1 00:00:00,880 --> 00:00:05,040 S1: Unsupervised Learning is a podcast about trends and ideas in cybersecurity, 2 00:00:05,080 --> 00:00:09,960 S1: national security, AI, technology and society, and how best to 3 00:00:10,000 --> 00:00:17,680 S1: upgrade ourselves to be ready for what's coming. There's a 4 00:00:17,680 --> 00:00:20,800 S1: ton of discussion everywhere about AGI and ASI and whether 5 00:00:20,800 --> 00:00:24,040 S1: or not they're possible to achieve. I think they are. 6 00:00:24,520 --> 00:00:26,759 S1: And I want to talk about one way we could 7 00:00:26,760 --> 00:00:31,480 S1: possibly pursue that. So I'm going to step through definitions 8 00:00:31,480 --> 00:00:34,600 S1: of AGI and ASI, why we should care about them, 9 00:00:34,600 --> 00:00:41,000 S1: and a system for pursuing them. First, on the definitions themselves, 10 00:00:41,000 --> 00:00:44,600 S1: I think a big problem with AGI and ASI definitions 11 00:00:44,640 --> 00:00:48,280 S1: are really around AI at all is that they're too technical. 12 00:00:48,400 --> 00:00:51,640 S1: They tend to be too technical and therefore not usable. 13 00:00:52,159 --> 00:00:55,800 S1: Not really useful in conversation. I think the best definition 14 00:00:55,800 --> 00:00:59,760 S1: for these things needs to be something that's very human centric. 15 00:01:00,250 --> 00:01:02,210 S1: It should be obvious, and I think we should use 16 00:01:02,210 --> 00:01:05,450 S1: this as a benchmark. Why should I care? We should 17 00:01:05,450 --> 00:01:08,369 S1: be able to look at these definitions and know why 18 00:01:08,370 --> 00:01:10,490 S1: we should care, or at least have a hint towards 19 00:01:10,490 --> 00:01:13,610 S1: why we should care. And I think if we can't 20 00:01:13,610 --> 00:01:16,130 S1: get that from the definition, then it's probably not a 21 00:01:16,130 --> 00:01:20,170 S1: very good one. So my definition for AGI is an 22 00:01:20,170 --> 00:01:24,209 S1: AI system that's able to perform most or all cognitive tasks, 23 00:01:24,530 --> 00:01:28,929 S1: as well as an average US based knowledge worker from 2022. 24 00:01:29,370 --> 00:01:33,130 S1: And I say a US based knowledge worker, because most 25 00:01:33,130 --> 00:01:35,730 S1: people probably won't doubt that there's some kind of base 26 00:01:35,730 --> 00:01:39,770 S1: level smart at doing lots of different tasks, which is 27 00:01:40,050 --> 00:01:46,370 S1: the general in AGI, right? AGI is artificial general intelligence. 28 00:01:46,370 --> 00:01:50,050 S1: So it's general tasks that you do in knowledge work. 29 00:01:50,370 --> 00:01:52,370 S1: And I think if someone's making, you know, a decent 30 00:01:52,370 --> 00:01:56,690 S1: salary as a US based knowledge worker, aren't too many 31 00:01:56,690 --> 00:02:00,110 S1: people that are are going to say that this person 32 00:02:00,110 --> 00:02:04,150 S1: doesn't have general intelligence. So we're using humans as the 33 00:02:04,150 --> 00:02:10,590 S1: baseline for having true general intelligence. And I say before 2023, 34 00:02:11,310 --> 00:02:14,590 S1: because that's when modern AI kicked off. And we don't 35 00:02:14,590 --> 00:02:17,710 S1: want to have the definition keep shifting because humans get 36 00:02:17,710 --> 00:02:20,510 S1: more and more augmented with AI. So so the bar 37 00:02:20,510 --> 00:02:23,350 S1: keeps moving, right. So we want to lock that in place. 