Hacker News Reader

Comment by willsmith72 | Hacker News Reader

I still don't really know how they're measuring AGI. They all seem very sure about it, does someone have a good definition?

OpenAI just say "AI systems that are generally smarter than humans". That's not measurable.

What will it look like when AGI is achieved?

View on HN

When people start running out of tests to benchmark it against humans. But that won't be as scary as when the agent is given (by humans) the ability to execute tasks outside of a sandbox. A lot of dumb animals are scary, after all.

OpenAI has a short (though difficult to evaluate) definition in their charter: "OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity." https://openai.com/charter

Something that can consider a variety of goals and take practical steps to achieve them.

The goals have to be similarly variable to humans, not just 3 or 4 types of goals. If it only supports a select few types of goals, it is not very general.

The steps it takes have to be similarly practical to those taken by humans, if the steps are just a random walk it is not very intelligent.

It will be well beyond general purpose and well beyond human level and there will be an inflection point where most people finally accept it and then everyone panics and completely overreacts but also possibly too late.

DeepMind wrote a paper categorising AGI: https://arxiv.org/pdf/2311.02462.pdf

Recursive self-improvement.

Strange definition. Humans can't recursively self improve, at least not in the way I'm assuming you to mean. That's more like definition of the singularity

Well, humans aren't machines. Why would a definition of AGI need to apply to humans? On the other hand, as we gain the ability to edit our DNA, I think recursive self-improvement over generations is on the table for our species.

I guess it would be useful to have a definition of weak AGI, but after reading Bostrom's Superintelligence, I struggle to imagine an AGI without a singularity or intelligence explosion. It seems like wishful thinking.

Everyone seems to make up their own bizarre and personal definition what involves Godlike abilities then claim everyone obviously agrees with it.

For me I think gpt4 is clearly generalized intelligence.

Reply on news.ycombinator.com