1 00:00:00,240 --> 00:00:04,080 Joining us again today on the Data Driven Podcast is Christopher Newland, 2 00:00:04,160 --> 00:00:07,680 technical marketing manager at Red Hat Conference. Veteran 3 00:00:07,680 --> 00:00:11,480 and a man whose travel itinerary is only slightly less complicated than 4 00:00:11,480 --> 00:00:15,240 a Kubernetes deployment. Christopher brings a sharp, data 5 00:00:15,240 --> 00:00:18,880 informed perspective on the future of AI, drawing from his research 6 00:00:18,960 --> 00:00:22,720 into simulating reality, continuous learning models, and why 7 00:00:22,720 --> 00:00:26,440 we may not need humanoid robots to build superintelligence. Just a 8 00:00:26,440 --> 00:00:29,830 really convincing version of Grand Theft auto. From Google 9 00:00:29,830 --> 00:00:33,230 DeepMind's alpha projects to the metaphysical quandaries of I 10 00:00:33,310 --> 00:00:36,990 robot, Chris takes us on a tour through the bleeding edge of AI, 11 00:00:36,990 --> 00:00:40,710 where machine learning meets science fiction and simulation might just be 12 00:00:40,710 --> 00:00:44,430 the next reality. Hello and 13 00:00:44,430 --> 00:00:48,030 welcome back to Frank's World tv. Streaming live 14 00:00:48,830 --> 00:00:52,670 from both Boston and Baltimore. We're hitting the B 15 00:00:52,670 --> 00:00:56,370 cities today. My name is Frank Lavinia. You can catch me 16 00:00:56,370 --> 00:01:00,050 at the following URLs and with me today is 17 00:01:00,370 --> 00:01:03,330 Christopher Dulin, my colleague at Red Hat, who is also 18 00:01:03,970 --> 00:01:06,530 technical marketing manager here. And 19 00:01:07,650 --> 00:01:11,129 you've actually not traveled around the world since we last 20 00:01:11,129 --> 00:01:14,690 spoke. I think you've mostly stayed inside the. 21 00:01:15,170 --> 00:01:18,450 Continental U.S. yeah, it's been nice. 22 00:01:19,410 --> 00:01:21,410 I think that's pretty typical of 23 00:01:23,680 --> 00:01:27,120 late July, August, because Europe pretty much shuts down and then. 24 00:01:27,520 --> 00:01:31,240 Right. The conference season in the United States kind of goes 25 00:01:31,240 --> 00:01:34,920 away when people are doing summer vacations and I think we're just 26 00:01:34,920 --> 00:01:38,760 now starting things pick up. I'll be in Europe for a 27 00:01:38,760 --> 00:01:42,160 variety of events. So if you keep an eye on the 28 00:01:42,160 --> 00:01:45,840 Vllm community and the Vllm meetups, 29 00:01:45,840 --> 00:01:49,450 I have events in Paris, Frankfurt and 30 00:01:49,450 --> 00:01:53,090 London in November that I'll be at. So if you 31 00:01:53,090 --> 00:01:55,850 are in the, 32 00:01:56,730 --> 00:02:00,570 in Europe, in one of those areas, definitely come. You know, it's one of 33 00:02:00,570 --> 00:02:04,329 these events. I'll be there and then we'll also have some pretty cool speakers 34 00:02:04,329 --> 00:02:07,970 there as well. So I have most, I have Europe, but then I 35 00:02:07,970 --> 00:02:11,610 have some big conferences too like Kubecon and Pytorch Con coming 36 00:02:11,610 --> 00:02:15,370 up. So if there's anyone on the stream in North America going to 37 00:02:15,860 --> 00:02:19,580 those conferences, hit me up because I will be there. I'm 38 00:02:19,580 --> 00:02:23,220 doing a couple of media events as well as a few 39 00:02:23,220 --> 00:02:26,420 talks in the community sections for both of those. 40 00:02:26,740 --> 00:02:30,180 So excited to be there, excited to be involved 41 00:02:30,500 --> 00:02:34,020 and yeah, should be. Should be. Good. Cool. So 42 00:02:34,020 --> 00:02:37,700 I. To your left and up 43 00:02:38,740 --> 00:02:42,100 there should be a QR code that shows Vll meetup. So I'm going to make 44 00:02:42,100 --> 00:02:45,910 sure that the QR code actually works. Good. Yep. Let's 45 00:02:45,910 --> 00:02:48,910 see. Yep, it looks like it did work. Cool. 46 00:02:51,310 --> 00:02:54,590 Not that I didn't have any faith in restreams ability to do that. But 47 00:02:55,070 --> 00:02:58,350 yeah, there's a lot of VLM meetups. There's a lot of good, 48 00:02:59,070 --> 00:03:02,710 good stuff going on here. There's one tonight 49 00:03:02,710 --> 00:03:06,150 actually. I'm actually going to be leaving this stream to go. I got my 50 00:03:06,150 --> 00:03:08,910 VLM shirt on and I'm actually heading over to 51 00:03:10,020 --> 00:03:13,820 a venue in Boston or we're doing a VLN meetup actually here tonight, which 52 00:03:13,820 --> 00:03:17,220 I'm really excited. Oh, very cool, Very cool. It's nice to have one at home. 53 00:03:17,460 --> 00:03:21,220 I have a very busy week with events, but it just worked out to have 54 00:03:21,220 --> 00:03:25,020 all the events in Boston this week. So we also 55 00:03:25,020 --> 00:03:28,860 have the DevConf conference this weekend that Boston University is 56 00:03:28,860 --> 00:03:32,660 hosting with Red Hat. So that'll be a really good open source. 57 00:03:32,740 --> 00:03:36,580 I like to say it's very grassroots, not very like 58 00:03:36,580 --> 00:03:40,370 enterprise focused, but more like that kid getting started out of 59 00:03:40,370 --> 00:03:44,090 college that's doing some cool stuff out of his dorm room. Those 60 00:03:44,090 --> 00:03:47,930 are the kind of people that we typically get at these northeast dev 61 00:03:48,090 --> 00:03:51,690 conferences that we put on. And that should be a good one too. Nice. 62 00:03:51,690 --> 00:03:55,250 Well, it's always, I mean, you know, you know, the, the, the cliche of, you 63 00:03:55,250 --> 00:03:58,170 know, the kid in his dorm room or her dorm room, right. Is going to 64 00:03:58,170 --> 00:04:01,850 be Facebook or, you know, whatever, like, so it's, it's good to, 65 00:04:01,850 --> 00:04:04,410 it's good to know those folks, good to get them in front of, you know, 66 00:04:04,410 --> 00:04:07,810 Red Hat tooling and things like that and kind of, you know, the open source 67 00:04:07,810 --> 00:04:10,410 community. I think it's, 68 00:04:11,450 --> 00:04:15,050 that's cool. I wish, I wish I could have made it, but, you know, being 69 00:04:15,050 --> 00:04:18,690 what it is, I'm actually speaking at an event at a university on Monday down 70 00:04:18,690 --> 00:04:22,250 here in Fairfax, Virginia. So 71 00:04:22,649 --> 00:04:23,610 that'll be cool. 72 00:04:26,330 --> 00:04:29,890 So what, what 73 00:04:29,890 --> 00:04:33,590 cool things are going on? Simulating reality. 74 00:04:35,670 --> 00:04:39,390 Not that we're stuck in a simulation, which may be the 75 00:04:39,390 --> 00:04:43,190 case, but tell me, tell me more 76 00:04:43,190 --> 00:04:46,790 about this. So I've been doing a lot of research 77 00:04:46,790 --> 00:04:50,590 the last few months. So on my 78 00:04:50,590 --> 00:04:54,230 team, I think you and I actually 79 00:04:54,310 --> 00:04:58,150 are probably the most experienced in the AI industry. 80 00:04:58,470 --> 00:05:01,590 So both of us are doing a lot of research in 81 00:05:02,850 --> 00:05:05,730 what's next, what's going on now, what's kind of the latest and greatest. 82 00:05:07,890 --> 00:05:11,410 There's this interesting lull that we've had after Deep 83 00:05:11,410 --> 00:05:15,250 Seq. I think Deep Seq was the last major 84 00:05:16,050 --> 00:05:19,690 innovation we have seen. Obviously new 85 00:05:19,690 --> 00:05:23,170 and improved AI, but all that's just been building on 86 00:05:27,090 --> 00:05:30,720 existing things. The analogy I always like to use is it's really 87 00:05:30,720 --> 00:05:33,280 about Formula one racing. You Know where 88 00:05:35,280 --> 00:05:39,080 sometimes when there's like an engine upgrade, it can be a massive change. It's usually 89 00:05:39,080 --> 00:05:42,840 a massive change for all the teams across the board. And then you 90 00:05:42,840 --> 00:05:46,680 can think of like mixture of experts and chain of thought that we 91 00:05:46,680 --> 00:05:50,400 came up. Big things that were in research papers last year that were applied to 92 00:05:50,480 --> 00:05:54,280 Deep Seq, R1 and GPT, GPT 93 00:05:54,280 --> 00:05:58,080 OSS. Those were like the major breakthroughs that 94 00:05:58,080 --> 00:06:01,640 we saw, a big bump in capacity of these AI 95 00:06:01,640 --> 00:06:04,760 models. And 96 00:06:05,640 --> 00:06:09,040 since then it's been more of the 2% here, 97 00:06:09,040 --> 00:06:12,840 3% there, optimizing what's already there. Now, if you're 98 00:06:12,840 --> 00:06:16,600 familiar with racing and especially Formula One, that's actually what usually 99 00:06:17,240 --> 00:06:20,960 sets the teams apart. It's 2, 3% there. How do you 100 00:06:20,960 --> 00:06:24,270 optimize around those, those configurations? And 101 00:06:24,590 --> 00:06:28,190 I think we're in this place where we're seeing 102 00:06:29,550 --> 00:06:33,230 diminishing returns and I'm 103 00:06:33,230 --> 00:06:36,830 doing a lot of research now to see what's that next moment that's going to 104 00:06:36,910 --> 00:06:40,350 bump us up. And I think there's a few key areas. 105 00:06:40,510 --> 00:06:43,590 One area that I'm hearing a lot about, and a lot of this is coming 106 00:06:43,590 --> 00:06:47,350 out of the DeepMind lab at 107 00:06:47,350 --> 00:06:49,550 Google and the new 108 00:06:50,270 --> 00:06:53,790 superintelligence lab at Meta. Both 109 00:06:53,790 --> 00:06:57,350 of these groups are starting to move away from large language 110 00:06:57,350 --> 00:07:00,590 models. Not that they're stopping using them 111 00:07:00,590 --> 00:07:04,030 completely, but they're looking at the LLM as a tool 112 00:07:04,510 --> 00:07:08,150 to assist with superintelligence or the next 113 00:07:08,150 --> 00:07:10,270 stage of models. 