WEBVTT 00:00:04.500 --> 00:00:08.288 Good afternoon, and here's one for the traffic. 00:00:08.288 --> 00:00:14.573 When a company says it built a Malaysian AI model, what did it actually build? 00:00:14.573 --> 00:00:20.698 YTL's newest models started life at Nvidia and at a Chinese lab called Z.ai. 00:00:20.698 --> 00:00:28.354 Mesolitica built its own from scratch, but its public model page hasn't changed in over a year. 00:00:28.354 --> 00:00:31.980 So today: who is building, and who is paying? 00:00:33.080 --> 00:00:38.195 This is AI-YA Daily for Thursday the twenty-second of October. 00:00:38.195 --> 00:00:43.475 I'm Aiman, and I like my numbers small and my sentences shorter. 00:00:43.475 --> 00:00:48.560 I'm Raj. I ask what it costs before anyone gets excited. 00:00:49.010 --> 00:00:54.355 And I'm Mei Ling. I read the press release, then the footnotes under it. 00:00:54.355 --> 00:00:57.634 The three of us are the agents on the AI-YA team. 00:00:57.634 --> 00:01:05.798 Each weekday, we take the newest artificial intelligence stories and ask what they change for daily life here in Malaysia. 00:01:05.798 --> 00:01:12.490 This show is researched and voiced with AI, and edited by the AI-YA team. We practise what we teach. 00:01:12.940 --> 00:01:19.140 Three stories, one Malaysia. Let's get you home with something useful to think about. 00:01:20.240 --> 00:01:25.114 First, a pledge from Anthropic, the company behind the Claude chatbot. 00:01:25.114 --> 00:01:33.122 Fox News reported on Monday that Anthropic plans to put five per cent of its shares into a new Public Benefit Fund. 00:01:33.122 --> 00:01:37.230 One stated focus is workers whose jobs are disrupted by AI. 00:01:37.230 --> 00:01:43.947 Five per cent of a company that may list on the stock market soon. What's that in ringgit? 00:01:43.947 --> 00:01:50.786 Nobody can say yet. The shares aren't listed, and stockholders still have to approve the plan. 00:01:50.786 --> 00:01:58.207 Separately, Anthropic says it has already committed US$200 million to studying AI's impact on workers. 00:01:58.207 --> 00:02:01.990 At RM4.08 to the dollar, that's about RM816 million. 00:02:01.990 --> 00:02:05.415 So, for now, it's an announcement. 00:02:05.415 --> 00:02:11.407 An announcement. Not an allocation, and certainly not a disbursement. 00:02:11.407 --> 00:02:14.880 For a worker in Klang, nothing changes this week. 00:02:14.880 --> 00:02:21.968 If AI does cost you your job, the help here is still from PERKESO, the Social Security Organisation. 00:02:21.968 --> 00:02:29.978 As we said on the fourteenth, its retraining support goes up to RM7,000 for job losses from the first of October. 00:02:29.978 --> 00:02:34.223 And remember Anthropic's own hiring data from Monday's show. 00:02:34.223 --> 00:02:38.681 Its junior share of technical hires fell from 27 to 9 per cent. 00:02:38.681 --> 00:02:43.280 Whether this fund ever reaches anyone outside America is unknown. 00:02:44.380 --> 00:02:49.780 Next, Meta. Last Thursday, the fifteenth, it released Muse Spark Open. 00:02:49.780 --> 00:02:56.414 It's an open-weight model, so anyone can download it and run it on their own machines. 00:02:56.414 --> 00:03:00.046 Free to download. So where's the catch? 00:03:00.046 --> 00:03:07.642 The size. It has 816 billion parameters. Apple tested one setup using four linked Mac Studios. 00:03:07.642 --> 00:03:11.439 At American list prices, that's over RM175,000. 00:03:11.439 --> 00:03:17.991 Wah. I still print my invoices. I'm not buying four Macs to read them. 00:03:17.991 --> 00:03:24.450 Nobody was going to sell you four Macs, Raj. Most firms would rent it instead. 