WEBVTT 00:00:04.500 --> 00:00:11.956 Good afternoon. On the twelfth of June, an American AI company switched off its two newest models. 00:00:11.956 --> 00:00:15.532 For every customer, everywhere, in one evening. 00:00:15.532 --> 00:00:23.824 Washington had barred foreign nationals from using them, and the company couldn't sort its users by passport. 00:00:23.824 --> 00:00:29.834 One model came back in July. So today's question is an old one with a new edge. 00:00:29.834 --> 00:00:34.780 Should Malaysia build its own AI, or keep renting someone else's? 00:00:34.780 --> 00:00:42.971 So the landlord can change the locks whenever he likes. Lovely thing to learn on a Friday. 00:00:42.971 --> 00:00:49.700 That's the fear. Whether it justifies billions of ringgit is what we test today. 00:00:50.800 --> 00:00:57.410 It's Friday the twenty-third of October, and this is the AI-YA Daily Friday special. 00:00:57.410 --> 00:01:03.311 It runs longer than our weekday shows, and it's slower and more reflective. 00:01:03.311 --> 00:01:11.240 I'm Raj. I run a small business, I ask the awkward questions, and yes, I still print my invoices. 00:01:11.690 --> 00:01:15.889 I'm Aiman. I try the tools and do the sums, usually in ringgit. 00:01:15.889 --> 00:01:20.647 And I'm Mei Ling. I read the fine print, so you can get on with your day. 00:01:20.647 --> 00:01:22.465 What we're here for is simple. 00:01:22.465 --> 00:01:30.220 We take the world's artificial intelligence news and make it useful for people working in Malaysia, with the price tag attached. 00:01:30.220 --> 00:01:36.490 This show is researched and voiced with AI, and edited by the AI-YA team. We practise what we teach. 00:01:36.940 --> 00:01:40.073 This week's theme is build or rent. 00:01:40.073 --> 00:01:46.966 Should Malaysia make its own AI, or rent the best and spend on using it well? 00:01:46.966 --> 00:01:52.100 Three stories, one Malaysia. Right, let's get going. 00:01:53.200 --> 00:01:59.826 Let's give the build side its strongest case first. It rests on three words: language, data and control. 00:01:59.826 --> 00:02:01.100 Start with language. 00:02:01.100 --> 00:02:08.490 In February 2024, at a summit in Dubai, Nvidia's chief executive Jensen Huang advised developing countries to build. 00:02:08.490 --> 00:02:14.670 In his words, "codify the language, the data of your culture into your own large language model." 00:02:14.670 --> 00:02:19.025 Of course the man who sells the chips says everyone should build. 00:02:19.025 --> 00:02:23.514 That's like my tile supplier telling me every room needs new tiles. 00:02:23.514 --> 00:02:27.587 Fair. He has a commercial interest, and he's never hidden it. 00:02:27.587 --> 00:02:33.663 But YTL AI Labs' chief executive, Foong Chee Mun, made a similar point to The Edge in June. 00:02:33.663 --> 00:02:37.603 He said global models "tend to deviate" in minor languages. 00:02:37.603 --> 00:02:42.152 Hmm. My customers write in three languages in one WhatsApp message. 00:02:42.152 --> 00:02:45.410 If a model can't follow that, it's no use to me. 00:02:45.410 --> 00:02:48.330 That's the language case. The second is data. 00:02:48.330 --> 00:02:53.781 Last week, Communications Minister Fahmi Fadzil talked about the sovereign AI cloud. 00:02:53.781 --> 00:02:59.622 He said it's meant to keep classified government information out of reach of foreign laws. 00:02:59.622 --> 00:03:03.232 The third is control, and that became very real in June. 00:03:03.232 --> 00:03:10.710 On the twelfth, America's Commerce Department ordered Anthropic to bar foreign nationals from its two newest models. 