1 00:00:03,619 --> 00:00:05,820 Speaker 1: You're listening to a CNA podcast. 2 00:00:08,659 --> 00:00:10,989 Speaker 1: So hey everyone, welcome back to another episode of Deep 3 00:00:10,989 --> 00:00:15,500 Speaker 1: Dive with me, Steven and Christina. So Krispy, what's on 4 00:00:15,500 --> 00:00:18,149 Speaker 1: the table today? What everybody's been talking about for more 5 00:00:18,149 --> 00:00:21,469 Speaker 1: than a week now, Steve. Deep Seek. Ah, we're gonna 6 00:00:21,469 --> 00:00:24,430 Speaker 1: do a deep dive into deep seek. Ah, I've got 7 00:00:24,430 --> 00:00:27,549 Speaker 1: a sinking feeling already. OK, I like the way the 8 00:00:27,549 --> 00:00:28,069 Speaker 1: Financial Times. 9 00:00:28,389 --> 00:00:31,620 Speaker 1: described it. OK, it said a little known Chinese hedge 10 00:00:31,620 --> 00:00:34,930 Speaker 1: fund has thrown a grenade into the world of AI 11 00:00:35,340 --> 00:00:38,659 Speaker 1: with a model that in effect matches market leader open 12 00:00:38,659 --> 00:00:42,778 Speaker 1: AI at a fraction of the cost. It's blown everyone away, 13 00:00:42,939 --> 00:00:45,419 Speaker 1: not so much what it can do, that's impressive, but 14 00:00:45,418 --> 00:00:46,900 Speaker 1: how much it costs. So it's a bit like when 15 00:00:46,900 --> 00:00:47,619 Speaker 1: people find out. 16 00:00:48,168 --> 00:00:50,849 Speaker 1: Hey, I can pay $5 for a bowl of noodles 17 00:00:50,848 --> 00:00:54,970 Speaker 1: that I've been paying $15 for and it's just as good. Exactly, right? Well, 18 00:00:55,049 --> 00:00:57,729 Speaker 1: that's part of it and why people are also calling 19 00:00:57,729 --> 00:01:00,330 Speaker 1: it a deep freak out, right? All the tech folks 20 00:01:00,330 --> 00:01:04,489 Speaker 1: are hyperventilating, the major tech stocks took a nosedive and 21 00:01:04,489 --> 00:01:06,410 Speaker 1: now there are all sorts of questions like who are 22 00:01:06,410 --> 00:01:09,050 Speaker 1: these people who came up with deep see? What does 23 00:01:09,050 --> 00:01:11,730 Speaker 1: it mean for the AI race and what does it 24 00:01:11,730 --> 00:01:14,120 Speaker 1: mean for us here in Little Singapore? 25 00:01:18,459 --> 00:01:21,730 Speaker 1: So we've got two very smart folks who are deep 26 00:01:21,730 --> 00:01:24,580 Speaker 1: in the space and also deep in thought. I hope 27 00:01:24,580 --> 00:01:26,769 Speaker 1: that's why they're here and we want to welcome them 28 00:01:26,769 --> 00:01:30,330 Speaker 1: to studio. We've got Doctor Leslie Teo, he's a senior 29 00:01:30,330 --> 00:01:33,220 Speaker 1: director of AI products at AI Singapore. Thanks for having 30 00:01:33,220 --> 00:01:36,860 Speaker 1: me here. AI Singapore is a national program. Our main 31 00:01:36,860 --> 00:01:40,139 Speaker 1: role is to help catalyze the AI ecosystem in Singapore. 32 00:01:40,529 --> 00:01:43,869 Speaker 1: OK, one of the distinctions about my team is we 33 00:01:43,870 --> 00:01:47,190 Speaker 1: built large language models like DeepSa and Sea Lion is 34 00:01:47,190 --> 00:01:49,510 Speaker 1: the model that we built for Southeast Asia. Ah, so 35 00:01:49,510 --> 00:01:51,669 Speaker 1: you're in the business of building the AI as well. OK. 36 00:01:51,889 --> 00:01:54,629 Speaker 1: That's great. J Lee, who is the founder and CEO 37 00:01:54,629 --> 00:01:57,989 Speaker 1: of Momentum Works, a Singapore headquartered venture outfit. Yeah, we 38 00:01:57,989 --> 00:02:01,220 Speaker 1: call ourselves a venture outfit. We do a venture building 39 00:02:01,220 --> 00:02:04,069 Speaker 1: and do a lot of insights covering different areas of 40 00:02:04,069 --> 00:02:07,339 Speaker 1: new technologies. I personally studied computer engineering for NTU. 41 00:02:07,800 --> 00:02:10,970 Speaker 1: And lots of my classmates back they ended up being 42 00:02:10,970 --> 00:02:14,728 Speaker 1: a CTOs of AI companies in Beijing, chief technology officers, 43 00:02:14,869 --> 00:02:18,289 Speaker 1: so they are having a very hard time with the deep. OK, 44 00:02:18,330 --> 00:02:19,970 Speaker 1: let's find out why they're having such a hard time. So, 45 00:02:20,000 --> 00:02:22,538 Speaker 1: so your first reaction when you heard about Desek, I mean, 46 00:02:22,940 --> 00:02:25,380 Speaker 1: There have been many other AIs that have been on 47 00:02:25,380 --> 00:02:27,500 Speaker 1: the market. They come out every day, but none have 48 00:02:27,500 --> 00:02:32,660 Speaker 1: made so much noise as Deepsek has. Why, Leslie? Maybe 49 00:02:32,660 --> 00:02:36,500 Speaker 1: I start off with a couple of observations. First of all, 50 00:02:36,619 --> 00:02:39,339 Speaker 1: Deepeek is version 3. Yeah, they have been at this 51 00:02:39,339 --> 00:02:40,699 Speaker 1: for a while. When was version 1? 52 00:02:40,764 --> 00:02:45,225 Speaker 1: Version 112 was, I think they started in 201. Version 53 00:02:45,225 --> 00:02:48,945 Speaker 1: three came out before Christmas. Nobody cared. We can talk 54 00:02:48,945 --> 00:02:51,634 Speaker 1: about why, but I think it's good to understand what 55 00:02:51,633 --> 00:02:55,543 Speaker 1: is the facts. OK, what are the facts. The first 56 00:02:55,544 --> 00:02:59,383 Speaker 1: fact is in AI there's a lot of innovation and 57 00:02:59,383 --> 00:03:04,375 Speaker 1: this is just another innovation in a whole string of innovations. 