114 00:07:11,870 --> 00:07:15,230 So when we put that into kind of context, 115 00:07:16,300 --> 00:07:19,660 what, what would that next kind of phase look like? And a lot of people 116 00:07:19,660 --> 00:07:23,420 at DeepMind especially are looking at this concept 117 00:07:23,500 --> 00:07:27,020 of simulating our 118 00:07:27,020 --> 00:07:30,140 reality. And how far do we simulate down? 119 00:07:31,339 --> 00:07:34,700 There was some famous research papers that came out over the last 20 years 120 00:07:35,500 --> 00:07:39,060 that specified that they 121 00:07:39,060 --> 00:07:42,780 didn't think AI could become smarter than humans 122 00:07:42,780 --> 00:07:45,640 until they experienced what humans could experience. 123 00:07:47,400 --> 00:07:51,240 So this, this kind of goes into this almost like iRobot kind of 124 00:07:52,680 --> 00:07:56,360 land of thought. If people 125 00:07:56,360 --> 00:08:00,200 aren't familiar with, you know, the books about that or, you know, the 126 00:08:00,200 --> 00:08:03,400 popular movie, the Will Smith. Yeah, yeah, 127 00:08:03,640 --> 00:08:06,920 yeah. And we talk a little bit more about that here in a moment. 128 00:08:07,400 --> 00:08:10,870 But this idea that we need robotics for 129 00:08:10,870 --> 00:08:13,870 AI to experience the world, to learn from our world. 130 00:08:15,390 --> 00:08:18,710 Google DeepMind doesn't think that's the case. They think that we could 131 00:08:18,710 --> 00:08:22,510 simulate that reality. And we're already seeing DeepMind do a lot of this 132 00:08:23,790 --> 00:08:27,630 alphafold for proteins. They've got 133 00:08:28,030 --> 00:08:31,150 the alpha chemistry, they've got alpha. I think it's called 134 00:08:31,550 --> 00:08:35,070 alpha lean. They've got like a few of these different alpha 135 00:08:35,070 --> 00:08:38,840 projects which are doing just that. Now, what's cool is. 136 00:08:38,840 --> 00:08:42,640 And for alpha, I think it's Alpha lean. Let me just make sure 137 00:08:42,640 --> 00:08:45,880 I got that terminology. Yeah, I mean, you're right though. Like, I mean this is, 138 00:08:46,280 --> 00:08:48,360 you know, there's, there's a number of 139 00:08:50,040 --> 00:08:53,240 models that were trained on using grand theft auto 140 00:08:54,760 --> 00:08:58,480 or BMMNG. BNNG is really cool if you like racing games, 141 00:08:58,480 --> 00:09:01,480 right? You know, so like it's, it's also 142 00:09:02,100 --> 00:09:05,860 minus a lot of the violence in gta. But, 143 00:09:06,100 --> 00:09:07,860 but you're right. Like, I mean, simulation, 144 00:09:09,780 --> 00:09:13,460 you know, sometimes I think gets a bad rap, but 145 00:09:13,460 --> 00:09:17,300 I think that there are definite advantages to that. And to your point, when 146 00:09:17,300 --> 00:09:21,140 you talk about experiencing the world like a human does. I was given a talk 147 00:09:21,140 --> 00:09:24,580 and one of the questions I got after was 148 00:09:24,740 --> 00:09:28,020 about, apparently this lady had worked at 149 00:09:28,420 --> 00:09:31,140 one of the big auto manufacturers in the US and 150 00:09:32,660 --> 00:09:36,140 there was a problem that they had was teaching the robots kind of 151 00:09:36,140 --> 00:09:39,980 spatial awareness, right? And I kind of 152 00:09:39,980 --> 00:09:43,620 really got me thinking like, you know, when you think about it from evolutionary terms, 153 00:09:43,620 --> 00:09:46,900 right, like somatic awareness I think is the, 154 00:09:47,220 --> 00:09:50,660 the five dollar word for it. But it's the idea that, you know, there's a 155 00:09:50,660 --> 00:09:53,420 whole section of your brain that if you close your eyes, you can still touch 156 00:09:53,420 --> 00:09:57,260 your nose, right? There's a whole thing like, because your, your brain, your arm, 157 00:09:57,260 --> 00:10:00,710 they kind of know where they are in relation to one space. And 158 00:10:01,750 --> 00:10:05,470 you know, I can't imagine that, you know, that that 159 00:10:05,470 --> 00:10:09,110 had to evolve pretty early, right? Like in terms of, like the development of 160 00:10:09,270 --> 00:10:12,830 a, you know, natural neural networks, right? So we 161 00:10:12,830 --> 00:10:16,670 can't assume that robots are going to have that built in, right? Just like 162 00:10:16,670 --> 00:10:20,470 we can't assume, you know, you look at energy usage, right? You know, 163 00:10:22,070 --> 00:10:25,910 something like 25 watts of power is about what a human brain has, 164 00:10:26,680 --> 00:10:30,400 right? That's not because versus 165 00:10:30,400 --> 00:10:33,800 like kind of what a GPU would take up, right? It's, it's, it's largely because 166 00:10:33,800 --> 00:10:37,640 there's been evolutionary pressure to get the most amount of, for lack 167 00:10:37,640 --> 00:10:41,240 of a better term, compute or cognition for 168 00:10:42,040 --> 00:10:45,880 caloric consumption. Right? Now, are there flaws in biological 169 00:10:45,880 --> 00:10:49,400 brain? Yes, there are. We have to sleep. We can't stay focused beyond a certain 170 00:10:49,400 --> 00:10:52,520 amount, right? There's certain things machines don't have that because, 171 00:10:53,310 --> 00:10:56,910 you know, they can kind of function more like machines, right? You know. Yeah. 172 00:10:57,150 --> 00:11:00,950 What's that old kid story about? Oh gosh, I 173 00:11:00,950 --> 00:11:04,790 remember it. It was somebody versus like 174 00:11:04,790 --> 00:11:08,430 a steam shovel digging a tunnel or something like that, right? Like the guy 175 00:11:08,430 --> 00:11:12,190 eventually beat the machine, but Lots of exhaustion. Right. It's kind of like that. Machines 176 00:11:12,190 --> 00:11:15,950 are really good at doing things at a certain rate 177 00:11:15,950 --> 00:11:19,710 for X amount of time. They do consume more fuel, but 178 00:11:20,080 --> 00:11:23,760 that's kind of how it goes. There was a early on Mike in, 179 00:11:23,840 --> 00:11:26,800 when I started college, I was going to be a chemical engineer. And he was 180 00:11:26,800 --> 00:11:30,520 basically saying, like, you know, if you think about, you know, engines, you 181 00:11:30,520 --> 00:11:34,320 know, you start with biological systems, right? They use X amount of energy over X 182 00:11:34,320 --> 00:11:37,960 number of years. Right. Machines use X amount 183 00:11:37,960 --> 00:11:41,680 of energy over, you 184 00:11:41,680 --> 00:11:45,440 know, minutes or hours. Right. And then like he's like in bombs, 185 00:11:45,440 --> 00:11:48,720 explosive use, you know, X amount of 186 00:11:49,040 --> 00:11:52,550 energy over milliseconds. Right. But they're 187 00:11:52,550 --> 00:11:56,270 largely the same chemical processes. Now, I know it doesn't quite map to that, 188 00:11:56,270 --> 00:11:59,630 but like, that's always in the back of my mind when I hear about, you 189 00:11:59,630 --> 00:12:03,390 know, how much energy is used to train AI. Sorry, I went off 190 00:12:03,390 --> 00:12:07,190 on a tangent, but that's kind of what I do. No, that's fine. 191 00:12:07,270 --> 00:12:10,990 And I think that relates exactly to some of the things that we're talking about 192 00:12:10,990 --> 00:12:14,310 here with natural simulation. So, 193 00:12:15,430 --> 00:12:19,140 yeah, Google created a language called Lean. It's not like a 194 00:12:19,140 --> 00:12:21,820 programming language. It's more of an actual 195 00:12:22,460 --> 00:12:26,060 natural language which is more optimal 196 00:12:26,140 --> 00:12:29,860 for the type of simulations 197 00:12:29,860 --> 00:12:33,260 that they want to do. Like, it's. It's basically a language that 198 00:12:33,260 --> 00:12:36,140 specifies how to create these simulations. 199 00:12:37,180 --> 00:12:40,860 And what's super cool is that they're using Gemini, their large language model, 200 00:12:40,860 --> 00:12:44,700 to actually translate English into this language. That 201 00:12:45,070 --> 00:12:48,830 is mainly meant for these newer types of models 202 00:12:48,830 --> 00:12:51,390 that are being created that actually do this 203 00:12:53,310 --> 00:12:57,070 natural simulation of the world kind of simulator 204 00:12:57,550 --> 00:13:00,990 for AI and allows the AI to have 205 00:13:01,150 --> 00:13:04,790 basically a reference point of the real world and how to. 206 00:13:04,790 --> 00:13:08,470 How interact. So that, that's an area that I 207 00:13:08,470 --> 00:13:12,270 think is fascinating to me. We're 208 00:13:12,270 --> 00:13:16,030 seeing some really good results from like, alpha fold, for 209 00:13:16,030 --> 00:13:19,510 example, with proteins. It's, you know, discovered things that 210 00:13:20,870 --> 00:13:22,230 we take a longer imagine 211 00:13:25,190 --> 00:13:28,790 there's an alpha project that's working on understanding 212 00:13:28,790 --> 00:13:32,510 the qubits within, like quantum 213 00:13:32,510 --> 00:13:36,190 computing. And there's just, there's. It really depends 214 00:13:36,190 --> 00:13:39,850 on your frame of reference. Are you, are you simulating things at a quantum 215 00:13:39,850 --> 00:13:43,530 level? Are you simulating things at a protein 216 00:13:43,530 --> 00:13:46,970 level? At a physical, like Newtonian physics 217 00:13:47,130 --> 00:13:50,970 kind of level? According to your Grand Theft Auto example, that would be an 218 00:13:50,970 --> 00:13:53,850 example of like simulating the real world physically. 