1 00:00:01,240 --> 00:00:04,629 Ray, how are you this afternoon or morning your time? 2 00:00:04,629 --> 00:00:05,952 How are you this morning? 3 00:00:05,984 --> 00:00:09,155 Yeah, it's morning time and yeah, hey Ted, feeling really good? 4 00:00:09,155 --> 00:00:09,997 Yeah. 5 00:00:10,094 --> 00:00:10,844 Good, man. 6 00:00:10,844 --> 00:00:11,775 Good. 7 00:00:11,775 --> 00:00:13,755 It's good to have you on the show. 8 00:00:14,033 --> 00:00:17,417 I enjoy reading your content on LinkedIn. 9 00:00:17,974 --> 00:00:35,735 I was looking at your global AI regulation tracker and knew we had to have a conversation because, and how timely, uh we are in the midst of all sorts of regulatory movement, both 10 00:00:35,735 --> 00:00:36,839 US and China. 11 00:00:36,839 --> 00:00:38,056 So we'll get into that. 12 00:00:38,056 --> 00:00:39,086 But before 13 00:00:39,086 --> 00:00:43,691 Before we do, let's get you introduced. 14 00:00:43,691 --> 00:00:49,167 you're a lawyer and a tech lawyer and a developer, and you don't see that a lot. 15 00:00:49,167 --> 00:00:57,386 Why don't you tell us a little bit about your background and how you ended up being a tech lawyer and application developer. 16 00:00:57,624 --> 00:00:58,755 Yeah, yeah. 17 00:00:58,755 --> 00:00:59,835 So... 18 00:01:00,156 --> 00:01:05,067 My story's gonna probably be one of those that starts off really random. 19 00:01:05,067 --> 00:01:09,093 You don't know where it's going, or eventually fall into place, makes sense. 20 00:01:09,093 --> 00:01:10,044 So... 21 00:01:10,244 --> 00:01:14,386 It really started when I was 5 or 6 years old. 22 00:01:14,494 --> 00:01:21,613 I was really into shows like Thomas the Tank Engine, Astro Boy, Iron Man. 23 00:01:22,154 --> 00:01:27,328 Anything that's got to do with machines, you know, coming alive and doing cool stuff. 24 00:01:27,470 --> 00:01:30,852 That has been my fascination even till today. 25 00:01:31,112 --> 00:01:36,996 And so, you know, throughout primary school and high school, I've always liked building things on the side. 26 00:01:36,996 --> 00:01:40,058 So in primary school, I would like to build my own toys. 27 00:01:40,058 --> 00:01:47,322 And in high school, you know, I started playing computer games and then started to pick up coding to learn how to build my own computer games. 28 00:01:47,322 --> 00:01:50,704 And what really helped was that my friends were also into the same thing. 29 00:01:50,704 --> 00:01:54,676 So we're all motivating each other and just doing cool things together. 30 00:01:55,390 --> 00:02:07,096 I think it was then during high school that I started reading, going to the library, and then I was also interested in books around, know, all of the detective genres, and also 31 00:02:07,096 --> 00:02:10,137 legal thrillers, like the John Grisham series. 32 00:02:10,518 --> 00:02:19,902 And that's where I got into, I really enjoy stories that talk about uh evidence and, 33 00:02:20,398 --> 00:02:22,790 trying to connect pieces together to uncover the truth. 34 00:02:22,790 --> 00:02:25,941 And even though along those lines, I find really fascinating. 35 00:02:26,042 --> 00:02:41,292 And so when it got to the end of high school, what I want to do for further study, I was really looking at either law or computer science, given my ongoing interests. 36 00:02:41,393 --> 00:02:49,678 And it came down to a really simple conclusion, which is that I can make coding a hobby, but I can't make law a hobby. 37 00:02:49,824 --> 00:02:52,145 So why not make law my career? 38 00:02:52,145 --> 00:02:56,097 And I was, you know, continuing my coding on the side. 39 00:02:56,398 --> 00:02:59,409 And so all my friends did computer science. 40 00:02:59,409 --> 00:03:09,065 I was the only one who went on to do law, but I still, you know, continued building games and websites just to, you know, as a way to still connect to my friends. 41 00:03:09,385 --> 00:03:14,888 And yeah, throughout uni, I was doing my legal studies and building random maps. 42 00:03:15,089 --> 00:03:17,902 I think, what was like a 43 00:03:17,902 --> 00:03:34,262 A big turning point was I was looking for work as a paralegal and I wasn't able to find one until I started to just look into internships in non-traditional law firms. 44 00:03:34,502 --> 00:03:39,922 So this was around 2017-18. 45 00:03:39,922 --> 00:03:44,582 I applied for an internship at a legal tech startup called Law Path. 46 00:03:44,926 --> 00:03:47,638 and I was doing as a media intern. 47 00:03:47,638 --> 00:03:52,521 So I was writing blog articles around, I think back then it was blockchain and smart contracts. 48 00:03:52,521 --> 00:03:54,453 That was the craze back then. 49 00:03:54,453 --> 00:03:56,674 So I was writing a lot of articles on that. 50 00:03:56,674 --> 00:04:02,538 And then around that time, I also did a hackathon, one of the world's first legal tech hackathon. 51 00:04:02,898 --> 00:04:05,620 And I was representing my university. 52 00:04:05,740 --> 00:04:07,732 I was the developer within my team. 53 00:04:07,732 --> 00:04:14,446 So I was building a prototype app, which basically helps streamline payments between the client 54 00:04:14,510 --> 00:04:18,930 and the Barrister, which is a very, you know, UK, Australia unique thing. 55 00:04:19,130 --> 00:04:25,450 We don't have the Barrister and sister that merge as an attorney like in the US, but basically it's just an app. 56 00:04:25,450 --> 00:04:27,010 It's like a payment app, right? 57 00:04:27,010 --> 00:04:36,470 And then when I pitched the app, it turns out that one of the judges was the boss of that legal tech startup that I was interning at. 58 00:04:36,470 --> 00:04:38,478 And so when I came back, the boss said, 59 00:04:38,478 --> 00:04:40,978 Hey Ray, I didn't know you could code like that. 60 00:04:40,978 --> 00:04:46,438 Why don't you move into our engineering department rather than the comms media department? 61 00:04:46,438 --> 00:04:48,018 And I was like, yeah, sure. 62 00:04:48,038 --> 00:04:54,298 And then that's how I got my actual job as like a legal engineer at that startup. 63 00:04:54,298 --> 00:04:57,598 And I was helping build out the document automation platform. 64 00:04:57,598 --> 00:05:06,758 And so that was like the first serious, I guess, gig that made me think, oh, wow, okay, tech and law can be combined in some way into some actual. 65 00:05:06,818 --> 00:05:17,943 meaningful career and since then when I was applying for graduate roles and you know Becoming a lawyer it has always been my path to sort of go down that tech lawyer sort of 66 00:05:17,943 --> 00:05:28,188 route and that's where I am right now I'm currently a practicing tech lawyer at Herbert Smith Freehills Kramer So and uh which is a global law firm and I do a lot of stuff within 67 00:05:28,188 --> 00:05:33,830 the AI uh legal space but at the same time, you know uh 68 00:05:33,922 --> 00:05:36,673 So that's that's like how my career sort of evolved. 69 00:05:36,673 --> 00:05:49,057 But on the side, as I said, I'm still building things and the AR regulation tracker is just one of those sort of uh projects that really developed as I was like, got really 70 00:05:49,057 --> 00:05:54,519 interested in AR ethics and robots, but I'm sure we can talk more about that one. 71 00:05:54,519 --> 00:05:57,410 But that's basically the rundown of my career today. 72 00:05:57,472 --> 00:05:58,342 Interesting. 73 00:05:58,342 --> 00:06:06,485 And yeah, so let's talk about the global AI regulation tracker that kind of started off as a personal project for you. 74 00:06:06,485 --> 00:06:13,148 And then it's kind of got in traction and now is an industry resource. 75 00:06:13,248 --> 00:06:16,179 like take us back to the origin story around that. 76 00:06:16,179 --> 00:06:27,243 think the last time you and I spoke, was you were procrastinating a reading about trolley problems and machine ethics and tell us about how this thing came about. 77 00:06:27,342 --> 00:06:31,662 Yeah, basically it started with procrastination. 78 00:06:31,662 --> 00:06:36,422 I think during uni there were some subjects which were a bit, you know, a bit of a bludge. 79 00:06:36,502 --> 00:06:41,422 was just, I'm a person who likes history and geography. 80 00:06:41,422 --> 00:06:46,742 So I like reading, you know, random articles on geopolitics and history. 81 00:06:46,742 --> 00:06:56,134 And I came across this YouTube documentary and this was around a time where self-driving vehicles were being trialed, experimented. 82 00:06:56,778 --> 00:06:59,899 and know, accidents, unfortunate accidents were happening. 83 00:06:59,899 --> 00:07:03,781 And then there videos just talking about the ethics in self-driving vehicles. 84 00:07:03,781 --> 00:07:15,085 So if a car were to, if it can't stop and can't change tracks, you, and it was approaching either an elderly person or a baby, which one should the car hit to minimize like damage? 85 00:07:15,085 --> 00:07:25,960 And that's such a tough question that it just opened up the whole door of AR ethics, which I never thought was a theme, but you know, the more I read into it, you know, reading 86 00:07:25,960 --> 00:07:26,562 about 87 00:07:26,562 --> 00:07:30,845 the three laws of robots and then the Trotty problem. 88 00:07:30,845 --> 00:07:39,531 And it was around that time where the EU government was also thinking about the EU AI Act, which today is an actual thing in effect. 89 00:07:39,531 --> 00:07:40,692 But like, this is 2019. 90 00:07:40,692 --> 00:07:44,754 This is just an idea just being floated around in government. 91 00:07:44,754 --> 00:07:49,978 And the whole idea of regulating AI back then was so alien. 92 00:07:49,978 --> 00:07:54,841 It was such a foreign concept that, wow, I never thought this could be an actual thing. 93 00:07:55,262 --> 00:07:56,140 And so, 94 00:07:56,140 --> 00:08:00,432 And as I was reading, I like to sort of write notes in a notebook. 95 00:08:00,633 --> 00:08:08,216 I don't know, like I didn't really have any particular reason why I just thought, you know, if there's anything interesting, I'll just write it down in the diary. 96 00:08:08,397 --> 00:08:11,598 And it got to a point where my diary started filling out. 97 00:08:11,959 --> 00:08:18,002 And then I was having the conversation with friends and also classmates and eventually colleagues. 98 00:08:18,002 --> 00:08:25,484 And I realized that, well, I actually have sort of quite a lot of ideas and insight over the past, you know, 99 00:08:25,484 --> 00:08:36,410 or threes of reading randomly and I thought you know maybe I could share this on LinkedIn so um when I first started becoming a lawyer on the side you know it was also during COVID 100 00:08:36,410 --> 00:08:47,406 so I started my career during the COVID lockdown so there's a bit of a lot of quite a lot of spare time to just play around so I was just writing stuff on LinkedIn initially like 101 00:08:47,406 --> 00:08:54,650 it only hit a very niche audience so I'll write updates around AR regulation and ideas 102 00:08:54,894 --> 00:08:57,675 um I did that for like a year or so. 103 00:08:58,275 --> 00:09:02,076 Modest engagement, but I had a lot of fun writing. 104 00:09:02,076 --> 00:09:12,579 And then when ChaiGPD came out, that's where everything changed because all of sudden, total proof of AI and AI regulation became like really popular and my posts started 105 00:09:12,579 --> 00:09:14,019 getting bit more attention. 106 00:09:14,019 --> 00:09:15,920 So that encouraged me to write more. 107 00:09:15,920 --> 00:09:18,781 I was also writing all around the world. 108 00:09:18,781 --> 00:09:22,322 not just one particular country, but as many countries as possible. 109 00:09:22,322 --> 00:09:23,726 It got to a point where 110 00:09:23,726 --> 00:09:30,706 my LinkedIn had all these posts in different countries and I thought, no, let's try to organize into one hub. 111 00:09:30,866 --> 00:09:32,566 And so I already had a website myself. 112 00:09:32,566 --> 00:09:38,726 So I just thought, you know, why not just add a new page to our website that's categorizes all my LinkedIn posts per country. 113 00:09:38,766 --> 00:09:42,906 So initially it was like a simple table, but then I thought that's so boring. 114 00:09:42,906 --> 00:09:44,626 Let's just take this a step further. 115 00:09:44,626 --> 00:09:45,786 I create upside. 116 00:09:45,786 --> 00:09:50,526 I put like a map on it and click on the country to show my LinkedIn posts. 117 00:09:50,526 --> 00:09:51,594 Then I thought, 118 00:09:51,594 --> 00:09:53,645 Why stop on my own LinkedIn page? 119 00:09:53,645 --> 00:09:58,979 This is actually prepare an uh encyclopedia summary to each country. 120 00:09:59,239 --> 00:10:04,333 And yeah, that's where I started summarizing each country's AI policies and regulations. 121 00:10:04,333 --> 00:10:09,987 I initially started with the G20 countries, but then I've sort of expanded. 122 00:10:09,987 --> 00:10:14,070 I've been running this project now for three months all by myself. 123 00:10:14,070 --> 00:10:21,024 And now it took me two and a half years to now cover every country uh and territory in the world. 124 00:10:21,110 --> 00:10:22,131 So over like 200. 125 00:10:22,131 --> 00:10:31,135 yeah, so I think it's been really great to be, I guess, one of the early ones building this sort of tool. 126 00:10:31,135 --> 00:10:37,298 And then a lot of people were really supportive and yeah, just a lot of encouragement from out the global industry. 127 00:10:37,298 --> 00:10:44,352 And that really helps me complete the map and also add new features to it to make it more user friendly and developer friendly. 128 00:10:44,352 --> 00:10:45,302 yeah. 129 00:10:46,103 --> 00:10:51,673 Yeah, so is your does your firm leverage the research that you've compiled? 130 00:10:53,471 --> 00:10:55,052 not, not directly. 131 00:10:55,052 --> 00:10:58,363 I think this is something that I do like in my own personal time. 132 00:10:58,363 --> 00:11:09,760 Um, and just everyone around the industry is mostly targeted to less say, you know, small businesses, academics, researchers, and developers, especially because there's now a new 133 00:11:09,760 --> 00:11:16,133 API that developers can now link the apps on top of it to sort of run the own monitoring tools, whatever. 134 00:11:16,133 --> 00:11:20,606 So it's more targeted at that sort of grassroots smaller end. 135 00:11:20,606 --> 00:11:21,386 Yeah. 136 00:11:21,985 --> 00:11:24,890 Do you have any plans for this? 137 00:11:24,890 --> 00:11:31,601 to get funding or, you know, either through a grant or private funding to help get some help with it? 138 00:11:31,601 --> 00:11:32,663 This sounds like a lot of work. 139 00:11:32,663 --> 00:11:34,096 uh 140 00:11:34,096 --> 00:11:40,189 yeah, it's um, it's it sounds like work, but it's actually like, it's not that much work. 141 00:11:40,189 --> 00:11:48,042 Because I because I progressively updated every day only takes like five minutes of my time each day just to like monitor updates. 142 00:11:48,042 --> 00:11:52,024 I've got a lot of tools in the background to help curate news items for me. 143 00:11:52,024 --> 00:11:54,395 So it's not a lot of work per day. 144 00:11:54,395 --> 00:11:55,865 But we put it all together. 145 00:11:55,865 --> 00:11:57,816 Sounds like it's quite a lot. 146 00:11:58,036 --> 00:11:59,637 in terms of your other question. 147 00:11:59,637 --> 00:12:00,207 Yeah, sure. 148 00:12:00,207 --> 00:12:02,274 Like I'm always on the, you know, 149 00:12:02,274 --> 00:12:06,139 on the lookout for opportunities, I'm also keeping myself open-minded. 