38 00:02:25,230 --> 00:02:27,710 S1: ASI is a bit harder and a bit easier to 39 00:02:27,710 --> 00:02:30,829 S1: define at the same time. It's a little more intuitive 40 00:02:30,870 --> 00:02:35,630 S1: because it should be super or above human, but it's 41 00:02:35,630 --> 00:02:40,350 S1: also harder to think about because unlike human level generality, 42 00:02:40,350 --> 00:02:43,910 S1: we've never actually seen anything that's smarter than us. So 43 00:02:43,910 --> 00:02:48,710 S1: you have to actively imagine that. And I think both 44 00:02:48,710 --> 00:02:51,790 S1: of these definitions here are simple enough, and it's obvious 45 00:02:51,790 --> 00:02:55,390 S1: by looking at them why you should care for AGI. 46 00:02:55,470 --> 00:02:57,870 S1: It could replace knowledge workers, which is going to affect 47 00:02:57,870 --> 00:03:00,960 S1: the economy massively. And for ASI you could do a 48 00:03:00,960 --> 00:03:05,320 S1: whole lot more than that. So the next thing is, 49 00:03:05,320 --> 00:03:09,200 S1: why do we care about AGI and ASI? Like what 50 00:03:09,200 --> 00:03:12,920 S1: are they actually going to produce as output. I think 51 00:03:12,919 --> 00:03:16,400 S1: the most important output, or at least the most tangible one, 52 00:03:16,400 --> 00:03:23,839 S1: is invention. Like coming up with. Net new things, ideas, concepts, products, services, 53 00:03:23,880 --> 00:03:27,919 S1: whatever in the same way that humans do. And whenever 54 00:03:27,919 --> 00:03:30,160 S1: I think of that, I have one main question. Well, 55 00:03:30,160 --> 00:03:34,120 S1: how do humans do it? Like what is that actual methodology? 56 00:03:34,600 --> 00:03:37,560 S1: And I saw a recent episode of Lex Fridman's podcast. 57 00:03:37,600 --> 00:03:40,640 S1: He had an evolutionary biologist on and he was talking 58 00:03:40,640 --> 00:03:43,520 S1: about during the enlightenment, there were people meeting and sharing 59 00:03:43,520 --> 00:03:47,960 S1: ideas and like different shops and salons and whatever, wine bars. 60 00:03:48,080 --> 00:03:51,600 S1: I'm not sure where they actually went, but they would 61 00:03:51,800 --> 00:03:53,920 S1: take their ideas, they would share their ideas, and they 62 00:03:53,920 --> 00:03:56,240 S1: would try to copy each other's ideas. But sometimes they 63 00:03:56,240 --> 00:04:00,690 S1: would make mistakes and those mistakes would make even better ideas. 64 00:04:01,090 --> 00:04:05,490 S1: But this idea exchange is like the natural way that 65 00:04:05,490 --> 00:04:09,690 S1: we had tons of innovation during the enlightenment. And this 66 00:04:09,690 --> 00:04:12,490 S1: tracks for me because I've always seen innovation as like 67 00:04:12,530 --> 00:04:16,570 S1: bombarding your brain like a particle accelerator with ideas from 68 00:04:16,570 --> 00:04:19,810 S1: multiple sources, right? You talk with your your smart friends, 69 00:04:19,810 --> 00:04:22,289 S1: you talk about cool ideas, you read a whole bunch 70 00:04:22,290 --> 00:04:24,450 S1: of books, you watch a whole bunch of videos. Whatever 71 00:04:24,450 --> 00:04:27,770 S1: you do, and all these ideas like go into your 72 00:04:27,770 --> 00:04:32,050 S1: brain getting bombarded by other ideas that may be different 73 00:04:32,050 --> 00:04:35,169 S1: or the same or whatever, and they just kind of 74 00:04:35,210 --> 00:04:38,250 S1: percolate in there and kind of reproduce in there. And 75 00:04:38,250 --> 00:04:40,330 S1: then as you sleep and you dream and you think 76 00:04:40,330 --> 00:04:42,450 S1: about other things and work on other things, all of 77 00:04:42,450 --> 00:04:44,969 S1: a sudden you'll be like, wait a minute and you'll 78 00:04:44,970 --> 00:04:50,290 S1: have like these moments where actual innovation happens. So the 79 00:04:50,290 --> 00:04:53,849 S1: idea here is really simple. Let's copy how humans do 80 00:04:53,850 --> 00:04:58,110 S1: this right. How do humans do this at an individual scale? 