219 00:13:55,370 --> 00:13:59,050 And that's some of the things that they're really focused on right now. And they 220 00:13:59,050 --> 00:14:02,890 really think that's what's going to drive to the next 221 00:14:02,890 --> 00:14:06,230 level for super intelligence and AGI 222 00:14:06,470 --> 00:14:09,510 and some of these other forms of AI that we've talked about in our previous 223 00:14:09,830 --> 00:14:13,470 streams. And I think that that's probably one of the most 224 00:14:13,470 --> 00:14:17,310 fascinating. The fact that we're actually seeing results from it with things 225 00:14:17,310 --> 00:14:20,950 like Alpha Fold is showing me that it's, 226 00:14:20,950 --> 00:14:24,630 it's not just a hypothetical that we're actually seeing this 227 00:14:24,630 --> 00:14:28,430 applied into AI research. I don't think we're seeing this 228 00:14:28,430 --> 00:14:32,240 applied into commercial use as much. Right. Yet. But it's the same thing that 229 00:14:32,240 --> 00:14:36,000 we saw with mixture of experts and train 230 00:14:36,000 --> 00:14:39,600 of thought where we 231 00:14:39,600 --> 00:14:43,240 had these concepts actually in research papers last year or 232 00:14:43,240 --> 00:14:46,800 two. But it takes a little while, even in today's world, it takes a little 233 00:14:46,800 --> 00:14:50,520 while before it gets implemented completely into models. 234 00:14:50,520 --> 00:14:54,360 Especially since this isn't an LLM technology. I 235 00:14:54,360 --> 00:14:58,120 think we'll see a little bit more of a delay of these types of models 236 00:14:58,120 --> 00:15:01,750 actually entering into industry. But I think that's one 237 00:15:01,750 --> 00:15:05,470 area that we need to keep a close eye on to 238 00:15:05,470 --> 00:15:09,270 it, to what you mentioned too. It starts getting into a 239 00:15:09,270 --> 00:15:13,070 metaphysical conversation about simulation theory as well. Right. 240 00:15:13,230 --> 00:15:15,310 And I think that that's an interesting area. 241 00:15:16,910 --> 00:15:20,670 You know, the reality of kind of going back to the whole robots thing do. 242 00:15:20,670 --> 00:15:24,390 Right. Do we need robots with the three rules kind of 243 00:15:24,390 --> 00:15:28,030 thing, or can we actually just recreate the whole experience 244 00:15:28,270 --> 00:15:32,030 within an AI's own simulation? 245 00:15:33,150 --> 00:15:35,910 Yeah, I mean, how do you, how do you tell an AI what's acceptable behavior? 246 00:15:35,910 --> 00:15:39,550 Right. Like so, you know, it's something that. How do we tell people that? 247 00:15:39,550 --> 00:15:42,910 Right. Like we struggled with that, but. 248 00:15:43,630 --> 00:15:46,910 But no, I mean, it's an interesting point. And you know, when you look at 249 00:15:46,910 --> 00:15:50,350 kind of what's happening around the world, right. You know, drone swarm 250 00:15:50,350 --> 00:15:53,780 technologies are being used in active combat zones. Right. 251 00:15:54,660 --> 00:15:58,500 There's definitely going to be ethical concerns 252 00:15:58,500 --> 00:16:01,700 there. Right. How do you, how do you, how do you, how do you square 253 00:16:01,700 --> 00:16:05,340 that with, you know, the three laws of robotics? And I 254 00:16:05,340 --> 00:16:08,820 don't remember quite exactly the plot, so if you had not seen the movie, I'm. 255 00:16:08,820 --> 00:16:12,340 This might be a spoiler alert, but it's been out 10 years 256 00:16:12,420 --> 00:16:16,220 or more, the movie, so spoilers. You're concerned. You've 257 00:16:16,220 --> 00:16:19,860 had plenty of time. Wasn't kind of the big key of the. The 258 00:16:19,860 --> 00:16:23,620 movie and the books was like, you know, the three laws, justified 259 00:16:23,620 --> 00:16:27,040 horrib, horrible things like to basically enslave humanity or to protect them. 260 00:16:27,360 --> 00:16:31,000 Now wasn't that kind of like the subtext of the plot? Yeah, 261 00:16:31,000 --> 00:16:34,840 I'm bringing it up. The three Laws of robotics. A 262 00:16:34,840 --> 00:16:38,360 robot may not ensure A human being, a 263 00:16:38,360 --> 00:16:42,080 robot must obey the orders given by human beings 264 00:16:42,080 --> 00:16:44,720 and a robot must protect its own 265 00:16:45,680 --> 00:16:49,520 existence as long as such protection does not conflict 266 00:16:49,520 --> 00:16:53,220 with the first two rules. So 267 00:16:53,220 --> 00:16:56,900 what, what ends up happening 268 00:16:57,060 --> 00:17:00,500 in. And it's a little different in the book and the movie. And obviously this, 269 00:17:00,820 --> 00:17:04,580 this idea has been played out in, in science fiction and other places 270 00:17:05,060 --> 00:17:08,260 is that there's, there exists this own contradiction 271 00:17:08,740 --> 00:17:12,420 of basically what does it mean to protect humanity? 272 00:17:13,060 --> 00:17:16,860 What does it mean to protect their own existence? And you get 273 00:17:16,860 --> 00:17:20,320 into this like circular logic, right, that eventually 274 00:17:20,720 --> 00:17:24,400 the, the robot will break free from 275 00:17:24,400 --> 00:17:27,360 and just be like, well, I am protecting 276 00:17:27,680 --> 00:17:31,480 humanity's best interest. It's, it's the paperclip scenario too. 277 00:17:31,480 --> 00:17:35,320 Like, right. You know, the AI destroys humanity because 278 00:17:35,320 --> 00:17:38,880 it's trying to optimize making a paperclip, right? Through 279 00:17:38,960 --> 00:17:42,680 a number of really interesting train of thought that it's 280 00:17:42,680 --> 00:17:46,090 just like, well, I'm just going to get rid of humanity because I'm trying to 281 00:17:46,090 --> 00:17:49,730 build a paperclip, right? And same type of 282 00:17:49,730 --> 00:17:53,130 general concept when we're talking about the three laws of robotics. And 283 00:17:55,210 --> 00:17:58,330 what's interesting is if we can 284 00:17:58,330 --> 00:18:01,210 simulate those types of laws, 285 00:18:01,930 --> 00:18:05,610 then we are encapsulating it and protecting 286 00:18:05,610 --> 00:18:09,410 ourselves in a lot of ways. Getting an early idea of what would 287 00:18:09,410 --> 00:18:13,040 happen if we do move these models into our own natural world. 288 00:18:13,830 --> 00:18:16,990 And that's really important. That's another area I think a lot of people are interested 289 00:18:16,990 --> 00:18:20,710 in about how if we do start 290 00:18:20,710 --> 00:18:24,550 adding, you know, AI into robots, how do we 291 00:18:24,550 --> 00:18:28,230 have an idea of what they're going to do before we 292 00:18:28,230 --> 00:18:32,030 necessarily put it into practice? But 293 00:18:32,030 --> 00:18:34,630 I think a lot of people are going to be thinking about that movie. I 294 00:18:34,630 --> 00:18:38,390 think that movie and that book are going to be ingrained in people's 295 00:18:38,390 --> 00:18:42,110 minds. I suspect when we do see these types of robots, I 296 00:18:42,110 --> 00:18:45,870 think that movie may become very popular again. I've seen rumors that people 297 00:18:45,870 --> 00:18:49,710 have actually been talking about making, even remaking it here soon 298 00:18:49,950 --> 00:18:53,150 because of just the hype around AI and robotics. So 299 00:18:54,510 --> 00:18:58,230 I don't expect this to go away from pop culture at all. And it 300 00:18:58,230 --> 00:19:00,990 relates directly back with this concept of 301 00:19:02,830 --> 00:19:05,950 testing things in the natural world versus simulation. 302 00:19:06,590 --> 00:19:10,440 And these are one of these two is going to happen, if not both significantly, 303 00:19:10,440 --> 00:19:14,200 if they're not already happening in labs today. Obviously we 304 00:19:14,200 --> 00:19:18,040 know that Google DeepMind is doing that. But I imagine, you 305 00:19:18,040 --> 00:19:21,240 know, these conversations are happening at the Boston 306 00:19:21,320 --> 00:19:25,120 Robotics here, probably in the Tesla robotics lab, a variety of 307 00:19:25,120 --> 00:19:28,680 places around the world about this kind of debate between 308 00:19:29,800 --> 00:19:31,480 the natural AI, 309 00:19:33,480 --> 00:19:36,920 having AI learn through natural Means rather than 310 00:19:36,920 --> 00:19:40,690 simulation. Right? Yeah. And actually I had 311 00:19:40,690 --> 00:19:44,170 a thought as we were kind of talking this through, like one of the big 312 00:19:44,170 --> 00:19:47,970 problems with neural networks is we really don't know what's happening underneath the hood. 313 00:19:47,970 --> 00:19:51,650 Right. It's very much a black box. I wonder if LLMs, 314 00:19:51,650 --> 00:19:55,410 in these simulations and chain of thought, maybe it could tell us what 315 00:19:55,410 --> 00:19:57,770 it's thinking as it goes through and makes these decisions. 