150 00:12:06,139 --> 00:12:08,842 I'm also not desperate for it. 151 00:12:08,842 --> 00:12:15,029 It's something that's nice to have at end of the day for me, just to learn about the world. 152 00:12:15,029 --> 00:12:16,230 eh 153 00:12:16,811 --> 00:12:17,431 Interesting. 154 00:12:17,431 --> 00:12:22,493 yeah, and you're, you're, LinkedIn posts have, have gotten a lot of traction. 155 00:12:22,493 --> 00:12:24,103 mean, you're on the other side of the world. 156 00:12:24,103 --> 00:12:25,814 um And I've seen your stuff. 157 00:12:25,814 --> 00:12:31,505 You're a top voice, which, uh know, that's a, that's a hard designation to get. 158 00:12:31,505 --> 00:12:36,367 um You really have to put in some effort and some work around that. 159 00:12:36,547 --> 00:12:40,118 Well, you've had, so we're recording this in the beginning of August. 160 00:12:40,118 --> 00:12:43,831 This will probably come out towards the later part of the month. 161 00:12:43,831 --> 00:12:46,762 But you've had a busy week or so, right? 162 00:12:46,762 --> 00:12:57,545 Because we've had major developments with the Trump administration announcing their AI action plan and then a very quick turnaround on a China response. 163 00:12:57,545 --> 00:13:01,086 um what is your, yeah. 164 00:13:01,086 --> 00:13:04,766 And you had some great posts that I thought were interesting. 165 00:13:05,427 --> 00:13:12,969 specifically on the China side, there's a lot on the U S side too, but you, let's start with, with China first. 166 00:13:12,969 --> 00:13:13,889 So. 167 00:13:13,985 --> 00:13:20,831 You kind of zeroed in on some nuance around language um with the China plan. 168 00:13:20,831 --> 00:13:26,785 Like, and if I read your post correctly, it was like 95 % of this is not new. 169 00:13:26,785 --> 00:13:30,128 Um, but the 5 % that is, is interesting. 170 00:13:30,128 --> 00:13:36,546 um tell us what your kind of take is on the, the, the China piece first. 171 00:13:36,546 --> 00:13:38,967 Yeah, yeah, yeah, of course. 172 00:13:39,347 --> 00:13:46,190 I guess I'll just first lay out sort of the macro context behind China's AI policy thinking. 173 00:13:46,190 --> 00:13:57,074 And this is based on, you know, both research and also being in the country talking to like the big tech, like companies are driving this sort of change. 174 00:13:57,655 --> 00:14:05,378 So I think really it comes down to one or two things, which is China has a very strong push for 175 00:14:05,496 --> 00:14:08,248 what they call like, know, sovereignty in AI. 176 00:14:08,248 --> 00:14:11,450 So being self-sufficient in the full stack. 177 00:14:11,811 --> 00:14:23,639 And part of this has been driven because of the pressure from US export controls, limiting access to the necessary chips that are required to build really advanced AI systems. 178 00:14:23,639 --> 00:14:33,696 So this whole central theme around self-sufficiency, having control of the full stack, that is like the major theme and that sort of... 179 00:14:33,814 --> 00:14:44,180 Manifesting itself in other smaller sub themes around, know, where with which certain industries will require investment and trade policies and all that. 180 00:14:44,380 --> 00:14:53,085 And the second thing is like China is also trying to uh lead in standards, especially for the global South. 181 00:14:53,325 --> 00:15:01,934 There's all part of like the whole BRICS initiative, all part of the Belt and Road project, which has been ongoing for a decade already. 182 00:15:01,934 --> 00:15:14,214 So there's that also that mindset just really to set the standard because standards are important because if you think about the internet, the internet's built on US led 183 00:15:14,214 --> 00:15:21,534 standards and that has given the US a lot of leverage over how the internet ecosystem should operate. 184 00:15:21,594 --> 00:15:29,094 And it's one of those things where it's a hugely contested front and actually it has a huge role in geopolitics. 185 00:15:29,094 --> 00:15:30,766 So when it comes to the new 186 00:15:30,766 --> 00:15:40,326 breakthrough technology like AI being like the next, the current general purpose technology that is as big or even bigger than the internet, yeah, obviously that's where 187 00:15:40,326 --> 00:15:44,046 countries start thinking about, okay, let's be the ones to set the standard. 188 00:15:44,046 --> 00:15:46,446 So that's the macro context in mind. 189 00:15:46,446 --> 00:16:00,014 And so when the AI action plan came out from China, and to be accurate, when we translate in English, it's called like the AI action plan, but the actual Chinese like, 190 00:16:00,014 --> 00:16:04,294 text, it's actually called the global AI governance sort of action plan. 191 00:16:04,294 --> 00:16:10,574 So it's like a, it has a global sort of mindset embedded into it. 192 00:16:10,574 --> 00:16:24,014 And the 95%, which I said was not new, that's basically the sort of stuff that we've seen in previous papers and also in government representative speeches around, know, as I said, 193 00:16:24,014 --> 00:16:28,654 securing the full stack, investing in green and sustainable ways of 194 00:16:28,654 --> 00:16:31,415 powering models, all that sort of stuff. 195 00:16:31,735 --> 00:16:39,858 The 5 % which I thought was sort of was interesting and highlighted was around open source. 196 00:16:40,259 --> 00:16:49,863 So I think when I say open source in China, people often think of DeepSeek, which that is really the big um milestone that we saw. 197 00:16:49,863 --> 00:16:58,712 um So, but before DeepSeek, it has always been like this sort of uh strategy of 198 00:16:58,712 --> 00:17:01,704 tech clients to release open source products. 199 00:17:01,945 --> 00:17:12,424 And there's a lot of reason why open source is such a huge theme in China is, but there's all of like, I think fundamentally it's because the internal domestic competition is so 200 00:17:12,424 --> 00:17:13,434 fierce. 201 00:17:13,695 --> 00:17:21,061 People often talk about the US and China competition as like the first layer of competition, who are actually in China. 202 00:17:21,441 --> 00:17:26,958 Companies care more about the competition with their next door neighbor, which is like the domestic. 203 00:17:26,958 --> 00:17:36,220 competition and so fierce, there is sort of a race to the bottom in terms of who can produce the highest quality model for the lowest price. 204 00:17:36,501 --> 00:17:45,863 And initially there was a race towards like who can provide the lowest API options until some of the big companies were like, actually, like, not just open source it? 205 00:17:45,863 --> 00:17:49,284 That's that's technically zero, zero dollars for free. 206 00:17:49,284 --> 00:17:54,866 So you basically beat everyone on the on the price front and just provide a very powerful model. 207 00:17:54,866 --> 00:17:56,846 And so already for like 208 00:17:56,846 --> 00:18:07,046 For a year or two all the big tech companies in China were releasing their own open source AR models What made DeepSea quite special is that it found like a new ways to make the 209 00:18:07,046 --> 00:18:15,906 training process even more cheaper and That caused a lot of headlines in the West as well as well of attention has been brought to DeepSea even though it's part of a it's only a 210 00:18:15,906 --> 00:18:25,336 small part of the bigger open source picture But what this plan So even though open source has been a long thing what this plan was quite different was that 211 00:18:25,336 --> 00:18:30,059 that its choice of language was uh very selective. 212 00:18:30,059 --> 00:18:38,054 And when it comes to Chinese policies, like the language itself probably says more about the story than the actual message. 213 00:18:38,375 --> 00:18:44,318 Certain words are selected to convey a certain sentiment. 214 00:18:44,419 --> 00:18:54,115 And one of the things to look out for is which phrases are being repeated, like which mantras and which combination of words are repeated throughout the paper. 215 00:18:54,115 --> 00:18:55,118 That's often the 216 00:18:55,118 --> 00:18:58,899 indicative of government thinking. 217 00:18:59,259 --> 00:19:05,731 like for like I say, for the past year, let me just bring out my notes. 218 00:19:05,731 --> 00:19:20,145 For the past year, there was like a um particular phrasing that government will use and it was called, to translate directly into English, is, know, uh safe, reliable and 219 00:19:20,145 --> 00:19:20,925 controllable. 220 00:19:20,925 --> 00:19:25,280 So these are the, that's a typical trio of words that we see 221 00:19:25,280 --> 00:19:28,351 in speeches and it's a very systems focused view. 222 00:19:28,351 --> 00:19:35,134 So it's all about trying to make any particular use case safe, reliable, controllable. 223 00:19:35,134 --> 00:19:48,820 But since then we start to see the repetitive slogan expanding more into broader terms to now what we call, again, direct translation, inclusive, open, sustainable, fair, secure 224 00:19:48,820 --> 00:19:52,221 and reliable, digital and intelligent future for all. 225 00:19:52,221 --> 00:19:55,212 So it's a mouthful when I say in English, but it's only like eight. 226 00:19:55,212 --> 00:19:56,583 eight characters in Chinese, right? 227 00:19:56,583 --> 00:20:04,948 So, that stuff repeats a lot throughout the policy and a much more global rather than system specific focus. 228 00:20:05,329 --> 00:20:08,201 And how that relates to open source? 229 00:20:08,201 --> 00:20:23,471 Well, in the actual paragraph that mentions open source, we in English, we call it open source, but in Chinese, it's technically open sharing of resources and nowhere in the 230 00:20:23,471 --> 00:20:24,922 actual text 231 00:20:25,228 --> 00:20:31,380 have I seen the words open source code, open source software, or open source models? 232 00:20:31,641 --> 00:20:40,164 Now, again, in English, when we say open source, we tend to mean that one thing, is, know, putting your stuff on GitHub, everyone can see the code and you can download it. 233 00:20:40,164 --> 00:20:46,347 But in Chinese, open source has so many different ways of expressing that one concept. 234 00:20:46,347 --> 00:20:52,830 And if you wanna talk about open source models, open source code, there's an actual literal direct way of saying that. 235 00:20:52,830 --> 00:20:54,094 So it's not a draft. 236 00:20:54,094 --> 00:20:59,414 I don't think it's a draft in oversight because there's so many different ways of expressing that one thing. 237 00:20:59,414 --> 00:21:07,574 There's got to be some conscious effort behind why it's only open sharing of resources compared to, let's say, open source models or code. 238 00:21:07,574 --> 00:21:11,914 And as I said, when it comes to Chinese policies, you have to read into the language. 239 00:21:11,914 --> 00:21:15,234 it's not, I'm not, I don't think I'm reading too deeply into it. 240 00:21:15,234 --> 00:21:17,774 I think it's meant to be read in that way. 241 00:21:17,994 --> 00:21:22,402 And taking that interpretation, if we're only talking about sharing, 242 00:21:22,402 --> 00:21:29,448 tech documentation, manuals, like the surface layer documentation instead of the actual code. 243 00:21:29,549 --> 00:21:42,019 What I'm thinking is that this is such a clever policy balance by China to sort of influence global standards, but also keeping the secret source back at home, which is a 244 00:21:42,340 --> 00:21:44,412 very subtle and clever sort of balance. 245 00:21:44,412 --> 00:21:46,734 So that's what I noticed in this policy. 246 00:21:46,734 --> 00:21:49,046 And again, it will take another few... 247 00:21:49,058 --> 00:21:53,220 policies or papers in the future to see if that message is being reinforced. 248 00:21:53,220 --> 00:22:03,316 But until then, this is like the first one that I think might be that slight pivot towards that selective open sharing technique. 249 00:22:03,351 --> 00:22:11,103 You know, what's interesting is us in the West, like open China is like an oxymoron. 250 00:22:11,103 --> 00:22:14,224 We don't think of China and open anything. 251 00:22:14,224 --> 00:22:21,226 We think of very, you know, closed, controlled, um, not open. 252 00:22:21,466 --> 00:22:31,409 And, this seems like a, again, from a Westerner standpoint, it seems like a departure from what we would expect. 253 00:22:31,427 --> 00:22:47,289 from China, which again, the great Chinese firewall and just how um there's also been uh issues around intellectual property rights within China. 254 00:22:47,289 --> 00:22:57,937 um I guess, again, from a Westerner standpoint, it seems surprising that China wants to have an open policy. 255 00:22:58,079 --> 00:22:58,891 around this. 256 00:22:58,891 --> 00:23:01,518 oh What about on your side of the globe? 257 00:23:01,518 --> 00:23:07,311 Is this surprising or does this line up exactly the direction you thought they'd head? 258 00:23:09,002 --> 00:23:20,820 For me, it's not surprising because I think the commercial drivers sort of explain the story why there's a drive towards um open source in the say from the Western definition 259 00:23:20,820 --> 00:23:21,331 standpoint. 260 00:23:21,331 --> 00:23:28,566 As I said, the domestic competition is already so fierce that really the only real way to stand out is to be open source. 261 00:23:28,566 --> 00:23:37,688 And actually, if you talk to local developers in China, if an AI company doesn't have an open source version of their product, they're not going to be considered 262 00:23:37,688 --> 00:23:45,975 by the developers in the tech stack because even if you don't use the open source tool, it's sort of like a fashion statement, right? 263 00:23:45,975 --> 00:23:53,232 Saying that, okay, like we're doing open source so that we know that we are within the top band of the market. 264 00:23:53,232 --> 00:23:56,295 If we don't do open source, it means that why you're hiding, right? 265 00:23:56,295 --> 00:23:59,468 That's sort of the suspicion that you get from the local developer base. 266 00:23:59,468 --> 00:24:03,411 So I think in China, open source is the market expectation. 267 00:24:03,411 --> 00:24:04,742 It's the market standard. 268 00:24:04,974 --> 00:24:08,715 um Unlike in the West, I think there is a slight pivot today. 269 00:24:08,715 --> 00:24:14,137 I think especially OpenAI doing open weights now, but I think in China is a different story. 270 00:24:14,137 --> 00:24:22,674 yeah, it's just a matter of like, so open source is always gonna be the direction. 271 00:24:22,674 --> 00:24:26,380 I think the question is what extent is open? 272 00:24:26,380 --> 00:24:34,478 And that's where you get really specific with is it open source documents or code or weights or models or the whole thing. 273 00:24:34,478 --> 00:24:40,698 I think that's the question that China over there is still figuring out from a policy perspective. 274 00:24:41,017 --> 00:24:41,877 Interesting. 275 00:24:41,877 --> 00:24:53,197 And, I don't know how, if this is different in China, but here in the U S the lawmakers don't have a clue, um, about the tech. 276 00:24:53,197 --> 00:25:04,437 mean, I, um, until recently, Donald Trump didn't know, know who, uh, Jensen Wang was and, yeah, a $4 trillion company. 277 00:25:04,437 --> 00:25:10,577 And, uh, our president doesn't didn't know who the CEO was, um, by his own admission. 278 00:25:10,701 --> 00:25:24,731 And apparently there's now a lot of dialogue going on and he's surrounded himself with advisors like David Sachs and people who really understand the technology and the, if we 279 00:25:24,731 --> 00:25:30,705 can pivot to the U S for a minute, um, the commentary I've heard, and I haven't read the entire thing. 