logoalt Hacker News

nomel • today at 8:34 PM • 0 replies • view on HN

> To me (and most people?) AGI is simply an AI system that is reasonably capable of general work.

I think the ability to learn, in a "modify the latent space" sense (weight change) rather than a "point to a different existing latent space" (context result), is part of most people's definition.

But, even then, you could claim these models are doing that, just in very slow time scales. Human somewhat do it awake, but mostly do it while sleeping. LLM do it between version releases, while dreaming about past user conversations. ;) I've seen some people claim that they meet some definition of "sentience" during these training sessions, because of this, which probably has some merit.

Regardless, it's all clearly some scale, rather than a binary. The proof is in the pudding, so it seems silly to argue about any of this. We'll know it's far past AGI, meeting everyones mushy, opinionated, definitions, when these conversations stop happening.