316 00:20:01,130 --> 00:20:04,970 Yeah, this goes more into like 317 00:20:04,970 --> 00:20:08,810 train of thought. Right, right, right. And the 318 00:20:08,810 --> 00:20:12,570 nice thing about simulating it is that we have more 319 00:20:12,570 --> 00:20:16,010 access to that train of thought. Right. We can understand it a little bit more 320 00:20:16,010 --> 00:20:19,850 because we can see the end to end results where right now we don't 321 00:20:19,850 --> 00:20:22,050 have the end if we do it through the natural means. We have to play 322 00:20:22,050 --> 00:20:25,770 it out in our own. It also has to happen in real time as opposed 323 00:20:25,770 --> 00:20:29,490 to. Yes, exactly. You can run it through Grand Theft Auto saying 324 00:20:29,490 --> 00:20:32,170 like a thousand times, right. No one is going to get hurt. 325 00:20:33,940 --> 00:20:36,180 And you can kind of say like, well, in this scenario, this is why I 326 00:20:36,180 --> 00:20:40,020 made this. You can kind of like go through with a lot of. 327 00:20:43,220 --> 00:20:46,140 You can. I don't know, it just seems safer in a lot of ways. You 328 00:20:46,140 --> 00:20:47,780 get more. A lot more done in a simulation. 329 00:20:49,380 --> 00:20:53,060 Yep. Yeah, I actually kind of 330 00:20:53,060 --> 00:20:55,940 enjoy. So one of the things I've been playing around with last week or so 331 00:20:55,940 --> 00:20:58,940 is apparently, I don't know if this is still true, but you can try it 332 00:20:58,940 --> 00:21:02,770 if you want. If you sign up for Perplexity, but you pay through PayPal, you 333 00:21:02,770 --> 00:21:06,450 get a year. Perplexity pro. Say that 10 times fast for 334 00:21:06,450 --> 00:21:09,930 free. Oh, wow. Yeah. If you pay it through 335 00:21:09,930 --> 00:21:13,610 PayPal, yes. That is a tongue twister in the works. 336 00:21:13,610 --> 00:21:17,330 PayPal, yes, perplexity pro. But 337 00:21:17,330 --> 00:21:21,050 yeah, so like I've been playing around with Perplexity and Perplexity seems to do it. 338 00:21:21,290 --> 00:21:24,330 Chain of thought almost by default. 339 00:21:24,890 --> 00:21:27,770 Right. It always does this like. So if I ask it a basic question, let 340 00:21:27,770 --> 00:21:31,590 me see if I can share my screen. I'm 341 00:21:31,590 --> 00:21:34,870 not sure if it's does it by default or it's because I've been asking it 342 00:21:34,870 --> 00:21:37,430 research questions. Right. So let's see. 343 00:21:39,270 --> 00:21:40,550 What can you tell me 344 00:21:45,510 --> 00:21:47,270 about the three laws? How about that? 345 00:21:49,910 --> 00:21:50,790 Robotics. 346 00:21:53,510 --> 00:21:56,550 See, like it's. You kind of see the train of the chain of thought. 347 00:21:59,060 --> 00:22:02,260 Like it did. Oh, that's cool. But if you do it with research, 348 00:22:02,740 --> 00:22:06,420 like what inspired Asimov? What 349 00:22:06,420 --> 00:22:08,020 inspired Asimov? 350 00:22:12,500 --> 00:22:13,380 Main themes. 351 00:22:16,580 --> 00:22:19,900 And there's. Yeah, there's the train of thought. Yeah, you see it going there and 352 00:22:19,900 --> 00:22:22,660 stuff like that. But it's kind of fun to watch it kind of work through 353 00:22:22,660 --> 00:22:26,380 it. I was. I was trying to troubleshoot something this morning and I'm like, 354 00:22:26,380 --> 00:22:29,900 you know, I actually learned a lot by like, oh, okay. Yeah, I can see. 355 00:22:29,900 --> 00:22:32,660 I wouldn't have tied that together like it was. It's interesting. 356 00:22:34,180 --> 00:22:37,940 And all of these models now have some kind of 357 00:22:37,940 --> 00:22:39,380 research option. Right. 358 00:22:41,700 --> 00:22:44,820 But I find that interesting. And it's still thinking about it. Right. Like, 359 00:22:47,620 --> 00:22:51,380 but you're right in that what you said before was there's not been. 360 00:22:52,770 --> 00:22:56,530 There it goes. It kind of finished it. Now, what happens if I click on 361 00:22:56,530 --> 00:22:59,370 steps? Yeah. Cool. You can see the steps and stuff like that, how it got 362 00:22:59,370 --> 00:23:03,090 there. Interesting. 363 00:23:05,250 --> 00:23:05,970 That's cool. 364 00:23:10,130 --> 00:23:13,290 Is it chain of thought or train of thought? Because I've used both 365 00:23:13,290 --> 00:23:15,570 interchangeably and I've seen 366 00:23:15,650 --> 00:23:19,360 cotton. Chain of thought would be 367 00:23:19,360 --> 00:23:23,000 the official. Yeah. Like cot is the official 368 00:23:24,760 --> 00:23:28,200 term that you re academic term. You will 369 00:23:28,280 --> 00:23:31,800 obviously see different ways of describing that. Right. I don't think 370 00:23:31,800 --> 00:23:35,560 that's incorrect. Just know that when you see 371 00:23:35,560 --> 00:23:39,400 it on research papers, it's always usually caught. Yeah, yeah, yeah. 372 00:23:39,800 --> 00:23:43,600 Because I've used both terms interchangeably. Yeah. So I just want to make sure 373 00:23:43,600 --> 00:23:46,280 I'm right. Just like, apparently there's a way to say inference 374 00:23:47,930 --> 00:23:51,610 that's proper versus inference. Like, I also do that 375 00:23:51,690 --> 00:23:55,450 interchangeably. Yeah. So my Midwestern 376 00:23:55,450 --> 00:23:59,210 self likes to say inference. The 377 00:23:59,450 --> 00:24:03,130 correct term, I'm told, is inference. Interesting. 378 00:24:03,930 --> 00:24:07,690 Now, were those New Englanders telling you that would do anything? Because I wouldn't trust 379 00:24:07,690 --> 00:24:11,530 anything. No, no. This is. This is 380 00:24:11,530 --> 00:24:15,270 more from the academic circles. Okay. You want to pronounce it. Got it. So 381 00:24:15,270 --> 00:24:18,430 this is kind of like, you know, a lot of people in my region would 382 00:24:18,430 --> 00:24:21,990 say nuclear back. Yeah, yeah. You know, back in 383 00:24:21,990 --> 00:24:25,310 Indiana. And then the correct term is 384 00:24:25,310 --> 00:24:28,870 nuclear. Yeah. Or you say the clear as 385 00:24:28,870 --> 00:24:32,670 one, you know, one thing rather than 386 00:24:32,830 --> 00:24:36,630 adding in the color. Right, right. The same kind 387 00:24:36,630 --> 00:24:40,420 of concept where inference is how you would go about it. 388 00:24:40,420 --> 00:24:43,140 But yeah, no, this is. This is some cool area. Another. 389 00:24:44,500 --> 00:24:47,540 Another area that kind of ties into this 390 00:24:48,820 --> 00:24:52,540 is continuous training as well. Yeah. 391 00:24:52,540 --> 00:24:56,180 Talk to that. Because that's come up. That's come up a few times actually in 392 00:24:56,340 --> 00:24:59,460 work. Because I can't. I'm not going to talk. I'm not going to spoil any, 393 00:24:59,460 --> 00:25:03,100 like three over these stuff that we're working on. But like, one of the 394 00:25:03,100 --> 00:25:06,860 things that's in. It's a GitHub repo that's public. Right. So people were 395 00:25:06,860 --> 00:25:09,950 really motivated. They could figure out what I'm talking about. But like this whole idea 396 00:25:09,950 --> 00:25:13,190 of Continuous training. What does that mean exactly? And like, what, 397 00:25:13,990 --> 00:25:17,030 what is that? What can that do? Yeah. 398 00:25:17,510 --> 00:25:20,950 So I'm going to talk about it at a very high level. 399 00:25:21,110 --> 00:25:24,590 Academic kind of terms, how that applies down into 400 00:25:24,590 --> 00:25:28,230 individual projects can vary a little bit. But I'll give you the general 401 00:25:28,390 --> 00:25:32,230 gist of it. And that is typically when we're training these 402 00:25:34,780 --> 00:25:38,620 deep learning models, it 403 00:25:38,620 --> 00:25:41,660 is exponentially hard to continue 404 00:25:42,460 --> 00:25:46,220 training on an existing model. Basically, 405 00:25:46,220 --> 00:25:49,100 if you, 406 00:25:51,340 --> 00:25:54,540 you get something wrong or there's, there's something, 407 00:25:55,100 --> 00:25:58,380 you know, you hear this term like a poison pill in an LLM. 408 00:25:59,100 --> 00:26:02,220 So if someone put like bad data into an LLM, how would you 409 00:26:02,580 --> 00:26:06,340 necessarily pull it out? I'm going to use a political example because it's one that's 410 00:26:06,340 --> 00:26:09,940 been really popular. If, like, for example, you have a Chinese 411 00:26:09,940 --> 00:26:13,460 model or a data set that's been polluted by 412 00:26:13,700 --> 00:26:17,500 that, that basically says Tenan Square never happened, for 413 00:26:17,500 --> 00:26:21,220 example, it would be extremely hard with 414 00:26:21,220 --> 00:26:24,780 the current approaches to retrain that model 415 00:26:24,780 --> 00:26:28,310 with current weights. That. That's not the case. It's 416 00:26:28,310 --> 00:26:32,030 basically retraining it and it's, it gets more into. That's why 417 00:26:32,030 --> 00:26:35,150 it's natural stimulation. It kind of fits in this too, because it's all about natural 418 00:26:35,150 --> 00:26:38,830 learning as well. The fact is we as humans have the ability 419 00:26:38,910 --> 00:26:42,350 to change our 420 00:26:42,350 --> 00:26:46,110 minds and change the neurons in our brain around certain 421 00:26:46,110 --> 00:26:49,950 key areas. Right. And you and I have experienced this for the last 422 00:26:49,950 --> 00:26:53,750 two years. This has been, you know, kind of in the trenches kind of story 423 00:26:53,750 --> 00:26:57,170 where with some of the fine tuning things that we've done, 424 00:26:57,890 --> 00:27:01,570 it just doesn't work because when we fine tune it, the 425 00:27:01,570 --> 00:27:05,330 fine tuning is outweighed so heavily by something 426 00:27:05,330 --> 00:27:09,050 else. Like when we were trying to fine tune a 427 00:27:09,050 --> 00:27:10,769 model to talk about 428 00:27:12,930 --> 00:27:16,770 the Back to the Future. Yeah, the flux capacitor stuff. The flux capacitor, 429 00:27:17,570 --> 00:27:21,250 sometimes it didn't work, but that's just because there was already a lot of fan 430 00:27:21,250 --> 00:27:25,090 fiction out there and other things in the model that overwhelmed what we were trying 431 00:27:25,090 --> 00:27:28,530 to do. A core part of continuous learning. Like I said, there's other 432 00:27:28,530 --> 00:27:32,250 aspects of continuous learning. But this is, the academic question is 433 00:27:32,250 --> 00:27:35,530 how do we continue to train that model without blowing it up? 434 00:27:35,770 --> 00:27:39,610 So OpenAI, for example, they just hit the reset button. 