280 00:25:30,705 --> 00:25:39,120 I've read excerpts, but on the U S side, it sounds like it's written from an informed perspective. 281 00:25:39,121 --> 00:25:40,041 So. 282 00:25:40,178 --> 00:25:56,952 it, whether or not you agree with, cause there's some, there are some, there are some, uh, controversial words in the U S policy around, you know, control over kind of the tone and 283 00:25:56,972 --> 00:26:05,739 wokeness and DEI and all those sorts of things that are very politically charged, um, topics of conversation here in the U S. 284 00:26:05,739 --> 00:26:08,241 But what was your take on 285 00:26:08,267 --> 00:26:15,319 And if I, if I recall correctly, the, U S, action plan came out and China's came out within 48 hours. 286 00:26:15,319 --> 00:26:16,632 was boom, boom. 287 00:26:16,632 --> 00:26:22,241 Um, but what, what, what is your take on kind of the U S's action plan on AI? 288 00:26:24,428 --> 00:26:29,402 Yeah, I wasn't surprised by the action points in the plan. 289 00:26:29,402 --> 00:26:36,158 I can give credit in the sense that they've been consistent in what they're going to do. 290 00:26:36,158 --> 00:26:44,615 just a quick recap, the action plan is a lot, but it's been manifested through three key executive orders. 291 00:26:44,615 --> 00:26:49,659 So the first order is around energy infrastructure, promoting that. 292 00:26:49,659 --> 00:26:53,622 Second order is around the export controls layer. 293 00:26:54,100 --> 00:27:05,306 and the third order is really around what they call trying to regulate, well not trying to regulate, but ensuring that language models do not generate quote unquote work or biased 294 00:27:05,306 --> 00:27:06,017 material. 295 00:27:06,017 --> 00:27:11,079 So these are the three uh key themes of the action plan. 296 00:27:11,079 --> 00:27:15,882 And it's not the first, it's not complete surprise that these were covered. 297 00:27:15,882 --> 00:27:23,626 I think it's been a consistent policy of the government since the change of government in 298 00:27:23,626 --> 00:27:24,927 start of the year. 299 00:27:25,888 --> 00:27:35,636 For me, my focus because I'm more interested in the global dynamics, the second one was the most interesting for me, which is around the export controls. 300 00:27:36,077 --> 00:27:45,464 And so ever since the Biden administration, the US has been tightening export controls around chips. 301 00:27:45,505 --> 00:27:47,520 And to get even more specific, 302 00:27:47,520 --> 00:27:52,643 Initially, was only restrictions around just the actual final chips themselves. 303 00:27:52,643 --> 00:27:58,007 So the final product before it's being shipped to select countries. 304 00:27:58,007 --> 00:28:08,103 the export controls really target like China, Russia, Iran, and there's other sort of like what the USC's as the competitor nations. 305 00:28:08,925 --> 00:28:17,390 But what I found really interesting in the action plan is that there is, it's only a small paragraph within the actual plan, but it mentions of 306 00:28:17,390 --> 00:28:32,190 the government or requiring the DOC to really look into targeted export controls around semiconductor manufacturing components. 307 00:28:33,070 --> 00:28:36,670 So if you think about the full stack, there is the actual chip itself. 308 00:28:36,670 --> 00:28:42,270 take your Nvidia chip or AMD chip, the final full package. 309 00:28:42,270 --> 00:28:44,470 Within that, there's a lot of different components. 310 00:28:45,050 --> 00:28:54,277 So existing controls target both the final one and also recently the components within that chip because there's been a growing sort of recognition within government that, 311 00:28:54,277 --> 00:29:04,003 actually maybe the chip is designed and exported from US, but all the different parts are being imported across the world, right? 312 00:29:04,003 --> 00:29:08,166 And so you've got to mostly make sure you're accountable for these different sub components. 313 00:29:08,666 --> 00:29:14,600 But now they're looking above the value chain as well, which is the actual manufacturing. 314 00:29:15,018 --> 00:29:15,629 of the chips. 315 00:29:15,629 --> 00:29:19,622 So there are already existing controls on manufacturing equipment. 316 00:29:19,622 --> 00:29:23,025 So we're talking about these big lithographic machines. 317 00:29:23,025 --> 00:29:35,276 So just a quick, uh quick explainer, like how these chips are built, like they're very small, like it's, basically a very intricate design on a silicon wafer, right? 318 00:29:35,417 --> 00:29:36,708 But it's really small. 319 00:29:36,708 --> 00:29:40,846 It's as small as like some like adamants is measured in nanometers, right? 320 00:29:40,846 --> 00:29:42,146 How the heck do people do that? 321 00:29:42,146 --> 00:29:52,006 Well, it's all done because there's like this big machine that's sole purpose is to fire a very specific beam of light through a bunch of mirrors. 322 00:29:52,006 --> 00:29:59,026 And that's what carves out these very small designs on a very small piece of silicon at a nanometer sort of level. 323 00:29:59,026 --> 00:30:00,786 That's how these things are created. 324 00:30:00,786 --> 00:30:10,694 Now, these big machines, like these lithographic machines, can only be built by one company, which is ASML, which is a base, a company based in the Netherlands, right? 325 00:30:11,042 --> 00:30:15,124 and that stuff's been bought by companies across the world to build these chips. 326 00:30:15,124 --> 00:30:24,932 So we're talking about the final big machine, but that machine itself, I think reports have said it's built from 700,000 components. 327 00:30:24,932 --> 00:30:31,777 They have components from Germany, from Spain, from France, even from China, from like Southeast Asia, everywhere. 328 00:30:31,777 --> 00:30:36,160 We're talking about the most complex supply chain in history. 329 00:30:36,300 --> 00:30:38,894 And I reckon it builds the... 330 00:30:38,894 --> 00:30:42,254 I reckon it's in the Guinness World Records or something for that complexity. 331 00:30:42,254 --> 00:30:52,374 But anyways, the action plan is considering not just targeting semiconductor manufacturing equipment, but also components within that manufacturing equipment. 332 00:30:52,374 --> 00:31:03,894 Now it might just be two words on the paper, but those two words could have massive complications because if we're gonna export control components within that big machine, 333 00:31:04,074 --> 00:31:06,654 that could potentially cover the whole world. 334 00:31:06,654 --> 00:31:08,258 Like these export controls could 335 00:31:08,258 --> 00:31:14,861 basically target every single supplier that pitches in into that one big machine. 336 00:31:14,962 --> 00:31:24,057 And it's really hard to evaluate like what's the actual specific impact, but all I know is that it's gonna make supply chains really complicated. 337 00:31:24,057 --> 00:31:33,882 And a lot of the costs, the supply side inflation that's happening in the world right now, partly due to oil price, but there's also a lot due to just cheap prices in generally 338 00:31:33,882 --> 00:31:37,502 making the cost of anything tech, any 339 00:31:37,502 --> 00:31:51,548 any good basically that's that's digitized that's they already quite expensive on themselves just from the current um export controls but with with further controls around 340 00:31:51,548 --> 00:32:02,793 the manufacturing components then yeah that yeah i'm just we're gonna brace ourselves for that i'm sure like again the action plan is only indicating that this is an area that the 341 00:32:02,793 --> 00:32:06,092 relevant departments have to look at so just to be clear it's not 342 00:32:06,092 --> 00:32:10,036 is not a direct action, it's just telling the agencies to look into that question. 343 00:32:10,036 --> 00:32:15,463 So I'm sure there'll be a lot of expert analysis into that, but that's one thing just to get your heads up for. 344 00:32:15,463 --> 00:32:16,693 Yeah. 345 00:32:17,725 --> 00:32:24,909 And speaking of export controls, so I banged on DeepSeek a little bit and lately Kimi. 346 00:32:25,129 --> 00:32:34,974 And uh I'm really impressed with, I was really impressed with DeepSeek uh R1 when it came out and I'm really impressed with Kimi. 347 00:32:35,895 --> 00:32:47,437 Are these export controls maybe having an unintended consequence of forcing these uh countries that have constraints? 348 00:32:47,437 --> 00:32:57,887 to become more efficient and creative and engineer more uh interesting solutions to these problems? 349 00:32:57,887 --> 00:33:00,311 Is it having that unintended consequence? 350 00:33:00,856 --> 00:33:09,814 Yeah, it's like kind of like the whole, you know, the famous quotes like, you know, necessity is the model of invention or that sort of thinking. 351 00:33:09,814 --> 00:33:13,617 And actually I'll show you something in Chinese internet meme culture. 352 00:33:13,617 --> 00:33:23,742 Like there's a lot of, it's a pretty popular meme that among Chinese netizens that Trump and Biden are the founders of China or the. 353 00:33:23,742 --> 00:33:28,015 or that it's called nation building fathers of China. 354 00:33:28,015 --> 00:33:29,666 That's like the joke, right? 355 00:33:29,666 --> 00:33:40,012 Because the reason is that their export controls have pressured Chinese industries, have limited their resources so much in a way that they just have to find new ways to build 356 00:33:40,012 --> 00:33:41,134 models. 357 00:33:41,134 --> 00:33:44,056 as you mentioned, Kimmy, but also notably DeepSeek, right? 358 00:33:44,056 --> 00:33:47,978 The fact that, so it's just a quick rundown, like. 359 00:33:49,106 --> 00:33:57,950 The traditional thinking is that in order to build a powerful AR model, you need a lot of labelled data and a lot of processing power to train on the labelled data. 360 00:33:57,950 --> 00:34:09,335 But DeepSeek, based on their paper, says that you can actually build a powerful model with less labelled data, but a lot more from reinforcement learning. 361 00:34:09,335 --> 00:34:14,517 And reinforcement learning, the advantage is that you don't need labelled data to do reinforcement learning. 362 00:34:14,517 --> 00:34:17,037 You have some to get it up to like a 363 00:34:17,037 --> 00:34:19,037 particular sort of like head start. 364 00:34:19,037 --> 00:34:28,279 But from there on, that's like the first 10%, but the rest, the 90%, the model just plays by itself and just learns from its own mistakes and then reapplies them. 365 00:34:28,279 --> 00:34:32,260 then that's the beauty of reinforcement learning. 366 00:34:32,560 --> 00:34:38,601 And then apparently they were able to do it on like older like chips, like H20 chips. 367 00:34:38,601 --> 00:34:43,622 But I think that claim is still being tested by independent experts. 368 00:34:44,002 --> 00:34:55,434 That's an example of really stretching the boundaries of existing legacy tech and finding new software layer, new algorithms to make the most out of your hardware components. 369 00:34:55,434 --> 00:34:59,878 So yeah, I'm sure there is that effect. 370 00:35:00,233 --> 00:35:15,821 Yeah, and like speaking of other policy effects like you know the didn't meta recently refused to sign the EU's uh You know acknowledgement around around their policies am I 371 00:35:15,821 --> 00:35:17,151 correct on that? 372 00:35:17,666 --> 00:35:22,049 Yeah, I think you're referring to the general purpose AI code. 373 00:35:22,049 --> 00:35:24,290 Yeah, it's been uh a... 374 00:35:24,891 --> 00:35:35,638 It's been one of those hotly contentious policy documents in the industry and even caused some sort of division among the big tech companies themselves. 375 00:35:35,638 --> 00:35:39,391 I think it's all been finalized since last week. 376 00:35:39,391 --> 00:35:41,744 So I think there is a list of signatories. 377 00:35:41,744 --> 00:35:43,344 You can see who signed and who's not. 378 00:35:43,344 --> 00:35:47,316 But yeah, I think in the weeks lead up to it, yeah, certain... 379 00:35:47,448 --> 00:35:50,254 Companies have said they'll sign on, some say they won't. 380 00:35:50,254 --> 00:35:51,345 Yeah. 381 00:35:51,928 --> 00:35:52,598 Yeah. 382 00:35:52,598 --> 00:36:04,093 And you know, I think another tone from the U S action plan is there's going to be very little regard for, um, environmental concerns. 383 00:36:04,093 --> 00:36:06,444 It's, you know, build baby build. 384 00:36:06,544 --> 00:36:17,158 And, know, again, this seems like the, this seems a little bit like a, a, a uh flipping of the script, you know, um, historically the picture has been painted that China has had 385 00:36:17,158 --> 00:36:18,849 less regard for. 386 00:36:18,881 --> 00:36:26,005 know, green initiatives and really the West has been putting that more in focus. 387 00:36:26,005 --> 00:36:38,672 And it seems to me the tone of the, you know, this administration's policy is that is in the way in the backseat, maybe in the trunk. 388 00:36:38,672 --> 00:36:46,557 uh First and foremost is about establishing global dominance around AI and maintaining the lead. 389 00:36:46,650 --> 00:36:48,657 Is that the way you read it as well? 390 00:36:50,082 --> 00:36:51,543 Yeah, interesting. 391 00:36:52,044 --> 00:37:05,595 It depends on how you define lead, because I hear lot of commentary around the whole AI race and who's leading what, but it's a very, I personally find it's very simplified view 392 00:37:05,595 --> 00:37:09,358 of how this whole ecosystem works. 393 00:37:09,358 --> 00:37:12,600 First of all, you have to divide it within like layers, right? 394 00:37:12,841 --> 00:37:15,543 And it depends on which layer you're looking at. 395 00:37:15,543 --> 00:37:18,744 So at the app layer, I'd say like, 396 00:37:18,744 --> 00:37:27,339 both China and US have equally diverse and widely used apps at the application layer. 397 00:37:27,739 --> 00:37:38,956 And as you go deeper within the stack, so I think maybe just to really simplify, at the first layer, it's really a question around diversity and USichu has the most diverse and 398 00:37:38,956 --> 00:37:40,867 used ecosystem. 399 00:37:40,867 --> 00:37:42,508 Then we get to the model layer. 400 00:37:42,508 --> 00:37:45,720 That's okay, that's why I can see the race concept being. 401 00:37:45,720 --> 00:37:51,354 being true because it's really a race to who can build the smallest and cheapest model. 402 00:37:51,354 --> 00:37:54,916 That's really what I see the races and it's different approaches, right? 403 00:37:54,916 --> 00:38:06,643 So from a US perspective, it's really driven by the private sector, private sector and also the inter competition trying to produce the cheapest sort of APIs for powerful 404 00:38:06,643 --> 00:38:07,250 models. 405 00:38:07,250 --> 00:38:12,086 Whereas in China is also that domestic competition from an open source level. 406 00:38:13,076 --> 00:38:19,846 At the infrastructure hardware chips layer, I didn't really see it as a race. 407 00:38:19,846 --> 00:38:21,218 It's more like... 408 00:38:23,062 --> 00:38:27,024 you find your own adventure to building self-sufficiency. 409 00:38:27,024 --> 00:38:28,485 That's how I see it. 410 00:38:28,845 --> 00:38:35,509 And you can either do it from a constructive or deconstructive uh approach. 411 00:38:35,509 --> 00:38:44,133 Constructive being, so when I say that, constructive means, for example, subsidies, government investments, promoting trade. 412 00:38:44,174 --> 00:38:46,715 So stuff that kind of helps grow. 413 00:38:46,775 --> 00:38:52,608 And deconstructive, is export controls, tariffs, or other um 414 00:38:52,608 --> 00:38:57,741 anti-free trade policies that try to stifle what your competitors are doing. 