58 00:03:04,585 --> 00:03:06,014 Speaker 1: Many of us in the field have said. 59 00:03:06,369 --> 00:03:10,600 Speaker 1: There's so much rapid change and innovation here. OK. Just 60 00:03:10,600 --> 00:03:12,770 Speaker 1: look at Chat GPT. If you use Chat GPT or 61 00:03:12,770 --> 00:03:15,410 Speaker 1: even Deepy today and you used it 6 months ago, 62 00:03:15,490 --> 00:03:16,250 Speaker 1: it's not the same. 63 00:03:16,600 --> 00:03:20,470 Speaker 1: Yeah, so we see it's improving a lot and deep 64 00:03:20,470 --> 00:03:26,109 Speaker 1: is part of the story. It's not unique. It's not, yeah, but, OK, 65 00:03:26,550 --> 00:03:31,788 Speaker 1: so why why why. OK, the three things that Deepk did. 66 00:03:31,910 --> 00:03:34,869 Speaker 1: The first is they had certain techniques that improved the 67 00:03:34,869 --> 00:03:37,259 Speaker 1: use of the GPU. The second is the model itself 68 00:03:37,259 --> 00:03:40,190 Speaker 1: is very efficient to use. And the third is the 69 00:03:40,190 --> 00:03:41,789 Speaker 1: training to make it reason well. 70 00:03:42,022 --> 00:03:45,311 Speaker 1: And to make it transparent in the reasoning, that's the 71 00:03:45,311 --> 00:03:48,501 Speaker 1: R1 model, right? So these are facts, these are real. 72 00:03:48,671 --> 00:03:51,621 Speaker 1: It made it cheaper, it shows that you could make 73 00:03:51,621 --> 00:03:56,511 Speaker 1: the performance better, it's more transparent. Why does it have 74 00:03:56,511 --> 00:03:59,822 Speaker 1: such a big impact? Well, before this, the view was 75 00:04:00,231 --> 00:04:04,951 Speaker 1: only people with 100,000s of GPUs, not 10,000, but 100 76 00:04:04,951 --> 00:04:07,031 Speaker 1: thousands of GPUs. 77 00:04:07,164 --> 00:04:10,654 Speaker 1: Millions of dollars could do it. And here we have 78 00:04:10,654 --> 00:04:14,744 Speaker 1: an example that shows it's not needed. Many people in 79 00:04:14,744 --> 00:04:19,993 Speaker 1: the field probably knew this beforehand, right? Because most of 80 00:04:19,993 --> 00:04:23,593 Speaker 1: AI innovation is software. You don't need a PhD and 81 00:04:23,834 --> 00:04:29,074 Speaker 1: fabrication lab like you would in semiconductor or quantum. The 82 00:04:29,074 --> 00:04:32,073 Speaker 1: math is not that complicated, but when you have millions 83 00:04:32,074 --> 00:04:32,394 Speaker 1: of 84 00:04:32,446 --> 00:04:36,835 Speaker 1: People trying different things, you get that organic innovation, right, 85 00:04:37,235 --> 00:04:40,716 Speaker 1: and many people in hugging face. So, do you agree 86 00:04:40,716 --> 00:04:42,825 Speaker 1: with that? Is that true? And why were they able 87 00:04:42,825 --> 00:04:45,096 Speaker 1: to do it so much cheaper? A year and a 88 00:04:45,096 --> 00:04:47,664 Speaker 1: half ago I was talking to some friends in Beijing 89 00:04:47,665 --> 00:04:50,955 Speaker 1: came out, we spent like the whole weekend, the 5 90 00:04:50,955 --> 00:04:52,955 Speaker 1: bars talking in a group like nonstop for like 24 91 00:04:52,955 --> 00:04:55,276 Speaker 1: hours a day we're testing that you and all the 92 00:04:55,276 --> 00:04:57,115 Speaker 1: tech guys on Reddit. Yeah, I'm a geek. 93 00:04:58,380 --> 00:05:00,529 Speaker 1: So, a year and a half ago, one of them 94 00:05:00,529 --> 00:05:02,679 Speaker 1: told me that he said, I mean, somebody in China 95 00:05:02,678 --> 00:05:05,579 Speaker 1: will eventually uh figured out how to do this cheaply, 96 00:05:05,988 --> 00:05:08,470 Speaker 1: and he said uh it will not be one of 97 00:05:08,470 --> 00:05:11,750 Speaker 1: the venture invested ones, because the venture invested ones, they 98 00:05:11,750 --> 00:05:14,549 Speaker 1: will feel the pressure to deliver the metrics that investors 99 00:05:14,549 --> 00:05:17,670 Speaker 1: are asking for. So it doesn't give them enough leeway 100 00:05:17,670 --> 00:05:19,670 Speaker 1: and time to figure out how to do the incremental 101 00:05:19,670 --> 00:05:20,630 Speaker 1: innovation needed. 102 00:05:21,010 --> 00:05:23,149 Speaker 1: But at that time, that guy was saying that he 103 00:05:23,149 --> 00:05:26,309 Speaker 1: thought most likely the person or the organization which will 104 00:05:26,309 --> 00:05:29,070 Speaker 1: figure this out would be one of the crypto miners 105 00:05:29,070 --> 00:05:31,989 Speaker 1: because they had lots of GPUs and the crypto miner 106 00:05:31,988 --> 00:05:34,380 Speaker 1: is banned in China. So they have to deploy elsewhere, 107 00:05:34,510 --> 00:05:37,149 Speaker 1: but he didn't expect that to bed sick. So not 108 00:05:37,149 --> 00:05:39,279 Speaker 1: entirely new to anyway for those of you in the 109 00:05:39,279 --> 00:05:40,719 Speaker 1: industry to find someone who's doing it. 110 00:05:41,529 --> 00:05:49,459 Speaker 1: So much cheaper and just as effective, uh, cheaper, not 111 00:05:49,459 --> 00:05:52,099 Speaker 1: just by the way, this trend will continue. Today is deep. 112 00:05:52,609 --> 00:05:59,369 Speaker 1: We will wait for, you know, they're going to stop, right? So, 113 00:05:59,399 --> 00:06:01,729 Speaker 1: so what you're saying, Leslie, is that it was just 114 00:06:01,730 --> 00:06:04,729 Speaker 1: a matter of time before somebody did it. It just 115 00:06:04,730 --> 00:06:08,359 Speaker 1: happened that Deep Seek did it before anyone else. 116 00:06:09,559 --> 00:06:13,630 Speaker 1: We have narratives that were punctured because of this, you 117 00:06:13,630 --> 00:06:17,899 Speaker 1: need a pocket money. The confluence of geopolitics because again 118 00:06:17,899 --> 00:06:20,869 Speaker 1: it's like chip ban couldn't make this possible, so you, 119 00:06:20,950 --> 00:06:23,670 Speaker 1: you had narratives, people saying, how did this become possible? 120 00:06:23,709 --> 00:06:25,869 Speaker 1: They must have used chips that they were not allowed 121 00:06:25,869 --> 00:06:28,549 Speaker 1: to use, right? But the truth is when you stop 122 00:06:28,549 --> 00:06:32,309 Speaker 1: people from doing something, they innovate and in a way 123 00:06:32,309 --> 00:06:36,269 Speaker 1: they did innovate. Did they use these Nvidia chips? We 124 00:06:36,269 --> 00:06:37,589 Speaker 1: don't know, but certainly. 