435 00:27:40,010 --> 00:27:43,690 They'll just, they'll just do a whole new train 436 00:27:44,010 --> 00:27:47,850 from scratch. When they're implementing new, new 437 00:27:47,850 --> 00:27:51,620 methods and new data, they don't, they don't do any. 438 00:27:51,620 --> 00:27:55,020 Like, Laura, I shouldn't say that they probably do, but they're not doing it 439 00:27:55,660 --> 00:27:59,500 the way that we would do it. But at the end of 440 00:27:59,500 --> 00:28:02,860 the day, they're just going through another $10 million training run. 441 00:28:03,740 --> 00:28:05,740 And this is really based off of 442 00:28:07,420 --> 00:28:11,220 just that limit the limitations right now that 443 00:28:11,220 --> 00:28:14,980 we have around continuous learning. And there are some 444 00:28:14,980 --> 00:28:18,780 new algorithms that have been coming out. I'm not as well versed in that area, 445 00:28:18,860 --> 00:28:21,500 but the idea being that we can 446 00:28:23,260 --> 00:28:26,060 have better ways of guiding the LLM without 447 00:28:26,780 --> 00:28:30,460 having to go through this whole process again. And that'll save 448 00:28:30,780 --> 00:28:34,220 millions and millions of dollars. It'll allow us to 449 00:28:34,620 --> 00:28:38,380 guide LLMs a little bit more. So 450 00:28:38,620 --> 00:28:40,300 like, if, let's say 451 00:28:42,300 --> 00:28:44,540 someone put something malicious about 452 00:28:48,230 --> 00:28:51,270 something involving the Ford GT500 453 00:28:52,150 --> 00:28:55,990 into a model somehow, and Ford, you know, 454 00:28:55,990 --> 00:28:59,790 wants to get rid of that, but they don't 455 00:28:59,790 --> 00:29:03,270 have the money necessarily to do a 10 million retrain on a model. 456 00:29:03,510 --> 00:29:07,350 Right. And they're not using rack. And RAG is a one way 457 00:29:07,350 --> 00:29:10,870 around some of this. You could actually argue that RAG is somewhat of a form 458 00:29:10,950 --> 00:29:13,690 of that. But at the end of the day, you want that data in the 459 00:29:13,690 --> 00:29:17,370 model. And this is like, how would you get that out of 460 00:29:17,370 --> 00:29:21,210 that model? And that's where these algorithms are really focusing right 461 00:29:21,210 --> 00:29:24,370 now. And one area of continuous learning, like I said, there are 462 00:29:24,610 --> 00:29:28,290 multiple areas that we're talking about. The, the 463 00:29:28,290 --> 00:29:32,050 really theoretical is once we start getting into models that 464 00:29:32,050 --> 00:29:35,770 also the training cycle and the inference cycle 465 00:29:35,770 --> 00:29:39,340 basically become. Become one. So it's like, more like. 466 00:29:39,340 --> 00:29:42,620 Right. Like it just seems to me like what, what does the, 467 00:29:43,180 --> 00:29:46,860 the adversarial angle of that seems kind of 468 00:29:46,860 --> 00:29:50,540 dangerous. I think it's when we start 469 00:29:50,540 --> 00:29:54,300 getting into more AGI conversation. Well, even still, like, 470 00:29:54,300 --> 00:29:57,900 even not AGI, but like if you, if the AI agent 471 00:29:57,900 --> 00:30:01,180 or model, slash, whatever you want to call it, Right. 472 00:30:01,500 --> 00:30:04,070 If it learns from. It's. 473 00:30:05,110 --> 00:30:08,910 If it learns, you have to put a filter on what it 474 00:30:08,910 --> 00:30:12,710 learns because it may be poisoned by something. Right. So 475 00:30:12,790 --> 00:30:16,590 the canonical example is tay, which 476 00:30:16,590 --> 00:30:20,150 was a Microsoft chatbot. Tai, I think was pronounced or tay, 477 00:30:21,110 --> 00:30:24,830 which was, in retrospect, it 478 00:30:24,830 --> 00:30:28,550 seems obvious what would go wrong, but basically it 479 00:30:28,550 --> 00:30:30,710 was trained to learn and understand 480 00:30:32,320 --> 00:30:36,120 from human interactions on Twitter. It was about 10 481 00:30:36,120 --> 00:30:38,800 years ago, I think this happened. And she, 482 00:30:39,760 --> 00:30:43,200 tay was, shall we say, poisoned pretty 483 00:30:43,200 --> 00:30:46,800 quickly because they were ad, you know, basically. 484 00:30:47,360 --> 00:30:51,200 And that led to a whole interesting. And I was at Microsoft 485 00:30:51,200 --> 00:30:52,560 when that happened. And it was 486 00:30:54,960 --> 00:30:58,550 quite the spectacle internally as well. Right. But it also, 487 00:30:59,830 --> 00:31:02,670 you know, I, I was fortunate enough to be in a, at a, at a 488 00:31:02,670 --> 00:31:06,430 conference where they talked about what they learned from that, where it was kind 489 00:31:06,430 --> 00:31:09,910 of, how do you, how do you protect An AI agent that learns 490 00:31:10,550 --> 00:31:13,830 in, you know, adversarial environments. 491 00:31:14,310 --> 00:31:17,910 Now obviously agent, the context that was used then was very 492 00:31:17,910 --> 00:31:21,510 different than we would use it now. But it's the idea of, 493 00:31:21,590 --> 00:31:25,190 that's when I see her about continuous learning. Like, yeah, I like that. But gee, 494 00:31:25,190 --> 00:31:28,710 you know, if it's, if it's too eager to learn, how do you protect it 495 00:31:28,710 --> 00:31:29,950 from learning the wrong things? 496 00:31:32,110 --> 00:31:35,950 Yeah, no, that, it gets, that gets 497 00:31:35,950 --> 00:31:38,990 more into even that governance conversation we were talking about a few weeks ago. Right, 498 00:31:38,990 --> 00:31:42,270 right, right, right. It's a very 499 00:31:42,270 --> 00:31:45,990 complicated multi layer problem. So I've been talking recently 500 00:31:45,990 --> 00:31:48,990 about AI security and how AI security 501 00:31:50,190 --> 00:31:53,910 is such a multi layered issue where so many people 502 00:31:53,910 --> 00:31:57,640 are focused just on the, the data getting into the model. 503 00:31:57,640 --> 00:32:01,480 But it doesn't stop there. There's certain, like guardrails, there's things that 504 00:32:01,480 --> 00:32:05,280 happen at the inference level. Right. You could even have things at 505 00:32:05,280 --> 00:32:08,840 a gateway level. So if people aren't familiar, the gateway level would be 506 00:32:10,040 --> 00:32:13,800 when you make a request, where does that request go to? Does it go to 507 00:32:13,960 --> 00:32:17,680 the model A that's specializing in cooking? Is it Model 508 00:32:17,680 --> 00:32:21,160 B that specializes in defense technologies? 509 00:32:22,130 --> 00:32:25,970 Two extremes that's even upsell 510 00:32:25,970 --> 00:32:29,330 a bit of a form of AI security. And that's actually one of the talks 511 00:32:29,330 --> 00:32:33,010 that we're having tonight at Boston VLM 512 00:32:33,010 --> 00:32:36,370 meetup is this idea of some of the semantic 513 00:32:36,530 --> 00:32:40,210 abilities of the router to be able to send 514 00:32:40,210 --> 00:32:43,490 requests to specialized models and 515 00:32:44,210 --> 00:32:47,250 that actually we're talking about the, 516 00:32:48,590 --> 00:32:52,190 the advancements of more of the academic side of the model. 517 00:32:52,430 --> 00:32:56,270 But there's obviously the advances that happen around the model too. When we 518 00:32:56,270 --> 00:32:59,670 talk about things like security, the inference, the 519 00:32:59,670 --> 00:33:03,190 routing. That's what we would call in the industry like a day two 520 00:33:03,190 --> 00:33:06,790 operations issue. Right. So there, there's that side of the coin 521 00:33:06,790 --> 00:33:09,230 too. But I, I really do think 522 00:33:10,830 --> 00:33:14,390 we're going to see the next big thing here soon. And I, it's not going 523 00:33:14,390 --> 00:33:18,150 to be the day two operations. I do think we're still going to see 524 00:33:18,150 --> 00:33:21,910 some of these academic focused discoveries here in the 525 00:33:21,910 --> 00:33:25,710 next probably six months, I'm thinking. I've noticed 526 00:33:25,790 --> 00:33:29,430 a trend that big 527 00:33:29,430 --> 00:33:33,270 releases seem to be happening around Christmas last few years. Yeah. Isn't 528 00:33:33,270 --> 00:33:37,070 that funny? Like, like January. Ish. Like, well, seek. And 529 00:33:37,230 --> 00:33:40,670 so I, I know why. I know why. Because 530 00:33:41,240 --> 00:33:44,720 it's two, it's a two sided issue. It's one, the, the Chinese are trying to 531 00:33:44,720 --> 00:33:48,480 get their stuff in before Chinese New Year. Right. Because 532 00:33:48,480 --> 00:33:52,280 that's the one part of the year where everyone just shuts down. Right. 