415 00:38:57,741 --> 00:39:02,164 And it's not that you can only be constructive, it can't be deconstructive. 416 00:39:02,164 --> 00:39:03,915 You're gonna have a balance between those two, right? 417 00:39:03,915 --> 00:39:05,576 That's how policy works, right? 418 00:39:05,576 --> 00:39:16,732 So I think that layer, it's really, again, choose your own adventure, but your policy mix depends on your current country circumstances. 419 00:39:16,932 --> 00:39:18,033 So that's how I see it. 420 00:39:18,033 --> 00:39:20,134 um 421 00:39:21,230 --> 00:39:34,701 In terms of more broadly, I think it is true that both, I think since 2022 and whole January AI, I think before that, the whole AI policy debate was really around just the 422 00:39:34,701 --> 00:39:38,705 typical safety and reliability, all that stuff. 423 00:39:38,705 --> 00:39:43,708 Since 2022, AI is gonna become a more geopolitical topic. 424 00:39:43,869 --> 00:39:50,254 And so the idea of like, so as I said, the idea of leading is more around 425 00:39:50,444 --> 00:39:57,056 establishing, I think like just who has more influence on standards. 426 00:39:57,056 --> 00:40:03,288 I think that's one specific angle that I can see where there's that strong competition, as I said before. 427 00:40:03,288 --> 00:40:07,719 Once you set the standards, your whole ecosystem becomes sticky. 428 00:40:07,719 --> 00:40:12,750 And when your system becomes sticky, people have to use it, revenue comes in, your GDP booms. 429 00:40:12,750 --> 00:40:16,802 That's like, that's how, I think that's sort of the more long meta strategy. 430 00:40:16,802 --> 00:40:18,442 So I see that. 431 00:40:18,994 --> 00:40:24,198 And you also mentioned the whole green and uh energy thing. 432 00:40:24,499 --> 00:40:28,302 That's also a big part of it because AI consumes lot of power. 433 00:40:28,582 --> 00:40:38,431 I think this is also where it's important not to see AI policy in isolation, but how it interconnects with every other domestic policy of a country. 434 00:40:38,431 --> 00:40:47,638 So AI crosses over into energy policy, also crosses over into land policy, because the amount of land you have to dedicate to data centers. 435 00:40:47,680 --> 00:40:49,622 it crosses over into like tax. 436 00:40:49,622 --> 00:40:50,933 That's a huge thing, right? 437 00:40:50,933 --> 00:40:54,316 Tax incentive and all that to incentivize development. 438 00:40:54,316 --> 00:40:56,670 You to see all this in one big picture. 439 00:40:56,670 --> 00:41:07,647 I think where it's true, at least from the objective stats, is that China does have a huge head start in this space because they have a lot of capacity, like in terms of electric 440 00:41:07,647 --> 00:41:14,172 generation, there's a lot of land that's still being underdeveloped that can be turned into electric plants. 441 00:41:14,172 --> 00:41:16,802 There is a stronger central 442 00:41:16,802 --> 00:41:19,783 government push in energy. 443 00:41:20,083 --> 00:41:23,985 That's been quite a, it's been consistent for many, many years. 444 00:41:24,345 --> 00:41:32,608 The green tech industry over there has a lot of state support and also a of private sector activity. 445 00:41:32,729 --> 00:41:34,780 It's also a very popular STEM subject. 446 00:41:34,780 --> 00:41:40,552 We talked to Chinese developers, a lot of them want to go into energy tech as their engineering field. 447 00:41:40,552 --> 00:41:45,454 So there's that sort of capacity that's all there, that they're just making use of that. 448 00:41:45,454 --> 00:41:47,634 and part of it's going towards AI. 449 00:41:47,634 --> 00:41:51,114 And the US is also now focusing the same efforts now. 450 00:41:51,114 --> 00:41:56,034 I think it's just a good, I think it's a consistent challenge with the West around energy. 451 00:41:56,774 --> 00:42:05,434 There's a lot of debate around which sources you have to use and it doesn't, each particular energy source has its own big policy debate. 452 00:42:05,434 --> 00:42:09,514 But I think in China, they sort of have, they sort of just have that one set. 453 00:42:09,514 --> 00:42:15,318 Like they just go with that one source and then go, so not one source, they go with a certain mix of sources. 454 00:42:15,318 --> 00:42:16,210 And just run with that. 455 00:42:16,210 --> 00:42:21,970 They kind of skip the whole policy debate in the beginning, just go straight to implementation. 456 00:42:22,541 --> 00:42:22,831 Yeah. 457 00:42:22,831 --> 00:42:27,993 And then, we're almost out of time and we've been taught this is, could talk about this stuff all day. 458 00:42:27,993 --> 00:42:33,806 Um, I geek out on, you know, AI in general, but bringing it back to legal. 459 00:42:34,026 --> 00:42:48,332 So what do you, what do you envision and, know, on what sort of timeline do you see legal work and the resources, the inputs, which are mostly human capital today? 460 00:42:48,432 --> 00:42:51,157 When do you see that being disrupted? 461 00:42:51,157 --> 00:43:01,384 where we will see material impact to law firm revenue, law firm headcount, inside uh council processes. 462 00:43:01,384 --> 00:43:06,147 Like today, there's a lot of experimentation and I think there is some impact. 463 00:43:06,147 --> 00:43:10,349 But when do you see real disruption taking place in the legal space? 464 00:43:11,086 --> 00:43:24,766 Yeah, I could draw a graph here, but I feel like it's just a function of the more standardized and the lower risk value of the work, the more prone it is to automation. 465 00:43:25,006 --> 00:43:37,726 And not just AI automation, but just any form of automation, like even your traditional boring, algorithmic sort of if-else statements, that stuff can also act as automation. 466 00:43:38,574 --> 00:43:40,794 as a standardized low risk. 467 00:43:40,794 --> 00:43:53,474 the reason why I say that is because obviously standardized is a consistent process that's really easy to encode into code and low risk being that if something goes wrong, the loss, 468 00:43:53,474 --> 00:43:55,954 the chance of harm is still going to be quite low. 469 00:43:55,954 --> 00:44:04,374 And there's, and it's also one of those like if something goes wrong, it's still easy or still practical for someone to just jump in and fix things. 470 00:44:04,814 --> 00:44:05,966 like, yeah. 471 00:44:05,966 --> 00:44:16,786 Usual suspects that people talk about as like, you know, low value real estate transactions, like mortgages, conveyancing, that's to the extent that's still done by 472 00:44:16,786 --> 00:44:17,866 lawyers. 473 00:44:18,606 --> 00:44:24,326 Some aspects of like loan, like loan contracts, equity, that's like all stock standard terms. 474 00:44:24,786 --> 00:44:30,086 Certain tech contracts, software contracts, again, that's anything that's got to do stock standard terms. 475 00:44:30,086 --> 00:44:31,086 Yeah, definitely. 476 00:44:31,086 --> 00:44:34,866 I mean, that's like the primary, lot of these legal tech startups are targeting. 477 00:44:36,300 --> 00:44:41,353 I'd say that's something that's already in the process of, since the past four, five years. 478 00:44:41,413 --> 00:44:49,038 The next two, three years or so, we'll start to target the more, still relatively standardized and still relatively low risk. 479 00:44:49,038 --> 00:44:55,762 But I'd say this, I think maybe 80 % standardized and an extra 10 % in risk. 480 00:44:55,762 --> 00:44:57,403 That's like the medium state level. 481 00:44:57,403 --> 00:45:01,646 um That's where we start to see stronger. 482 00:45:01,772 --> 00:45:07,444 reasoning capabilities of these models to be able to tackle these sort of semi standardized problems. 483 00:45:07,444 --> 00:45:17,319 So they've still got some consistency in a problem, but there's also a level of customization or nuance thinking that these models have to have to sort of recognize, but 484 00:45:17,319 --> 00:45:21,290 not too nuanced that it sort of confuses the model. 485 00:45:21,791 --> 00:45:29,954 So that's probably where we're getting to again, the same errors I just identified, but the bit more complex like problems. 486 00:45:29,986 --> 00:45:39,444 We also start to see areas like, I say, crime, like sudden, I uh don't know what's the right word to use, but petty crime. 487 00:45:39,444 --> 00:45:42,436 I think that's where you can start using for petty crime. 488 00:45:42,596 --> 00:45:49,181 Also, um yeah, a lot more areas of commercial law, so commercial contracting. 489 00:45:50,403 --> 00:45:59,510 What everyone's really excited about is in the next, I say, eight or 10 years, where we really start tackling highly nuanced legal problems. 490 00:45:59,968 --> 00:46:04,560 And actually, this is where I honestly don't really know what will be the end outcome. 491 00:46:04,560 --> 00:46:11,663 As a practicing lawyer myself, when I say highly nuanced problems, I do mean they're highly nuanced. 492 00:46:11,663 --> 00:46:21,027 I think the common misconception that people have is that like contracts or reading laws, it's all based on what's written on the paper. 493 00:46:21,027 --> 00:46:24,198 As long as you know what's on paper, you can interpret that. 494 00:46:24,198 --> 00:46:26,909 You basically have the full answer. 495 00:46:26,909 --> 00:46:29,614 Actually, no, like the text. 496 00:46:29,614 --> 00:46:33,254 let's say probably only addresses 30 % of your problem. 497 00:46:33,274 --> 00:46:40,874 The 60 % is actually understanding your client's needs, the problem at hand, and also the market. 498 00:46:41,054 --> 00:46:45,274 And the question is, how do you encode all of that into numbers? 499 00:46:45,274 --> 00:46:48,894 That's ultimately what developers have to do. 500 00:46:49,034 --> 00:46:53,814 Encode the legal problem into numbers that can be read by a machine. 501 00:46:53,814 --> 00:46:55,594 That's what you have to do at the end of the day. 502 00:46:55,594 --> 00:46:58,272 How do you encode client interests 503 00:46:58,272 --> 00:47:03,794 marker standard, marker practices, to extent that they're not written down in words or standards. 504 00:47:03,794 --> 00:47:07,606 We're just talking about conversation dialogues and all that. 505 00:47:07,606 --> 00:47:15,669 How do you encode that in a consistent manner that a model can reliably reference for XYZ problems? 506 00:47:15,669 --> 00:47:17,490 I've tried doing that myself. 507 00:47:17,490 --> 00:47:19,591 It's really hard, right? 508 00:47:19,851 --> 00:47:20,601 But who knows? 509 00:47:20,601 --> 00:47:24,223 Maybe at that point, the whole architecture will change. 510 00:47:24,223 --> 00:47:27,192 We're currently still on a transformer architecture. 511 00:47:27,192 --> 00:47:32,482 which is very much a predict the next word, predict next token. 512 00:47:32,482 --> 00:47:34,376 Obviously there's a lot of layers around that. 513 00:47:34,376 --> 00:47:37,808 It's not just that, but fundamentally that's still what happens. 514 00:47:37,808 --> 00:47:45,612 Who knows, it might be a new mainstream model, like the state space model that might allow us to do a way more nuanced reasoning. 515 00:47:46,113 --> 00:47:50,335 Right now, all of the reasoning models are just limited to like chain of thought. 516 00:47:51,135 --> 00:47:53,417 But I think chain of thought is just level one. 517 00:47:53,417 --> 00:47:56,158 There's like way more levels down the line. 518 00:47:56,526 --> 00:48:01,186 which I don't know yet because I'm not within the research centers themselves. 519 00:48:01,186 --> 00:48:08,786 yeah, really, I'm probably, I'm gonna be one of those people who are really optimistic around disruption in law. 520 00:48:08,786 --> 00:48:14,846 It's weird for a lawyer to say that, but I feel like it's gonna be amazing because it'll free up a lot of our time. 521 00:48:14,846 --> 00:48:17,286 I think laws would just be happier in general. 522 00:48:17,286 --> 00:48:19,606 We don't wanna be bogged down with boring work. 523 00:48:19,606 --> 00:48:23,246 We wanna do cool, more strategic work and there'll be new types of. 524 00:48:23,246 --> 00:48:27,805 industries and work coming out of that as well that we can't conceive of today. 525 00:48:28,026 --> 00:48:40,946 If you think about the idea of a corporation, like when the Dutch Empire wanted to expand, that's when they created this idea of a corporation as a vehicle to collect private funds 526 00:48:40,946 --> 00:48:42,926 to fund expansion. 527 00:48:43,046 --> 00:48:52,526 that's when you have the idea of a corporation that created the idea of shares, which then created the whole stock market, which then created the whole securities law. 528 00:48:53,058 --> 00:48:58,261 commercial law, corporate law, all of that just came from one new abstract idea. 529 00:48:58,261 --> 00:49:08,908 Who knows one day there'll be a new abstract idea that we can't conceive of today, but it will be there in the future and that will create a whole new area of law that's way above 530 00:49:08,908 --> 00:49:13,931 the pay grade of AI models and we humans have to navigate through that. 531 00:49:14,312 --> 00:49:16,193 So I'm very optimistic. 532 00:49:16,193 --> 00:49:18,034 Yeah, I'm actually so keen for it. 533 00:49:18,143 --> 00:49:26,475 Yeah, I mean, it's as a legal tech CEO, I am really enjoying myself. 534 00:49:26,475 --> 00:49:30,166 It doesn't come without uh heartburn. 535 00:49:30,166 --> 00:49:33,877 You know, we are solely dependent on law firms for our business. 536 00:49:33,877 --> 00:49:38,519 And, you know, I see a lot of complacency. 537 00:49:38,519 --> 00:49:44,750 um And I also see firms that are being aggressive and going out and hiring talent and making investment. 538 00:49:44,750 --> 00:49:47,211 So I see kind of all ends of the spectrum. 539 00:49:47,245 --> 00:49:52,968 But I worry that things, I don't know how it is in Australia, but here in the US, it's a very fragmented law. 540 00:49:52,968 --> 00:50:02,412 AmLaw 200, mean, 200 law firm, the AmLaw 100 is, if you add up all the revenue, they'd be like Fortune 150. 541 00:50:02,412 --> 00:50:08,936 So it um does concern me, but I'm optimistic as well. 542 00:50:08,936 --> 00:50:11,877 And yeah, this has been a fantastic conversation. 543 00:50:12,010 --> 00:50:20,402 Before we wrap up, how do people find out more about the work that you're doing with your regulation tracker or any other projects that you're working on? 544 00:50:20,960 --> 00:50:26,012 Yeah, so simply I have a website that links everything. 545 00:50:26,072 --> 00:50:35,816 So it's like www.techcareer.com or you can also just search for me on LinkedIn Raymond Sun and yeah, they have all links in there. 546 00:50:35,816 --> 00:50:40,193 But yeah, just start with these two and yeah, hopefully you find my content fun. 547 00:50:40,193 --> 00:50:45,036 Yeah, we'll include links in the show notes um so people can get to you. 548 00:50:45,036 --> 00:50:50,660 Well, Ray, I really appreciate you taking a little bit of time out of your morning to have a conversation with me. 549 00:50:50,660 --> 00:50:52,320 This has been a lot of fun. 550 00:50:52,821 --> 00:50:55,302 let's keep doing the work that you're doing, man. 551 00:50:55,302 --> 00:50:59,266 uh We're all benefited from it, so we appreciate it. 552 00:50:59,266 --> 00:51:00,169 Yeah, likewise, Ted. 553 00:51:00,169 --> 00:51:04,133 Thank you very much for bringing me on board, and I always love chatting with you, especially on these topics. 554 00:51:04,133 --> 00:51:05,536 Yeah, thank you. 555 00:51:05,536 --> 00:51:06,017 All right. 556 00:51:06,017 --> 00:51:07,418 Have a good afternoon. 557 00:51:07,821 --> 00:51:08,781 Thanks. 