125 00:06:37,859 --> 00:06:40,440 Speaker 1: What was legal and they made what was legal very 126 00:06:40,440 --> 00:06:43,809 Speaker 1: efficient and they did it because they couldn't use the 127 00:06:43,809 --> 00:06:46,928 Speaker 1: bigger chips, isn't it funny? It's almost like because you're 128 00:06:46,928 --> 00:06:49,890 Speaker 1: broke and because you only have this box of tools 129 00:06:49,890 --> 00:06:51,609 Speaker 1: to work with, you just make it work with what 130 00:06:51,609 --> 00:06:55,130 Speaker 1: you have, right? Well, necessity is the mother of invention, actually, 131 00:06:55,170 --> 00:06:59,529 Speaker 1: to Jean's point, uh, this group were quants, right? Their 132 00:06:59,529 --> 00:07:04,450 Speaker 1: origins was in maybe they were not directly crypto, who 133 00:07:04,450 --> 00:07:06,100 Speaker 1: had an incentive to. 134 00:07:06,299 --> 00:07:09,058 Speaker 1: Reprogram the chips to be more efficient. And also I 135 00:07:09,059 --> 00:07:12,429 Speaker 1: think these guys are very good at the sensing what's 136 00:07:12,429 --> 00:07:14,589 Speaker 1: out there, what kind of information would cause what kind 137 00:07:14,589 --> 00:07:17,109 Speaker 1: of effect. So I do think that knowledge will probably 138 00:07:17,109 --> 00:07:20,019 Speaker 1: help them propagate the timing of all these releases and stuff. 139 00:07:20,190 --> 00:07:23,679 Speaker 1: So a bit of marketing, you mean? I think they 140 00:07:23,679 --> 00:07:27,359 Speaker 1: know how the market react to certain news and I 141 00:07:27,359 --> 00:07:30,149 Speaker 1: say manipulate, but I think they certainly leveraged all these 142 00:07:30,149 --> 00:07:32,750 Speaker 1: buzz in the media very well. Somebody was actually saying 143 00:07:32,750 --> 00:07:34,230 Speaker 1: that it's not just a complete. 144 00:07:35,010 --> 00:07:38,130 Speaker 1: But we don't know. One thing that some observers have said, 145 00:07:38,170 --> 00:07:40,809 Speaker 1: it's not new for the Chinese to make things which 146 00:07:40,809 --> 00:07:44,190 Speaker 1: are cheap and good. We talk about cars, batteries, talk 147 00:07:44,190 --> 00:07:48,519 Speaker 1: about vacuum cleaners. In a sense, that's in their DNA, right? 148 00:07:48,730 --> 00:07:51,450 Speaker 1: Would you agree? Kaiful Lee, uh, long time ago, he 149 00:07:51,450 --> 00:07:54,410 Speaker 1: wrote a book where in the book, he observes that 150 00:07:54,410 --> 00:07:57,010 Speaker 1: innovation in the US in AI will be at the 151 00:07:57,010 --> 00:08:00,010 Speaker 1: frontier in theory, right? In China, the innovation will be 152 00:08:00,010 --> 00:08:02,649 Speaker 1: driven by lots of data and lots of experimentation. 153 00:08:02,940 --> 00:08:06,839 Speaker 1: In a way, this is, that's what's happening right now, right? 154 00:08:06,850 --> 00:08:12,010 Speaker 1: And that's what broadly is how many things move intact. 155 00:08:12,730 --> 00:08:16,609 Speaker 1: You have the stepwise new idea, like a transformer, that's 156 00:08:16,609 --> 00:08:20,769 Speaker 1: from 2018. Then you have the constant like little bit, 157 00:08:20,890 --> 00:08:23,209 Speaker 1: little bit, push down, push down. 158 00:08:23,609 --> 00:08:27,380 Speaker 1: And that comes from many, many people trying many things, uh, 159 00:08:27,420 --> 00:08:30,140 Speaker 1: and scale of China. I mean, you would imagine that 160 00:08:30,140 --> 00:08:32,619 Speaker 1: two years since Chat GBT first came out was all 161 00:08:32,619 --> 00:08:35,299 Speaker 1: the buzz, right? You would have lots of people trying 162 00:08:35,299 --> 00:08:37,969 Speaker 1: different things and Chat GBT came out, the first version 163 00:08:37,969 --> 00:08:41,880 Speaker 1: in November 2022. Within a week, I know at least 164 00:08:41,880 --> 00:08:42,340 Speaker 1: 3 or 4. 165 00:08:42,460 --> 00:08:45,049 Speaker 1: Companies of my friends, they have complete pivot from their 166 00:08:45,049 --> 00:08:47,750 Speaker 1: existing business models. They said, OK, there's no way our 167 00:08:47,750 --> 00:08:50,369 Speaker 1: existing business model will work. We take the remaining money, 168 00:08:50,409 --> 00:08:52,909 Speaker 1: we rest for investors, we try new things. We multiply 169 00:08:52,909 --> 00:08:55,569 Speaker 1: that by, I know hundreds of companies, hundreds of teams, 170 00:08:55,729 --> 00:08:57,729 Speaker 1: someone's going to figure out. It's a bit of a race, 171 00:08:57,770 --> 00:08:59,989 Speaker 1: like an arms race is an AI race who gets 172 00:08:59,989 --> 00:09:02,609 Speaker 1: out first, right? In a way, have they set a 173 00:09:02,609 --> 00:09:05,250 Speaker 1: new benchmark because by coming out and saying, hey, is 174 00:09:05,250 --> 00:09:07,289 Speaker 1: this cheap? Now the next time an investor goes to 175 00:09:07,289 --> 00:09:07,770 Speaker 1: a company. 176 00:09:08,059 --> 00:09:10,349 Speaker 1: I'm not going to give you $50 million when I 177 00:09:10,349 --> 00:09:12,669 Speaker 1: know some guy did it for $5. Investors are very 178 00:09:12,669 --> 00:09:15,468 Speaker 1: anxious because the thing about AI is that when you 179 00:09:15,469 --> 00:09:17,478 Speaker 1: invest in a company, how do you value this company 180 00:09:17,479 --> 00:09:19,989 Speaker 1: this company continues to need more money, you can't use 181 00:09:19,989 --> 00:09:21,750 Speaker 1: the same metric of how many users you have or 182 00:09:21,750 --> 00:09:23,950 Speaker 1: how much revenue you're generating. It just doesn't work that 183 00:09:23,950 --> 00:09:26,270 Speaker 1: way because of how fast it's moving. Yeah, you know, 184 00:09:26,349 --> 00:09:29,228 Speaker 1: it's a whole topic in itself, but like where do 185 00:09:29,229 --> 00:09:32,349 Speaker 1: you make money, right? If you are the shovel provider 186 00:09:32,349 --> 00:09:34,909 Speaker 1: like Nvidia, you don't really care, more people use it. 187 00:09:35,900 --> 00:09:38,130 Speaker 1: But if you're a model builder. 188 00:09:38,599 --> 00:09:40,858 Speaker 1: The challenge for you is you try to create a 189 00:09:40,859 --> 00:09:43,890 Speaker 1: mode by being closed, like OpenAI or entropic, then you 190 00:09:43,890 --> 00:09:47,289 Speaker 1: have these open models that are roughly free. How do 191 00:09:47,289 --> 00:09:50,169 Speaker 1: you charge for your model, right? That's what we were 192 00:09:50,169 --> 00:09:54,369 Speaker 1: gonna come to because this is the proprietary, that means 193 00:09:54,369 --> 00:09:56,690 Speaker 1: I'm building something and I charge you for it, right? 