Earlier this year I did a presentation discussing AGI, what it means and where we are currently at. Personally I’m an AI timeline sceptic. I believe we will reach AI with pretty much any definition given enough time, but the wild predictions from Sam Altman and others like him suffer from a lack of agreed definition of what AGI means.
To start with, let’s agree on one thing. We are not creating Intelligent machines that somehow mimics ourselves. We are not after equals. We are after tools to support us in our quest to understand the world around us and ourselves, to automate tasks and to remove friction in our lives.
There are of course many tests nowadays with very strict limitations so that we can compare models, but I don’t think this necessarily equates to AGI. In the table I’ve collected some of these tests within categories. It shows, at least to me, that we need to assess across many different levels.
Let’s look at this from a more practical standpoint. Say I want the robot of the future. The one Elon Musk promised in his We Robot event (see here: )
A robot that can help out with chores, cook, be a teacher, help the kids with their homework, babysit, mow the lawn, get the groceries, be your friend!
To be able to do all those things, and more, it needs:
to be a general-purpose intelligence with the ability to learn.
ability to interpret its ever-changing surroundings and the fine motor skills to interact safely with it.
ability to reason, act on commands and perform actions both inside and outside without supervision.
at least a basic understanding of human emotions and ability to respond to them.
ability to teach subjects in an engaging way
ability to do all these things at the same speed (or faster) as a human while operating in safe way for the people and pets around it.
to be affordable within reason
It needs to score well in most of the above test categories. Interestingly the category that is least important is the “how knowledgeable is it?” None of the tasks above requires a Nobel prize winning intellect or the ability to answer all Trivial Pursuit questions. What it does need is other skills, such as ability to assess human emotion, to operate safely and have a reasoning function that can take more ethical concerns into account.
So colour me sceptic when it comes to AI timelines.
Comments