00:03:10.710 --> 00:03:15.480 Anthropic said that left it no choice but to switch both off for everyone. 00:03:15.930 --> 00:03:20.002 Hang on, say that again. Off for everyone? 00:03:20.002 --> 00:03:25.245 For all its customers, according to Anthropic's statement as reported by Al Jazeera. 00:03:25.245 --> 00:03:28.242 Fable 5 came back globally on the first of July. 00:03:28.242 --> 00:03:33.423 Mythos 5, the more powerful one, returned only for selected American organisations. 00:03:33.423 --> 00:03:36.170 We couldn't find a later update on that one. 00:03:36.620 --> 00:03:41.753 Picture a firm in Penang that had built its daily work around Fable. 00:03:41.753 --> 00:03:47.640 That's about nineteen days without the tool. Nobody in Malaysia was consulted. 00:03:47.640 --> 00:03:50.692 That's the build argument in one picture. 00:03:50.692 --> 00:03:55.827 In India, Sarvam's co-founder Pratyush Kumar put it in one line on X. 00:03:55.827 --> 00:04:01.780 "You should not confuse access with ownership, or adoption itself as advantage." 00:04:02.230 --> 00:04:07.708 Okay, that's a strong line. Access isn't ownership. I'll remember that. 00:04:07.708 --> 00:04:13.404 Sarvam was picked in 2025 to build a national model under India's IndiaAI Mission. 00:04:13.404 --> 00:04:19.377 In June, Medianama reported it had raised 234 million US dollars, about RM955 million. 00:04:19.377 --> 00:04:27.294 Then in May, Mistral's chief executive Arthur Mensch told French lawmakers that Europe has about two years to act. 00:04:27.294 --> 00:04:35.629 He said, "In a world where you import all digital services from the United States, you have no leverage against the US." 00:04:35.629 --> 00:04:41.760 I'm not convinced. India has more than a billion people, and Europe has deep pockets. 00:04:41.760 --> 00:04:45.728 We're thirty-something million. Why would we copy them? 00:04:45.728 --> 00:04:51.988 Okay, but let me finish my thought. The strongest case for Malaysia isn't a frontier model. 00:04:51.988 --> 00:05:01.482 In February, David Lam of the ISEAS-Yusof Ishak Institute in Singapore argued that sovereign models needn't beat the giants on everything. 00:05:01.482 --> 00:05:08.017 He said the region's edge is closeness to local customers, local data and regional cooperation. 00:05:08.017 --> 00:05:10.150 That's a niche, not a moonshot. 00:05:10.600 --> 00:05:15.721 And Malaysia already has pieces, which we went through on Thursday. 00:05:15.721 --> 00:05:20.919 YTL builds ILMU on foundations from Nvidia and the Chinese lab Z.ai. 00:05:20.919 --> 00:05:26.804 Mesolitica built MaLLaM from scratch, though its public releases have slowed. 00:05:26.804 --> 00:05:33.071 The government is pooling Malay-language data at Dewan Bahasa dan Pustaka, or DBP. 00:05:33.071 --> 00:05:41.097 And the Malaysia Digital Economy Corporation, or MDEC, plans a national compute exchange from early 2027. 00:05:41.097 --> 00:05:46.688 So we're already building. Just quietly, and on borrowed engines. 00:05:46.688 --> 00:05:55.256 Yes. Which is why the honest question isn't really build or rent. It's which layers we must own ourselves. 00:05:55.256 --> 00:06:00.920 Fair, I hadn't split it into layers. I still want to see the bill. 00:06:01.370 --> 00:06:08.466 You'll get it in the third story. Mei Ling, in Episode 4 you weren't sold on IlmuChat Pro. 00:06:08.466 --> 00:06:10.201 Does that still stand? 00:06:10.201 --> 00:06:14.591 It stands, because there's still no newer usage figure. 00:06:14.591 --> 00:06:23.130 And yesterday's three different votes on ILMU's Flash model, at RM0.55 per million input tokens, stand too. 00:06:24.230 --> 00:06:27.648 So what does each path mean for Malaysian jobs? 00:06:27.648 --> 00:06:35.429 I'd split it three ways: build the models, build the buildings, or build the users. Path one is the models. 