194 00:09:56,700 --> 00:09:59,289 Speaker 1: $200 or whatever for you to use. And then this 195 00:09:59,289 --> 00:10:03,049 Speaker 1: is open stuff, maybe my own engineers total around with 196 00:10:03,049 --> 00:10:05,289 Speaker 1: it and then come up with their own product, customize 197 00:10:05,289 --> 00:10:06,489 Speaker 1: it to my needs, right? 198 00:10:06,909 --> 00:10:10,030 Speaker 1: That's a huge difference, isn't it, in terms of the 199 00:10:10,030 --> 00:10:16,309 Speaker 1: way this whole thing that we actually have some answers 200 00:10:16,309 --> 00:10:18,710 Speaker 1: from the tech changes, right? 201 00:10:19,119 --> 00:10:22,169 Speaker 1: So one answer is you actually cannot make money from 202 00:10:22,169 --> 00:10:26,729 Speaker 1: these models, they are commodities. Sorry to say, it takes 203 00:10:26,729 --> 00:10:30,880 Speaker 1: billions of dollars for a front leader to show it's possible. 204 00:10:31,570 --> 00:10:34,539 Speaker 1: Then the rest of us will just follow and make 205 00:10:34,539 --> 00:10:37,289 Speaker 1: it very cheap. That's not where the mode is. So 206 00:10:37,289 --> 00:10:39,530 Speaker 1: there are investors. How are they gonna make money? One 207 00:10:39,530 --> 00:10:42,429 Speaker 1: strategy is invest in the shovels, right? The people who 208 00:10:42,429 --> 00:10:43,010 Speaker 1: make the shovels. 209 00:10:43,359 --> 00:10:46,449 Speaker 1: The other is the people who make the applications. They 210 00:10:46,450 --> 00:10:49,090 Speaker 1: control the data, they control the use case, they know 211 00:10:49,090 --> 00:10:51,409 Speaker 1: the customer, they can make the money. But you know 212 00:10:51,409 --> 00:10:53,950 Speaker 1: what do you see that AI's trying to do that. 213 00:10:56,429 --> 00:10:58,409 Speaker 1: Remember in the early days, they would say we are 214 00:10:58,409 --> 00:11:01,400 Speaker 1: not going to compete with you. You build the application, 215 00:11:01,609 --> 00:11:04,489 Speaker 1: we are just here for you to use. So why 216 00:11:04,489 --> 00:11:07,080 Speaker 1: would I invest in if I'm a company. 217 00:11:07,469 --> 00:11:12,159 Speaker 1: In open AI products, if I could just use someone 218 00:11:12,159 --> 00:11:14,830 Speaker 1: like Deep Sea, which has his innards on the outside, 219 00:11:15,039 --> 00:11:17,159 Speaker 1: so to speak, right? Now, would you put your money 220 00:11:17,159 --> 00:11:19,900 Speaker 1: in there? Because now they are selling their products, you know, 221 00:11:20,039 --> 00:11:23,919 Speaker 1: for enterprise users, for personal use, etc. yeah. Why would 222 00:11:23,919 --> 00:11:26,520 Speaker 1: I pay for that that much of money. All my friends, 223 00:11:26,599 --> 00:11:28,960 Speaker 1: the ones who still have customers outside China where they 224 00:11:28,960 --> 00:11:33,590 Speaker 1: are physically inside China, they still use OpenAI for enterprise 225 00:11:33,590 --> 00:11:34,549 Speaker 1: applications because 226 00:11:35,010 --> 00:11:36,449 Speaker 1: At the end of the day, I mean, DeepSpe is 227 00:11:36,450 --> 00:11:39,909 Speaker 1: still very new, it's not stable even for enterprise usage, 228 00:11:40,260 --> 00:11:43,130 Speaker 1: so they can't put full trust on that yet, but 229 00:11:43,130 --> 00:11:44,890 Speaker 1: this might fleet, we don't know yet. That's just a 230 00:11:44,890 --> 00:11:48,039 Speaker 1: matter of time, right? Possibly, but there are other complications. Yeah, 231 00:11:48,250 --> 00:11:51,609 Speaker 1: so actually the most widely used operating software in the 232 00:11:51,609 --> 00:11:56,449 Speaker 1: world is arguably open source. It's Linux, which Mac is 233 00:11:56,450 --> 00:11:57,080 Speaker 1: part of it. 234 00:11:57,309 --> 00:12:00,900 Speaker 1: Android is part of it, but Linux has actually enterprise 235 00:12:00,900 --> 00:12:04,329 Speaker 1: version for which people will pay even though it's open source, right? Yeah. So, 236 00:12:04,539 --> 00:12:07,130 Speaker 1: so there are models like in many things that we use, 237 00:12:07,500 --> 00:12:09,940 Speaker 1: the underlying technology may be open source, but when it 238 00:12:09,940 --> 00:12:13,218 Speaker 1: comes to a business or government, you pay somebody to 239 00:12:13,219 --> 00:12:18,010 Speaker 1: make it more safe, more supported. So that's another model, right? 240 00:12:18,099 --> 00:12:20,799 Speaker 1: The second level is that you can download these models, 241 00:12:21,059 --> 00:12:23,939 Speaker 1: but it takes a bit to set it up. And 242 00:12:23,940 --> 00:12:26,140 Speaker 1: so even if you're a developer, instead of downloading. 243 00:12:26,366 --> 00:12:28,966 Speaker 1: it and it's free, you don't mind paying somebody to 244 00:12:28,966 --> 00:12:31,434 Speaker 1: set it up for you and that's called an API 245 00:12:31,434 --> 00:12:33,395 Speaker 1: to make your life easier, say I want to do this, 246 00:12:33,486 --> 00:12:35,325 Speaker 1: but you help me get there by figuring out the 247 00:12:35,325 --> 00:12:39,205 Speaker 1: whole thing in between. And just to also complications, even 248 00:12:39,205 --> 00:12:41,165 Speaker 1: like Desick, there are 3 ways at least I can 249 00:12:41,166 --> 00:12:43,966 Speaker 1: use Deep Sick. I can call the Desk API that's 250 00:12:43,966 --> 00:12:47,245 Speaker 1: going to China. OK. Then there are these services like 251 00:12:47,245 --> 00:12:51,236 Speaker 1: Amazon or others Huawei cloud has put it on the system, 252 00:12:51,486 --> 00:12:54,325 Speaker 1: then I can use that, right? Lastly, I can do 253 00:12:54,325 --> 00:12:55,005 Speaker 1: it myself. 254 00:12:55,122 --> 00:12:59,192 Speaker 1: All have different costs and implications. OK. Do you think 255 00:12:59,192 --> 00:13:02,622 Speaker 1: that the OpenAI is used by enterprises more because it 256 00:13:02,622 --> 00:13:05,081 Speaker 1: has this time advantage, right? It was forced to reckon 257 00:13:05,081 --> 00:13:07,831 Speaker 1: with enterprise customers earlier, so that's why they have built 258 00:13:07,831 --> 00:13:11,252 Speaker 1: the system earlier. It's like a first mover advantage, right? 