Hacker News Reader

It does once you have a human level AGI, it should be trivial to scale it up to a superhuman level.

I'm wondering if scaling up is trivial. Ofc, it depends on how much computational resources a working AGI needs. And if at that point they are capable of optimizing themselves further. Or optimizing production of more resources.

Still, scaling up might not be simple if we look at all the human resources currently poured in software and hardware.

You can task the human-level AGI to optimize itself. You can create an army of copies of itself to work towards that. It's self fulfilling.

Most models take orders of magnitude more compute to train than to run.

So you could pretty easily duplicate the result a couple order of magnitude times.

Assuming it does not get stuck at some local maximum.

Reply on news.ycombinator.com