00:00:04,629 Ray, how are you this afternoon or morning your time? 2 00:00:04,629 --> 00:00:05,952 How are you this morning? 3 00:00:05,984 --> 00:00:09,155 Yeah, it's morning time and yeah, hey Ted, feeling really good? 4 00:00:09,155 --> 00:00:09,997 Yeah. 5 00:00:10,094 --> 00:00:10,844 Good, man. 6 00:00:10,844 --> 00:00:11,775 Good. 7 00:00:11,775 --> 00:00:13,755 It's good to have you on the show. 8 00:00:14,033 --> 00:00:17,417 I enjoy reading your content on LinkedIn. 9 00:00:17,974 --> 00:00:35,735 I was looking at your global AI regulation tracker and knew we had to have a conversation because, and how timely, uh we are in the midst of all sorts of regulatory movement, both 10 00:00:35,735 --> 00:00:36,839 US and China. 11 00:00:36,839 --> 00:00:38,056 So we'll get into that. 12 00:00:38,056 --> 00:00:39,086 But before 13 00:00:39,086 --> 00:00:43,691 Before we do, let's get you introduced. 14 00:00:43,691 --> 00:00:49,167 you're a lawyer and a tech lawyer and a developer, and you don't see that a lot. 15 00:00:49,167 --> 00:00:57,386 Why don't you tell us a little bit about your background and how you ended up being a tech lawyer and application developer. 16 00:00:57,624 --> 00:00:58,755 Yeah, yeah. 17 00:00:58,755 --> 00:00:59,835 So... 18 00:01:00,156 --> 00:01:05,067 My story's gonna probably be one of those that starts off really random. 19 00:01:05,067 --> 00:01:09,093 You don't know where it's going, or eventually fall into place, makes sense. 20 00:01:09,093 --> 00:01:10,044 So... 21 00:01:10,244 --> 00:01:14,386 It really started when I was 5 or 6 years old. 22 00:01:14,494 --> 00:01:21,613 I was really into shows like Thomas the Tank Engine, Astro Boy, Iron Man. 23 00:01:22,154 --> 00:01:27,328 Anything that's got to do with machines, you know, coming alive and doing cool stuff. 24 00:01:27,470 --> 00:01:30,852 That has been my fascination even till today. 25 00:01:31,112 --> 00:01:36,996 And so, you know, throughout primary school and high school, I've always liked building things on the side. 26 00:01:36,996 --> 00:01:40,058 So in primary school, I would like to build my own toys. 27 00:01:40,058 --> 00:01:47,322 And in high school, you know, I started playing computer games and then started to pick up coding to learn how to build my own computer games. 28 00:01:47,322 --> 00:01:50,704 And what really helped was that my friends were also into the same thing. 29 00:01:50,704 --> 00:01:54,676 So we're all motivating each other and just doing cool things together. 30 00:01:55,390 --> 00:02:07,096 I think it was then during high school that I started reading, going to the library, and then I was also interested in books around, know, all of the detective genres, and also 31 00:02:07,096 --> 00:02:10,137 legal thrillers, like the John Grisham series. 32 00:02:10,518 --> 00:02:19,902 And that's where I got into, I really enjoy stories that talk about uh evidence and, 33 00:02:20,398 --> 00:02:22,790 trying to connect pieces together to uncover the truth. 34 00:02:22,790 --> 00:02:25,941 And even though along those lines, I find really fascinating. 35 00:02:26,042 --> 00:02:41,292 And so when it got to the end of high school, what I want to do for further study, I was really looking at either law or computer science, given my ongoing interests. 36 00:02:41,393 --> 00:02:49,678 And it came down to a really simple conclusion, which is that I can make coding a hobby, but I can't make law a hobby. 37 00:02:49,824 --> 00:02:52,145 So why not make law my career? 38 00:02:52,145 --> 00:02:56,097 And I was, you know, continuing my coding on the side. 39 00:02:56,398 --> 00:02:59,409 And so all my friends did computer science. 40 00:02:59,409 --> 00:03:09,065 I was the only one who went on to do law, but I still, you know, continued building games and websites just to, you know, as a way to still connect to my friends. 41 00:03:09,385 --> 00:03:14,888 And yeah, throughout uni, I was doing my legal studies and building random maps. 42 00:03:15,089 --> 00:03:17,902 I think, what was like a 43 00:03:17,902 --> 00:03:34,262 A big turning point was I was looking for work as a paralegal and I wasn't able to find one until I started to just look into internships in non-traditional law firms. 44 00:03:34,502 --> 00:03:39,922 So this was around 2017-18. 45 00:03:39,922 --> 00:03:44,582 I applied for an internship at a legal tech startup called Law Path. 46 00:03:44,926 --> 00:03:47,638 and I was doing as a media intern. 47 00:03:47,638 --> 00:03:52,521 So I was writing blog articles around, I think back then it was blockchain and smart contracts. 48 00:03:52,521 --> 00:03:54,453 That was the craze back then. 49 00:03:54,453 --> 00:03:56,674 So I was writing a lot of articles on that. 50 00:03:56,674 --> 00:04:02,538 And then around that time, I also did a hackathon, one of the world's first legal tech hackathon. 51 00:04:02,898 --> 00:04:05,620 And I was representing my university. 52 00:04:05,740 --> 00:04:07,732 I was the developer within my team. 53 00:04:07,732 --> 00:04:14,446 So I was building a prototype app, which basically helps streamline payments between the client 54 00:04:14,510 --> 00:04:18,930 and the Barrister, which is a very, you know, UK, Australia unique thing. 55 00:04:19,130 --> 00:04:25,450 We don't have the Barrister and sister that merge as an attorney like in the US, but basically it's just an app. 56 00:04:25,450 --> 00:04:27,010 It's like a payment app, right? 57 00:04:27,010 --> 00:04:36,470 And then when I pitched the app, it turns out that one of the judges was the boss of that legal tech startup that I was interning at. 58 00:04:36,470 --> 00:04:38,478 And so when I came back, the boss said, 59 00:04:38,478 --> 00:04:40,978 Hey Ray, I didn't know you could code like that. 60 00:04:40,978 --> 00:04:46,438 Why don't you move into our engineering department rather than the comms media department? 61 00:04:46,438 --> 00:04:48,018 And I was like, yeah, sure. 62 00:04:48,038 --> 00:04:54,298 And then that's how I got my actual job as like a legal engineer at that startup. 63 00:04:54,298 --> 00:04:57,598 And I was helping build out the document automation platform. 64 00:04:57,598 --> 00:05:06,758 And so that was like the first serious, I guess, gig that made me think, oh, wow, okay, tech and law can be combined in some way into some actual. 65 00:05:06,818 --> 00:05:17,943 meaningful career and since then when I was applying for graduate roles and you know Becoming a lawyer it has always been my path to sort of go down that tech lawyer sort of 66 00:05:17,943 --> 00:05:28,188 route and that's where I am right now I'm currently a practicing tech lawyer at Herbert Smith Freehills Kramer So and uh which is a global law firm and I do a lot of stuff within 67 00:05:28,188 --> 00:05:33,830 the AI uh legal space but at the same time, you know uh 68 00:05:33,922 --> 00:05:36,673 So that's that's like how my career sort of evolved. 69 00:05:36,673 --> 00:05:49,057 But on the side, as I said, I'm still building things and the AR regulation tracker is just one of those sort of uh projects that really developed as I was like, got really 70 00:05:49,057 --> 00:05:54,519 interested in AR ethics and robots, but I'm sure we can talk more about that one. 71 00:05:54,519 --> 00:05:57,410 But that's basically the rundown of my career today. 72 00:05:57,472 --> 00:05:58,342 Interesting. 73 00:05:58,342 --> 00:06:06,485 And yeah, so let's talk about the global AI regulation tracker that kind of started off as a personal project for you. 74 00:06:06,485 --> 00:06:13,148 And then it's kind of got in traction and now is an industry resource. 75 00:06:13,248 --> 00:06:16,179 like take us back to the origin story around that. 76 00:06:16,179 --> 00:06:27,243 think the last time you and I spoke, was you were procrastinating a reading about trolley problems and machine ethics and tell us about how this thing came about. 77 00:06:27,342 --> 00:06:31,662 Yeah, basically it started with procrastination. 78 00:06:31,662 --> 00:06:36,422 I think during uni there were some subjects which were a bit, you know, a bit of a bludge. 79 00:06:36,502 --> 00:06:41,422 was just, I'm a person who likes history and geography. 80 00:06:41,422 --> 00:06:46,742 So I like reading, you know, random articles on geopolitics and history. 81 00:06:46,742 --> 00:06:56,134 And I came across this YouTube documentary and this was around a time where self-driving vehicles were being trialed, experimented. 82 00:06:56,778 --> 00:06:59,899 and know, accidents, unfortunate accidents were happening. 83 00:06:59,899 --> 00:07:03,781 And then there videos just talking about the ethics in self-driving vehicles. 84 00:07:03,781 --> 00:07:15,085 So if a car were to, if it can't stop and can't change tracks, you, and it was approaching either an elderly person or a baby, which one should the car hit to minimize like damage? 85 00:07:15,085 --> 00:07:25,960 And that's such a tough question that it just opened up the whole door of AR ethics, which I never thought was a theme, but you know, the more I read into it, you know, reading 86 00:07:25,960 --> 00:07:26,562 about 87 00:07:26,562 --> 00:07:30,845 the three laws of robots and then the Trotty problem. 88 00:07:30,845 --> 00:07:39,531 And it was around that time where the EU government was also thinking about the EU AI Act, which today is an actual thing in effect. 89 00:07:39,531 --> 00:07:40,692 But like, this is 2019. 90 00:07:40,692 --> 00:07:44,754 This is just an idea just being floated around in government. 91 00:07:44,754 --> 00:07:49,978 And the whole idea of regulating AI back then was so alien. 92 00:07:49,978 --> 00:07:54,841 It was such a foreign concept that, wow, I never thought this could be an actual thing. 93 00:07:55,262 --> 00:07:56,140 And so, 94 00:07:56,140 --> 00:08:00,432 And as I was reading, I like to sort of write notes in a notebook. 95 00:08:00,633 --> 00:08:08,216 I don't know, like I didn't really have any particular reason why I just thought, you know, if there's anything interesting, I'll just write it down in the diary. 96 00:08:08,397 --> 00:08:11,598 And it got to a point where my diary started filling out. 97 00:08:11,959 --> 00:08:18,002 And then I was having the conversation with friends and also classmates and eventually colleagues. 98 00:08:18,002 --> 00:08:25,484 And I realized that, well, I actually have sort of quite a lot of ideas and insight over the past, you know, 99 00:08:25,484 --> 00:08:36,410 or threes of reading randomly and I thought you know maybe I could share this on LinkedIn so um when I first started becoming a lawyer on the side you know it was also during COVID 100 00:08:36,410 --> 00:08:47,406 so I started my career during the COVID lockdown so there's a bit of a lot of quite a lot of spare time to just play around so I was just writing stuff on LinkedIn initially like 101 00:08:47,406 --> 00:08:54,650 it only hit a very niche audience so I'll write updates around AR regulation and ideas 102 00:08:54,894 --> 00:08:57,675 um I did that for like a year or so. 103 00:08:58,275 --> 00:09:02,076 Modest engagement, but I had a lot of fun writing. 104 00:09:02,076 --> 00:09:12,579 And then when ChaiGPD came out, that's where everything changed because all of sudden, total proof of AI and AI regulation became like really popular and my posts started 105 00:09:12,579 --> 00:09:14,019 getting bit more attention. 106 00:09:14,019 --> 00:09:15,920 So that encouraged me to write more. 107 00:09:15,920 --> 00:09:18,781 I was also writing all around the world. 108 00:09:18,781 --> 00:09:22,322 not just one particular country, but as many countries as possible. 109 00:09:22,322 --> 00:09:23,726 It got to a point where 110 00:09:23,726 --> 00:09:30,706 my LinkedIn had all these posts in different countries and I thought, no, let's try to organize into one hub. 111 00:09:30,866 --> 00:09:32,566 And so I already had a website myself. 112 00:09:32,566 --> 00:09:38,726 So I just thought, you know, why not just add a new page to our website that's categorizes all my LinkedIn posts per country. 113 00:09:38,766 --> 00:09:42,906 So initially it was like a simple table, but then I thought that's so boring. 114 00:09:42,906 --> 00:09:44,626 Let's just take this a step further. 115 00:09:44,626 --> 00:09:45,786 I create upside. 116 00:09:45,786 --> 00:09:50,526 I put like a map on it and click on the country to show my LinkedIn posts. 117 00:09:50,526 --> 00:09:51,594 Then I thought, 118 00:09:51,594 --> 00:09:53,645 Why stop on my own LinkedIn page? 119 00:09:53,645 --> 00:09:58,979 This is actually prepare an uh encyclopedia summary to each country. 120 00:09:59,239 --> 00:10:04,333 And yeah, that's where I started summarizing each country's AI policies and regulations. 121 00:10:04,333 --> 00:10:09,987 I initially started with the G20 countries, but then I've sort of expanded. 122 00:10:09,987 --> 00:10:14,070 I've been running this project now for three months all by myself. 123 00:10:14,070 --> 00:10:21,024 And now it took me two and a half years to now cover every country uh and territory in the world. 124 00:10:21,110 --> 00:10:22,131 So over like 200. 125 00:10:22,131 --> 00:10:31,135 yeah, so I think it's been really great to be, I guess, one of the early ones building this sort of tool. 126 00:10:31,135 --> 00:10:37,298 And then a lot of people were really supportive and yeah, just a lot of encouragement from out the global industry. 127 00:10:37,298 --> 00:10:44,352 And that really helps me complete the map and also add new features to it to make it more user friendly and developer friendly. 128 00:10:44,352 --> 00:10:45,302 yeah. 129 00:10:46,103 --> 00:10:51,673 Yeah, so is your does your firm leverage the research that you've compiled? 130 00:10:53,471 --> 00:10:55,052 not, not directly. 131 00:10:55,052 --> 00:10:58,363 I think this is something that I do like in my own personal time. 132 00:10:58,363 --> 00:11:09,760 Um, and just everyone around the industry is mostly targeted to less say, you know, small businesses, academics, researchers, and developers, especially because there's now a new 133 00:11:09,760 --> 00:11:16,133 API that developers can now link the apps on top of it to sort of run the own monitoring tools, whatever. 134 00:11:16,133 --> 00:11:20,606 So it's more targeted at that sort of grassroots smaller end. 135 00:11:20,606 --> 00:11:21,386 Yeah. 136 00:11:21,985 --> 00:11:24,890 Do you have any plans for this? 137 00:11:24,890 --> 00:11:31,601 to get funding or, you know, either through a grant or private funding to help get some help with it? 138 00:11:31,601 --> 00:11:32,663 This sounds like a lot of work. 139 00:11:32,663 --> 00:11:34,096 uh 140 00:11:34,096 --> 00:11:40,189 yeah, it's um, it's it sounds like work, but it's actually like, it's not that much work. 141 00:11:40,189 --> 00:11:48,042 Because I because I progressively updated every day only takes like five minutes of my time each day just to like monitor updates. 142 00:11:48,042 --> 00:11:52,024 I've got a lot of tools in the background to help curate news items for me. 143 00:11:52,024 --> 00:11:54,395 So it's not a lot of work per day. 144 00:11:54,395 --> 00:11:55,865 But we put it all together. 145 00:11:55,865 --> 00:11:57,816 Sounds like it's quite a lot. 146 00:11:58,036 --> 00:11:59,637 in terms of your other question. 147 00:11:59,637 --> 00:12:00,207 Yeah, sure. 148 00:12:00,207 --> 00:12:02,274 Like I'm always on the, you know, 149 00:12:02,274 --> 00:12:06,139 on the lookout for opportunities, I'm also keeping myself open-minded. 