Mind Readings: AGI Part 1: Demystifying AGI – What It Is and What It Isn’t

In this five-part series this week, we're gonna tackle artificial general intelligence. This is a term that you've heard a lot in the news, in uh all sorts of places, and it's not really clear to a lot of people what this means. Partly because, well, the folks who make systems uh keep moving the goalposts because of various things that pretty clear that they're not gonna be able to achieve. So we're gonna tackle what artificial general intelligence is, how we get there, uh, when we'll get there, the what we can expect from it, assuming that we are able to actually create this thing, uh, the unintended consequences and risks of it, and sort of ethics governance and and what it all means in the five-part series this week. So let's get started with part one, which is what the heck is this stuff?

Why why do we why are we even talking about this? Artificial general intelligence is a sort of an umbrella term that basically means computers that have the ability to have general intelligence. What is general intelligence? It is when you approach a problem that you haven't solved before, and you may not have a lot of domain knowledge about it, but you give it your best shot, and because you have general flexible, agile intelligence, you might be able to solve it. So you are walking through the woods and you see a plant, and you're not sure whether the plant's edible or not.

So you look around, maybe you sniff it, maybe you take a little taste of it. If it's bitter, you're like, yeah, nope, bitter things generally are bad. Uh brightly colored uh things like ride reds, generally a kind of a warning. And so you would apply general reasoning to a new situation, situation you've never tackled before. You encounter a new animal, and you're like, well, it's got legs, so it's clearly walking around, doesn't look like it have wings, so it probably can't fly.

That's general intelligence. Human beings and many animals have general intelligence, the ability to deal with a wide range of situations. Think about, for example, when you are in your basement or you're in your apartment somewhere and something's not working, an appliance isn't working, uh, maybe there's a strange sound from coming somewhere. You may not be a plumber, you may not be an electrician, but you can at least get a sense of, oh, I think I know what's wrong. That's a generalization, general intelligence.

Today's AI is very narrow intelligence. It is good at a few things, really good, uh, for example, at say language, but terrible at music or not great at symbolic logic, really good at math if it's classical AI, not good at math if it's generative AI. That narrow intelligence really limits what a tool can do because it can't just adapt to new situations. It needs training, it needs prompting, it needs a lot of data to work with. We've been talking forever now about how to improve your generative AI capabilities using better and better prompts and different prompt structures and things.

With a general intelligence system, you wouldn't need to prompt it as much, right? I always say in my keynotes, imagine generative AI like ChatGPT as the world's smartest intern, right? The intern is still clueless on day one. There, uh they don't even know where the coffee machine is, and you got to tell them how to do everything. An artificial general intelligence system would be instead of hiring uh an intern with 500 PhDs, you're hiring a senior executive with 500 PhDs, where they can be a lot more autonomous, they can be a lot more adaptable, they can be uh they can apply knowledge across a bunch of different domains.

And so in employment terms, it's like let it's like laddering up. Okay, you're hiring a manager now instead of an intern. And that's that's how to think about AGI. It does not necessarily mean terminators, for example, or commander data from Star Trek. We're not talking about sentient machines, although early early conversation about AGI implied sentience.

Um that's not where these tools are going. Intelligence does not necessarily mean consciousness. There's there's a difference between the two. You can have intelligence knowledge and capability in a system that isn't alive. It's not alive, it's not aware, it has no agency.

It is unlikely anytime soon we're going to have a live, living systems, uh AGI systems, uh, just because the compute power to do that is really, really, really expensive. Um, and it's beyond what we can do today. So those are the three key characteristics of artificial general intelligence. It's general, it's flexible, it's adaptable, and it is semi-autonomous in that it can it can make some decisions without constant human supervision. If you've watched recent uh Trusted Insights live streams, we've shown how you write prompts that are semi-autonomous that can do a bunch of things without needing to be told every single step of the way, AGI systems will be able to do more than that.

Some other things that people get wrong about AGI, that whole idea of a superintelligence is pretty far away, right? Uh the idea of a machine that is so good at everything. The reason why is because so much of what is real intelligence is rooted in more than just say language or images. You and I have vision, hearing, taste, touch, and smell. We have five senses.

We have very, very complex brains that are billions and billions of neurons linked together into the into a massive parallel computer. I mean, that's the thing that we forget. This is a computer up here. This is a computer. It's an analog computer, it's biological, it's electrical, it's chemical, but it is a computer that has massive parallel processing.

And that complexity is how we have things like consciousness. Silicon computers, even though they're very, very capable, are not, they don't have the same parallelism, right? This device here can do crazy, crazy things. But it's it's compute powers like that, like an ant in terms of of that parallelism. Because to build a piece of technology that has that much parallelism, you would need giant, giant data centers the size of you know city blocks and then some.

And even then, it's you the complexity of interlocking those systems is just so it's so far away. So far away. There are pathways to get there. Things like quantum computing or bioneural computing, where you have computers that have living tissue inside them, but we're still far away from that. The whole idea of Terminator and Skynet, probably not likely.

Because again, without that agency, AGI is is just more tooling. It's it will be very flexible tooling, but still just more tooling. And there's this belief that you'll have a singularity, right? Uh this is something that Ray Kurtzwell talked about years ago of a machine that is super intelligent and takes over the world again, Skynet. Our current architecture doesn't support that.

And our current our current developments don't support that. Could it happen? Sure. In theory, once you get to a machine that has agency of its own, but we're really far away from that. So there are plenty of tools that have that are starting to get there, like chat GPT.

But in the next part, in part two, we're going to talk about sort of what that ladder of progression is, but we're we're still pretty far away. So this is part one. Artificial general intelligence is a spectrum. It is something we're trying to achieve. Machines that can do general tasks with little to no additional training very well, as well as a human being, or possibly better.

We're not there yet, and we've got a long way to go. So stay tuned for part two. Thanks for tuning in. I'll see you on the next one. If you enjoyed this video, please hit the like button.

Subscribe to my channel if you haven't already. And if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.