WEBVTT - A Possible Path to ASI 0:00:00.880 --> 0:00:05.040 Unsupervised Learning is a podcast about trends and ideas in cybersecurity, 0:00:05.080 --> 0:00:09.960 national security, AI, technology and society, and how best to 0:00:10.000 --> 0:00:17.680 upgrade ourselves to be ready for what's coming. There's a 0:00:17.680 --> 0:00:20.800 ton of discussion everywhere about AGI and ASI and whether 0:00:20.800 --> 0:00:24.040 or not they're possible to achieve. I think they are. 0:00:24.520 --> 0:00:26.759 And I want to talk about one way we could 0:00:26.760 --> 0:00:31.480 possibly pursue that. So I'm going to step through definitions 0:00:31.480 --> 0:00:34.600 of AGI and ASI, why we should care about them, 0:00:34.600 --> 0:00:41.000 and a system for pursuing them. First, on the definitions themselves, 0:00:41.000 --> 0:00:44.600 I think a big problem with AGI and ASI definitions 0:00:44.640 --> 0:00:48.280 are really around AI at all is that they're too technical. 0:00:48.400 --> 0:00:51.640 They tend to be too technical and therefore not usable. 0:00:52.159 --> 0:00:55.800 Not really useful in conversation. I think the best definition 0:00:55.800 --> 0:00:59.760 for these things needs to be something that's very human centric. 0:01:00.250 --> 0:01:02.210 It should be obvious, and I think we should use 0:01:02.210 --> 0:01:05.450 this as a benchmark. Why should I care? We should 0:01:05.450 --> 0:01:08.369 be able to look at these definitions and know why 0:01:08.370 --> 0:01:10.490 we should care, or at least have a hint towards 0:01:10.490 --> 0:01:13.610 why we should care. And I think if we can't 0:01:13.610 --> 0:01:16.130 get that from the definition, then it's probably not a 0:01:16.130 --> 0:01:20.170 very good one. So my definition for AGI is an 0:01:20.170 --> 0:01:24.209 AI system that's able to perform most or all cognitive tasks, 0:01:24.530 --> 0:01:28.929 as well as an average US based knowledge worker from 2022. 0:01:29.370 --> 0:01:33.130 And I say a US based knowledge worker, because most 0:01:33.130 --> 0:01:35.730 people probably won't doubt that there's some kind of base 0:01:35.730 --> 0:01:39.770 level smart at doing lots of different tasks, which is 0:01:40.050 --> 0:01:46.370 the general in AGI, right? AGI is artificial general intelligence. 0:01:46.370 --> 0:01:50.050 So it's general tasks that you do in knowledge work. 0:01:50.370 --> 0:01:52.370 And I think if someone's making, you know, a decent 0:01:52.370 --> 0:01:56.690 salary as a US based knowledge worker, aren't too many 0:01:56.690 --> 0:02:00.110 people that are are going to say that this person 0:02:00.110 --> 0:02:04.150 doesn't have general intelligence. So we're using humans as the 0:02:04.150 --> 0:02:10.590 baseline for having true general intelligence. And I say before 2023, 0:02:11.310 --> 0:02:14.590 because that's when modern AI kicked off. And we don't 0:02:14.590 --> 0:02:17.710 want to have the definition keep shifting because humans get 0:02:17.710 --> 0:02:20.510 more and more augmented with AI. So so the bar 0:02:20.510 --> 0:02:23.350 keeps moving, right. So we want to lock that in place. 