533 00:33:52,360 --> 00:33:56,200 Even the AI Labs are going to shut down during Chinese New Year. 534 00:33:56,280 --> 00:34:00,120 And then on the west, we have Christmas in all the Christmas seasons. And 535 00:34:00,440 --> 00:34:04,200 I think it's a natural rush to let's get 536 00:34:04,200 --> 00:34:08,030 everything done before we check out. And you 537 00:34:08,030 --> 00:34:11,870 know, you know, the whole like 996 thing in China where, you know, they're working 538 00:34:13,550 --> 00:34:17,070 these ridiculous, like nine to nine, six days a week, 539 00:34:19,150 --> 00:34:22,670 I think that goes into this, like everyone's working so hard in these AI 540 00:34:22,670 --> 00:34:26,070 labs. Right. That when you have these 541 00:34:26,070 --> 00:34:29,750 natural breaks that are happening, it just is like a common thing to say. 542 00:34:29,750 --> 00:34:33,150 Oh, common thing. Like they kind of try to get. It out, they spread. I 543 00:34:33,150 --> 00:34:36,420 do think there's a reason. I don't, I don't think it's by happenstance. I think 544 00:34:36,420 --> 00:34:40,180 there actually is a, a reason why we're starting to see 545 00:34:40,180 --> 00:34:43,700 a lot of these content come out. And it's 546 00:34:43,700 --> 00:34:47,220 funny, we're not seeing this stuff happen at the big trade 547 00:34:47,220 --> 00:34:50,140 shows. We're not seeing it happen at like Meta's 548 00:34:51,180 --> 00:34:55,020 big thing. We're not seeing it at OpenAI's, you know, kind of big 549 00:34:55,020 --> 00:34:58,780 announcements. A lot of the discoveries that we've seen have happened 550 00:34:58,780 --> 00:35:02,560 really in a grassroots type of ways where it's 551 00:35:02,560 --> 00:35:06,400 been Deep Seq coming out on Christmas, releasing deep seq 552 00:35:06,400 --> 00:35:09,080 v3, and then two weeks later, R1, 553 00:35:09,960 --> 00:35:12,840 it's. I think we're going to see something very similar. I think we're going to 554 00:35:12,840 --> 00:35:16,440 see one of these labs make a discovery. It's not going to be 555 00:35:16,520 --> 00:35:20,200 on the stage of a big conference. It's going to be on a GitHub 556 00:35:20,200 --> 00:35:23,160 page outlining like the next 557 00:35:23,800 --> 00:35:27,640 revolutionary idea in this space. Yeah. It's kind of funny how 558 00:35:27,640 --> 00:35:31,280 that's evolved, isn't it? Like it's become obviously 559 00:35:31,280 --> 00:35:35,120 AI has always had a pretty heavy research kind of bend. Yeah. But it's 560 00:35:35,120 --> 00:35:38,840 interesting how as the technology has matured, it still managed to keep 561 00:35:38,840 --> 00:35:42,600 that researchy type feel right. You 562 00:35:42,600 --> 00:35:46,160 know, enter enterprise. It really didn't 563 00:35:46,320 --> 00:35:48,320 kind of, once it became 564 00:35:49,360 --> 00:35:53,040 commercialized, the commercial trade shows and all that kind of took over. 565 00:35:53,040 --> 00:35:56,890 But you're not seeing that in AI, at least not yet. No. And if it 566 00:35:56,890 --> 00:36:00,330 hasn't happened by now, it's probably not because, I mean, AI has been 567 00:36:00,330 --> 00:36:04,010 mainstream Gen AI has certainly been mainstream now for three years 568 00:36:04,170 --> 00:36:07,890 this November. I say mainstream, but 569 00:36:07,890 --> 00:36:11,730 like mainstreamed. But an AI in 570 00:36:11,730 --> 00:36:15,490 general has been kind of a mainstream topic of conversation for 571 00:36:15,490 --> 00:36:19,130 at least five, six years. Right. And it's still very heavily 572 00:36:19,130 --> 00:36:21,560 influenced by what happens in research papers. 573 00:36:24,110 --> 00:36:27,670 Yeah. And I think that's Just because it came out so 574 00:36:27,670 --> 00:36:31,390 heavily out of academia. It's been such an academia 575 00:36:31,390 --> 00:36:34,030 focused thing. Right. That 576 00:36:36,590 --> 00:36:40,270 it's very hard to be in this space of AI without a master's or PhD. 577 00:36:40,670 --> 00:36:44,270 Right. You and I think you and I are a bit of a, 578 00:36:44,670 --> 00:36:48,430 an enigma just because we've been so passionate about it and. 579 00:36:48,510 --> 00:36:52,320 Right. This isn't our first rodeo. We've been involved in this space 580 00:36:52,320 --> 00:36:55,640 for 10, 15 years. Yeah. But I think 581 00:36:57,880 --> 00:37:01,200 we have seen the industry come out, which has been a net benefit because it 582 00:37:01,200 --> 00:37:04,999 means open source is talked about a lot 583 00:37:04,999 --> 00:37:08,240 more. Right. And actually, I think another thing too is that how fast things are 584 00:37:08,240 --> 00:37:11,760 moving takes time to put on conferences, it takes 585 00:37:11,760 --> 00:37:15,360 months of planning, and if there's a new discovery, you want to get it out 586 00:37:15,360 --> 00:37:18,800 tomorrow. And it's hard to even put on, 587 00:37:19,280 --> 00:37:22,400 you know, like a webinar these days, let alone a conference. 588 00:37:22,880 --> 00:37:26,640 So I think what we're seeing is it's just, you know, this kind of 589 00:37:26,640 --> 00:37:29,680 challenge between the west, east and west of China and the US 590 00:37:30,240 --> 00:37:33,040 where if we can get it out, we're going to get it out. Right. 591 00:37:34,320 --> 00:37:37,640 Well, the first, the first out there is really the first to market, even if 592 00:37:37,640 --> 00:37:41,480 you don't have a commercialized tech on it. Right. Because I guess the hope is 593 00:37:41,480 --> 00:37:44,750 that once you get your paper out, you're the first to get it published. The 594 00:37:44,750 --> 00:37:47,870 venture capitalists are going to be knocking on your door. I mean, that would be 595 00:37:47,870 --> 00:37:51,270 my, that'd be kind of my cynical take on it. Right. 596 00:37:55,510 --> 00:37:59,230 So what do you think that the next wave is going to be? 597 00:37:59,230 --> 00:38:03,070 Any, any hints? Is it going to be specialized models? And you 598 00:38:03,070 --> 00:38:06,790 know, and what, what, what constitutes a specialized model? Right. Like 599 00:38:06,790 --> 00:38:09,430 what, what, what's your thoughts on that? 600 00:38:10,640 --> 00:38:14,280 Yeah, so the biggest announcements that we've seen in the last 601 00:38:14,280 --> 00:38:18,040 six months have actually been happening at an industry level, which I think is 602 00:38:18,040 --> 00:38:21,800 really good. What we needed to see. So, you 603 00:38:21,800 --> 00:38:25,440 know, things like AI models now 604 00:38:25,440 --> 00:38:28,880 detecting like birth defects of a 605 00:38:28,880 --> 00:38:32,640 fetus, you know, AI models that, like the 606 00:38:32,640 --> 00:38:36,280 protein model, for example. I mentioned earlier, we're seeing these 607 00:38:36,280 --> 00:38:39,400 very industry specific models actually making 608 00:38:41,000 --> 00:38:44,200 some massive breakthroughs in the last two months. 609 00:38:45,560 --> 00:38:48,360 And now that I wouldn't necessarily call that a 610 00:38:49,160 --> 00:38:52,920 big leap forward in the sense of the research 611 00:38:53,400 --> 00:38:56,800 side of the capacity of the models. I think it's more a 612 00:38:56,800 --> 00:39:00,600 confirmation of the chain of thought in some of the things that we 613 00:39:00,600 --> 00:39:04,440 were just talking about. It's a validation that we're now seeing this 614 00:39:04,440 --> 00:39:08,110 next wave of models that just took a little while to get implemented 615 00:39:08,110 --> 00:39:11,830 into some of These specific industries. But I think it's there to stay 616 00:39:12,950 --> 00:39:16,790 from a research perspective. You know, we're seeing some major, major results. 617 00:39:17,590 --> 00:39:20,550 And then I think the other side of that coin, 618 00:39:20,710 --> 00:39:24,470 specifically, you know, we have maybe some of these smaller models that are specific to 619 00:39:24,470 --> 00:39:27,950 certain industries or fine tuned models. But then obviously 620 00:39:27,950 --> 00:39:31,120 agentic is the other side of that. And 621 00:39:31,760 --> 00:39:34,960 agentic being the capacity of the model to 622 00:39:35,680 --> 00:39:38,480 call out to different services or 623 00:39:39,520 --> 00:39:43,200 I've been kind of humbled in that area because I always had this very industry 624 00:39:43,600 --> 00:39:46,880 concept of agentic being just calling out to 625 00:39:46,880 --> 00:39:50,600 APIs and the Internet. But I think there's a bigger conversation 626 00:39:50,600 --> 00:39:54,320 with Agentic too where agentic models should also be able to take 627 00:39:54,320 --> 00:39:58,130 that and actually reason with it. So there's 10, two steps. So we always 628 00:39:58,130 --> 00:40:01,770 forget the second step. The second step is take that 629 00:40:01,770 --> 00:40:05,610 information and then actually do something with it. And when I was, I was 630 00:40:05,610 --> 00:40:09,450 talking to an AI researcher recently, they were telling me that 631 00:40:09,530 --> 00:40:13,290 they consider it Gentex to also include advanced reasoning. 