4 ms·

> As per the current terms, when OpenAI creates AGI - defined as a "highly autonomous system that outperforms humans at most economically valuable work" - Micro

> As per the current terms, when OpenAI creates AGI - defined as a "highly autonomous system that outperforms humans at most economically valuable work" - Microsoft's access to such a technology would be void. That’s their definition of AGI? What does that have to do with actual intelligence?

bbor 2y ago
It’s the so called “economic” definition. You’re absolutely right to call out that it’s not very philosophically valuable, but they were trying to be pragmatic — and it was kinda hard to imagine, anyway. Of course it’s being thrown out the window now that money’s on the line… drives one to cynicism! No wonder Altman is a paranoid prepper.
heresie-dabord 2y ago
> "highly autonomous system that outperforms humans at most economically valuable work" As you note, this has nothing to do with actual intelligence, but with the for-profit calculus. The disturbing aspect of such a definition is that economically-valuable work/activity may be grossly unsustainable and directly detrimental.
etna_ramequin 2y ago
That’s a self-defeating definition. Work that can be automated at no marginal cost is economically worthless. Edit for clarity: “worthless” as in “no value add”, its worth is only capital investment and energy input.
- kelseyfrog 2y ago
  Work that can be automated is worth exactly at the price it's sold at, not at what it costs. Besides, the marginal cost of AGI isn't zero - servers still cost electricity to run.
fixprix 2y ago
Imagine someone from 4 years ago reading this comment and being like… what?
paxys 2y ago
By that definition we invented AGI during the industrial revolution.
xelxebar 2y ago
How would you, personally, operationalize intelligence? Outperforming humans across a broad range of tasks does sound like general competence. That said, we do want to make sure the tasks are 1) not stupid, and 2) not cherry picked. Requiring outputs to generate profit is a well-established proxy for human value, even if it's imperfect, and letting the economy pick tasks makes it hard for OpenAI to cherry pick. I don't know about you, but a reliable ability to generate high value sounds like the kind of intelligence I care about. On the meta-level, I think disagreements like this about "actual intelligence" and "really conscious" do a good job of showing how out of distribution we are on our human wetware classifiers. The terms really are fuzzy and ill-suited for generating understanding of current AI model behavior, IMHO.

bbor 2y ago

It’s the so called “economic” definition. You’re absolutely right to call out that it’s not very philosophically valuable, but they were trying to be pragmatic — and it was kinda hard to imagine, anyway. Of course it’s being thrown out the window now that money’s on the line… drives one to cynicism! No wonder Altman is a paranoid prepper.

heresie-dabord 2y ago

> "highly autonomous system that outperforms humans at most economically valuable work" As you note, this has nothing to do with actual intelligence, but with the for-profit calculus. The disturbing aspect of such a definition is that economically-valuable work/activity may be grossly unsustainable and directly detrimental.

etna_ramequin 2y ago

That’s a self-defeating definition. Work that can be automated at no marginal cost is economically worthless. Edit for clarity: “worthless” as in “no value add”, its worth is only capital investment and energy input.

kelseyfrog 2y ago
Work that can be automated is worth exactly at the price it's sold at, not at what it costs. Besides, the marginal cost of AGI isn't zero - servers still cost electricity to run.

fixprix 2y ago

Imagine someone from 4 years ago reading this comment and being like… what?

paxys 2y ago

By that definition we invented AGI during the industrial revolution.

xelxebar 2y ago

How would you, personally, operationalize intelligence? Outperforming humans across a broad range of tasks does sound like general competence. That said, we do want to make sure the tasks are 1) not stupid, and 2) not cherry picked. Requiring outputs to generate profit is a well-established proxy for human value, even if it's imperfect, and letting the economy pick tasks makes it hard for OpenAI to cherry pick. I don't know about you, but a reliable ability to generate high value sounds like the kind of intelligence I care about. On the meta-level, I think disagreements like this about "actual intelligence" and "really conscious" do a good job of showing how out of distribution we are on our human wetware classifiers. The terms really are fuzzy and ill-suited for generating understanding of current AI model behavior, IMHO.