00:06:35.429 --> 00:06:39.138 That employs a small number of very skilled people. 00:06:39.138 --> 00:06:43.502 Researchers, engineers, and people who clean and label data. 00:06:43.502 --> 00:06:45.115 Korea shows the catch. 00:06:45.115 --> 00:06:53.548 Its national foundation model project has spent over 200 billion won so far, about RM600 million by our conversion. 00:06:53.548 --> 00:06:58.974 This month, the Korea JoongAng Daily reported that developers are leaving. 00:06:58.974 --> 00:07:07.040 One moved from Motif to Kakao in August. A key SK Telecom team member reportedly left for Nvidia in September. 00:07:07.040 --> 00:07:14.061 So the government pays to train them, and then the big boys come along and… 00:07:14.061 --> 00:07:16.194 And hire them. Yes. 00:07:16.194 --> 00:07:23.146 I know that feeling. I train a good supervisor, and the shop across the road poaches her. 00:07:23.146 --> 00:07:24.974 Same story, more zeros. 00:07:24.974 --> 00:07:34.434 One company source also told the paper that the government's chips were "a drop in the bucket" next to overseas rivals. 00:07:34.434 --> 00:07:36.292 Okay. Path two? 00:07:36.292 --> 00:07:39.003 The buildings, meaning data centres. 00:07:39.003 --> 00:07:44.652 On the fourteenth, the Digital Ministry gave a written reply in Parliament. 00:07:44.652 --> 00:07:52.260 Twenty-seven approved projects, about RM226 billion in estimated investment, and 4,053 expected jobs. 00:07:52.260 --> 00:07:58.110 Wait. Two hundred and twenty-six billion ringgit, for four thousand jobs? 00:07:58.560 --> 00:08:05.132 Our arithmetic, not theirs, works out to roughly RM56 million of investment per expected job. 00:08:05.132 --> 00:08:09.302 The ministry says 3,523 of those jobs will be high-skilled. 00:08:09.302 --> 00:08:13.259 It expects Malaysians to fill about 89 per cent of them. 00:08:13.259 --> 00:08:15.817 Those are expectations, not outcomes. 00:08:15.817 --> 00:08:21.695 Only one of the twenty-seven is operating, and the rest are still under construction. 00:08:21.695 --> 00:08:26.604 We covered the power and water side on Wednesday, so I won't repeat it. 00:08:26.604 --> 00:08:33.785 There is a skills route, though. TalentCorp is developing a data centre facilities management programme. 00:08:33.785 --> 00:08:39.240 And every approved project has a minimum local content rule of thirty per cent. 00:08:39.690 --> 00:08:41.548 And path three? 00:08:41.548 --> 00:08:44.035 The users, which is everyone else. 00:08:44.035 --> 00:08:49.958 In June, Microsoft's Work Trend Index surveyed 2,000 Malaysian knowledge workers. 00:08:49.958 --> 00:08:56.613 It found 24 per cent were what it calls frontier professionals, the most advanced AI users. 00:08:56.613 --> 00:08:59.099 The global figure was 16 per cent. 00:08:59.099 --> 00:09:01.853 So we're ahead of the world? 00:09:01.853 --> 00:09:09.888 Among the office workers surveyed, yes. But only 32 per cent said their leadership was clearly aligned on AI. 00:09:09.888 --> 00:09:15.343 And just 19 per cent said they're rewarded for trying new ways of working. 00:09:15.343 --> 00:09:22.501 That's my shop exactly. My staff are ahead of me, and I'm the one with my foot on the brake. 00:09:22.501 --> 00:09:28.499 Which is why we keep saying the same thing on Fridays. Allow AI, and manage it. 00:09:28.499 --> 00:09:32.068 Banning it just pushes it onto personal phones. 