Mind Readings: AGI Part 2: The Path to AGI – Where We Are and How We Get There

In this part, part two of our series on artificial general intelligence, we're going to talk about the path to artificial general intelligence or AGI, where things are and where we could be going. Now, to recap, real quick, artificial general intelligence means machines that can do any general task without being specifically built for it. So chat GPT, for example, is a tool that does language. And to some degree now it does things like images and stuff. And you have these multimodal models like Gemini and things that can do images and process video and audio.

But they're not going to fix a leaky toilet for you. General intelligence is being able to tackle any problem you throw at it in the same way that you humans do, you and I do. So let's talk about sort of five layers of how you get to artificial general intelligence. The first layer, well, the layer zero would be no AI, right? Just computers, spreadsheets, where we've been for the last 50 years.

Level one is AI that it has a very narrow focus. Specific single purpose tools. The AI that has been in place in your inbox for the last 25 years, helping you decide spam or not spam, that's an example of very narrow AI. It's good at one thing, can't do anything else. And so that's your first rung on the ladder towards artificial general intelligence.

Single purpose AI. Second level is general within a specific domain. So we're talking about chat GPT. You can ask it to write a poem. You can ask it to write uh a term paper, you can ask it to write an academic study.

We're still talking about generation, generative, but we're talking about things where you don't have to necessarily pre-train it just to do the specific task you want it to do. If you want it to rephrase a memo, you don't have to spend an hour and a half giving it examples of memos and how to rewrite them properly. It's baked into the model's knowledge. And so this AI is sort of a consultant is sort of level two on the on the ladder towards general intelligence. Level three is things like machines as collaborators, right?

Where they can do some of the tasks that you give them potentially somewhat autonomously. So this would be an example where you have an an AI agent in your Slack or your Discord for your company, and it's just doing stuff for you. It attends a meeting, it takes notes, it sends you the notes, it puts things on your calendar. These are not intellectually rigorous tasks, but they are collaborative. They are they're assistants, right?

They so you go from a uh a passive consultant that's just sitting there waiting to do something, waiting for you to prompt it, to an assistant that can actually take action and do stuff. Um level four is when you start getting into AI expertise, we start to get to human and beyond human levels of intelligence. An example of this would be where you have, for example, a you go to the doctor, and the the machine ingests your medical records and your scans and stuff, and it provides you with sort of that first-level diagnosis, like, hey, it looks like your cholesterol is a little high, buddy. You know, lay off the the quarter pounders, and then the human essentially confirms that says, Yep, that's that's that's a sensible conclusion. So the AI is acting as an expert semi-autonomously.

It's it's it's in place to take in a bunch of very generalized situations, like, hey, you've got this, you've got some high cholesterol and a genetic history of this and things, and it's at the point where it's gonna perform better than the human. To a degree, we're we're at that point today, to a degree for stuff that requires minimal competency. So summarizing meeting notes, right? If you were to look at the spectrum of executive assistance and administrative assistance from day one to 30 year veteran, somewhere in the middle, people have been in the profession, you know, five years, ten years, whatever, they're good at what they do. Machines today are at that point where they can do the work as well as an average person.

They can't do the work as well as that 30 year expert who just anticipates and knows this is exactly what's needed at this moment. They've got the life experience. They can do the job way better than that person who just graduated college. They've got their English degree, and they're like, I don't even know how to operate this phone system. So that's where tools today, like some of the transcription software is can do these these summarizations.

But when you get to that level four, that expertise, you're talking about things that are beyond human capabilities today, because machines can just hold more in their memories than we can. If you think about Google's Gemini, the current Gemini 1.5 model, as of the date of this recording, it has a two million token context window. Put in practical terms, it can hold two of these in one prompt, right? You and I can't do this. We cannot recite word for word the entirety of William Shakespeare accurately in seconds.

Just can't do it. But machines can hold this much information. And so, again, this sort of scale or these this ladder of towards artificial general intelligence is is how things are likely to play out. And the top level is truly autonomous artificial superintelligence, machines that are self-directed, machines that have understand the world, machines that could be fully autonomous. You say, hey, here is this computer network.

Optimize it. Optimize for maximum performance. And it decides what maximum performance means. It looks at all the data and starts making changes on its own. We're not there yet.

We are not there yet. And we will not be there for quite some time because just the compute power alone would be insurmountable as an issue to deal with that. Now, how do we get there? What are the technologies? Well, a lot of the stuff that the tooling we need exists to say, we don't have the compute power.

We have deep learning and reinforcement learning with human feedback and all and symbolic artificial intelligence. We have all these fancy technologies that essentially allow a machine to learn. You give it an example, it learns from the example, it can do a better job with the example. That's all of AI. It's all of machine learning, hence the name machine learning.

But we're running into scale issues. The amount of power that AI consumes is an issue. The amount of memory that a video card can hold to do inference is an issue. And so it's going to be very challenging to overcome those issues, but if we can, with the variety of technologies that are available, we will be able to make our way up the ladder. So we are at level, we're at level two today, right?

So level zero is no AI. Level one is narrow purpose tools, level two is broader tools within a domain, chat GPT, Anthropic Claude, et cetera. We are not at level three yet. Those collaborators, those those semi-autonomous agents, we're not there yet. We're getting there.

And level three is probably achievable in the next year or two as models get more sophisticated. Level four of AI superhuman experts, we're not there yet. It's gonna be a long time. Years if not decades. Uh level five, decades, probably for a truly autonomous self-directed machine.

We're a long way away from them. So in part three, we're gonna talk about what we can expect as we start moving up this ladder, as we start getting towards collaborative AI and then expert AI. So stick around for that. We'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.