259 00:13:11,461 --> 00:13:14,290 Speaker 1: And all the data they've gathered to make their systems, 260 00:13:14,992 --> 00:13:17,661 Speaker 1: but also don't underestimate this right, for we as users 261 00:13:17,660 --> 00:13:21,151 Speaker 1: or even people who are developer, we want all the 262 00:13:21,151 --> 00:13:23,911 Speaker 1: extra fuels, right, like more power, sounds good. 263 00:13:24,280 --> 00:13:29,929 Speaker 1: But for enterprise wants predictability and consistency, right? So it 264 00:13:29,929 --> 00:13:32,728 Speaker 1: might be like this 5 people build the model to 265 00:13:32,729 --> 00:13:36,210 Speaker 1: service the enterprise, you need 200 people. So, so you 266 00:13:36,210 --> 00:13:41,239 Speaker 1: can imagine the difference here, right? Again, deep, 20 people reportedly. 267 00:13:41,489 --> 00:13:44,049 Speaker 1: Now if they want to build a business, suddenly they 268 00:13:44,049 --> 00:13:47,369 Speaker 1: have to build this organization. They can't use their interns and. 269 00:13:47,465 --> 00:13:56,694 Speaker 1: PhD students will not trust profile. OK, let's talk about impact, right? 270 00:13:56,765 --> 00:13:59,895 Speaker 1: I spoke to April Chin. She's managing partner and CEO 271 00:13:59,895 --> 00:14:02,494 Speaker 1: of a local startup called Rosario, and here's something she said, 272 00:14:02,534 --> 00:14:04,494 Speaker 1: and I want you guys to take a listen so 273 00:14:04,494 --> 00:14:05,924 Speaker 1: that you can share your thoughts. 274 00:14:06,015 --> 00:14:06,174 Speaker 2: What 275 00:14:06,174 --> 00:14:10,614 Speaker 2: is worrying from the perspective of whether or not. 276 00:14:10,669 --> 00:14:15,598 Speaker 2: That technology can ultimately be used for public good is 277 00:14:15,599 --> 00:14:19,260 Speaker 2: in the context of that competition right between US and 278 00:14:19,260 --> 00:14:24,250 Speaker 2: China and the fact that large language models are trained 279 00:14:24,250 --> 00:14:32,090 Speaker 2: on content and content reflects worldviews. Content reflects political priorities 280 00:14:32,299 --> 00:14:33,650 Speaker 2: and these models were. 281 00:14:33,895 --> 00:14:37,505 Speaker 2: Neces have guardrails. You want to make sure, for example, 282 00:14:37,594 --> 00:14:40,094 Speaker 2: that it's not teaching you how to create a bomb, 283 00:14:40,304 --> 00:14:43,094 Speaker 2: but as we can see with deep seek, it will 284 00:14:43,385 --> 00:14:46,815 Speaker 2: tell you things about other political systems, but it won't comment. 285 00:14:47,195 --> 00:14:50,505 Speaker 2: For example, about the Chinese political system. And I guess 286 00:14:50,505 --> 00:14:55,304 Speaker 2: the question is moving forward, if we have different foundation 287 00:14:55,304 --> 00:15:01,864 Speaker 2: models reflecting different worldviews, what ultimately does that mean for 288 00:15:02,190 --> 00:15:05,940 Speaker 2: Enterprises and also public sector that's trying to. 289 00:15:06,590 --> 00:15:08,419 Speaker 2: Build AI technologies. 290 00:15:08,590 --> 00:15:11,349 Speaker 1: What are your views about this? What are the facts? 291 00:15:11,469 --> 00:15:14,469 Speaker 1: Whoever builds the model gets to determine what goes into 292 00:15:14,469 --> 00:15:17,669 Speaker 1: the model. This is true. So one of the things 293 00:15:17,669 --> 00:15:20,090 Speaker 1: that is very true is that models, even deep see, 294 00:15:20,349 --> 00:15:23,630 Speaker 1: have a bias because they are built by people using 295 00:15:23,630 --> 00:15:27,030 Speaker 1: public internet data for which mostly reflects the Western world. 296 00:15:27,190 --> 00:15:29,630 Speaker 1: No one's intending to do it, but it's it's the 297 00:15:29,630 --> 00:15:31,590 Speaker 1: way it is and in a way this is why 298 00:15:31,590 --> 00:15:33,469 Speaker 1: we do sea lion, right? So just to give you 299 00:15:33,469 --> 00:15:34,669 Speaker 1: a comparison, the bias. 300 00:15:34,830 --> 00:15:37,341 Speaker 1: It was quite bad. When things were more open, you 301 00:15:37,341 --> 00:15:39,661 Speaker 1: could see that Southeast Asian data was only like less 302 00:15:39,661 --> 00:15:42,411 Speaker 1: than 1%. Sea lion, we have like tried to get 303 00:15:42,411 --> 00:15:45,370 Speaker 1: to 20% to balance this out, right? I don't take 304 00:15:45,370 --> 00:15:47,731 Speaker 1: the view that it is a perfect answer. We just 305 00:15:47,731 --> 00:15:50,700 Speaker 1: know that whatever data you put in will have an 306 00:15:50,700 --> 00:15:53,861 Speaker 1: impact on how it behaves. The second is that how 307 00:15:53,861 --> 00:15:56,051 Speaker 1: do you decide whether this is a good thing or not? 308 00:15:56,380 --> 00:15:58,661 Speaker 1: I tend to take a pragmatic view. It depends on 309 00:15:58,661 --> 00:16:01,820 Speaker 1: the use case. We shouldn't say that I want my 310 00:16:01,820 --> 00:16:03,341 Speaker 1: model to give me every answer. 311 00:16:03,932 --> 00:16:06,691 Speaker 1: correctly, the way I define it, when actually I'm just 312 00:16:06,692 --> 00:16:09,411 Speaker 1: using it to do simple math. I, I don't know 313 00:16:09,411 --> 00:16:12,811 Speaker 1: if we need that kind of standard. The use case 314 00:16:12,812 --> 00:16:16,452 Speaker 1: determines whether it's appropriate or not. People will say, OK, 315 00:16:16,611 --> 00:16:18,651 Speaker 1: deep see, by the way, if you use the China 316 00:16:18,651 --> 00:16:21,932 Speaker 1: version or the app, it has something called a system prom. 317 00:16:22,002 --> 00:16:24,721 Speaker 1: System prom is the instructions that I give the machine 318 00:16:24,971 --> 00:16:27,882 Speaker 1: as a base instruction, right? The system prom will say, 319 00:16:27,892 --> 00:16:30,642 Speaker 1: say things like don't say anything negative about China. Even 320 00:16:30,642 --> 00:16:31,931 Speaker 1: if I ask how is the economy. 321 00:16:32,112 --> 00:16:36,752 Speaker 1: in China doing? It'll just give you because I told it. 322 00:16:36,922 --> 00:16:40,002 Speaker 1: Now depends on what you're using it for and the context, right? 