0:02:25.230 --> 0:02:27.710 ASI is a bit harder and a bit easier to 0:02:27.710 --> 0:02:30.829 define at the same time. It's a little more intuitive 0:02:30.870 --> 0:02:35.630 because it should be super or above human, but it's 0:02:35.630 --> 0:02:40.350 also harder to think about because unlike human level generality, 0:02:40.350 --> 0:02:43.910 we've never actually seen anything that's smarter than us. So 0:02:43.910 --> 0:02:48.710 you have to actively imagine that. And I think both 0:02:48.710 --> 0:02:51.790 of these definitions here are simple enough, and it's obvious 0:02:51.790 --> 0:02:55.390 by looking at them why you should care for AGI. 0:02:55.470 --> 0:02:57.870 It could replace knowledge workers, which is going to affect 0:02:57.870 --> 0:03:00.960 the economy massively. And for ASI you could do a 0:03:00.960 --> 0:03:05.320 whole lot more than that. So the next thing is, 0:03:05.320 --> 0:03:09.200 why do we care about AGI and ASI? Like what 0:03:09.200 --> 0:03:12.920 are they actually going to produce as output. I think 0:03:12.919 --> 0:03:16.400 the most important output, or at least the most tangible one, 0:03:16.400 --> 0:03:23.839 is invention. Like coming up with. Net new things, ideas, concepts, products, services, 0:03:23.880 --> 0:03:27.919 whatever in the same way that humans do. And whenever 0:03:27.919 --> 0:03:30.160 I think of that, I have one main question. Well, 0:03:30.160 --> 0:03:34.120 how do humans do it? Like what is that actual methodology? 0:03:34.600 --> 0:03:37.560 And I saw a recent episode of Lex Fridman's podcast. 0:03:37.600 --> 0:03:40.640 He had an evolutionary biologist on and he was talking 0:03:40.640 --> 0:03:43.520 about during the enlightenment, there were people meeting and sharing 0:03:43.520 --> 0:03:47.960 ideas and like different shops and salons and whatever, wine bars. 0:03:48.080 --> 0:03:51.600 I'm not sure where they actually went, but they would 0:03:51.800 --> 0:03:53.920 take their ideas, they would share their ideas, and they 0:03:53.920 --> 0:03:56.240 would try to copy each other's ideas. But sometimes they 0:03:56.240 --> 0:04:00.690 would make mistakes and those mistakes would make even better ideas. 0:04:01.090 --> 0:04:05.490 But this idea exchange is like the natural way that 0:04:05.490 --> 0:04:09.690 we had tons of innovation during the enlightenment. And this 0:04:09.690 --> 0:04:12.490 tracks for me because I've always seen innovation as like 0:04:12.530 --> 0:04:16.570 bombarding your brain like a particle accelerator with ideas from 0:04:16.570 --> 0:04:19.810 multiple sources, right? You talk with your your smart friends, 0:04:19.810 --> 0:04:22.289 you talk about cool ideas, you read a whole bunch 0:04:22.290 --> 0:04:24.450 of books, you watch a whole bunch of videos. Whatever 0:04:24.450 --> 0:04:27.770 you do, and all these ideas like go into your 0:04:27.770 --> 0:04:32.050 brain getting bombarded by other ideas that may be different 0:04:32.050 --> 0:04:35.169 or the same or whatever, and they just kind of 0:04:35.210 --> 0:04:38.250 percolate in there and kind of reproduce in there. And 0:04:38.250 --> 0:04:40.330 then as you sleep and you dream and you think 0:04:40.330 --> 0:04:42.450 about other things and work on other things, all of 0:04:42.450 --> 0:04:44.969 a sudden you'll be like, wait a minute and you'll 0:04:44.970 --> 0:04:50.290 have like these moments where actual innovation happens. So the 0:04:50.290 --> 0:04:53.849 idea here is really simple. Let's copy how humans do 0:04:53.850 --> 0:04:58.110 this right. How do humans do this at an individual scale? 