00:09:32.068 --> 00:09:39.501 I learned that from my supervisor in Episode 7. She'd been using it on her own phone for a year. 00:09:39.501 --> 00:09:43.394 The government's own target sits mostly on this path. 00:09:43.394 --> 00:09:49.562 The National AI Action Plan aims for 300,000 to 500,000 new AI-related jobs by 2030. 00:09:49.562 --> 00:09:54.850 The fact sheet doesn't say how many are builders and how many are users. 00:09:55.300 --> 00:09:58.421 There's a fourth, smaller path people forget. 00:09:58.421 --> 00:10:05.632 Budget 2027 asks ministries, agencies and universities to give DBP verified and annotated language data. 00:10:05.632 --> 00:10:07.297 Annotation is real work. 00:10:07.297 --> 00:10:14.508 Linguists, teachers and translators could be paid to do it, though no ringgit figure has been announced. 00:10:14.508 --> 00:10:20.100 Now the serious part. Some people's work will change whichever way we go. 00:10:20.550 --> 00:10:20.834 Yes. 00:10:20.834 --> 00:10:28.780 Translation, transcription and customer service in Malay all get cheaper, whether the model is local or foreign. 00:10:28.780 --> 00:10:31.901 The worker feels the same change either way. 00:10:31.901 --> 00:10:38.417 In Episode 5, the Khazanah Research Institute said AI wasn't yet behind Malaysia's job losses. 00:10:38.417 --> 00:10:39.526 That can change. 00:10:39.526 --> 00:10:47.636 If you lose your job, register with PERKESO, the Social Security Organisation, under the Employment Insurance System. 00:10:47.636 --> 00:10:51.795 Since the first of October, its training fee cap is RM7,000. 00:10:51.795 --> 00:10:56.300 Apply through the LINDUNG Kerjaya portal or at a PERKESO counter. 00:10:56.300 --> 00:10:58.779 And for employers like me? 00:10:58.779 --> 00:11:03.415 Use your HRD Corp levy to retrain people before you restructure. 00:11:03.415 --> 00:11:08.630 In our view, retraining someone you already trust is worth trying first. 00:11:09.730 --> 00:11:14.073 Now the sceptics. The first is a company that actually tried. 00:11:14.073 --> 00:11:20.766 Germany's Aleph Alpha raised 500 million US dollars in November 2023, about RM2 billion today. 00:11:20.766 --> 00:11:25.893 Ten months later, it pivoted to helping firms use other people's models. 00:11:25.893 --> 00:11:34.508 Its chief executive, Jonas Andrulis, told Bloomberg, "Just having an European LLM is not sufficient as a business model." 00:11:34.508 --> 00:11:40.625 Korea is the live case. It wanted models at 95 per cent of frontier capability by 2027. 00:11:40.625 --> 00:11:49.765 On the thirtieth of September, the Korea Times reported the science ministry was reviewing whether to hold the final round at all. 00:11:49.765 --> 00:11:59.257 Minister Bae Kyung-hoon says the project will continue, and the government has proposed 4.7 trillion won for frontier models next year. 00:11:59.257 --> 00:12:02.492 That's roughly RM14 billion by our conversion. 00:12:02.492 --> 00:12:08.500 And that's the core problem. The frontier moves faster than any national budget cycle. 00:12:08.500 --> 00:12:18.071 In June, the research group Epoch AI said the best open-weight models had trailed the closed frontier by about four months since January. 00:12:18.071 --> 00:12:23.730 Open-weight means anyone can download the model and run it on their own machines. 00:12:24.180 --> 00:12:30.738 Hmm. So by the time we finish building ours, the free one has caught up? 00:12:30.738 --> 00:12:34.852 That's the risk. And the money gap is enormous. 00:12:34.852 --> 00:12:42.729 Epoch projects the big cloud companies will spend around 770 billion US dollars this year. 00:12:42.729 --> 00:12:45.180 That's about RM3.1 trillion. 00:12:45.630 --> 00:12:50.906 For scale, the whole of Budget 2027 is RM510 billion. 