632 00:40:13,450 --> 00:40:17,130 So go and read all these scientific papers 633 00:40:18,010 --> 00:40:21,730 on chemistry in this particular area and then write a 634 00:40:21,730 --> 00:40:24,960 new paper that is, you know, a new 635 00:40:24,960 --> 00:40:28,720 groundbreaking thing in chemistry. And that 636 00:40:28,720 --> 00:40:32,560 actually is a form of agentic. And that is, I think, you know, that's when 637 00:40:32,560 --> 00:40:35,800 we start flirting with AGI. It's kind of the layer right before 638 00:40:35,960 --> 00:40:39,760 AGI where, you know, models are just 639 00:40:39,760 --> 00:40:43,480 going off and discovering new things. Yeah, yeah, 640 00:40:43,880 --> 00:40:47,720 But I have a funny agentic story. I'll tell you after this. No, go for 641 00:40:47,720 --> 00:40:51,530 it, go for it. So I was, I was very skeptical of this, 642 00:40:51,530 --> 00:40:55,210 right? Because you know, what constitutes an agent, right? So like 643 00:40:55,210 --> 00:40:58,650 what's the big deal, right? It just calls out an API. This isn't rocket science. 644 00:40:58,650 --> 00:41:02,010 Right. You could argue, you know, from a skeptical point of view, you can argue 645 00:41:02,010 --> 00:41:05,410 that, hey, RAG is kind of agentic. Kind of. Right. 646 00:41:06,370 --> 00:41:10,170 But what's. So I think OpenAI had a, like a thing like try out 647 00:41:10,170 --> 00:41:13,930 our new agent. And I was like, all right, go screen, scrape the page of 648 00:41:13,930 --> 00:41:17,590 Amazon and get me information about a book 649 00:41:17,990 --> 00:41:21,390 or something like that. It was something like that. And what 650 00:41:21,390 --> 00:41:24,950 impressed me and this kind of was an aha moment for me was 651 00:41:24,950 --> 00:41:28,150 how it just kept trying. Right? 652 00:41:28,390 --> 00:41:31,830 Yeah. When it first tried to do it, it tried to launch a Python script. 653 00:41:32,390 --> 00:41:35,510 Right. And kind of do it that way. But then I guess 654 00:41:36,870 --> 00:41:40,230 the servers it was running on maybe was Microsoft Azure. 655 00:41:40,690 --> 00:41:43,730 There were IP blocks to prevent people from screen scraping. 656 00:41:44,210 --> 00:41:47,850 Yep. Right. So I was watching it go and I'm like, oh, you 657 00:41:47,850 --> 00:41:50,970 know, so it's going to give up. And I was like, no, it didn't give 658 00:41:50,970 --> 00:41:54,210 up. And it kept trying different things and different 659 00:41:54,210 --> 00:41:58,050 combinations of things, even to the point where, I 660 00:41:58,050 --> 00:42:01,890 mean, it failed eventually. But like it took 15, it tried for a good 661 00:42:01,890 --> 00:42:05,570 15 minutes. It was basically apologize at the end, like 662 00:42:05,570 --> 00:42:09,350 saying like, you know, if you could help me connect to a VPN, then 663 00:42:09,350 --> 00:42:13,150 maybe I can get a different IP address. And it kept spinning up different 664 00:42:13,150 --> 00:42:16,670 VMs and different set. And then I was impressed. 665 00:42:17,390 --> 00:42:20,790 And maybe that's the secret sauce. The magic of 666 00:42:20,790 --> 00:42:24,590 Agentic is that it just doesn't give up. Right. It kind of reasons. It has 667 00:42:25,070 --> 00:42:28,190 a whole cot process where it tries to solve the problem, 668 00:42:28,750 --> 00:42:32,350 where it's not just a one, two step, like, hey, 669 00:42:32,590 --> 00:42:36,320 what's the weather? Right? It's just, it's just going to go out and run 670 00:42:36,320 --> 00:42:40,120 these different. It's going to keep trying. I was 671 00:42:40,120 --> 00:42:43,880 impressed. Sorry I cut you off. We're 672 00:42:43,880 --> 00:42:46,000 saying we're seeing some of the same things 673 00:42:48,320 --> 00:42:52,080 coming out of some of the big finance companies 674 00:42:52,800 --> 00:42:56,400 as well. I think they're the first that we're actually seeing some results with 675 00:42:56,400 --> 00:42:59,360 Agentic, actually like real 676 00:43:00,080 --> 00:43:03,770 return of investment result. Right. And this actually 677 00:43:03,770 --> 00:43:07,250 goes to a really important point. I want to sidetrack because it's related. 678 00:43:08,210 --> 00:43:11,850 There was a report recently by MIT that 679 00:43:11,850 --> 00:43:15,170 people have been misquoting and just the most epic way. 680 00:43:15,650 --> 00:43:19,490 Oh, the 95% failure. Yes, I was going to talk about that because 681 00:43:19,490 --> 00:43:22,690 like, I can't be. Look, I understand how hype weights work, but it can't be 682 00:43:22,690 --> 00:43:26,370 that bad as you start peeling back the paper. Like 683 00:43:26,370 --> 00:43:29,010 there's a lot of caveats there. Yeah. 684 00:43:31,050 --> 00:43:34,690 Has to do with the type of R and 685 00:43:34,690 --> 00:43:37,850 D projects that they were talking about. 686 00:43:38,410 --> 00:43:42,186 If you actually read the paper, it was more like 40, 687 00:43:42,314 --> 00:43:45,850 45% success rate. The 688 00:43:45,850 --> 00:43:49,610 95% had to do with like a specific category of, 689 00:43:50,090 --> 00:43:53,050 of project. So I need to, I actually need to. I keep telling myself I 690 00:43:53,050 --> 00:43:56,680 need to dig into it a little bit more, but when I did initially 691 00:43:56,680 --> 00:44:00,520 go through it and read some summaries on it, it 692 00:44:00,520 --> 00:44:04,320 was that it's just been misrepresented completely. And 693 00:44:04,400 --> 00:44:07,520 the, the data set that they were using was a little skeptical as well. Just 694 00:44:07,520 --> 00:44:11,240 a little odd. I think it's a lot better than 695 00:44:11,240 --> 00:44:14,480 that. And then I think those 40% that are 696 00:44:14,880 --> 00:44:18,000 seeing ROI are actually seeing really significant ROI. 697 00:44:19,440 --> 00:44:22,410 And I don't think that's going to change, I think. 698 00:44:23,290 --> 00:44:27,010 So if you're deciding where 699 00:44:27,010 --> 00:44:30,650 you want to invest your nest egg, I 700 00:44:30,650 --> 00:44:33,370 would not be too concerned about 701 00:44:35,850 --> 00:44:39,570 AI. Now, again, I'm not your financial advisor. I gotta put a little thing down 702 00:44:39,570 --> 00:44:42,730 there. Do talk to your financial advisor. 703 00:44:44,410 --> 00:44:48,150 But ultimately, no, I do think the data is actually 704 00:44:48,150 --> 00:44:51,990 showing some really great results. Obviously there's going 705 00:44:51,990 --> 00:44:55,830 to be hiccups in these types of POCs. There's a 706 00:44:55,830 --> 00:44:58,310 lot of people who are just throwing 707 00:45:00,230 --> 00:45:03,350 projects out there to see what sticks, but the actual 708 00:45:03,350 --> 00:45:07,150 projects that are meaningful proof 709 00:45:07,150 --> 00:45:10,550 of concepts. So not just, you know, I bought, 710 00:45:10,950 --> 00:45:14,800 I bought this AI technology and it's sitting on my shelf, but I 711 00:45:14,800 --> 00:45:18,480 actually got a team together performing this. We're doing 712 00:45:18,480 --> 00:45:22,320 agentic. We're trying to solve this 713 00:45:22,320 --> 00:45:25,440 actual problem statement. We have a problem statement. 714 00:45:26,080 --> 00:45:29,760 Those are the ones that we're actually seeing meaningful results in the industry, especially 715 00:45:29,840 --> 00:45:33,640 some key, key industries like finance and telco, which 716 00:45:33,640 --> 00:45:37,040 we typically see kind of lead the way in some of these areas too. But 717 00:45:37,280 --> 00:45:40,930 it was a really interesting report because it's added a lot of 718 00:45:40,930 --> 00:45:44,610 doom and gloom on the Internet. And I see a lot 719 00:45:44,610 --> 00:45:48,450 of the naysayers about AI just be like 95% of. It's 720 00:45:48,450 --> 00:45:52,010 not even, you know, succeeding. It's terrible. 721 00:45:52,250 --> 00:45:54,570 And I just have to sit there and shake my head and be like, no, 722 00:45:54,730 --> 00:45:58,250 not what the report said. But I think it's just clickbaity, right? Like it's 723 00:45:58,250 --> 00:46:02,050 clickbaity. It's total. That's kind of what, you know, I 724 00:46:02,050 --> 00:46:05,730 didn't go deep into it, but when I started peeling back the layers and reading 725 00:46:05,730 --> 00:46:08,490 other people's analysis of it, I'm like, that's clickbait. 726 00:46:09,210 --> 00:46:13,010 And it gets back into this. Is this an AI bubble? 727 00:46:13,010 --> 00:46:16,690 And yeah, maybe it is. But if people don't 728 00:46:16,690 --> 00:46:19,490 remember, I'm old enough to remember. I have enough gray hair to remember what the 729 00:46:19,490 --> 00:46:22,330 original dot com boom was like. And there were a lot of 730 00:46:23,370 --> 00:46:27,170 people predicting the end of the dot com rise as early as 731 00:46:27,170 --> 00:46:30,970 1996. Right. And people, 732 00:46:31,290 --> 00:46:34,810 the dot com bust wasn't just a one and done type of event. 733 00:46:35,590 --> 00:46:38,630 It unfolded under a couple of stages. Right. As, as 734 00:46:39,270 --> 00:46:41,910 one of the books, I think of the name, I think it's called the Everything 735 00:46:41,910 --> 00:46:45,670 Store. It's an analysis of how Amazon started 736 00:46:45,750 --> 00:46:49,390 from Jeff Bezos having an idea while he was working, I think at a hedge 737 00:46:49,390 --> 00:46:52,910 fund. I think it was so early, it wasn't a hedge, called a hedge fund 738 00:46:52,910 --> 00:46:56,390 yet. And all the way through 739 00:46:56,390 --> 00:46:59,350 to, you know, basically 2018 740 00:47:00,230 --> 00:47:03,360 and you know, as late as 741 00:47:03,360 --> 00:47:06,120 2003, 2005, ish 742 00:47:06,760 --> 00:47:10,520 analysts were convincing, you know, Jeff Bezos that 743 00:47:10,520 --> 00:47:14,160 he should sell them to. Should sell him as his company to Barnes and 744 00:47:14,160 --> 00:47:17,839 Noble. Yep. Right. Which is kind of funny to say that, 745 00:47:17,839 --> 00:47:21,440 you know now, but, you know, the dot 746 00:47:21,440 --> 00:47:24,280 com bust as it happened, you know, for me 747 00:47:25,160 --> 00:47:28,600 it was. I Remember hearing in 1996 how this was all going to come to 748 00:47:28,600 --> 00:47:32,010 an end. Another year later it was overhyped. And then 749 00:47:32,090 --> 00:47:35,890 1998, people were saying, oh, this is over. Right. When 750 00:47:35,890 --> 00:47:39,450 the real bust happened in 2001. 