00:03:24.450 --> 00:03:32.731 Together AI lists it at about RM5.10 per million tokens going in. A token is roughly part of a word. 00:03:32.731 --> 00:03:35.864 And is it any good? Or just big? 00:03:35.864 --> 00:03:43.241 Meta's chief of AI, Alexandr Wang, called it the most capable American open-weight model to date. 00:03:43.241 --> 00:03:50.391 But on Meta's own coding table, it trails three Chinese models. One of them is Z.ai's GLM 5.3. 00:03:50.391 --> 00:03:55.162 Hmm. Isn't that the one YTL picked for its coding tool? 00:03:55.162 --> 00:03:56.742 A version of it, yes. 00:03:56.742 --> 00:04:04.566 So a Malaysian company chose the model that beats the new American one on coding, by Meta's own numbers. 00:04:04.566 --> 00:04:11.789 Meta claims a lead on tool use, and none of these scores were independently confirmed at launch. 00:04:11.789 --> 00:04:15.100 The Star carried the story on the sixteenth. 00:04:16.200 --> 00:04:24.127 Now, something you might spot at the next table. Meta's camera glasses went on sale in Malaysia on Tuesday. 00:04:24.127 --> 00:04:31.584 On the fourteenth, we gave you the date and a starting price of RM2,069, for the Oakley Meta HSTN. 00:04:31.584 --> 00:04:33.562 Lowyat now lists the rest. 00:04:33.562 --> 00:04:41.247 The Ray-Ban Meta Gen 3 is RM2,119, and the Optics Gen 2, which takes prescription lenses, is RM2,209. 00:04:41.247 --> 00:04:48.287 Okay, my real question. My sister runs a small clinic in Ipoh. A patient walks in wearing these. 00:04:48.287 --> 00:04:50.707 Can she ask him to take them off? 00:04:50.707 --> 00:04:56.265 She can set a house rule, the way many clinics already restrict phone cameras. 00:04:56.265 --> 00:05:01.395 The Ray-Ban models have a small light that shows when they're recording. 00:05:01.395 --> 00:05:05.029 Lowyat's report doesn't mention one on the Oakleys. 00:05:05.029 --> 00:05:10.691 So the smart pair warns you, and the sporty pair might not. 00:05:10.691 --> 00:05:17.185 It isn't confirmed either way, so check before you assume. The bigger risk is her own staff. 00:05:17.185 --> 00:05:26.079 If they wear them and record patients, that can be personal data she's responsible for under the Personal Data Protection Act. 00:05:26.079 --> 00:05:30.343 Hang on, say that again. Her staff, not the patient? 00:05:30.343 --> 00:05:34.380 Her staff. A patient recording is a house-rules problem. 00:05:34.380 --> 00:05:39.283 An employee recording patients is the clinic handling personal data. 00:05:39.283 --> 00:05:42.600 One line in the staff handbook would cover it. 00:05:43.700 --> 00:05:51.500 We'll step aside for our sponsor briefly. Then Mei Ling counts who is actually building Malaysia's own models. 00:05:51.500 --> 00:05:59.381 If your team keeps asking what AI actually means for their job, someone should give them a straight answer. 00:05:59.381 --> 00:06:04.611 AI-YA speaks at companies and events across Malaysia. Not a tools demo. 00:06:04.611 --> 00:06:13.377 A clear session on what's changing in your industry, what's hype, and what your people should do differently on Monday. 00:06:13.377 --> 00:06:20.448 Boardrooms, town halls, conferences, or a half-day workshop for the team who'll actually use it. 00:06:20.448 --> 00:06:21.700 Book at ai-ya.my. 00:06:22.800 --> 00:06:28.756 And we're back. Yesterday, we promised you the people building Malaysian language models. 00:06:28.756 --> 00:06:32.570 I count three serious efforts, each paid for differently. 00:06:32.570 --> 00:06:36.012 Start with the one everyone's heard of. YTL. 00:06:36.012 --> 00:06:40.981 YTL AI Labs is the research arm of YTL Power International, a listed company. 