81 00:04:58,470 --> 00:05:02,309 S1: And let's use automation and AI to orchestrate and scale 82 00:05:02,310 --> 00:05:06,030 S1: that process, which looks, I think, something like this. So 83 00:05:06,029 --> 00:05:08,710 S1: you have your own ideas. Ideas from books, ideas from 84 00:05:08,710 --> 00:05:12,310 S1: other people, ideas from wherever. And you basically put that 85 00:05:12,310 --> 00:05:15,190 S1: into an idea repository. And you could look at this 86 00:05:15,550 --> 00:05:18,430 S1: project right here called substrate, which I put together a 87 00:05:18,430 --> 00:05:23,150 S1: couple of years ago. And it's basically crowdsourced ideas, crowdsourced problems, 88 00:05:23,150 --> 00:05:27,390 S1: crowdsourced solutions. This is a way for us to pull 89 00:05:27,430 --> 00:05:31,830 S1: together ideas and solutions and problems all into a place 90 00:05:31,830 --> 00:05:35,110 S1: that we can crowdsource them and see them and work 91 00:05:35,110 --> 00:05:38,029 S1: on them. And most importantly, we can now hand this 92 00:05:38,029 --> 00:05:41,830 S1: to AI to start thinking about them all together. Then 93 00:05:41,830 --> 00:05:45,270 S1: you have this idea of an idea combination system, and 94 00:05:45,270 --> 00:05:48,750 S1: this is where you combine ideas. You vary them slightly, 95 00:05:48,750 --> 00:05:52,390 S1: change them in a subtle way, add randomness, whatever, and 96 00:05:52,390 --> 00:05:57,200 S1: then fold those back into the idea store. and so 97 00:05:57,520 --> 00:06:00,760 S1: the list of ideas just keeps growing. And then you 98 00:06:00,760 --> 00:06:03,920 S1: have the testing stuff. This testing stuff is absolutely critical. 99 00:06:03,920 --> 00:06:08,880 S1: And the most difficult actually, where you actually test the 100 00:06:08,880 --> 00:06:12,000 S1: ideas against the problems and you need to have a 101 00:06:12,000 --> 00:06:14,800 S1: way to experiment, right? And this is why so many 102 00:06:14,800 --> 00:06:18,599 S1: startups are actually spinning up labs like material science labs 103 00:06:18,600 --> 00:06:22,240 S1: or bio labs, where you can actually build molecules and 104 00:06:22,240 --> 00:06:25,359 S1: test them against living tissue. Right. And you have to 105 00:06:25,360 --> 00:06:27,760 S1: be able to do this. Otherwise you can't know whether 106 00:06:27,760 --> 00:06:30,520 S1: or not the idea worked or not. Uh, in some 107 00:06:30,520 --> 00:06:33,360 S1: cases you can in some like digital cases, you could 108 00:06:33,360 --> 00:06:35,760 S1: do like a B testing or something like that, and 109 00:06:35,760 --> 00:06:37,560 S1: you could say, yes, this is good enough to say 110 00:06:37,560 --> 00:06:41,520 S1: this actually worked. But in a lot of cases it's 111 00:06:41,520 --> 00:06:44,200 S1: hard science, it's hard reality. You actually have to have 112 00:06:44,200 --> 00:06:47,440 S1: a lab to do that. But what you do once 113 00:06:47,440 --> 00:06:49,920 S1: you have all these components, the ideas, the problems, the 114 00:06:49,920 --> 00:06:55,099 S1: idea combination engine and then the experimentation engine, You. Now 115 00:06:55,100 --> 00:06:58,620 S1: just run through this. You iterate through this. So we 116 00:06:58,660 --> 00:07:03,300 S1: have taken the human system of trying these different things, 117 00:07:03,500 --> 00:07:06,540 S1: and we've sort of broken it into its components of 118 00:07:06,540 --> 00:07:11,500 S1: the scientific method. And we are scaling it with AI, 119 00:07:11,980 --> 00:07:17,460 S1: with crowdsourcing and automation, you know, using pure tech to 120 00:07:17,500 --> 00:07:22,940 S1: scale the crap out of an already awesome human process. 121 00:07:24,260 --> 00:07:26,300 S1: And keep in mind, this is not just for like 122 00:07:26,340 --> 00:07:28,500 S1: a new type of keyboard or a better car battery 123 00:07:28,500 --> 00:07:31,180 S1: or something like that. The list of problems could be 124 00:07:31,180 --> 00:07:34,540 S1: anything from like marketing campaigns to figuring out better ways 125 00:07:34,540 --> 00:07:38,380 S1: to connect with kids who need to learn math or whatever. 126 00:07:38,860 --> 00:07:43,740 S1: We could put all of humanity's problems into these problem buckets, right? 