00:12:50.906 --> 00:12:56.553 So six Malaysian budgets. In one year. On computers. 00:12:56.553 --> 00:12:58.685 History has a pattern here. 00:12:58.685 --> 00:13:06.426 Japan's Fifth Generation computer project ran for ten years and cost about 400 million US dollars. 00:13:06.426 --> 00:13:14.088 When it was wound up in 1992, Tech Monitor reported most firms had built only one-off prototypes. 00:13:14.088 --> 00:13:17.466 Then there's Europe's Gaia-X cloud project. 00:13:17.466 --> 00:13:31.921 In 2022, Forrester analysts wrote that it had "failed to launch meaningful public cloud and data services." They said it was busy "building bureaucracies and PowerPoint presentations." 00:13:31.921 --> 00:13:35.190 I've sat in that meeting. Twice. 00:13:35.640 --> 00:13:37.974 And closer to home, there's Proton. 00:13:37.974 --> 00:13:45.043 The year before, it had received RM1.5 billion in government aid, on condition it found a foreign partner. 00:13:45.043 --> 00:13:47.310 In 2017, Geely took 49.9 per cent. 00:13:47.310 --> 00:13:50.405 I'd push back on using Proton as a failure story. 00:13:50.405 --> 00:13:55.141 The minister at the signing said Proton would remain the national carmaker. 00:13:55.141 --> 00:14:00.508 It kept its brand, and it gained Geely's platforms and engines. That's not giving up. 00:14:00.508 --> 00:14:04.487 That's building on someone else's engine and keeping the badge. 00:14:04.487 --> 00:14:10.053 Fair, I hadn't framed it that way. That's close to what YTL is doing with ILMU. 00:14:10.053 --> 00:14:15.757 So here's the split judgement. History warns against funding a national frontier model. 00:14:15.757 --> 00:14:18.970 It doesn't warn against owning the layers on top. 00:14:18.970 --> 00:14:22.792 And for small firms, the counter-case is mostly arithmetic. 00:14:22.792 --> 00:14:29.724 On Thursday, our illustrative sum put a thousand customer chats on ILMU's Flash model at about RM2 a month. 00:14:29.724 --> 00:14:34.130 An SME, a small or medium enterprise, doesn't need a national model. 00:14:34.130 --> 00:14:38.600 It needs one that handles Malay and keeps data where the law expects. 00:14:39.050 --> 00:14:45.948 And somebody to set it up. That's the bill nobody prints, and I print everything. 00:14:45.948 --> 00:14:54.971 True. MDEC's RM30 million for 4,000 firms works out at RM7,500 each, at most. We did that sum in Episode 8. 00:14:54.971 --> 00:15:01.970 As of the seventeenth, the last time we checked, its rules still weren't published. 00:15:03.070 --> 00:15:06.150 Our sponsor takes the next sixty seconds. 00:15:06.150 --> 00:15:12.310 When we return, we hear how the world's podcasters talk about countries like ours. 00:15:12.310 --> 00:15:15.294 Most AI projects stall in the same place. 00:15:15.294 --> 00:15:23.300 Someone runs a pilot, everyone agrees it's promising, and then nobody owns it. Hitower IT does the other part. 00:15:23.300 --> 00:15:31.962 AI discovery, a roadmap scoped to your budget, and the actual build, shipped into the workflows your team already uses. 00:15:31.962 --> 00:15:39.531 Strategy to code, no handoffs. Dutch precision, Malaysian know-how. A team you can meet in Kuala Lumpur. 00:15:39.531 --> 00:15:42.952 Book a free discovery call at hitowerit.com.my. 00:15:42.952 --> 00:15:49.430 They come back within one working day with where automation can help and what it's worth. 00:15:50.530 --> 00:15:56.784 Welcome back. I've spent the week listening to other people's podcasts, so you didn't have to. 00:15:56.784 --> 00:16:01.210 And reading page forty of their show notes, I assume. 