Hacker News Reader

Comment by mdp2021 | Hacker News Reader

mdp2021

They use it like a marketing term, but it is unnecessary.

I am not aware of a complete operational definition, but it would not be that difficult to build a good tentative one. I have mentioned many different requirements several times.

One non-abstract definition is in the "10 points" here: https://garymarcus.substack.com/p/where-will-ai-be-at-the-en...

Edit: "definition" (the 10 points above) that is cannot be that of "AGI", really. AGI just needs to actually implement Intelligence and requires very many features that are not there today. The list from Marcus is about "Artificial Super General Intelligence" - not the same thing.

You can implement a replacement for Intellligence - that is AI;

you can implement Intelligence - that is AGI;

you can implement Intelligence of very high value output: that is "ASGI" (or similar).

View on HN

From the 10 tasks of Gary Marcus' list:

"With little or no human involvement, come up with paradigm-shifting, Nobel-caliber scientific discoveries."

Indeed far above the capabilities of almost all humans. I agree we should have a separate category for this. "ASGI" is still a bit unintuitive, imho. Maybe Artificial Excellent Intelligence.

With AGI explicitly meaning matching average humans, I would say that we really have already reached AGI. I can't think of any intelligence-only task where I'd think an average human would do significantly better than frontier AI provided with correct context.

> "ASGI" is still a bit unintuitive

It seems that "ASI" ("Artificial Super Intelligence") is in use, maybe it will get further traction. It should imply AGI.

> AGI ... matching average humans

Very bad definition. AGI has to implement Intelligence. And to be Intelligent it has to contain the missing pieces. Intelligence is not there - the implementation is unfinished.

> It seems that "ASI" ("Artificial Super Intelligence") is in use, maybe it will get further traction. It should imply AGI.

That's why I suggested AEI.

> Very bad definition. AGI has to implement Intelligence. And to be Intelligent it has to contain the missing pieces.

Are you saying that the average human is missing pieces and thus can't be considered intelligent?

> the average human is missing pieces

No, I was saying that current technologies with emergent simil-intelligent behaviour miss properties that are necessary to have proper Intelligence.

The average human does have atrophied abilities, compared of course to proficient humans. And in fact, we certainly do not aim to simulate average humans. We aim to simulate or actually implement Intelligence.

I see. For which intelligence only tasks do you think an average human performs better than current frontier AI (if any)?

??? This question makes no sense. We utterly disregard the "average human". There is no comparison (no competition) between the "average human" and an accomplished human. This makes using it as a parameter nonsensical.

A human with an average IQ then. Please try to go with the spirit of the question rather than trying to find some escape from it.

Well, explain said "spirit of the question" then, because "human with an average IQ" gets the same reply as above (49606153).

Justify why we should have any consideration for the average human etc. in comparison to software systems - I have written we disregard it as a nonsensical parameter. "Does Stable Diffusion paint better than John's aunt": we cannot care.

It's a very simple experiment: Create a set of intelligence-only tasks with answers that can be easily graded for correctness, then instruct one or more humans with an IQ of about 100 to do the tasks, then compare the performance to that of frontier AI on the same tasks.

We should care because "John's aunt" might rely on her paycheck as a professional painter to pay her bills. Many average humans have average jobs, some of which require intelligence (or creative production) only. Ignoring this reality is folly.

> Ignoring this reality

is part of other concerns.

So you meant: "we are already at a point in which many humans can be non-competitive in the professional market". Yes. We were still there some time ago already.

I did not mean that, but I'll take it as agreement that we've already achieved AGI if that is defined as being on average more intelligent than average humans.

If you wanted to be sure and date that: it happened around the end of 2024 and the beginning of 2025 - check Maxim Lott's trackingai.org .

But, again (and again), that which you propose cannot be a productive definition of AGI. In engineering we strive to understand "Fast", then to implement it, and then obtain "faster than a cheetah" - not "faster than a roadkill", which is really not a goal.

> more intelligent than average humans

Adequate language and ideas, please: one probably does not have actual intelligence, the other probably does not use it.

Reply on news.ycombinator.com