323 00:16:40,232 --> 00:16:46,322 Speaker 1: I it for government, for example, schools, for example, education, etc. 324 00:16:46,562 --> 00:16:49,283 Speaker 1: then it matters, right? But if I'm using it for 325 00:16:49,283 --> 00:16:52,203 Speaker 1: a chatbot in a different, it may not matter. Even 326 00:16:52,203 --> 00:16:54,523 Speaker 1: as a person, by the way, who's probably not a 327 00:16:54,523 --> 00:16:57,482 Speaker 1: fan of censorship, I just want to make the point 328 00:16:57,482 --> 00:16:59,643 Speaker 1: we have to be very careful and not absolute in 329 00:16:59,643 --> 00:16:59,872 Speaker 1: our view. 330 00:17:00,973 --> 00:17:03,674 Speaker 1: But I mean, already, already other platforms have found ways 331 00:17:03,674 --> 00:17:08,394 Speaker 1: to give you the uncensored, right? So, so if you 332 00:17:08,394 --> 00:17:10,713 Speaker 1: download the model and run it yourself, you don't have 333 00:17:10,713 --> 00:17:12,994 Speaker 1: the system problem. So you can get it to tell 334 00:17:12,994 --> 00:17:16,943 Speaker 1: you whatever is in public. You can. However, there's one 335 00:17:16,943 --> 00:17:19,713 Speaker 1: thing that people need to also realize, which is that 336 00:17:19,713 --> 00:17:22,832 Speaker 1: we don't know what data was used, which means that 337 00:17:22,833 --> 00:17:26,624 Speaker 1: I could erase all data about an event or person 338 00:17:26,624 --> 00:17:29,073 Speaker 1: and training on that. How, how is that different? 339 00:17:29,194 --> 00:17:31,314 Speaker 1: From the history books of the past, where certain events 340 00:17:31,314 --> 00:17:33,224 Speaker 1: were left out because it did not paint a good 341 00:17:33,224 --> 00:17:36,494 Speaker 1: light on the country, right? So yeah, the question is, 342 00:17:36,744 --> 00:17:40,214 Speaker 1: if we know that Deep Sea has certain biases for China, 343 00:17:40,425 --> 00:17:43,185 Speaker 1: we may know that Chad GPT has certain biases for 344 00:17:43,185 --> 00:17:45,864 Speaker 1: the US. Then it's just like watching Fox News or CNN, 345 00:17:45,905 --> 00:17:48,505 Speaker 1: you know that they both are biased to different political parties. 346 00:17:48,665 --> 00:17:50,864 Speaker 1: I observed this as a fact and we should know it. 347 00:17:51,185 --> 00:17:53,544 Speaker 1: I guess I don't get upset that this is happening. 348 00:17:56,025 --> 00:17:57,734 Speaker 1: So today when I was doing some work and 349 00:17:57,786 --> 00:18:00,176 Speaker 1: Had a few AI tools open at the same time 350 00:18:00,176 --> 00:18:02,495 Speaker 1: as Grok, so I think lots of people have lots 351 00:18:02,494 --> 00:18:06,975 Speaker 1: of opinions about Groks. I mean the Elon Musk ex-affiliated chatbot. 352 00:18:07,135 --> 00:18:13,895 Speaker 1: Elon Musk is a whole different podcasts, the anti-woke kind 353 00:18:13,895 --> 00:18:19,816 Speaker 1: of you know that it opens a different discussion. So 354 00:18:19,816 --> 00:18:21,656 Speaker 1: I tried that with different things and sometimes we will 355 00:18:21,656 --> 00:18:24,135 Speaker 1: give them the same task we can clearly see that 356 00:18:24,135 --> 00:18:26,206 Speaker 1: how the different models are trained differently. So for instance. 357 00:18:26,800 --> 00:18:28,890 Speaker 1: I was asking and the models to draw me a 358 00:18:28,890 --> 00:18:32,000 Speaker 1: picture of a politician with a business figure, I will 359 00:18:32,000 --> 00:18:34,520 Speaker 1: not name which ones, because we try to like create 360 00:18:34,520 --> 00:18:37,680 Speaker 1: a sort of a poster for something, and Chad GPT 361 00:18:37,680 --> 00:18:39,599 Speaker 1: would tell me that OK, because our content policy we 362 00:18:39,599 --> 00:18:40,189 Speaker 1: can't do it. 363 00:18:40,660 --> 00:18:42,889 Speaker 1: Deep seek would say that I don't draw pictures now. 364 00:18:43,180 --> 00:18:46,150 Speaker 1: Grok gave me 4 versions saying which one do you like? 365 00:18:46,180 --> 00:18:47,900 Speaker 1: I look at 4 versions I said, OK, there's no 366 00:18:47,900 --> 00:18:49,500 Speaker 1: way I'm going to post any of them or any 367 00:18:49,500 --> 00:18:53,619 Speaker 1: of my social media inappropriate. Uh, it can be interpreted 368 00:18:53,619 --> 00:18:57,300 Speaker 1: in very negative ways, yeah, right. As you can see that, 369 00:18:57,380 --> 00:19:00,260 Speaker 1: I mean, all these models are biased in different ways. 370 00:19:00,660 --> 00:19:03,619 Speaker 1: I would consider myself a relatively informed user, so, so 371 00:19:03,619 --> 00:19:05,770 Speaker 1: I check different things before I actually use it. 372 00:19:06,189 --> 00:19:08,750 Speaker 1: Uh, I do think that policymakers would probably be in 373 00:19:08,750 --> 00:19:11,520 Speaker 1: a better position, right, and then look at different possibilities 374 00:19:11,520 --> 00:19:13,510 Speaker 1: out there, they look at the biases and they can 375 00:19:13,510 --> 00:19:16,069 Speaker 1: make an informed decision because we all have biases as 376 00:19:16,069 --> 00:19:18,188 Speaker 1: human beings. We tell people too that when you look 377 00:19:18,189 --> 00:19:21,109 Speaker 1: at the media, don't just read one source, read several 378 00:19:21,109 --> 00:19:23,389 Speaker 1: news agencies and get a sense of the story because 379 00:19:23,390 --> 00:19:25,189 Speaker 1: everybody would tell the story a little bit differently. 380 00:19:25,609 --> 00:19:27,790 Speaker 1: But I'm afraid that most people are not like him, 381 00:19:27,869 --> 00:19:31,189 Speaker 1: you know, they're not having 5 different chat GBTs open. 382 00:19:31,270 --> 00:19:34,670 Speaker 1: I'm telling you, I mean you just. I'm telling you 383 00:19:34,670 --> 00:19:38,750 Speaker 1: the kid in university chat GTs to write his essays 384 00:19:39,189 --> 00:19:42,069 Speaker 1: is just gonna pick whatever it gives him. But that's 385 00:19:42,069 --> 00:19:44,119 Speaker 1: a separate point, but this is why we need to 386 00:19:44,119 --> 00:19:46,949 Speaker 1: help people understand how to use these AI models and 387 00:19:46,949 --> 00:19:48,420 Speaker 1: actually Steve's point is really important. 