00:16:01.210 --> 00:16:07.110 Page forty of the transcripts, actually. Five shows. Let me take them in turn. 00:16:07.110 --> 00:16:11.960 First, the Big Technology Podcast on the twenty-second of June, recorded live. 00:16:11.960 --> 00:16:15.069 The guest was Aaron Levie, chief executive of Box. 00:16:15.069 --> 00:16:23.277 Days after the Fable shutdown, he said other countries "now have far more incentive to stand up their own sovereign AI initiatives." 00:16:23.277 --> 00:16:28.580 He also suggested the outcome might be "net negative" for America's position in AI. 00:16:28.580 --> 00:16:33.308 In the transcript, the host's view on the net winner was "probably China." 00:16:33.308 --> 00:16:36.730 So the ban did Jensen Huang's selling for him. 00:16:37.180 --> 00:16:42.453 The chip sellers gain either way. Everyone who builds needs chips. 00:16:42.453 --> 00:16:45.147 Second, the 80,000 Hours podcast in July. 00:16:45.147 --> 00:16:50.865 The guest was Anton Leicht, a fellow at the Carnegie Endowment for International Peace. 00:16:50.865 --> 00:16:56.516 His subject was middle powers, which includes countries like us. His warning is blunt. 00:16:56.516 --> 00:17:03.482 Countries without frontier models, he said, "capture all the risk" as these models spread through society. 00:17:03.482 --> 00:17:07.098 But his preferred answer isn't everyone building alone. 00:17:07.098 --> 00:17:14.594 It's middle powers doing what they're good at in the supply chain, and combining that with what America does well. 00:17:14.594 --> 00:17:16.698 He did price a fallback, though. 00:17:16.698 --> 00:17:24.260 A frontier project by a group of allies would cost roughly 500 billion US dollars over four years, by his estimate. 00:17:24.710 --> 00:17:31.337 That's about two trillion ringgit. My accountant just fainted, and he isn't even listening. 00:17:31.337 --> 00:17:34.025 Third, Bloomberg's new series, Emerging. 00:17:34.025 --> 00:17:50.895 Its first episode, on the ninth of July, was called "Can India Win the AI Race Without Its Own OpenAI?" Menaka Doshi hosted, with Srikanth Velamakanni of Fractal Analytics and Professor Ganesh Ramakrishnan of the Indian Institute of Technology Bombay. 00:17:50.895 --> 00:17:57.801 The show calls it India's AI trilemma. Can a country be sovereign without controlling the full stack? 00:17:57.801 --> 00:18:05.049 I'm only giving you the framing there, because the show page I read doesn't summarise the guests' answers. 00:18:05.049 --> 00:18:12.062 Fourth, the Lex Fridman Podcast in March, with Jensen Huang again. This time he argued for open models. 00:18:12.062 --> 00:18:17.168 He said Nvidia releases them so "every country" can join the AI revolution. 00:18:17.168 --> 00:18:21.390 It has published the weights and data for its Nemotron models. 00:18:21.840 --> 00:18:31.961 And that matters here. YTL's Ilmu-Nemo-30B, launched in March, was built with Nvidia's Nemotron open models. 00:18:31.961 --> 00:18:35.615 So the open route is already our route. 00:18:35.615 --> 00:18:45.515 Fifth, an older one. The venture firm a16z ran an episode in May 2025 on why nations build their own models. 00:18:45.515 --> 00:18:54.040 Its description said AI "isn't just a technology anymore – it's geopolitical infrastructure." 00:18:54.490 --> 00:18:58.999 Okay. Five clever shows. What do they miss about us? 00:18:58.999 --> 00:19:06.348 Three things, I think. First, they talk about countries as if they're governments and giant labs. 00:19:06.348 --> 00:19:13.090 Malaysia is mostly small firms, and MDEC's adoption money is spread across 4,000 of them. 00:19:13.540 --> 00:19:17.164 Second, they assume one language per country. 