0:04:58.470 --> 0:05:02.309 And let's use automation and AI to orchestrate and scale 0:05:02.310 --> 0:05:06.030 that process, which looks, I think, something like this. So 0:05:06.029 --> 0:05:08.710 you have your own ideas. Ideas from books, ideas from 0:05:08.710 --> 0:05:12.310 other people, ideas from wherever. And you basically put that 0:05:12.310 --> 0:05:15.190 into an idea repository. And you could look at this 0:05:15.550 --> 0:05:18.430 project right here called substrate, which I put together a 0:05:18.430 --> 0:05:23.150 couple of years ago. And it's basically crowdsourced ideas, crowdsourced problems, 0:05:23.150 --> 0:05:27.390 crowdsourced solutions. This is a way for us to pull 0:05:27.430 --> 0:05:31.830 together ideas and solutions and problems all into a place 0:05:31.830 --> 0:05:35.110 that we can crowdsource them and see them and work 0:05:35.110 --> 0:05:38.029 on them. And most importantly, we can now hand this 0:05:38.029 --> 0:05:41.830 to AI to start thinking about them all together. Then 0:05:41.830 --> 0:05:45.270 you have this idea of an idea combination system, and 0:05:45.270 --> 0:05:48.750 this is where you combine ideas. You vary them slightly, 0:05:48.750 --> 0:05:52.390 change them in a subtle way, add randomness, whatever, and 0:05:52.390 --> 0:05:57.200 then fold those back into the idea store. and so 0:05:57.520 --> 0:06:00.760 the list of ideas just keeps growing. And then you 0:06:00.760 --> 0:06:03.920 have the testing stuff. This testing stuff is absolutely critical. 0:06:03.920 --> 0:06:08.880 And the most difficult actually, where you actually test the 0:06:08.880 --> 0:06:12.000 ideas against the problems and you need to have a 0:06:12.000 --> 0:06:14.800 way to experiment, right? And this is why so many 0:06:14.800 --> 0:06:18.599 startups are actually spinning up labs like material science labs 0:06:18.600 --> 0:06:22.240 or bio labs, where you can actually build molecules and 0:06:22.240 --> 0:06:25.359 test them against living tissue. Right. And you have to 0:06:25.360 --> 0:06:27.760 be able to do this. Otherwise you can't know whether 0:06:27.760 --> 0:06:30.520 or not the idea worked or not. Uh, in some 0:06:30.520 --> 0:06:33.360 cases you can in some like digital cases, you could 0:06:33.360 --> 0:06:35.760 do like a B testing or something like that, and 0:06:35.760 --> 0:06:37.560 you could say, yes, this is good enough to say 0:06:37.560 --> 0:06:41.520 this actually worked. But in a lot of cases it's 0:06:41.520 --> 0:06:44.200 hard science, it's hard reality. You actually have to have 0:06:44.200 --> 0:06:47.440 a lab to do that. But what you do once 0:06:47.440 --> 0:06:49.920 you have all these components, the ideas, the problems, the 0:06:49.920 --> 0:06:55.099 idea combination engine and then the experimentation engine, You. Now 0:06:55.100 --> 0:06:58.620 just run through this. You iterate through this. So we 0:06:58.660 --> 0:07:03.300 have taken the human system of trying these different things, 0:07:03.500 --> 0:07:06.540 and we've sort of broken it into its components of 0:07:06.540 --> 0:07:11.500 the scientific method. And we are scaling it with AI, 0:07:11.980 --> 0:07:17.460 with crowdsourcing and automation, you know, using pure tech to 0:07:17.500 --> 0:07:22.940 scale the crap out of an already awesome human process. 0:07:24.260 --> 0:07:26.300 And keep in mind, this is not just for like 0:07:26.340 --> 0:07:28.500 a new type of keyboard or a better car battery 0:07:28.500 --> 0:07:31.180 or something like that. The list of problems could be 0:07:31.180 --> 0:07:34.540 anything from like marketing campaigns to figuring out better ways 0:07:34.540 --> 0:07:38.380 to connect with kids who need to learn math or whatever. 0:07:38.860 --> 0:07:43.740 We could put all of humanity's problems into these problem buckets, right? 