Hacker News Reader

> New models can be trained completely on generated data.

How does that account for all the things that change in the world, but in ways only humans can observe?

How can AI discover that a beloved tourist destination has turned to crap, or that the best vacuum cleaner of 2022 has a new challenger, or that German tipping culture is shifting, or that the café down the road has great banana bread but is a little loud on Saturdays?

View on HN

The same way most humans do, from internet. Generated data can be the result of processing yesterday's new by old model. It can be multiplied, repeated from different angles. This will make it more likely to stick in new model. But the best way is to add latest data to some storage which can act as a long term memory. In this case even old model will look fresh and up to date. I'm sure we'll get it soon. RAG can be considered as a primitive form of it.

Multiplication requires something to multiply. My point again is that if you destroy any incentives to put useful things on the internet, we'll have nothing to train AI on.

Reply on news.ycombinator.com