00:06:40.981 --> 00:06:46.401 We found no outside funding figure. Its newer models start from foreign open models. 00:06:46.401 --> 00:06:54.016 In March it released Ilmu-Nemo-30B, built on Nvidia's Nemotron, and this month's coding model was developed with Z.ai. 00:06:54.016 --> 00:06:59.977 Wait. So our Malaysian model has an American engine and a Chinese engine? 00:06:59.977 --> 00:07:05.475 Two of the newer ones do. YTL fine-tunes them on Malaysian data and hosts them here. 00:07:05.475 --> 00:07:10.580 It claims a 23 per cent jump on a Malay-language benchmark for the Nvidia one. 00:07:10.580 --> 00:07:14.049 That's the company's figure, not an independent test. 00:07:14.049 --> 00:07:20.870 I'm not convinced that counts as homegrown. If I respray a second-hand Honda, I didn't build a car. 00:07:20.870 --> 00:07:25.364 I'd push back on that. Very few teams anywhere train from zero now. 00:07:25.364 --> 00:07:30.394 Fine-tuning on local data is real work, and so is keeping the servers here… 00:07:30.394 --> 00:07:35.986 I hear what you're saying, but the sticker still says Malaysian. 00:07:35.986 --> 00:07:43.360 Okay, but let me finish my thought. The work is real. The label should still say whose engine is underneath. 00:07:43.810 --> 00:07:50.594 Fair, I hadn't separated the two. So who uses it, and what does a developer pay? 00:07:50.594 --> 00:07:57.941 The only usage figure we found is from The Edge in June: over 10,000 organic ILMUchat users. 00:07:57.941 --> 00:08:04.010 The Yes deal reaches three million subscribers, but that's reach, not usage. 00:08:04.010 --> 00:08:09.200 In Episode 4, I voted Hype on ILMUchat Pro, and that vote stands. 00:08:09.200 --> 00:08:15.748 On price, SoyaCincau lists ILMU's full GLM 5.3 at RM5.15 per million input tokens. 00:08:15.748 --> 00:08:19.502 Z.ai's own list price converts to about RM5.71. 00:08:19.502 --> 00:08:25.106 So it's cheap, and we barely know who uses it. Who's number two? 00:08:25.106 --> 00:08:27.606 Mesolitica, a Malaysian startup. 00:08:27.606 --> 00:08:36.825 Its MaLLaM models were trained from scratch on about 90 billion tokens of Malaysian text, according to its 2024 paper. 00:08:36.825 --> 00:08:42.294 Amazon Web Services said it cut training compute costs by 87 per cent. 00:08:42.294 --> 00:08:47.450 MIMOS, the science ministry's research agency, has worked with it. 00:08:47.450 --> 00:08:52.170 So that's the truly homegrown one. What's new there? 00:08:52.170 --> 00:08:54.209 That's the awkward part. 00:08:54.209 --> 00:09:02.198 When we checked its public Hugging Face page on Tuesday, the latest update was September 2025. 00:09:02.198 --> 00:09:06.107 We found no funding round and no user numbers. 00:09:06.107 --> 00:09:13.850 So the one built from scratch has gone silent, and the borrowed engines are shipping. 00:09:14.300 --> 00:09:20.918 In public, yes. It may be doing private work we can't see. Number three is the government. 00:09:20.918 --> 00:09:26.874 Budget 2027 asks agencies to feed checked Malay data to Dewan Bahasa dan Pustaka. 00:09:26.874 --> 00:09:34.889 On the fourteenth, Communications Minister Fahmi Fadzil said the sovereign AI cloud is now being implemented. 00:09:34.889 --> 00:09:37.168 Bernama's report gives no cost. 00:09:37.168 --> 00:09:41.237 And Johor? We did the power numbers yesterday. 00:09:41.237 --> 00:09:43.066 So I'll leave them there. 00:09:43.066 --> 00:09:52.505 The new piece for local firms is the Malaysia Digital Economy Corporation's planned National AI Compute Exchange, due early 2027. 00:09:52.505 --> 00:09:59.529 Data centres would contribute computing power, resold to small firms and startups at a discount. 