.
Honestly not sure why PG is even thing on HN anymore given he rarely comments on HN, or if he does, it’s under pseudonyms: