A Possible Path to ASI

and a system for pursuing them. First, on the definitions themselves, I think a big problem with AGI and ASI definitions are really around AI at all is that they're too technical. They tend to be too technical and therefore not usable. Not really useful in conversation. I think the best definition for these things needs to be something that's very human centric. It should be obvious, and I think we should use

this as a benchmark. Why should I care? We should be able to look at these definitions and know why we should care, or at least have a hint towards why we should care. And I think if we can't get that from the definition, then it's probably not a very good one. So my definition for AGI is an AI system that's able to perform most or all cognitive tasks, as well as an average US based knowledge worker from 2022.

And I say a US based knowledge worker, because most people probably won't doubt that there's some kind of base level smart at doing lots of different tasks, which is the general in AGI, right? AGI is artificial general intelligence. So it's general tasks that you do in knowledge work. And I think if someone's making, you know, a decent salary as a US based knowledge worker, aren't too many people that are are going to say that this person

doesn't have general intelligence. So we're using humans as the baseline for having true general intelligence. And I say before 2023, because that's when modern AI kicked off. And we don't want to have the definition keep shifting because humans get more and more augmented with AI. So so the bar keeps moving, right. So we want to lock that in place. ASI is a bit harder and a bit easier to

define at the same time. It's a little more intuitive because it should be super or above human, but it's also harder to think about because unlike human level generality, we've never actually seen anything that's smarter than us. So you have to actively imagine that. And I think both of these definitions here are simple enough, and it's obvious by looking at them why you should care for AGI. It could replace knowledge workers, which is going to affect

the economy massively. And for ASI you could do a whole lot more than that. So the next thing is, why do we care about AGI and ASI? Like what are they actually going to produce as output. I think the most important output, or at least the most tangible one, is invention. Like coming up with. Net new things, ideas, concepts, products, services, whatever in the same way that humans do. And whenever I think of that, I have one main question. Well,

how do humans do it? Like what is that actual methodology? And I saw a recent episode of Lex Fridman's podcast. He had an evolutionary biologist on and he was talking about during the enlightenment, there were people meeting and sharing ideas and like different shops and salons and whatever, wine bars. I'm not sure where they actually went, but they would take their ideas, they would share their ideas, and they

would try to copy each other's ideas. But sometimes they would make mistakes and those mistakes would make even better ideas. But this idea exchange is like the natural way that we had tons of innovation during the enlightenment. And this tracks for me because I've always seen innovation as like bombarding your brain like a particle accelerator with ideas from multiple sources, right? You talk with your your smart friends, you talk about cool ideas, you read a whole bunch

of books, you watch a whole bunch of videos. Whatever you do, and all these ideas like go into your brain getting bombarded by other ideas that may be different or the same or whatever, and they just kind of percolate in there and kind of reproduce in there. And then as you sleep and you dream and you think about other things and work on other things, all of a sudden you'll be like, wait a minute and you'll have like these moments where actual innovation happens. So the

idea here is really simple. Let's copy how humans do this right. How do humans do this at an individual scale? And let's use automation and AI to orchestrate and scale that process, which looks, I think, something like this. So you have your own ideas. Ideas from books, ideas from other people, ideas from wherever. And you basically put that into an idea repository. And you could look at this project right here called substrate, which I put together a

couple of years ago. And it's basically crowdsourced ideas, crowdsourced problems, crowdsourced solutions. This is a way for us to pull together ideas and solutions and problems all into a place that we can crowdsource them and see them and work on them. And most importantly, we can now hand this to AI to start thinking about them all together. Then you have this idea of an idea combination system, and

this is where you combine ideas. You vary them slightly, change them in a subtle way, add randomness, whatever, and then fold those back into the idea store. and so the list of ideas just keeps growing. And then you have the testing stuff. This testing stuff is absolutely critical. And the most difficult actually, where you actually test the ideas against the problems and you need to have a

way to experiment, right? And this is why so many startups are actually spinning up labs like material science labs or bio labs, where you can actually build molecules and test them against living tissue. Right. And you have to be able to do this. Otherwise you can't know whether or not the idea worked or not. Uh, in some cases you can in some like digital cases, you could do like a B testing or something like that, and you could say, yes, this is good enough to say

this actually worked. But in a lot of cases it's hard science, it's hard reality. You actually have to have a lab to do that. But what you do once you have all these components, the ideas, the problems, the idea combination engine and then the experimentation engine, You. Now just run through this. You iterate through this. So we have taken the human system of trying these different things, and we've sort of broken it into its components of

the scientific method. And we are scaling it with AI, with crowdsourcing and automation, you know, using pure tech to scale the crap out of an already awesome human process. And keep in mind, this is not just for like a new type of keyboard or a better car battery or something like that. The list of problems could be anything from like marketing campaigns to figuring out better ways to connect with kids who need to learn math or whatever.

We could put all of humanity's problems into these problem buckets, right? And as we get better and better ways to test them, we accelerate, right? We accelerate this entire process of automating the scientific method. So this ends up being an algorithm for solving general problems and testing them. And instead of doing it at the scale of like the few universities that we have and the few researchers that we have,

we now can do this at AI scale. And with the bottleneck really only being, you know, how much testing we actually need to do in the real world. Uh, and I'm just really excited about this because, I mean, we're talking about, I don't know, five x ten x 100 x 1000 x million x. Whatever. Our current iterations, our current, you know, attempts on goal for doing the scientific method, but just scaling that to an insane level. So I don't think this system is actually needed for

AGI or ASI, to be clear. But this chart here I think, shows how it is actually just a continuum going from bottom to top. So you go from the bottom subhuman level of general intelligence or cognitive capability. You move up through AGI and then into AC at the top. But I do think a system like this that we've talked about is a way to actually make the transition from where we are into AGI and then beyond into AC. Now, my current guess, as I've sort of captured here in

this chart for AGI is 2027. And I think that's going to instantiate as a true knowledge worker replacement agent that actually you just hire as a company. It comes in, it basically logs in and starts doing onboarding. It reads the slack messages, it reads Confluence and Google Docs and basically onboards like a regular employee. And this will be our first instance of AGI will be like a commercial project like that Um, or a commercial product like that.

And again, I think that's going to be around 2027. My original range that I gave in 2023 was 25 to 28. So I'm, you know, well within those bounds. And then for ASI, I have a lot less strong of an intuition, but I'm guessing 2028 to 2030 for ASI. And hopefully this has been helpful. Cool way to sort of think about this uh scientific method algorithm. And we'll see you next time. Unsupervised learning is produced on Hindenburg

Pro using an SM seven B microphone. A video version of the podcast is available on the Unsupervised Learning YouTube channel, and the text version with full links and notes is available at Daniel Mysa.com newsletter. We'll see you next time.