388 00:19:49,030 --> 00:19:51,909 Speaker 1: You need to know the weaknesses and the strengths, right, 389 00:19:52,130 --> 00:19:56,550 Speaker 1: and don't believe that this thing is perfect. It's very powerful, 390 00:19:56,630 --> 00:19:58,589 Speaker 1: it's great. I, I'm a big fan, but it's not 391 00:19:58,589 --> 00:20:01,390 Speaker 1: perfect because depending on how you prompt it, you can 392 00:20:01,390 --> 00:20:04,030 Speaker 1: kind of get the results you want to. So you know, 393 00:20:04,189 --> 00:20:06,270 Speaker 1: I mean it's getting better, but used to be a 394 00:20:06,270 --> 00:20:08,069 Speaker 1: fun thing to do is you just, it gives you 395 00:20:08,069 --> 00:20:09,589 Speaker 1: an answer and you say, I think you're wrong, and 396 00:20:09,589 --> 00:20:10,899 Speaker 1: then the whole answer changes. 397 00:20:12,400 --> 00:20:14,560 Speaker 1: I see what you mean. Let me rethink it and 398 00:20:14,560 --> 00:20:18,680 Speaker 1: then it gives you another answer. Different AIs nowadays have 399 00:20:18,680 --> 00:20:20,920 Speaker 1: sort of different sort of trade because if someone gives 400 00:20:20,920 --> 00:20:22,760 Speaker 1: me a piece of work, I can clearly tell that 401 00:20:22,760 --> 00:20:25,969 Speaker 1: which AI helped that person write it. I mean once 402 00:20:25,969 --> 00:20:29,520 Speaker 1: you are exposed long enough, you see some patterns and 403 00:20:29,520 --> 00:20:32,209 Speaker 1: that's why sometimes when we use models, we also know, OK, 404 00:20:32,359 --> 00:20:32,679 Speaker 1: they use. 405 00:20:33,425 --> 00:20:36,125 Speaker 1: GPT4 to train this because it starts to sound like 406 00:20:36,125 --> 00:20:42,005 Speaker 1: the personality, the voice of the AI, you know, it 407 00:20:42,005 --> 00:20:44,875 Speaker 1: does make sense. We can't talk about this without talking 408 00:20:44,875 --> 00:20:47,484 Speaker 1: about how it will impact us, right, in terms of 409 00:20:47,484 --> 00:20:51,754 Speaker 1: jobs professionally, for example, we're all deep into this, you know, 410 00:20:51,844 --> 00:20:55,045 Speaker 1: every company I know is experimenting. 411 00:20:55,500 --> 00:20:57,988 Speaker 1: How is it making your work more efficient? How can 412 00:20:57,989 --> 00:21:00,859 Speaker 1: you do things better, and some people are saying it's 413 00:21:00,859 --> 00:21:05,640 Speaker 1: even displacing some very junior level jobs which Chad GPT 414 00:21:05,640 --> 00:21:08,319 Speaker 1: can do. How worried are you on a scale of 415 00:21:08,319 --> 00:21:14,699 Speaker 1: 1 to 10? Give me a 22, OK, not 1010, 416 00:21:14,709 --> 00:21:18,270 Speaker 1: you're worried. OK, now tell us why.2, I think that's 417 00:21:18,270 --> 00:21:20,698 Speaker 1: the kind of answer to my question, right, because if 418 00:21:20,699 --> 00:21:22,698 Speaker 1: you have people like Leslie who worry that at the 419 00:21:22,699 --> 00:21:24,300 Speaker 1: level of 10, then I don't have to worry. 420 00:21:25,359 --> 00:21:25,369 Speaker 1: But 421 00:21:29,150 --> 00:21:31,958 Speaker 1: You have to tell. OK, so you're optimistic that humans 422 00:21:31,959 --> 00:21:35,238 Speaker 1: will adapt. I'm optimistic for a long term. I'm a 423 00:21:35,239 --> 00:21:37,800 Speaker 1: little bit cautious about the short term because I do 424 00:21:37,800 --> 00:21:40,449 Speaker 1: see that when Cha GBT came out, when Mi came out, 425 00:21:40,560 --> 00:21:42,520 Speaker 1: I do see some friends, so my friends. 426 00:21:42,910 --> 00:21:46,510 Speaker 1: Gaming studios started laying off people because graphic design and 427 00:21:46,510 --> 00:21:49,229 Speaker 1: certain functions, they could do it more efficiently and they 428 00:21:49,229 --> 00:21:51,409 Speaker 1: need fewer people, and we can always say that OK, 429 00:21:51,510 --> 00:21:53,270 Speaker 1: the skills can be retrained, I mean, you will open 430 00:21:53,270 --> 00:21:56,349 Speaker 1: new opportunities, but there's always a gap between the time 431 00:21:56,349 --> 00:21:59,550 Speaker 1: that the jobs get displaced versus the time that new 432 00:21:59,550 --> 00:22:02,349 Speaker 1: jobs get created and it's painful for people, especially people 433 00:22:02,349 --> 00:22:05,030 Speaker 1: who have been trained as a specialist, and if their 434 00:22:05,030 --> 00:22:07,869 Speaker 1: jobs are at risk, it's much, much harder for that 435 00:22:07,869 --> 00:22:11,030 Speaker 1: to be replaced, right? But you're still confident that overall 436 00:22:11,030 --> 00:22:11,630 Speaker 1: humans will. 437 00:22:11,714 --> 00:22:14,665 Speaker 1: Prevail. I mean, it's not gonna really take away I'm 438 00:22:14,665 --> 00:22:17,665 Speaker 1: kind of sort of optimist. I mean, maybe it's not 439 00:22:17,665 --> 00:22:19,905 Speaker 1: right to say. I look at global warming, I don't 440 00:22:19,905 --> 00:22:22,025 Speaker 1: like it, but I'm not terribly worried about it. I'm 441 00:22:22,025 --> 00:22:25,324 Speaker 1: hoping that someone would find the technology to solve it. 442 00:22:25,464 --> 00:22:30,685 Speaker 1: I'm not worried because I'll be dead by then. I, 443 00:22:30,704 --> 00:22:33,584 Speaker 1: what about you? Why are you so worried? Technology is 444 00:22:33,584 --> 00:22:36,704 Speaker 1: generally a positive thing, economists are really positive about it, 445 00:22:36,775 --> 00:22:40,504 Speaker 1: but this speed is very fast, so it's a timing problem. 446 00:22:41,069 --> 00:22:43,989 Speaker 1: For me, it's good to be at 10 because I 447 00:22:43,989 --> 00:22:47,339 Speaker 1: want to make sure the technology is used for human good. 448 00:22:48,069 --> 00:22:52,478 Speaker 1: Policy makers will say AI is the old man, not replace. Unfortunately, 449 00:22:52,589 --> 00:22:55,310 Speaker 1: the profit motive is, what are you talking about? Let's 450 00:22:55,310 --> 00:22:58,910 Speaker 1: be truthful about this. It's to replace, and this is 451 00:22:58,910 --> 00:23:03,180 Speaker 1: happening so fast, right? I'm worried we can't absorb this actually. 