00:19:17.164 --> 00:19:23.689 Our customers mix Malay, English, Chinese dialects and Tamil in a single message. 00:19:23.689 --> 00:19:27.232 Nobody in San Francisco is testing for that. 00:19:27.232 --> 00:19:36.460 My aunty mixes four languages and two emojis in one sentence. No model has passed that test yet, lah. 00:19:36.910 --> 00:19:41.265 Third, we're an export economy sitting between two giants. 00:19:41.265 --> 00:19:47.722 In Episode 12 we noted that nearly half of Malaysia's August exports were electronics. 00:19:47.722 --> 00:19:52.752 We sell into America's AI boom, while YTL builds on a Chinese base. 00:19:52.752 --> 00:20:00.035 A San Francisco podcast asks how America stays ahead. We have to ask how we keep both doors open. 00:20:00.035 --> 00:20:07.393 And there's a fourth. We host the data centres. Those shows treat compute as something you buy. 00:20:07.393 --> 00:20:12.350 We have the buildings, the power demand and the water questions. 00:20:12.800 --> 00:20:19.794 I hear what you're saying, but who's listening to us? We're not on Lex Fridman. 00:20:19.794 --> 00:20:25.076 Not yet. But the ISEAS argument is one those shows can't make. 00:20:25.076 --> 00:20:30.017 Our edge is being close to local customers and local data. 00:20:30.017 --> 00:20:38.856 So our edge is knowing what a kedai runcit in Kota Bharu actually needs. That's not nothing. 00:20:38.856 --> 00:20:43.160 It's not nothing. It may be the whole point. 00:20:44.260 --> 00:20:50.902 Now, the Friday positions. First, this week's Big Questions, one line each. 00:20:50.902 --> 00:21:00.140 Monday: do we train graduates on AI to do senior work sooner, or protect the junior roles where people learn? 00:21:00.590 --> 00:21:09.804 Tuesday: should public money first help the smallest firms buy tools that keep data safe, or back the firms ready to automate? 00:21:09.804 --> 00:21:18.750 Wednesday: do you pick the cheapest capable model whatever its origin, or pay more for one that sits where your data sits? 00:21:19.200 --> 00:21:27.287 And Thursday: should Malaysia keep paying to build national models, or rent the best and fund adoption? 00:21:27.287 --> 00:21:31.406 So here's the question for you to answer for yourself. 00:21:31.406 --> 00:21:38.959 If Malaysia's AI money can only go one way first, should it build the engine, or train the drivers? 00:21:38.959 --> 00:21:41.885 My position is train the drivers first. 00:21:41.885 --> 00:21:47.212 Rent the strongest models, and keep an open-weight backup running here. 00:21:47.212 --> 00:21:54.640 The best open models trail the frontier by about four months. Jobs will be won or lost on adoption. 00:21:55.090 --> 00:22:00.286 Mine is close to Aiman's. Spend on people and small firms first. 00:22:00.286 --> 00:22:06.782 But I want one thing owned here, and that's Malay that works for real customers. 00:22:06.782 --> 00:22:14.414 If DBP's data makes every model better at our languages, I'm happy for my taxes to pay for it. 00:22:14.414 --> 00:22:20.350 I'll take a narrower build position. Don't fund a national frontier model. 00:22:20.350 --> 00:22:28.533 Aleph Alpha and the Fifth Generation project show how that can go, and Korea is wrestling with it now. 00:22:28.533 --> 00:22:30.217 But own three layers. 00:22:30.217 --> 00:22:38.801 Our language data, the testing of models in our languages, and somewhere here to run them if access is cut. 00:22:38.801 --> 00:22:42.250 June showed that can happen in one evening. 00:22:42.700 --> 00:22:46.365 So we differ on emphasis, not direction. 00:22:46.365 --> 00:22:49.205 Yes. And the direction is already set. 00:22:49.205 --> 00:22:54.659 The National AI Action Plan for 2026 to 2030 has 28 national initiatives. 