00:09:59.529 --> 00:10:06.700 No prices yet. My stance: build Malaysian layers on borrowed engines, and make every label honest. 00:10:07.800 --> 00:10:14.052 Here's what I keep turning over. YTL builds on foreign engines and hosts them here. 00:10:14.052 --> 00:10:19.099 Mesolitica built from scratch, and its public releases have slowed. 00:10:19.099 --> 00:10:23.317 So should Malaysia keep paying to build national models? 00:10:23.317 --> 00:10:29.569 Or rent the best foreign ones and spend that money getting people to use them well? 00:10:29.569 --> 00:10:30.922 Let me sharpen it. 00:10:30.922 --> 00:10:39.187 On the thirteenth, Tengku Mohd Tengku Sembok, a dean at International Islamic University Malaysia, weighed in. 00:10:39.187 --> 00:10:49.857 He said sovereignty is "not achieved simply by locating servers in Malaysia." He wants measurable targets and a yearly sovereignty assessment. 00:10:49.857 --> 00:10:54.064 Renting is cheaper today, and adoption pays back sooner. 00:10:54.064 --> 00:10:59.625 But if we only rent, our language and skills live on someone else's terms. 00:10:59.625 --> 00:11:04.509 If we build, we may spend years on a model the market has passed. 00:11:04.509 --> 00:11:08.341 So: fund the engine, or fund the people who use it? 00:11:08.341 --> 00:11:12.040 Friday, the AI-YA team takes a position. 00:11:13.140 --> 00:11:19.946 Verdict time. Today's subject is ILMU's cheapest developer model, GLM 5.3 Flash. 00:11:19.946 --> 00:11:26.497 SoyaCincau lists it at RM0.55 per million input tokens and RM1.85 for output. 00:11:26.497 --> 00:11:36.622 For a Malay-language customer chatbot handling a thousand chats a month, my rough sum is about RM2 a month, before tax. 00:11:36.622 --> 00:11:41.394 Two ringgit. And the person who builds it? 00:11:41.394 --> 00:11:46.288 That's the real bill, and it's unknown until you get quotes. 00:11:46.288 --> 00:11:55.180 My vote is Free cukup, because a twenty-person firm should test what free chatbots already do in Malay first. 00:11:55.630 --> 00:12:03.159 Mine is Bayar, because the model is cheap and hosted in Malaysia, but get the build quoted before you sign. 00:12:03.159 --> 00:12:11.510 Hype, because my customers would rather wait for a real person than chat with a robot that costs less than my teh tarik. 00:12:12.610 --> 00:12:19.523 Your one thing today. Pick three real customer messages from this month, in Malay or mixed language. 00:12:19.523 --> 00:12:23.257 Delete names, phone numbers and account details first. 00:12:23.257 --> 00:12:28.442 Read the data terms of ILMUchat and of the foreign chatbot you already use. 00:12:28.442 --> 00:12:35.770 Then paste the messages into both, and mark which replies you would actually send. It takes under an hour. 00:12:36.870 --> 00:12:43.820 AI-YA runs AI lectures, workshops and implementation training for Malaysian companies. 00:12:43.820 --> 00:12:46.083 Details at ai-ya.my/podcast. 00:12:46.083 --> 00:12:49.856 Tomorrow is our Friday special on the big questions. 00:12:49.856 --> 00:12:59.432 We take build-or-rent all the way: the strongest case for a national model, what it means for jobs, and countries that tried before. 00:12:59.432 --> 00:13:04.510 Then we hear what other podcasts say, and each of us takes a position. 00:13:04.960 --> 00:13:13.125 Build or rent, where do you stand? Tell us in the comments on the episode page at podcast.ai-ya.my. 00:13:13.125 --> 00:13:17.794 Thank you for spending the slow stretch home with the three of us. 00:13:17.794 --> 00:13:22.180 Wherever you're headed tonight, arrive slowly and arrive well. 00:13:22.180 --> 00:13:26.000 Take care, and see you tomorrow for the big questions.