452 00:23:03,770 --> 00:23:06,640 Speaker 1: So it's, it's better to be more alarmist in that 453 00:23:06,640 --> 00:23:09,550 Speaker 1: situation than to say it's OK. So in other words, 454 00:23:09,719 --> 00:23:12,459 Speaker 1: then we'll be forced to figure out how to protect 455 00:23:12,670 --> 00:23:16,180 Speaker 1: even the transition you are saying, right? I think is important. 456 00:23:16,349 --> 00:23:19,709 Speaker 1: People say the industrial revolution was a good thing. Yes, 457 00:23:19,790 --> 00:23:24,030 Speaker 1: it was, but two generations of workers suffered before the 458 00:23:24,030 --> 00:23:26,629 Speaker 1: good thing happened. Yeah, change is always hard. 459 00:23:26,989 --> 00:23:28,739 Speaker 1: But at the same time, you could look at it 460 00:23:28,739 --> 00:23:32,510 Speaker 1: that what has happened has forced us to think again, 461 00:23:32,670 --> 00:23:34,709 Speaker 1: Chad GPT came out and then after a while we're 462 00:23:34,709 --> 00:23:35,859 Speaker 1: all kind of like, OK. 463 00:23:36,199 --> 00:23:40,959 Speaker 1: Then suddenly another hello, wake up again. Actually the broader message, right? 464 00:23:41,000 --> 00:23:45,000 Speaker 1: The prices come down, it's even easier to use. Even 465 00:23:45,000 --> 00:23:50,400 Speaker 1: more innovations will happen and we have to move faster. Unfortunately, 466 00:23:50,439 --> 00:23:54,040 Speaker 1: it's not easy because technology is moving faster than society 467 00:23:54,040 --> 00:23:59,510 Speaker 1: can figure this out can meet it. Yeah, I'm glad 468 00:23:59,510 --> 00:24:01,959 Speaker 1: that I learned how to write when none of this 469 00:24:01,959 --> 00:24:04,478 Speaker 1: was happening, but now you don't need to write that 470 00:24:04,479 --> 00:24:05,959 Speaker 1: well because you can cheat, you can use. 471 00:24:06,385 --> 00:24:09,215 Speaker 1: But but then now all this becomes augmentation tools for you, 472 00:24:09,385 --> 00:24:11,505 Speaker 1: but for people who are fresh out of university, they 473 00:24:11,505 --> 00:24:14,176 Speaker 1: might not have that opportunity to train for that. So 474 00:24:14,176 --> 00:24:17,264 Speaker 1: the challenge is that many of us like good coder, 475 00:24:17,595 --> 00:24:21,105 Speaker 1: you know what's good coding and the GPT or Cloud 476 00:24:21,105 --> 00:24:23,906 Speaker 1: is really good at it, makes you a better coder, faster, 477 00:24:23,984 --> 00:24:28,104 Speaker 1: more productive, but the problem is the younger junior coder 478 00:24:28,105 --> 00:24:29,906 Speaker 1: never gets a chance now because I'm not going to 479 00:24:29,906 --> 00:24:33,706 Speaker 1: use a junior coder. I'm just GPT right. So you 480 00:24:33,705 --> 00:24:35,825 Speaker 1: don't get a chance to practice, right? How will you become. 481 00:24:35,952 --> 00:24:39,932 Speaker 1: Let's not be too pessimistic. The answer will be actually 482 00:24:39,932 --> 00:24:42,212 Speaker 1: the AI can help you become that, right? You can 483 00:24:42,212 --> 00:24:44,331 Speaker 1: become a master, but it's not that we don't know 484 00:24:44,332 --> 00:24:46,332 Speaker 1: how to solve it, but the profit motive may not 485 00:24:46,332 --> 00:24:48,771 Speaker 1: be there to do it as a business, I have 486 00:24:48,771 --> 00:24:51,491 Speaker 1: no interest in doing this. As someone who runs a business, 487 00:24:51,641 --> 00:24:54,621 Speaker 1: and you have this fear, right, if my competitors use 488 00:24:54,621 --> 00:24:58,612 Speaker 1: this to improve the productive, I'm forced to react. You 489 00:24:58,612 --> 00:25:00,650 Speaker 1: don't have the luxury, even if you want it, you 490 00:25:00,651 --> 00:25:03,212 Speaker 1: don't have the luxury, right? So everyone is forced to 491 00:25:03,212 --> 00:25:05,771 Speaker 1: just do things faster and be more reactive. 492 00:25:06,239 --> 00:25:08,630 Speaker 1: Yeah, I don't think we have that many big headlines 493 00:25:08,630 --> 00:25:10,869 Speaker 1: at least in Singapore, but you know, I'm beginning to 494 00:25:10,869 --> 00:25:14,310 Speaker 1: hear stories about like junior coders, you know, web designers, 495 00:25:15,189 --> 00:25:20,629 Speaker 1: content moderators, artists, that's what I'm hearing. I mean, even podcast, right, 496 00:25:20,790 --> 00:25:25,390 Speaker 1: sorry to say this, but yes, yes, we've heard of that. 497 00:25:25,989 --> 00:25:27,109 Speaker 1: We can do something else, right. 498 00:25:28,729 --> 00:25:29,520 Speaker 1: Well, 499 00:25:31,119 --> 00:25:34,270 Speaker 1: Not the same. I I I rather sell my voice 500 00:25:34,270 --> 00:25:37,409 Speaker 1: and then they can put whatever you have royalties forever 501 00:25:37,410 --> 00:25:39,359 Speaker 1: when I retire, you know, that's better. 502 00:25:39,650 --> 00:25:43,599 Speaker 1: So do you like my voice? Think again. OK, is 503 00:25:43,599 --> 00:25:47,198 Speaker 1: far from over, right? The race is going to become intense, 504 00:25:47,319 --> 00:25:50,599 Speaker 1: so I guess buckle up, everybody. That's right. We're in 505 00:25:50,599 --> 00:25:53,829 Speaker 1: for a long and could be pretty bumpy, right? That's right. 506 00:25:54,119 --> 00:25:56,040 Speaker 1: Thanks guys for coming in, sharing your insights with us. 507 00:25:56,280 --> 00:25:58,000 Speaker 1: Thanks to all of you who have been joining us. 508 00:25:58,160 --> 00:25:59,869 Speaker 1: Leave us your comments. Let us know what you think. 509 00:25:59,920 --> 00:26:01,520 Speaker 1: This is one topic that is going to be hanging 510 00:26:01,520 --> 00:26:02,439 Speaker 1: over our heads for, I think. 511 00:26:02,760 --> 00:26:05,739 Speaker 1: The next few months and years to come. OK, this 512 00:26:05,739 --> 00:26:09,729 Speaker 1: podcast has a very capable human team behind it thankfully. 513 00:26:11,030 --> 00:26:12,629 Speaker 1: I think some of them use AI to also come 514 00:26:12,630 --> 00:26:18,099 Speaker 1: up with our Tiffanyi Johari, Joan Chan and Hanida Ain. 515 00:26:18,189 --> 00:26:19,380 Speaker 1: Thank you and see you next week.