127 00:07:44,180 --> 00:07:47,140 S1: And as we get better and better ways to test them, 128 00:07:47,860 --> 00:07:53,990 S1: we accelerate, right? We accelerate this entire process of automating 129 00:07:53,990 --> 00:07:58,510 S1: the scientific method. So this ends up being an algorithm 130 00:07:58,870 --> 00:08:05,150 S1: for solving general problems and testing them. And instead of 131 00:08:05,150 --> 00:08:08,790 S1: doing it at the scale of like the few universities 132 00:08:08,790 --> 00:08:11,310 S1: that we have and the few researchers that we have, 133 00:08:11,630 --> 00:08:15,430 S1: we now can do this at AI scale. And with 134 00:08:15,430 --> 00:08:19,350 S1: the bottleneck really only being, you know, how much testing 135 00:08:19,350 --> 00:08:22,950 S1: we actually need to do in the real world. Uh, 136 00:08:23,830 --> 00:08:26,870 S1: and I'm just really excited about this because, I mean, 137 00:08:26,870 --> 00:08:30,350 S1: we're talking about, I don't know, five x ten x 138 00:08:30,390 --> 00:08:35,510 S1: 100 x 1000 x million x. Whatever. Our current iterations, 139 00:08:35,510 --> 00:08:39,510 S1: our current, you know, attempts on goal for doing the 140 00:08:39,510 --> 00:08:43,790 S1: scientific method, but just scaling that to an insane level. 141 00:08:45,510 --> 00:08:48,750 S1: So I don't think this system is actually needed for 142 00:08:48,790 --> 00:08:52,790 S1: AGI or ASI, to be clear. But this chart here 143 00:08:53,160 --> 00:08:56,120 S1: I think, shows how it is actually just a continuum 144 00:08:56,120 --> 00:08:58,920 S1: going from bottom to top. So you go from the 145 00:08:58,920 --> 00:09:04,200 S1: bottom subhuman level of general intelligence or cognitive capability. You 146 00:09:04,200 --> 00:09:07,960 S1: move up through AGI and then into AC at the top. 147 00:09:08,640 --> 00:09:10,760 S1: But I do think a system like this that we've 148 00:09:10,800 --> 00:09:14,400 S1: talked about is a way to actually make the transition 149 00:09:14,679 --> 00:09:19,840 S1: from where we are into AGI and then beyond into AC. Now, 150 00:09:19,880 --> 00:09:23,200 S1: my current guess, as I've sort of captured here in 151 00:09:23,200 --> 00:09:28,000 S1: this chart for AGI is 2027. And I think that's 152 00:09:28,000 --> 00:09:31,920 S1: going to instantiate as a true knowledge worker replacement agent 153 00:09:31,920 --> 00:09:35,320 S1: that actually you just hire as a company. It comes in, 154 00:09:35,320 --> 00:09:39,080 S1: it basically logs in and starts doing onboarding. It reads 155 00:09:39,080 --> 00:09:43,280 S1: the slack messages, it reads Confluence and Google Docs and 156 00:09:43,280 --> 00:09:45,760 S1: basically onboards like a regular employee. And this will be 157 00:09:45,760 --> 00:09:49,840 S1: our first instance of AGI will be like a commercial 158 00:09:49,840 --> 00:09:54,540 S1: project like that Um, or a commercial product like that. 159 00:09:55,500 --> 00:09:58,260 S1: And again, I think that's going to be around 2027. 160 00:09:58,820 --> 00:10:02,060 S1: My original range that I gave in 2023 was 25 161 00:10:02,059 --> 00:10:05,140 S1: to 28. So I'm, you know, well within those bounds. 162 00:10:05,780 --> 00:10:10,660 S1: And then for ASI, I have a lot less strong 163 00:10:10,700 --> 00:10:15,260 S1: of an intuition, but I'm guessing 2028 to 2030 for ASI. 164 00:10:15,940 --> 00:10:18,660 S1: And hopefully this has been helpful. Cool way to sort 165 00:10:18,700 --> 00:10:23,260 S1: of think about this uh scientific method algorithm. And we'll 166 00:10:23,260 --> 00:10:30,219 S1: see you next time. Unsupervised learning is produced on Hindenburg 167 00:10:30,220 --> 00:10:34,300 S1: Pro using an SM seven B microphone. A video version 168 00:10:34,300 --> 00:10:37,819 S1: of the podcast is available on the Unsupervised Learning YouTube channel, 169 00:10:38,380 --> 00:10:40,740 S1: and the text version with full links and notes is 170 00:10:40,740 --> 00:10:45,980 S1: available at Daniel Mysa.com newsletter. We'll see you next time.