150 00:12:06,139 --> 00:12:08,842 I'm also not desperate for it. 151 00:12:08,842 --> 00:12:15,029 It's something that's nice to have at end of the day for me, just to learn about the world. 152 00:12:15,029 --> 00:12:16,230 eh 153 00:12:16,811 --> 00:12:17,431 Interesting. 154 00:12:17,431 --> 00:12:22,493 yeah, and you're, you're, LinkedIn posts have, have gotten a lot of traction. 155 00:12:22,493 --> 00:12:24,103 mean, you're on the other side of the world. 156 00:12:24,103 --> 00:12:25,814 um And I've seen your stuff. 157 00:12:25,814 --> 00:12:31,505 You're a top voice, which, uh know, that's a, that's a hard designation to get. 158 00:12:31,505 --> 00:12:36,367 um You really have to put in some effort and some work around that. 159 00:12:36,547 --> 00:12:40,118 Well, you've had, so we're recording this in the beginning of August. 160 00:12:40,118 --> 00:12:43,831 This will probably come out towards the later part of the month. 161 00:12:43,831 --> 00:12:46,762 But you've had a busy week or so, right? 162 00:12:46,762 --> 00:12:57,545 Because we've had major developments with the Trump administration announcing their AI action plan and then a very quick turnaround on a China response. 163 00:12:57,545 --> 00:13:01,086 um what is your, yeah. 164 00:13:01,086 --> 00:13:04,766 And you had some great posts that I thought were interesting. 165 00:13:05,427 --> 00:13:12,969 specifically on the China side, there's a lot on the U S side too, but you, let's start with, with China first. 166 00:13:12,969 --> 00:13:13,889 So. 167 00:13:13,985 --> 00:13:20,831 You kind of zeroed in on some nuance around language um with the China plan. 168 00:13:20,831 --> 00:13:26,785 Like, and if I read your post correctly, it was like 95 % of this is not new. 169 00:13:26,785 --> 00:13:30,128 Um, but the 5 % that is, is interesting. 170 00:13:30,128 --> 00:13:36,546 um tell us what your kind of take is on the, the, the China piece first. 171 00:13:36,546 --> 00:13:38,967 Yeah, yeah, yeah, of course. 172 00:13:39,347 --> 00:13:46,190 I guess I'll just first lay out sort of the macro context behind China's AI policy thinking. 173 00:13:46,190 --> 00:13:57,074 And this is based on, you know, both research and also being in the country talking to like the big tech, like companies are driving this sort of change. 174 00:13:57,655 --> 00:14:05,378 So I think really it comes down to one or two things, which is China has a very strong push for 175 00:14:05,496 --> 00:14:08,248 what they call like, know, sovereignty in AI. 176 00:14:08,248 --> 00:14:11,450 So being self-sufficient in the full stack. 177 00:14:11,811 --> 00:14:23,639 And part of this has been driven because of the pressure from US export controls, limiting access to the necessary chips that are required to build really advanced AI systems. 178 00:14:23,639 --> 00:14:33,696 So this whole central theme around self-sufficiency, having control of the full stack, that is like the major theme and that sort of... 179 00:14:33,814 --> 00:14:44,180 Manifesting itself in other smaller sub themes around, know, where with which certain industries will require investment and trade policies and all that. 180 00:14:44,380 --> 00:14:53,085 And the second thing is like China is also trying to uh lead in standards, especially for the global South. 181 00:14:53,325 --> 00:15:01,934 There's all part of like the whole BRICS initiative, all part of the Belt and Road project, which has been ongoing for a decade already. 182 00:15:01,934 --> 00:15:14,214 So there's that also that mindset just really to set the standard because standards are important because if you think about the internet, the internet's built on US led 183 00:15:14,214 --> 00:15:21,534 standards and that has given the US a lot of leverage over how the internet ecosystem should operate. 184 00:15:21,594 --> 00:15:29,094 And it's one of those things where it's a hugely contested front and actually it has a huge role in geopolitics. 185 00:15:29,094 --> 00:15:30,766 So when it comes to the new 186 00:15:30,766 --> 00:15:40,326 breakthrough technology like AI being like the next, the current general purpose technology that is as big or even bigger than the internet, yeah, obviously that's where 187 00:15:40,326 --> 00:15:44,046 countries start thinking about, okay, let's be the ones to set the standard. 188 00:15:44,046 --> 00:15:46,446 So that's the macro context in mind. 189 00:15:46,446 --> 00:16:00,014 And so when the AI action plan came out from China, and to be accurate, when we translate in English, it's called like the AI action plan, but the actual Chinese like, 190 00:16:00,014 --> 00:16:04,294 text, it's actually called the global AI governance sort of action plan. 191 00:16:04,294 --> 00:16:10,574 So it's like a, it has a global sort of mindset embedded into it. 192 00:16:10,574 --> 00:16:24,014 And the 95%, which I said was not new, that's basically the sort of stuff that we've seen in previous papers and also in government representative speeches around, know, as I said, 193 00:16:24,014 --> 00:16:28,654 securing the full stack, investing in green and sustainable ways of 194 00:16:28,654 --> 00:16:31,415 powering models, all that sort of stuff. 195 00:16:31,735 --> 00:16:39,858 The 5 % which I thought was sort of was interesting and highlighted was around open source. 196 00:16:40,259 --> 00:16:49,863 So I think when I say open source in China, people often think of DeepSeek, which that is really the big um milestone that we saw. 197 00:16:49,863 --> 00:16:58,712 um So, but before DeepSeek, it has always been like this sort of uh strategy of 198 00:16:58,712 --> 00:17:01,704 tech clients to release open source products. 199 00:17:01,945 --> 00:17:12,424 And there's a lot of reason why open source is such a huge theme in China is, but there's all of like, I think fundamentally it's because the internal domestic competition is so 200 00:17:12,424 --> 00:17:13,434 fierce. 201 00:17:13,695 --> 00:17:21,061 People often talk about the US and China competition as like the first layer of competition, who are actually in China. 202 00:17:21,441 --> 00:17:26,958 Companies care more about the competition with their next door neighbor, which is like the domestic. 203 00:17:26,958 --> 00:17:36,220 competition and so fierce, there is sort of a race to the bottom in terms of who can produce the highest quality model for the lowest price. 204 00:17:36,501 --> 00:17:45,863 And initially there was a race towards like who can provide the lowest API options until some of the big companies were like, actually, like, not just open source it? 205 00:17:45,863 --> 00:17:49,284 That's that's technically zero, zero dollars for free. 206 00:17:49,284 --> 00:17:54,866 So you basically beat everyone on the on the price front and just provide a very powerful model. 207 00:17:54,866 --> 00:17:56,846 And so already for like 208 00:17:56,846 --> 00:18:07,046 For a year or two all the big tech companies in China were releasing their own open source AR models What made DeepSea quite special is that it found like a new ways to make the 209 00:18:07,046 --> 00:18:15,906 training process even more cheaper and That caused a lot of headlines in the West as well as well of attention has been brought to DeepSea even though it's part of a it's only a 210 00:18:15,906 --> 00:18:25,336 small part of the bigger open source picture But what this plan So even though open source has been a long thing what this plan was quite different was that 211 00:18:25,336 --> 00:18:30,059 that its choice of language was uh very selective. 212 00:18:30,059 --> 00:18:38,054 And when it comes to Chinese policies, like the language itself probably says more about the story than the actual message. 213 00:18:38,375 --> 00:18:44,318 Certain words are selected to convey a certain sentiment. 214 00:18:44,419 --> 00:18:54,115 And one of the things to look out for is which phrases are being repeated, like which mantras and which combination of words are repeated throughout the paper. 215 00:18:54,115 --> 00:18:55,118 That's often the 216 00:18:55,118 --> 00:18:58,899 indicative of government thinking. 217 00:18:59,259 --> 00:19:05,731 like for like I say, for the past year, let me just bring out my notes. 218 00:19:05,731 --> 00:19:20,145 For the past year, there was like a um particular phrasing that government will use and it was called, to translate directly into English, is, know, uh safe, reliable and 219 00:19:20,145 --> 00:19:20,925 controllable. 220 00:19:20,925 --> 00:19:25,280 So these are the, that's a typical trio of words that we see 221 00:19:25,280 --> 00:19:28,351 in speeches and it's a very systems focused view. 222 00:19:28,351 --> 00:19:35,134 So it's all about trying to make any particular use case safe, reliable, controllable. 223 00:19:35,134 --> 00:19:48,820 But since then we start to see the repetitive slogan expanding more into broader terms to now what we call, again, direct translation, inclusive, open, sustainable, fair, secure 224 00:19:48,820 --> 00:19:52,221 and reliable, digital and intelligent future for all. 225 00:19:52,221 --> 00:19:55,212 So it's a mouthful when I say in English, but it's only like eight. 226 00:19:55,212 --> 00:19:56,583 eight characters in Chinese, right? 227 00:19:56,583 --> 00:20:04,948 So, that stuff repeats a lot throughout the policy and a much more global rather than system specific focus. 228 00:20:05,329 --> 00:20:08,201 And how that relates to open source? 229 00:20:08,201 --> 00:20:23,471 Well, in the actual paragraph that mentions open source, we in English, we call it open source, but in Chinese, it's technically open sharing of resources and nowhere in the 230 00:20:23,471 --> 00:20:24,922 actual text 231 00:20:25,228 --> 00:20:31,380 have I seen the words open source code, open source software, or open source models? 232 00:20:31,641 --> 00:20:40,164 Now, again, in English, when we say open source, we tend to mean that one thing, is, know, putting your stuff on GitHub, everyone can see the code and you can download it. 233 00:20:40,164 --> 00:20:46,347 But in Chinese, open source has so many different ways of expressing that one concept. 234 00:20:46,347 --> 00:20:52,830 And if you wanna talk about open source models, open source code, there's an actual literal direct way of saying that. 235 00:20:52,830 --> 00:20:54,094 So it's not a draft. 236 00:20:54,094 --> 00:20:59,414 I don't think it's a draft in oversight because there's so many different ways of expressing that one thing. 237 00:20:59,414 --> 00:21:07,574 There's got to be some conscious effort behind why it's only open sharing of resources compared to, let's say, open source models or code. 238 00:21:07,574 --> 00:21:11,914 And as I said, when it comes to Chinese policies, you have to read into the language. 239 00:21:11,914 --> 00:21:15,234 it's not, I'm not, I don't think I'm reading too deeply into it. 240 00:21:15,234 --> 00:21:17,774 I think it's meant to be read in that way. 241 00:21:17,994 --> 00:21:22,402 And taking that interpretation, if we're only talking about sharing, 242 00:21:22,402 --> 00:21:29,448 tech documentation, manuals, like the surface layer documentation instead of the actual code. 243 00:21:29,549 --> 00:21:42,019 What I'm thinking is that this is such a clever policy balance by China to sort of influence global standards, but also keeping the secret source back at home, which is a 244 00:21:42,340 --> 00:21:44,412 very subtle and clever sort of balance. 245 00:21:44,412 --> 00:21:46,734 So that's what I noticed in this policy. 246 00:21:46,734 --> 00:21:49,046 And again, it will take another few... 247 00:21:49,058 --> 00:21:53,220 policies or papers in the future to see if that message is being reinforced. 248 00:21:53,220 --> 00:22:03,316 But until then, this is like the first one that I think might be that slight pivot towards that selective open sharing technique. 249 00:22:03,351 --> 00:22:11,103 You know, what's interesting is us in the West, like open China is like an oxymoron. 250 00:22:11,103 --> 00:22:14,224 We don't think of China and open anything. 251 00:22:14,224 --> 00:22:21,226 We think of very, you know, closed, controlled, um, not open. 252 00:22:21,466 --> 00:22:31,409 And, this seems like a, again, from a Westerner standpoint, it seems like a departure from what we would expect. 253 00:22:31,427 --> 00:22:47,289 from China, which again, the great Chinese firewall and just how um there's also been uh issues around intellectual property rights within China. 254 00:22:47,289 --> 00:22:57,937 um I guess, again, from a Westerner standpoint, it seems surprising that China wants to have an open policy. 255 00:22:58,079 --> 00:22:58,891 around this. 256 00:22:58,891 --> 00:23:01,518 oh What about on your side of the globe? 257 00:23:01,518 --> 00:23:07,311 Is this surprising or does this line up exactly the direction you thought they'd head? 258 00:23:09,002 --> 00:23:20,820 For me, it's not surprising because I think the commercial drivers sort of explain the story why there's a drive towards um open source in the say from the Western definition 259 00:23:20,820 --> 00:23:21,331 standpoint. 260 00:23:21,331 --> 00:23:28,566 As I said, the domestic competition is already so fierce that really the only real way to stand out is to be open source. 261 00:23:28,566 --> 00:23:37,688 And actually, if you talk to local developers in China, if an AI company doesn't have an open source version of their product, they're not going to be considered 262 00:23:37,688 --> 00:23:45,975 by the developers in the tech stack because even if you don't use the open source tool, it's sort of like a fashion statement, right? 263 00:23:45,975 --> 00:23:53,232 Saying that, okay, like we're doing open source so that we know that we are within the top band of the market. 264 00:23:53,232 --> 00:23:56,295 If we don't do open source, it means that why you're hiding, right? 265 00:23:56,295 --> 00:23:59,468 That's sort of the suspicion that you get from the local developer base. 266 00:23:59,468 --> 00:24:03,411 So I think in China, open source is the market expectation. 267 00:24:03,411 --> 00:24:04,742 It's the market standard. 268 00:24:04,974 --> 00:24:08,715 um Unlike in the West, I think there is a slight pivot today. 269 00:24:08,715 --> 00:24:14,137 I think especially OpenAI doing open weights now, but I think in China is a different story. 270 00:24:14,137 --> 00:24:22,674 yeah, it's just a matter of like, so open source is always gonna be the direction. 271 00:24:22,674 --> 00:24:26,380 I think the question is what extent is open? 272 00:24:26,380 --> 00:24:34,478 And that's where you get really specific with is it open source documents or code or weights or models or the whole thing. 273 00:24:34,478 --> 00:24:40,698 I think that's the question that China over there is still figuring out from a policy perspective. 274 00:24:41,017 --> 00:24:41,877 Interesting. 275 00:24:41,877 --> 00:24:53,197 And, I don't know how, if this is different in China, but here in the U S the lawmakers don't have a clue, um, about the tech. 276 00:24:53,197 --> 00:25:04,437 mean, I, um, until recently, Donald Trump didn't know, know who, uh, Jensen Wang was and, yeah, a $4 trillion company. 277 00:25:04,437 --> 00:25:10,577 And, uh, our president doesn't didn't know who the CEO was, um, by his own admission. 278 00:25:10,701 --> 00:25:24,731 And apparently there's now a lot of dialogue going on and he's surrounded himself with advisors like David Sachs and people who really understand the technology and the, if we 279 00:25:24,731 --> 00:25:30,705 can pivot to the U S for a minute, um, the commentary I've heard, and I haven't read the entire thing. 