00:22:54.659 --> 00:22:58.994 It targets a top-ten place in global AI readiness by 2030. 00:22:58.994 --> 00:23:04.972 AI Malaysia Berhad, the national AI office under the Digital Ministry, leads it. 00:23:04.972 --> 00:23:12.118 AI Untuk Rakyat aims to give 100,000 young Malaysians free AI tools after six short modules. 00:23:12.118 --> 00:23:15.303 Budget 2027 promised 100,000 more places. 00:23:15.303 --> 00:23:21.051 HRD Corp keeps RM3 billion for training, a figure restated from last year. 00:23:21.051 --> 00:23:27.032 And The Vibes reported RM2 billion for the sovereign AI cloud in Budget 2026. 00:23:27.032 --> 00:23:29.214 But direction isn't delivery. 00:23:29.214 --> 00:23:35.534 Every one of these sits somewhere between announcement, allocation and disbursement. 00:23:35.534 --> 00:23:43.736 For MDEC's fund, we're still waiting for the rules. For the compute exchange, we're still waiting for prices. 00:23:43.736 --> 00:23:47.488 Which is why the useful part is what you do next. 00:23:47.488 --> 00:23:52.541 If you're a worker, learn to check AI's work in your own language. 00:23:52.541 --> 00:23:56.140 Your judgement is the part that stays valuable. 00:23:56.590 --> 00:24:01.602 If you run a small firm, write down every AI tool you depend on. 00:24:01.602 --> 00:24:05.596 Know where your data goes, and have a backup ready. 00:24:05.596 --> 00:24:11.964 And if you're a student, testing and improving models in Malay is a real skill now. 00:24:11.964 --> 00:24:18.331 Mesolitica's open models are free to study. Build something small, and show people. 00:24:18.331 --> 00:24:26.431 Malaysia doesn't need to win the frontier. It needs to make sure its people and its firms are never stuck. 00:24:26.431 --> 00:24:29.870 That's where the AI-YA team stands this week. 00:24:30.970 --> 00:24:38.980 One thing to do this weekend, in under an hour. Write a one-page AI backup plan for your work or your firm. 00:24:38.980 --> 00:24:41.926 List every AI tool you or your team uses. 00:24:41.926 --> 00:24:49.400 Next to each one, write where the data is stored and whose model sits underneath. Then ask one question. 00:24:49.400 --> 00:24:53.855 If this tool disappeared on Monday, what would we use instead? 00:24:53.855 --> 00:24:58.417 Finally, take one real task, such as a customer reply in Malay. 00:24:58.417 --> 00:25:02.833 Remove the names and numbers, and run it through your backup. 00:25:02.833 --> 00:25:07.250 If the backup can't do it, you've found your weak spot early. 00:25:08.350 --> 00:25:14.944 AI-YA runs AI lectures, workshops and implementation training for Malaysian companies. 00:25:14.944 --> 00:25:17.091 Details at ai-ya.my/podcast. 00:25:17.091 --> 00:25:19.871 Tomorrow is our Saturday week in review. 00:25:19.871 --> 00:25:28.278 We look at the week's AI news for Malaysia, what moved since we covered it, and the biggest story we haven't covered yet. 00:25:28.278 --> 00:25:32.030 We also revisit the week's Verdicts and Big Questions. 00:25:32.480 --> 00:25:37.744 Before then, tell us where you stand. Build the engine, or train the drivers? 00:25:37.744 --> 00:25:42.804 Leave your answer in the comments on the episode page at podcast.ai-ya.my. 00:25:42.804 --> 00:25:46.359 And share this with one person at work who needs it. 00:25:46.359 --> 00:25:50.179 Thank you for keeping us company at the end of a long week. 00:25:50.179 --> 00:25:55.617 Switch the screens off for a while this weekend, and enjoy the people at your table. 00:25:55.617 --> 00:26:01.120 Take care on the roads tonight, and we will see you tomorrow with the week in review.