Mind Readings: AGI Part 3: The Promise of AGI – What We Can Expect

Welcome back. This is part three in our series on artificial general intelligence. We've talked about what it is general intelligence, the ability to solve problems that you haven't been trained to do. We've talked about where we are from level one being narrow purpose tools to level tool being broad tools within a domain, which is where we are now, tools like chat GPT, to level three collaborative tools that are semi-autonomous to level four expert tools that can do a better job than human experts in a given domain, and then level five self-directed, fully autonomous machines, which we are very far away from. In this part, in part three, let's talk about what does the world look like if we have artificial general intelligence.

What are the things that we could see as we move up this ladder? So we're I want to be clear that we're not talking about okay, once this thing arrives in three decades, the way here's the world will look. We will be making progress along that ladder through this time. So let's talk about some of those collaborative things. When you have a tool that is general and self-directed, you can give it an overall objective.

Like here is a known type of cancer. Figure out how to kill it without hurting the cells around it. And then given that overall objective, and we're starting to see this with agentic AI today, uh, AI that can take a task, break it down into subtasks and then processes, individual tasks. We are seeing this today. Agentic AI can look at that overall objective and say, okay, well, what are the what causes cancer and so on and so forth, and let's try to break the task down into little pieces.

As we look at things like science and medicine and things, you will I would expect to be able to see progress towards setting a machine on a specific goal. Here is Hodgkin's lymphoma. Here's everything we know about it. Here's the mechanism for it. Solve it.

And let the machine start working on that to see what it can come up with with parameters like, oh, you either you can't kill the host, right? Like, yes, technically, if you kill the host of cancer solved, that's not a viable solution. So here's the rules and the parameters to that task. General intelligence means a tool doesn't necessarily need to be pre-trained in that specific task to tackle it, can look at it. Um, another example, education is a general intelligence task because every student is different.

Every student has different educational needs. A machine that can semi-autonomously do a good rigorous assessment of where a student is and where their where their educational gaps are, and then build a curriculum and serve the curriculum to that student to patch those gaps and get feedback from the education process. Like, hey, I'm supposed to be helping you with statistics, but you're still scoring in the 70s, so let's figure out new ways of teaching this to you. That's an example of general intelligence being able to improve the quality of an outcome, given the outcome and the access to the right tools and data to be able to solve those problems. Another example would be in marketing, right?

We already have companies are working really hard on the on the idea of general intelligence within marketing to say, okay, I need to advertise to this audience and I need to sell them this thing. How do we do that? We have narrow examples of this in like ad optimization where tools can just create a thousand ads all at once, test them all on the market and see which ones succeed, and use that human feedback to get smarter. But that's a very narrow task. General intelligence would mean okay, I have ads, but I also have email, I have SEO, I have mobile, I have interactive apps, I have all these different options.

How do I orchestrate all of these options together to maximize the likelihood that somebody buys something? That's an example of how what general intelligence would be able to do. Whereas today you can do that, but you, the human, have to be the one orchestrating it all. You would run an ad optimizer and an email subject line optimizer and this and that and the other thing, and then bring all the data together and have uh language model, for example, do an analysis of the data. You, as the human are still the glue in that situation.

And if we have general intelligence, you the human can step out of that, have general intelligence figure out well, here's the things that are most likely to overall optimize for this particular situation. This is where general intelligence is going in those, like I mentioned. Level three is that collaborative nature where it can start taking on more of a task. Instead of for today, for example, today we have tools like Suno that can write a song and you know, put together the music and stuff, and it's okay. It's not gonna win a Grammy anytime soon.

But it's okay. Being able a general intelligence would have more knowledge, not just of song composition, but of how human beings reacted to a song, it would have data about the performance of that song and be able to simulate and synthesize and test to come up with a hit song that actually sounds good because it has the ability to hop across domains to not only say I can I know how what melody is and I know what harmony is, and I know what the circle of fifths is, and I know what lyrics are, to saying, hey, these people on YouTube are commenting about this this piece that's similar to the piece I made. What do they have in common, and how can I take lessons from that piece over there and incorporate them into my piece? That's general intelligence. That's what a human musician would do.

A human musician would say, Well, what makes a good pop song? Well, it's gonna have this lyric structure, it's gonna have this chord progression, it's gonna have this tempo, this key, etc. Even if you're not trying to actively copy, you know, Taylor Swift, you know what works as an expert human composer and general intelligence, your general intelligence would allow you to apply that same general set of boundaries and rules to a problem. That's what general intelligence will do. There are no shortage of problems that require general intelligence because they're too big for a specific tool, and they're too big for us.

Think about climate change for a second. Climate change is a massive problem, not because of the consequences, only it is because of the consequence, but because there's so many system inputs. There's carbon dioxide, there's methane, there's sea ice, there's the Atlantic meridional overturning circulation, there is solar activity and solar minimum, solar maximum, how much energy the earth receives. There's far infrared frequencies that can broadcast heat energy out into space. There's so much information within a topic like climate change that if you were to try and solve it with your head, your head would explode.

But a general intelligence could ingest all of that at scale and come up potentially with things that you haven't thought of yet. For example, we're starting to see that with today's language models to a much smaller degree. When a court case comes out and and uh the court publishes its opinion, you can take that 500-page opinion, stuff it in a language model and say, How does this impact me? How does this impact my business? How does this impact the way I do things?

You, as the human, yeah, you could read all 500 pages, you probably couldn't recall them with precision without a lot of reading. And you would struggle to keep in mind everything that was in there. Machine doesn't have that problem. And so it can act as an expert consultant on that specific topic. A general intelligence can do that without you having to preload it.

It will be able to go and find the information itself, pull it in, and come up with these conclusions for you. So that's sort of the promise of general intelligence. If if we can get it working, and as we move up that ladder from narrow use to broad use, interactive use to autonomous use, that's that's the things that this technology should be able to do. And some of it it will be able to do in the near term. So that's gonna do it for this episode.

Stay tuned for the next one. We talk about what could go wrong. 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.