280 00:25:30,705 --> 00:25:39,120 I've read excerpts, but on the U S side, it sounds like it's written from an informed perspective. 281 00:25:39,121 --> 00:25:40,041 So. 282 00:25:40,178 --> 00:25:56,952 it, whether or not you agree with, cause there's some, there are some, there are some, uh, controversial words in the U S policy around, you know, control over kind of the tone and 283 00:25:56,972 --> 00:26:05,739 wokeness and DEI and all those sorts of things that are very politically charged, um, topics of conversation here in the U S. 284 00:26:05,739 --> 00:26:08,241 But what was your take on 285 00:26:08,267 --> 00:26:15,319 And if I, if I recall correctly, the, U S, action plan came out and China's came out within 48 hours. 286 00:26:15,319 --> 00:26:16,632 was boom, boom. 287 00:26:16,632 --> 00:26:22,241 Um, but what, what, what is your take on kind of the U S's action plan on AI? 288 00:26:24,428 --> 00:26:29,402 Yeah, I wasn't surprised by the action points in the plan. 289 00:26:29,402 --> 00:26:36,158 I can give credit in the sense that they've been consistent in what they're going to do. 290 00:26:36,158 --> 00:26:44,615 just a quick recap, the action plan is a lot, but it's been manifested through three key executive orders. 291 00:26:44,615 --> 00:26:49,659 So the first order is around energy infrastructure, promoting that. 292 00:26:49,659 --> 00:26:53,622 Second order is around the export controls layer. 293 00:26:54,100 --> 00:27:05,306 and the third order is really around what they call trying to regulate, well not trying to regulate, but ensuring that language models do not generate quote unquote work or biased 294 00:27:05,306 --> 00:27:06,017 material. 295 00:27:06,017 --> 00:27:11,079 So these are the three uh key themes of the action plan. 296 00:27:11,079 --> 00:27:15,882 And it's not the first, it's not complete surprise that these were covered. 297 00:27:15,882 --> 00:27:23,626 I think it's been a consistent policy of the government since the change of government in 298 00:27:23,626 --> 00:27:24,927 start of the year. 299 00:27:25,888 --> 00:27:35,636 For me, my focus because I'm more interested in the global dynamics, the second one was the most interesting for me, which is around the export controls. 300 00:27:36,077 --> 00:27:45,464 And so ever since the Biden administration, the US has been tightening export controls around chips. 301 00:27:45,505 --> 00:27:47,520 And to get even more specific, 302 00:27:47,520 --> 00:27:52,643 Initially, was only restrictions around just the actual final chips themselves. 303 00:27:52,643 --> 00:27:58,007 So the final product before it's being shipped to select countries. 304 00:27:58,007 --> 00:28:08,103 the export controls really target like China, Russia, Iran, and there's other sort of like what the USC's as the competitor nations. 305 00:28:08,925 --> 00:28:17,390 But what I found really interesting in the action plan is that there is, it's only a small paragraph within the actual plan, but it mentions of 306 00:28:17,390 --> 00:28:32,190 the government or requiring the DOC to really look into targeted export controls around semiconductor manufacturing components. 307 00:28:33,070 --> 00:28:36,670 So if you think about the full stack, there is the actual chip itself. 308 00:28:36,670 --> 00:28:42,270 take your Nvidia chip or AMD chip, the final full package. 309 00:28:42,270 --> 00:28:44,470 Within that, there's a lot of different components. 310 00:28:45,050 --> 00:28:54,277 So existing controls target both the final one and also recently the components within that chip because there's been a growing sort of recognition within government that, 311 00:28:54,277 --> 00:29:04,003 actually maybe the chip is designed and exported from US, but all the different parts are being imported across the world, right? 312 00:29:04,003 --> 00:29:08,166 And so you've got to mostly make sure you're accountable for these different sub components. 313 00:29:08,666 --> 00:29:14,600 But now they're looking above the value chain as well, which is the actual manufacturing. 314 00:29:15,018 --> 00:29:15,629 of the chips. 315 00:29:15,629 --> 00:29:19,622 So there are already existing controls on manufacturing equipment. 316 00:29:19,622 --> 00:29:23,025 So we're talking about these big lithographic machines. 317 00:29:23,025 --> 00:29:35,276 So just a quick, uh quick explainer, like how these chips are built, like they're very small, like it's, basically a very intricate design on a silicon wafer, right? 318 00:29:35,417 --> 00:29:36,708 But it's really small. 319 00:29:36,708 --> 00:29:40,846 It's as small as like some like adamants is measured in nanometers, right? 320 00:29:40,846 --> 00:29:42,146 How the heck do people do that? 321 00:29:42,146 --> 00:29:52,006 Well, it's all done because there's like this big machine that's sole purpose is to fire a very specific beam of light through a bunch of mirrors. 322 00:29:52,006 --> 00:29:59,026 And that's what carves out these very small designs on a very small piece of silicon at a nanometer sort of level. 323 00:29:59,026 --> 00:30:00,786 That's how these things are created. 324 00:30:00,786 --> 00:30:10,694 Now, these big machines, like these lithographic machines, can only be built by one company, which is ASML, which is a base, a company based in the Netherlands, right? 325 00:30:11,042 --> 00:30:15,124 and that stuff's been bought by companies across the world to build these chips. 326 00:30:15,124 --> 00:30:24,932 So we're talking about the final big machine, but that machine itself, I think reports have said it's built from 700,000 components. 327 00:30:24,932 --> 00:30:31,777 They have components from Germany, from Spain, from France, even from China, from like Southeast Asia, everywhere. 328 00:30:31,777 --> 00:30:36,160 We're talking about the most complex supply chain in history. 329 00:30:36,300 --> 00:30:38,894 And I reckon it builds the... 330 00:30:38,894 --> 00:30:42,254 I reckon it's in the Guinness World Records or something for that complexity. 331 00:30:42,254 --> 00:30:52,374 But anyways, the action plan is considering not just targeting semiconductor manufacturing equipment, but also components within that manufacturing equipment. 332 00:30:52,374 --> 00:31:03,894 Now it might just be two words on the paper, but those two words could have massive complications because if we're gonna export control components within that big machine, 333 00:31:04,074 --> 00:31:06,654 that could potentially cover the whole world. 334 00:31:06,654 --> 00:31:08,258 Like these export controls could 335 00:31:08,258 --> 00:31:14,861 basically target every single supplier that pitches in into that one big machine. 336 00:31:14,962 --> 00:31:24,057 And it's really hard to evaluate like what's the actual specific impact, but all I know is that it's gonna make supply chains really complicated. 337 00:31:24,057 --> 00:31:33,882 And a lot of the costs, the supply side inflation that's happening in the world right now, partly due to oil price, but there's also a lot due to just cheap prices in generally 338 00:31:33,882 --> 00:31:37,502 making the cost of anything tech, any 339 00:31:37,502 --> 00:31:51,548 any good basically that's that's digitized that's they already quite expensive on themselves just from the current um export controls but with with further controls around 340 00:31:51,548 --> 00:32:02,793 the manufacturing components then yeah that yeah i'm just we're gonna brace ourselves for that i'm sure like again the action plan is only indicating that this is an area that the 341 00:32:02,793 --> 00:32:06,092 relevant departments have to look at so just to be clear it's not 342 00:32:06,092 --> 00:32:10,036 is not a direct action, it's just telling the agencies to look into that question. 343 00:32:10,036 --> 00:32:15,463 So I'm sure there'll be a lot of expert analysis into that, but that's one thing just to get your heads up for. 344 00:32:15,463 --> 00:32:16,693 Yeah. 345 00:32:17,725 --> 00:32:24,909 And speaking of export controls, so I banged on DeepSeek a little bit and lately Kimi. 346 00:32:25,129 --> 00:32:34,974 And uh I'm really impressed with, I was really impressed with DeepSeek uh R1 when it came out and I'm really impressed with Kimi. 347 00:32:35,895 --> 00:32:47,437 Are these export controls maybe having an unintended consequence of forcing these uh countries that have constraints? 348 00:32:47,437 --> 00:32:57,887 to become more efficient and creative and engineer more uh interesting solutions to these problems? 349 00:32:57,887 --> 00:33:00,311 Is it having that unintended consequence? 350 00:33:00,856 --> 00:33:09,814 Yeah, it's like kind of like the whole, you know, the famous quotes like, you know, necessity is the model of invention or that sort of thinking. 351 00:33:09,814 --> 00:33:13,617 And actually I'll show you something in Chinese internet meme culture. 352 00:33:13,617 --> 00:33:23,742 Like there's a lot of, it's a pretty popular meme that among Chinese netizens that Trump and Biden are the founders of China or the. 353 00:33:23,742 --> 00:33:28,015 or that it's called nation building fathers of China. 354 00:33:28,015 --> 00:33:29,666 That's like the joke, right? 355 00:33:29,666 --> 00:33:40,012 Because the reason is that their export controls have pressured Chinese industries, have limited their resources so much in a way that they just have to find new ways to build 356 00:33:40,012 --> 00:33:41,134 models. 357 00:33:41,134 --> 00:33:44,056 as you mentioned, Kimmy, but also notably DeepSeek, right? 358 00:33:44,056 --> 00:33:47,978 The fact that, so it's just a quick rundown, like. 359 00:33:49,106 --> 00:33:57,950 The traditional thinking is that in order to build a powerful AR model, you need a lot of labelled data and a lot of processing power to train on the labelled data. 360 00:33:57,950 --> 00:34:09,335 But DeepSeek, based on their paper, says that you can actually build a powerful model with less labelled data, but a lot more from reinforcement learning. 361 00:34:09,335 --> 00:34:14,517 And reinforcement learning, the advantage is that you don't need labelled data to do reinforcement learning. 362 00:34:14,517 --> 00:34:17,037 You have some to get it up to like a 363 00:34:17,037 --> 00:34:19,037 particular sort of like head start. 364 00:34:19,037 --> 00:34:28,279 But from there on, that's like the first 10%, but the rest, the 90%, the model just plays by itself and just learns from its own mistakes and then reapplies them. 365 00:34:28,279 --> 00:34:32,260 then that's the beauty of reinforcement learning. 366 00:34:32,560 --> 00:34:38,601 And then apparently they were able to do it on like older like chips, like H20 chips. 367 00:34:38,601 --> 00:34:43,622 But I think that claim is still being tested by independent experts. 368 00:34:44,002 --> 00:34:55,434 That's an example of really stretching the boundaries of existing legacy tech and finding new software layer, new algorithms to make the most out of your hardware components. 369 00:34:55,434 --> 00:34:59,878 So yeah, I'm sure there is that effect. 370 00:35:00,233 --> 00:35:15,821 Yeah, and like speaking of other policy effects like you know the didn't meta recently refused to sign the EU's uh You know acknowledgement around around their policies am I 371 00:35:15,821 --> 00:35:17,151 correct on that? 372 00:35:17,666 --> 00:35:22,049 Yeah, I think you're referring to the general purpose AI code. 373 00:35:22,049 --> 00:35:24,290 Yeah, it's been uh a... 374 00:35:24,891 --> 00:35:35,638 It's been one of those hotly contentious policy documents in the industry and even caused some sort of division among the big tech companies themselves. 375 00:35:35,638 --> 00:35:39,391 I think it's all been finalized since last week. 376 00:35:39,391 --> 00:35:41,744 So I think there is a list of signatories. 377 00:35:41,744 --> 00:35:43,344 You can see who signed and who's not. 378 00:35:43,344 --> 00:35:47,316 But yeah, I think in the weeks lead up to it, yeah, certain... 379 00:35:47,448 --> 00:35:50,254 Companies have said they'll sign on, some say they won't. 380 00:35:50,254 --> 00:35:51,345 Yeah. 381 00:35:51,928 --> 00:35:52,598 Yeah. 382 00:35:52,598 --> 00:36:04,093 And you know, I think another tone from the U S action plan is there's going to be very little regard for, um, environmental concerns. 383 00:36:04,093 --> 00:36:06,444 It's, you know, build baby build. 384 00:36:06,544 --> 00:36:17,158 And, know, again, this seems like the, this seems a little bit like a, a, a uh flipping of the script, you know, um, historically the picture has been painted that China has had 385 00:36:17,158 --> 00:36:18,849 less regard for. 386 00:36:18,881 --> 00:36:26,005 know, green initiatives and really the West has been putting that more in focus. 387 00:36:26,005 --> 00:36:38,672 And it seems to me the tone of the, you know, this administration's policy is that is in the way in the backseat, maybe in the trunk. 388 00:36:38,672 --> 00:36:46,557 uh First and foremost is about establishing global dominance around AI and maintaining the lead. 389 00:36:46,650 --> 00:36:48,657 Is that the way you read it as well? 390 00:36:50,082 --> 00:36:51,543 Yeah, interesting. 391 00:36:52,044 --> 00:37:05,595 It depends on how you define lead, because I hear lot of commentary around the whole AI race and who's leading what, but it's a very, I personally find it's very simplified view 392 00:37:05,595 --> 00:37:09,358 of how this whole ecosystem works. 393 00:37:09,358 --> 00:37:12,600 First of all, you have to divide it within like layers, right? 394 00:37:12,841 --> 00:37:15,543 And it depends on which layer you're looking at. 395 00:37:15,543 --> 00:37:18,744 So at the app layer, I'd say like, 396 00:37:18,744 --> 00:37:27,339 both China and US have equally diverse and widely used apps at the application layer. 397 00:37:27,739 --> 00:37:38,956 And as you go deeper within the stack, so I think maybe just to really simplify, at the first layer, it's really a question around diversity and USichu has the most diverse and 398 00:37:38,956 --> 00:37:40,867 used ecosystem. 399 00:37:40,867 --> 00:37:42,508 Then we get to the model layer. 400 00:37:42,508 --> 00:37:45,720 That's okay, that's why I can see the race concept being. 401 00:37:45,720 --> 00:37:51,354 being true because it's really a race to who can build the smallest and cheapest model. 402 00:37:51,354 --> 00:37:54,916 That's really what I see the races and it's different approaches, right? 403 00:37:54,916 --> 00:38:06,643 So from a US perspective, it's really driven by the private sector, private sector and also the inter competition trying to produce the cheapest sort of APIs for powerful 404 00:38:06,643 --> 00:38:07,250 models. 405 00:38:07,250 --> 00:38:12,086 Whereas in China is also that domestic competition from an open source level. 406 00:38:13,076 --> 00:38:19,846 At the infrastructure hardware chips layer, I didn't really see it as a race. 407 00:38:19,846 --> 00:38:21,218 It's more like... 408 00:38:23,062 --> 00:38:27,024 you find your own adventure to building self-sufficiency. 409 00:38:27,024 --> 00:38:28,485 That's how I see it. 410 00:38:28,845 --> 00:38:35,509 And you can either do it from a constructive or deconstructive uh approach. 411 00:38:35,509 --> 00:38:44,133 Constructive being, so when I say that, constructive means, for example, subsidies, government investments, promoting trade. 412 00:38:44,174 --> 00:38:46,715 So stuff that kind of helps grow. 413 00:38:46,775 --> 00:38:52,608 And deconstructive, is export controls, tariffs, or other um 414 00:38:52,608 --> 00:38:57,741 anti-free trade policies that try to stifle what your competitors are doing. 