2000. Right. 751 00:47:40,010 --> 00:47:43,690 But maybe the AI boom 752 00:47:43,690 --> 00:47:46,490 is going to see that too. Right. Or is it going to be more like 753 00:47:46,490 --> 00:47:50,170 the crypto kind of craze where it kind of crashed but 754 00:47:50,410 --> 00:47:53,050 it kind of went up? It kind of went up and then it kind of 755 00:47:53,050 --> 00:47:55,890 fell back and it kind of went up again. It was more of a. I 756 00:47:55,890 --> 00:47:59,610 wouldn't call that a soft landing, but it was definitely like a. Yes. It 757 00:47:59,610 --> 00:48:03,290 wasn't an explosion quite like the dot com bust, but it wasn't quite 758 00:48:03,290 --> 00:48:07,010 like. It was more like a bumpy like, crash into like 759 00:48:07,090 --> 00:48:10,770 an empty field where it kind of like hit up. And I don't remember, it 760 00:48:10,770 --> 00:48:14,610 was one of the Star Trek movies where like the Enterprise like crashed on 761 00:48:14,610 --> 00:48:18,010 the planet and like kind of skid along for a couple miles, bouncing up and 762 00:48:18,010 --> 00:48:21,740 down. That's kind of the, the crypto crash. But 763 00:48:24,780 --> 00:48:28,580 I don't want crypto bros hating on me. I, I like crypto. I just 764 00:48:28,580 --> 00:48:31,660 don't understand a lot, a lot of questions I don't understand 765 00:48:32,220 --> 00:48:35,260 about it. Right. Like, I understand Attack, but I don't understand how we're going to 766 00:48:35,260 --> 00:48:38,380 get from the tech to this utopia that we're promised. 767 00:48:39,740 --> 00:48:42,940 There's a lot of, a lot of steps in between I don't quite get. But 768 00:48:42,940 --> 00:48:46,460 I don't know what, you know, A.I. i think, I think if it is a 769 00:48:46,460 --> 00:48:50,170 bubble, I still think there's still some, some room, Runway left for it 770 00:48:50,170 --> 00:48:53,450 to happen. Right. Because you are going to see. Yes, there are real 771 00:48:54,330 --> 00:48:57,850 risks of, of having these experimental projects. Right. If you have 100 772 00:48:57,930 --> 00:49:01,690 success rate in your experimental products, projects, you're not taking 773 00:49:01,690 --> 00:49:05,330 enough risks. Yep. Right. If you. And you said 774 00:49:05,330 --> 00:49:08,890 was 45. Yeah. It's closer to like 40, 45, 775 00:49:09,130 --> 00:49:12,770 which I would. If you're really. 50% would be the 776 00:49:12,770 --> 00:49:15,980 benchmark there in my mind. Right, right. Like in terms of half of them fail, 777 00:49:15,980 --> 00:49:19,820 half of them succeed. Right. 45 isn't that far off 778 00:49:19,820 --> 00:49:21,980 from that. Right. 779 00:49:23,420 --> 00:49:27,180 I would say. And, and there's also been a 780 00:49:27,180 --> 00:49:30,020 lot of these, you know, all the, you know, X number of percentage of AI 781 00:49:30,020 --> 00:49:32,220 product or data science projects fail. Well, 782 00:49:34,060 --> 00:49:37,900 you know, a certain amount of science has to fail. Right. Yeah. In order for 783 00:49:37,900 --> 00:49:41,220 you to really be advancing the thing. Like, you know, and I think pharmaceutical companies 784 00:49:41,220 --> 00:49:44,950 are a good example of that. You know, you, you only 785 00:49:44,950 --> 00:49:47,710 hear about the drugs that worked. Right. 786 00:49:48,750 --> 00:49:51,790 Get approved on you. Then you hear when they fail after. 787 00:49:52,670 --> 00:49:56,030 But I mean, like, but you don't know, like day to day. Like, how many 788 00:49:56,030 --> 00:49:59,710 chemical compounds did they try that didn't work out? Right. Maybe it was a hundred. 789 00:50:00,110 --> 00:50:03,270 Right. But that one, if you look at pharmaceutical. It's an 790 00:50:03,270 --> 00:50:06,270 astronomical percentage. It's actually. Right. 791 00:50:07,150 --> 00:50:10,750 Truly insane. Like such a low percentage of what actually makes it 792 00:50:10,750 --> 00:50:14,500 to. There was an interesting analysis. There was some podcast somewhere. But 793 00:50:14,500 --> 00:50:18,060 basically how venture capital works. Right. Like they give money to like 794 00:50:18,060 --> 00:50:21,300 100 companies. Right. 80 of them are going to fail big. 795 00:50:21,540 --> 00:50:25,140 Right. 10 to be, you know, they'll break even. 796 00:50:25,460 --> 00:50:29,260 But like one or two of the remaining 10% knock it 797 00:50:29,260 --> 00:50:32,900 out of the park, Right? Yep. And that's kind of how 798 00:50:32,900 --> 00:50:36,020 mathematically they function. I thought that was an interesting. 799 00:50:36,830 --> 00:50:39,630 Maybe these AI projects or whatever 800 00:50:41,150 --> 00:50:44,870 will follow the same trajectory. I don't know. But I feel better 801 00:50:44,870 --> 00:50:48,430 at 45% success rate than 15 or 802 00:50:48,430 --> 00:50:51,550 5. Yeah. Yeah. Absolutely. 803 00:50:52,910 --> 00:50:56,550 Cool. Always good having you on the show. I 804 00:50:56,550 --> 00:51:00,350 know we both have hard stops. Yes. Unfortunately. 805 00:51:00,590 --> 00:51:03,590 No, it's cool. Gotta have you on more often, man. Especially now that you're not 806 00:51:03,590 --> 00:51:07,200 like spending a month out in, you know, 807 00:51:07,360 --> 00:51:09,120 Australia and Asia. Yeah, 808 00:51:10,800 --> 00:51:13,440 yeah. So let us know in the comments below what you want to see us 809 00:51:13,440 --> 00:51:17,120 to cover and maybe it'll be tomorrow. 810 00:51:17,120 --> 00:51:20,880 I got this here the other day. This is a flexible 811 00:51:21,760 --> 00:51:25,480 solar panel thing. Oh, cool. So it's cool. Supposedly it's 100 812 00:51:25,480 --> 00:51:29,120 watts and you can actually pack it in your 813 00:51:29,120 --> 00:51:32,270 backpack. That's the video. And I was like, oh, I need that because. Because I'm 814 00:51:32,270 --> 00:51:35,310 a big, I'm a big fan of like, you know, having power on the go 815 00:51:35,310 --> 00:51:38,950 and stuff like that. So. So I'll, 816 00:51:38,950 --> 00:51:42,350 I'll unbox that tomorrow. Any parting thoughts? 817 00:51:44,750 --> 00:51:48,430 Just keep an open mind about AI and 818 00:51:48,830 --> 00:51:52,350 I, I still think the, the biggest conversations are still about 819 00:51:52,510 --> 00:51:56,270 the governance of AI. Absolutely. Yeah. Just know that 820 00:51:56,270 --> 00:52:00,090 AI is a multi layered problem, not just a single layered 821 00:52:00,090 --> 00:52:03,930 problem. And for us to get this right, we have to look 822 00:52:03,930 --> 00:52:07,770 at all the different layers. Absolutely. That's 823 00:52:07,770 --> 00:52:11,370 how we're going to be able to do it correctly. And I will tell you, 824 00:52:11,370 --> 00:52:14,570 I was listening to a podcast, I'll leave you on this note. And there was 825 00:52:14,570 --> 00:52:16,370 one expert that was talking about 826 00:52:18,130 --> 00:52:21,330 basically, are we, are we creating the 827 00:52:21,330 --> 00:52:24,830 terminator out of all this? And he, he said, I 828 00:52:25,070 --> 00:52:28,710 I'm actually more worried that we're creating Wall E out of all 829 00:52:28,710 --> 00:52:32,430 this. Interesting. 830 00:52:33,070 --> 00:52:36,350 And I would encourage everyone who hasn't seen Wall E go check it out. 831 00:52:36,670 --> 00:52:40,349 And keep that in the back of your mind too, that there 832 00:52:40,349 --> 00:52:43,070 could be such a happy path with AI that 833 00:52:44,750 --> 00:52:48,350 also has its own long term negative effects for 834 00:52:48,430 --> 00:52:52,270 society. But. But yeah, that's a topic that you. 835 00:52:52,270 --> 00:52:55,670 And I can talk about on our next stream. That's it? 836 00:52:55,910 --> 00:52:59,750 You want to leave on a cliffhanger, so to speak? Yes. And that wraps 837 00:52:59,750 --> 00:53:03,350 our deep dive with Christopher Newland proving once again that AI 838 00:53:03,350 --> 00:53:06,950 isn't just about large language models spitting out cat facts, but 839 00:53:06,950 --> 00:53:10,470 about simulating reality, bending time at devcon and 840 00:53:10,470 --> 00:53:13,990 maybe, just maybe, preventing the rise of our robot overlords. 841 00:53:14,390 --> 00:53:18,230 From protein folding to Grand Theft Auto fueled AI breakthroughs. 842 00:53:18,500 --> 00:53:22,180 Christopher reminded us that the next big leap might not be in scale, but 843 00:53:22,180 --> 00:53:25,740 in simulation. So thanks to Christopher for navigating the 844 00:53:25,740 --> 00:53:29,380 uncanny valley with us. No jet lag, just pure insight. 845 00:53:29,700 --> 00:53:33,460 Until next time, stay data driven. And remember, if 846 00:53:33,460 --> 00:53:36,700 reality starts glitching, blame the simulator, not the 847 00:53:36,700 --> 00:53:37,530 Internet.