0:07:44.180 --> 0:07:47.140 And as we get better and better ways to test them, 0:07:47.860 --> 0:07:53.990 we accelerate, right? We accelerate this entire process of automating 0:07:53.990 --> 0:07:58.510 the scientific method. So this ends up being an algorithm 0:07:58.870 --> 0:08:05.150 for solving general problems and testing them. And instead of 0:08:05.150 --> 0:08:08.790 doing it at the scale of like the few universities 0:08:08.790 --> 0:08:11.310 that we have and the few researchers that we have, 0:08:11.630 --> 0:08:15.430 we now can do this at AI scale. And with 0:08:15.430 --> 0:08:19.350 the bottleneck really only being, you know, how much testing 0:08:19.350 --> 0:08:22.950 we actually need to do in the real world. Uh, 0:08:23.830 --> 0:08:26.870 and I'm just really excited about this because, I mean, 0:08:26.870 --> 0:08:30.350 we're talking about, I don't know, five x ten x 0:08:30.390 --> 0:08:35.510 100 x 1000 x million x. Whatever. Our current iterations, 0:08:35.510 --> 0:08:39.510 our current, you know, attempts on goal for doing the 0:08:39.510 --> 0:08:43.790 scientific method, but just scaling that to an insane level. 0:08:45.510 --> 0:08:48.750 So I don't think this system is actually needed for 0:08:48.790 --> 0:08:52.790 AGI or ASI, to be clear. But this chart here 0:08:53.160 --> 0:08:56.120 I think, shows how it is actually just a continuum 0:08:56.120 --> 0:08:58.920 going from bottom to top. So you go from the 0:08:58.920 --> 0:09:04.200 bottom subhuman level of general intelligence or cognitive capability. You 0:09:04.200 --> 0:09:07.960 move up through AGI and then into AC at the top. 0:09:08.640 --> 0:09:10.760 But I do think a system like this that we've 0:09:10.800 --> 0:09:14.400 talked about is a way to actually make the transition 0:09:14.679 --> 0:09:19.840 from where we are into AGI and then beyond into AC. Now, 0:09:19.880 --> 0:09:23.200 my current guess, as I've sort of captured here in 0:09:23.200 --> 0:09:28.000 this chart for AGI is 2027. And I think that's 0:09:28.000 --> 0:09:31.920 going to instantiate as a true knowledge worker replacement agent 0:09:31.920 --> 0:09:35.320 that actually you just hire as a company. It comes in, 0:09:35.320 --> 0:09:39.080 it basically logs in and starts doing onboarding. It reads 0:09:39.080 --> 0:09:43.280 the slack messages, it reads Confluence and Google Docs and 0:09:43.280 --> 0:09:45.760 basically onboards like a regular employee. And this will be 0:09:45.760 --> 0:09:49.840 our first instance of AGI will be like a commercial 0:09:49.840 --> 0:09:54.540 project like that Um, or a commercial product like that. 0:09:55.500 --> 0:09:58.260 And again, I think that's going to be around 2027. 0:09:58.820 --> 0:10:02.060 My original range that I gave in 2023 was 25 0:10:02.059 --> 0:10:05.140 to 28. So I'm, you know, well within those bounds. 0:10:05.780 --> 0:10:10.660 And then for ASI, I have a lot less strong 0:10:10.700 --> 0:10:15.260 of an intuition, but I'm guessing 2028 to 2030 for ASI. 0:10:15.940 --> 0:10:18.660 And hopefully this has been helpful. Cool way to sort 0:10:18.700 --> 0:10:23.260 of think about this uh scientific method algorithm. And we'll 0:10:23.260 --> 0:10:30.219 see you next time. Unsupervised learning is produced on Hindenburg 0:10:30.220 --> 0:10:34.300 Pro using an SM seven B microphone. A video version 0:10:34.300 --> 0:10:37.819 of the podcast is available on the Unsupervised Learning YouTube channel, 0:10:38.380 --> 0:10:40.740 and the text version with full links and notes is 0:10:40.740 --> 0:10:45.980 available at Daniel Mysa.com newsletter. We'll see you next time.