415 00:38:57,741 --> 00:39:02,164 And it's not that you can only be constructive, it can't be deconstructive. 416 00:39:02,164 --> 00:39:03,915 You're gonna have a balance between those two, right? 417 00:39:03,915 --> 00:39:05,576 That's how policy works, right? 418 00:39:05,576 --> 00:39:16,732 So I think that layer, it's really, again, choose your own adventure, but your policy mix depends on your current country circumstances. 419 00:39:16,932 --> 00:39:18,033 So that's how I see it. 420 00:39:18,033 --> 00:39:20,134 um 421 00:39:21,230 --> 00:39:34,701 In terms of more broadly, I think it is true that both, I think since 2022 and whole January AI, I think before that, the whole AI policy debate was really around just the 422 00:39:34,701 --> 00:39:38,705 typical safety and reliability, all that stuff. 423 00:39:38,705 --> 00:39:43,708 Since 2022, AI is gonna become a more geopolitical topic. 424 00:39:43,869 --> 00:39:50,254 And so the idea of like, so as I said, the idea of leading is more around 425 00:39:50,444 --> 00:39:57,056 establishing, I think like just who has more influence on standards. 426 00:39:57,056 --> 00:40:03,288 I think that's one specific angle that I can see where there's that strong competition, as I said before. 427 00:40:03,288 --> 00:40:07,719 Once you set the standards, your whole ecosystem becomes sticky. 428 00:40:07,719 --> 00:40:12,750 And when your system becomes sticky, people have to use it, revenue comes in, your GDP booms. 429 00:40:12,750 --> 00:40:16,802 That's like, that's how, I think that's sort of the more long meta strategy. 430 00:40:16,802 --> 00:40:18,442 So I see that. 431 00:40:18,994 --> 00:40:24,198 And you also mentioned the whole green and uh energy thing. 432 00:40:24,499 --> 00:40:28,302 That's also a big part of it because AI consumes lot of power. 433 00:40:28,582 --> 00:40:38,431 I think this is also where it's important not to see AI policy in isolation, but how it interconnects with every other domestic policy of a country. 434 00:40:38,431 --> 00:40:47,638 So AI crosses over into energy policy, also crosses over into land policy, because the amount of land you have to dedicate to data centers. 435 00:40:47,680 --> 00:40:49,622 it crosses over into like tax. 436 00:40:49,622 --> 00:40:50,933 That's a huge thing, right? 437 00:40:50,933 --> 00:40:54,316 Tax incentive and all that to incentivize development. 438 00:40:54,316 --> 00:40:56,670 You to see all this in one big picture. 439 00:40:56,670 --> 00:41:07,647 I think where it's true, at least from the objective stats, is that China does have a huge head start in this space because they have a lot of capacity, like in terms of electric 440 00:41:07,647 --> 00:41:14,172 generation, there's a lot of land that's still being underdeveloped that can be turned into electric plants. 441 00:41:14,172 --> 00:41:16,802 There is a stronger central 442 00:41:16,802 --> 00:41:19,783 government push in energy. 443 00:41:20,083 --> 00:41:23,985 That's been quite a, it's been consistent for many, many years. 444 00:41:24,345 --> 00:41:32,608 The green tech industry over there has a lot of state support and also a of private sector activity. 445 00:41:32,729 --> 00:41:34,780 It's also a very popular STEM subject. 446 00:41:34,780 --> 00:41:40,552 We talked to Chinese developers, a lot of them want to go into energy tech as their engineering field. 447 00:41:40,552 --> 00:41:45,454 So there's that sort of capacity that's all there, that they're just making use of that. 448 00:41:45,454 --> 00:41:47,634 and part of it's going towards AI. 449 00:41:47,634 --> 00:41:51,114 And the US is also now focusing the same efforts now. 450 00:41:51,114 --> 00:41:56,034 I think it's just a good, I think it's a consistent challenge with the West around energy. 451 00:41:56,774 --> 00:42:05,434 There's a lot of debate around which sources you have to use and it doesn't, each particular energy source has its own big policy debate. 452 00:42:05,434 --> 00:42:09,514 But I think in China, they sort of have, they sort of just have that one set. 453 00:42:09,514 --> 00:42:15,318 Like they just go with that one source and then go, so not one source, they go with a certain mix of sources. 454 00:42:15,318 --> 00:42:16,210 And just run with that. 455 00:42:16,210 --> 00:42:21,970 They kind of skip the whole policy debate in the beginning, just go straight to implementation. 456 00:42:22,541 --> 00:42:22,831 Yeah. 457 00:42:22,831 --> 00:42:27,993 And then, we're almost out of time and we've been taught this is, could talk about this stuff all day. 458 00:42:27,993 --> 00:42:33,806 Um, I geek out on, you know, AI in general, but bringing it back to legal. 459 00:42:34,026 --> 00:42:48,332 So what do you, what do you envision and, know, on what sort of timeline do you see legal work and the resources, the inputs, which are mostly human capital today? 460 00:42:48,432 --> 00:42:51,157 When do you see that being disrupted? 461 00:42:51,157 --> 00:43:01,384 where we will see material impact to law firm revenue, law firm headcount, inside uh council processes. 462 00:43:01,384 --> 00:43:06,147 Like today, there's a lot of experimentation and I think there is some impact. 463 00:43:06,147 --> 00:43:10,349 But when do you see real disruption taking place in the legal space? 464 00:43:11,086 --> 00:43:24,766 Yeah, I could draw a graph here, but I feel like it's just a function of the more standardized and the lower risk value of the work, the more prone it is to automation. 465 00:43:25,006 --> 00:43:37,726 And not just AI automation, but just any form of automation, like even your traditional boring, algorithmic sort of if-else statements, that stuff can also act as automation. 466 00:43:38,574 --> 00:43:40,794 as a standardized low risk. 467 00:43:40,794 --> 00:43:53,474 the reason why I say that is because obviously standardized is a consistent process that's really easy to encode into code and low risk being that if something goes wrong, the loss, 468 00:43:53,474 --> 00:43:55,954 the chance of harm is still going to be quite low. 469 00:43:55,954 --> 00:44:04,374 And there's, and it's also one of those like if something goes wrong, it's still easy or still practical for someone to just jump in and fix things. 470 00:44:04,814 --> 00:44:05,966 like, yeah. 471 00:44:05,966 --> 00:44:16,786 Usual suspects that people talk about as like, you know, low value real estate transactions, like mortgages, conveyancing, that's to the extent that's still done by 472 00:44:16,786 --> 00:44:17,866 lawyers. 473 00:44:18,606 --> 00:44:24,326 Some aspects of like loan, like loan contracts, equity, that's like all stock standard terms. 474 00:44:24,786 --> 00:44:30,086 Certain tech contracts, software contracts, again, that's anything that's got to do stock standard terms. 475 00:44:30,086 --> 00:44:31,086 Yeah, definitely. 476 00:44:31,086 --> 00:44:34,866 I mean, that's like the primary, lot of these legal tech startups are targeting. 477 00:44:36,300 --> 00:44:41,353 I'd say that's something that's already in the process of, since the past four, five years. 478 00:44:41,413 --> 00:44:49,038 The next two, three years or so, we'll start to target the more, still relatively standardized and still relatively low risk. 479 00:44:49,038 --> 00:44:55,762 But I'd say this, I think maybe 80 % standardized and an extra 10 % in risk. 480 00:44:55,762 --> 00:44:57,403 That's like the medium state level. 481 00:44:57,403 --> 00:45:01,646 um That's where we start to see stronger. 482 00:45:01,772 --> 00:45:07,444 reasoning capabilities of these models to be able to tackle these sort of semi standardized problems. 483 00:45:07,444 --> 00:45:17,319 So they've still got some consistency in a problem, but there's also a level of customization or nuance thinking that these models have to have to sort of recognize, but 484 00:45:17,319 --> 00:45:21,290 not too nuanced that it sort of confuses the model. 485 00:45:21,791 --> 00:45:29,954 So that's probably where we're getting to again, the same errors I just identified, but the bit more complex like problems. 486 00:45:29,986 --> 00:45:39,444 We also start to see areas like, I say, crime, like sudden, I uh don't know what's the right word to use, but petty crime. 487 00:45:39,444 --> 00:45:42,436 I think that's where you can start using for petty crime. 488 00:45:42,596 --> 00:45:49,181 Also, um yeah, a lot more areas of commercial law, so commercial contracting. 489 00:45:50,403 --> 00:45:59,510 What everyone's really excited about is in the next, I say, eight or 10 years, where we really start tackling highly nuanced legal problems. 490 00:45:59,968 --> 00:46:04,560 And actually, this is where I honestly don't really know what will be the end outcome. 491 00:46:04,560 --> 00:46:11,663 As a practicing lawyer myself, when I say highly nuanced problems, I do mean they're highly nuanced. 492 00:46:11,663 --> 00:46:21,027 I think the common misconception that people have is that like contracts or reading laws, it's all based on what's written on the paper. 493 00:46:21,027 --> 00:46:24,198 As long as you know what's on paper, you can interpret that. 494 00:46:24,198 --> 00:46:26,909 You basically have the full answer. 495 00:46:26,909 --> 00:46:29,614 Actually, no, like the text. 496 00:46:29,614 --> 00:46:33,254 let's say probably only addresses 30 % of your problem. 497 00:46:33,274 --> 00:46:40,874 The 60 % is actually understanding your client's needs, the problem at hand, and also the market. 498 00:46:41,054 --> 00:46:45,274 And the question is, how do you encode all of that into numbers? 499 00:46:45,274 --> 00:46:48,894 That's ultimately what developers have to do. 500 00:46:49,034 --> 00:46:53,814 Encode the legal problem into numbers that can be read by a machine. 501 00:46:53,814 --> 00:46:55,594 That's what you have to do at the end of the day. 502 00:46:55,594 --> 00:46:58,272 How do you encode client interests 503 00:46:58,272 --> 00:47:03,794 marker standard, marker practices, to extent that they're not written down in words or standards. 504 00:47:03,794 --> 00:47:07,606 We're just talking about conversation dialogues and all that. 505 00:47:07,606 --> 00:47:15,669 How do you encode that in a consistent manner that a model can reliably reference for XYZ problems? 506 00:47:15,669 --> 00:47:17,490 I've tried doing that myself. 507 00:47:17,490 --> 00:47:19,591 It's really hard, right? 508 00:47:19,851 --> 00:47:20,601 But who knows? 509 00:47:20,601 --> 00:47:24,223 Maybe at that point, the whole architecture will change. 510 00:47:24,223 --> 00:47:27,192 We're currently still on a transformer architecture. 511 00:47:27,192 --> 00:47:32,482 which is very much a predict the next word, predict next token. 512 00:47:32,482 --> 00:47:34,376 Obviously there's a lot of layers around that. 513 00:47:34,376 --> 00:47:37,808 It's not just that, but fundamentally that's still what happens. 514 00:47:37,808 --> 00:47:45,612 Who knows, it might be a new mainstream model, like the state space model that might allow us to do a way more nuanced reasoning. 515 00:47:46,113 --> 00:47:50,335 Right now, all of the reasoning models are just limited to like chain of thought. 516 00:47:51,135 --> 00:47:53,417 But I think chain of thought is just level one. 517 00:47:53,417 --> 00:47:56,158 There's like way more levels down the line. 518 00:47:56,526 --> 00:48:01,186 which I don't know yet because I'm not within the research centers themselves. 519 00:48:01,186 --> 00:48:08,786 yeah, really, I'm probably, I'm gonna be one of those people who are really optimistic around disruption in law. 520 00:48:08,786 --> 00:48:14,846 It's weird for a lawyer to say that, but I feel like it's gonna be amazing because it'll free up a lot of our time. 521 00:48:14,846 --> 00:48:17,286 I think laws would just be happier in general. 522 00:48:17,286 --> 00:48:19,606 We don't wanna be bogged down with boring work. 523 00:48:19,606 --> 00:48:23,246 We wanna do cool, more strategic work and there'll be new types of. 524 00:48:23,246 --> 00:48:27,805 industries and work coming out of that as well that we can't conceive of today. 525 00:48:28,026 --> 00:48:40,946 If you think about the idea of a corporation, like when the Dutch Empire wanted to expand, that's when they created this idea of a corporation as a vehicle to collect private funds 526 00:48:40,946 --> 00:48:42,926 to fund expansion. 527 00:48:43,046 --> 00:48:52,526 that's when you have the idea of a corporation that created the idea of shares, which then created the whole stock market, which then created the whole securities law. 528 00:48:53,058 --> 00:48:58,261 commercial law, corporate law, all of that just came from one new abstract idea. 529 00:48:58,261 --> 00:49:08,908 Who knows one day there'll be a new abstract idea that we can't conceive of today, but it will be there in the future and that will create a whole new area of law that's way above 530 00:49:08,908 --> 00:49:13,931 the pay grade of AI models and we humans have to navigate through that. 531 00:49:14,312 --> 00:49:16,193 So I'm very optimistic. 532 00:49:16,193 --> 00:49:18,034 Yeah, I'm actually so keen for it. 533 00:49:18,143 --> 00:49:26,475 Yeah, I mean, it's as a legal tech CEO, I am really enjoying myself. 534 00:49:26,475 --> 00:49:30,166 It doesn't come without uh heartburn. 535 00:49:30,166 --> 00:49:33,877 You know, we are solely dependent on law firms for our business. 536 00:49:33,877 --> 00:49:38,519 And, you know, I see a lot of complacency. 537 00:49:38,519 --> 00:49:44,750 um And I also see firms that are being aggressive and going out and hiring talent and making investment. 538 00:49:44,750 --> 00:49:47,211 So I see kind of all ends of the spectrum. 539 00:49:47,245 --> 00:49:52,968 But I worry that things, I don't know how it is in Australia, but here in the US, it's a very fragmented law. 540 00:49:52,968 --> 00:50:02,412 AmLaw 200, mean, 200 law firm, the AmLaw 100 is, if you add up all the revenue, they'd be like Fortune 150. 541 00:50:02,412 --> 00:50:08,936 So it um does concern me, but I'm optimistic as well. 542 00:50:08,936 --> 00:50:11,877 And yeah, this has been a fantastic conversation. 543 00:50:12,010 --> 00:50:20,402 Before we wrap up, how do people find out more about the work that you're doing with your regulation tracker or any other projects that you're working on? 544 00:50:20,960 --> 00:50:26,012 Yeah, so simply I have a website that links everything. 545 00:50:26,072 --> 00:50:35,816 So it's like www.techcareer.com or you can also just search for me on LinkedIn Raymond Sun and yeah, they have all links in there. 546 00:50:35,816 --> 00:50:40,193 But yeah, just start with these two and yeah, hopefully you find my content fun. 547 00:50:40,193 --> 00:50:45,036 Yeah, we'll include links in the show notes um so people can get to you. 548 00:50:45,036 --> 00:50:50,660 Well, Ray, I really appreciate you taking a little bit of time out of your morning to have a conversation with me. 549 00:50:50,660 --> 00:50:52,320 This has been a lot of fun. 550 00:50:52,821 --> 00:50:55,302 let's keep doing the work that you're doing, man. 551 00:50:55,302 --> 00:50:59,266 uh We're all benefited from it, so we appreciate it. 552 00:50:59,266 --> 00:51:00,169 Yeah, likewise, Ted. 553 00:51:00,169 --> 00:51:04,133 Thank you very much for bringing me on board, and I always love chatting with you, especially on these topics. 554 00:51:04,133 --> 00:51:05,536 Yeah, thank you. 555 00:51:05,536 --> 00:51:06,017 All right. 556 00:51:06,017 --> 00:51:07,418 Have a good afternoon. 557 00:51:07,821 --> 00:51:08,781 Thanks. -->