Hacker News new | past | comments | ask | show | jobs | submit login

>Obviously they can't just keep scaling up the number of parameters or working memory size.

If they manage to scale it to just a hundred times bigger, and the qualitative increase in "intelligence" is the same as that gained by the previous 100x increase in scale, that might be good enough for "real" intelligence. And 100x scaling is certainly feasible within five years, if people keep throwing resources at it.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: