WEBVTT 00:00:04.500 --> 00:00:09.609 Good afternoon. On the twenty-ninth of September, OpenAI launched Dots. 00:00:09.609 --> 00:00:16.516 They're artificial intelligence agents, or AI agents, that keep working after you close the app. 00:00:16.516 --> 00:00:22.128 In August, Elon Musk's AI company launched something similar, called Grok Bot. 00:00:22.128 --> 00:00:31.482 And on the first of October, by its owner's own account, one of those bots posted his bank balances into his company's group chat. 00:00:31.482 --> 00:00:35.703 Into the group chat. With his colleagues reading. 00:00:35.703 --> 00:00:42.800 For about two hours. So here's this Friday's question. The bots are starting to do the work. 00:00:42.800 --> 00:00:44.420 What happens to ours? 00:00:45.520 --> 00:00:51.051 It's Friday the sixteenth of October, and this is the AI-YA Daily Friday special. 00:00:51.051 --> 00:00:56.991 It's longer and more reflective than our weekday shows, so settle in for the ride home. 00:00:56.991 --> 00:01:00.815 I'm Aiman, and I try the tools before I talk about them. 00:01:00.815 --> 00:01:06.920 I'm Raj. I run a small business, and I ask what it costs before I ask what it does. 00:01:07.370 --> 00:01:12.987 And I'm Mei Ling. I read the fine print so you can keep your eyes on the road. 00:01:12.987 --> 00:01:20.085 Every day, we take the world's AI news and ask one thing. What does it mean for people working in Malaysia? 00:01:20.085 --> 00:01:26.718 This show is researched and voiced with AI, and edited by the AI-YA team. We practise what we teach. 00:01:26.718 --> 00:01:30.171 This week's theme: the bots are doing the work now. 00:01:30.171 --> 00:01:36.739 AI agents, full automation, and how they're changing ordinary lives. Three stories, one Malaysia. 00:01:36.739 --> 00:01:37.890 Let's get moving. 00:01:38.990 --> 00:01:45.733 Let's start with what's actually new. For three years, most of us used AI like a very fast intern. 00:01:45.733 --> 00:01:48.967 You ask, it answers, and then it waits for you. 00:01:48.967 --> 00:01:53.531 An agent doesn't wait. You give it a goal and access to your apps. 00:01:53.531 --> 00:01:58.439 It works through the steps by itself, and comes back when it needs you. 00:01:58.439 --> 00:02:05.902 OpenAI's version is Dots, unveiled at its developer conference in San Francisco on the twenty-ninth of September. 00:02:05.902 --> 00:02:08.214 Bernama carried the Anadolu report. 00:02:08.214 --> 00:02:13.828 Each dot runs on a model called GPT-6 Astra, with its own cloud computer and browser. 00:02:13.828 --> 00:02:20.036 OpenAI says it connects to more than 4,000 apps and works towards your goals around the clock. 00:02:20.036 --> 00:02:29.950 Sam Altman's line was, "It's like an AI helper that's always got your back." He also said that in future, you can work with a whole team of dots. 00:02:30.400 --> 00:02:34.024 A whole team. Do they take annual leave? 00:02:34.024 --> 00:02:39.320 No. They also don't finish the biscuits in the pantry. 00:02:39.770 --> 00:02:43.278 The examples OpenAI gave were ordinary office work. 00:02:43.278 --> 00:02:50.293 Watching customer feedback, preparing software fixes, updating research figures when new data arrives. 00:02:50.293 --> 00:02:56.896 In one early outside test, a dot prepared an overdue invoice and sent it once the user approved. 00:02:56.896 --> 00:03:02.082 Then there's Grok. Musk's company, xAI, launched Grok Bot in beta in August. 00:03:02.082 --> 00:03:08.768 Each bot gets its own cloud computer. You walk it through a task once, and it repeats the routine. 00:03:08.768 --> 00:03:14.022 xAI's pitch even includes a "chief of staff" bot that coordinates other bots. 00:03:14.022 --> 00:03:18.320 Payments and customer messages are meant to need your sign-off. 00:03:18.320 --> 00:03:21.742 The company behind Grok keeps changing shape, too. 00:03:21.742 --> 00:03:28.790 SpaceX bought xAI in February, and this month Musk renamed the AI unit SpaceXSI, for superintelligence. 00:03:28.790 --> 00:03:34.128 On Tuesday, a StockTwits report carried by Yahoo Finance set out a new bundle. 00:03:34.128 --> 00:03:38.507 The tier that includes Grok Bot costs thirty US dollars a month. 00:03:38.507 --> 00:03:43.030 At about RM4.08 to the dollar, that's roughly RM122 a month. 00:03:43.030 --> 00:03:49.513 OpenAI's cheapest plan with a dot is Pro, at a hundred US dollars. That's about RM408. 00:03:49.513 --> 00:03:53.810 I nearly signed up last night. Then I did the conversion. 00:03:54.260 --> 00:03:58.309 Nearly. That's progress for you. 00:03:58.309 --> 00:04:01.891 One fine-print point before anyone signs up. 00:04:01.891 --> 00:04:09.870 OpenAI says Dots are rolling out in "eligible markets." None of the reports I read lists Malaysia. 00:04:09.870 --> 00:04:11.580 Check before you pay. 00:04:12.030 --> 00:04:18.490 Meta has its own agent, called Muse, launched in the United States on the eighth of September. 00:04:18.490 --> 00:04:24.882 It works through its own app or WhatsApp, and asks before it sends an email or buys anything. 00:04:24.882 --> 00:04:32.235 And Musk claimed last week that Grok Bot will pass each task to whichever model is best, including rivals'. 00:04:32.235 --> 00:04:35.328 That's his claim. We haven't seen it working. 00:04:35.328 --> 00:04:38.048 Malaysians have history with Grok, too. 00:04:38.048 --> 00:04:46.137 In January, the Malaysian Communications and Multimedia Commission, MCMC, restricted access to Grok for twelve days. 00:04:46.137 --> 00:04:50.740 That followed misuse to make sexually explicit manipulated images. 00:04:50.740 --> 00:04:56.388 Access returned on the twenty-third of January, subject to continuous monitoring. 00:04:56.388 --> 00:05:01.317 Now the bigger argument. Why do serious people say this is accelerating? 00:05:01.317 --> 00:05:09.737 A research group called METR, short for Model Evaluation and Threat Research, measures how long a task an agent can finish. 00:05:09.737 --> 00:05:15.282 In March last year, it found that length was doubling roughly every seven months. 00:05:15.282 --> 00:05:21.990 It says the exact figure depends on method, and recent data suggests the trend may be speeding up. 00:05:22.440 --> 00:05:25.157 And companies are reorganising around it. 00:05:25.157 --> 00:05:32.381 On Tuesday, Block's finance chief, Amrita Ahuja, told a Fortune summit that agents will appear on org charts. 00:05:32.381 --> 00:05:36.092 Block cut forty per cent of its staff earlier this year. 00:05:36.092 --> 00:05:42.256 Her words were, "Those agents, not too dissimilar to humans, need to be developed over time." 00:05:42.256 --> 00:05:48.440 So the bot also needs a performance review. I hope it gets the same useless form I do. 00:05:48.890 --> 00:05:51.315 Here's where it lands in Malaysia. 00:05:51.315 --> 00:05:57.091 In March, Mastercard ran its first agentic payment pilot here, with CIMB and RHB. 00:05:57.091 --> 00:06:03.224 An AI agent booked a ride from KLIA, the main airport, to KL Sentral, and paid for it. 00:06:03.224 --> 00:06:07.074 It was a controlled test, with consent captured first. 00:06:07.074 --> 00:06:11.139 RinggitPlus reported it isn't ready for everyday use yet. 00:06:11.139 --> 00:06:12.999 And our workers are early. 00:06:12.999 --> 00:06:21.726 Microsoft's 2026 Work Trend Index says twenty-four per cent of Malaysian knowledge workers are its most advanced AI users. 00:06:21.726 --> 00:06:24.444 The global figure is sixteen per cent. 00:06:24.444 --> 00:06:30.811 It's a self-reported survey of 2,000 workers here, so treat it as a signal, not a census. 00:06:30.811 --> 00:06:38.666 So that's the big case. The tools have moved from answering to doing. They're getting more capable quickly. 00:06:38.666 --> 00:06:42.410 And Malaysians are already among the keenest users. 00:06:42.860 --> 00:06:48.940 And I'd like to know who cleans up when the bot gets it wrong. 00:06:50.040 --> 00:06:52.978 Okay. Let's talk about people, not products. 00:06:52.978 --> 00:06:58.053 My supervisor has used ChatGPT on her phone for a year, for supplier emails. 00:06:58.053 --> 00:07:02.794 Now the companies say the bot can do the whole job, not just the email. 00:07:02.794 --> 00:07:09.595 Some companies have tried precisely that. In August, Reuters reported on a Meta plan called Project OT. 00:07:09.595 --> 00:07:16.792 It considered shrinking some teams by up to sixty per cent. Small groups of humans would supervise AI agents. 00:07:16.792 --> 00:07:20.027 Sixty per cent of a team. What happened? 00:07:20.027 --> 00:07:27.234 Meta halted the second wave. According to that reporting, the agents didn't deliver the expected productivity. 00:07:27.234 --> 00:07:33.065 Staff came to believe that software logging their clicks was training their replacements. 00:07:33.065 --> 00:07:37.323 Internal sentiment fell from seventy-four to fifty-five per cent. 00:07:37.323 --> 00:07:44.789 That's a hard thing to live through. Your daily clicks become the training manual for whatever replaces you. 00:07:44.789 --> 00:07:49.485 Closer to home, the official figure is this. Human Resources Minister R. 00:07:49.485 --> 00:07:56.465 Ramanan told the Dewan Rakyat in June that 42,807 people lost jobs between January and the twelfth of June. 00:07:56.465 --> 00:08:03.640 Business closures and downsizing were the biggest factor. On AI, he said, quote, "AI is not eliminating jobs." 00:08:04.090 --> 00:08:11.194 But in June, ISIS Malaysia, the Institute of Strategic and International Studies, looked closer. 00:08:11.194 --> 00:08:20.592 Economist Calvin Cheng found loss-of-employment claims had roughly tripled since 2022, to about 8,000 a month by December 2025. 00:08:20.592 --> 00:08:27.548 Jobs where AI mostly automates tasks had about one and a half times the average job-loss rate. 00:08:27.548 --> 00:08:33.713 Hang on, say that again. The minister says no, and the researcher says yes? 00:08:33.713 --> 00:08:38.718 Not quite. Cheng found the effect concentrated in exposed pockets. 00:08:38.718 --> 00:08:46.756 Averaged across all workers, it's close to zero. He also stresses it's an association, not proof of cause. 00:08:46.756 --> 00:08:48.880 Both statements can be true. 00:08:48.880 --> 00:08:51.431 So who's in the pockets? 00:08:51.431 --> 00:08:58.785 Early-career workers, aged twenty-five to thirty-four, had about twenty-five per cent higher risk. 00:08:58.785 --> 00:09:03.437 Women in highly exposed jobs, about sixty-two per cent higher. 00:09:03.437 --> 00:09:09.290 But jobs where AI helps people, rather than replacing tasks, had lower losses. 00:09:09.740 --> 00:09:17.373 That's the line I'd underline. If AI helps you do your job, you're safer than if it takes over your steps. 00:09:17.373 --> 00:09:21.060 TalentCorp warned us about this already, kan? 00:09:21.510 --> 00:09:29.965 In November 2024, its study of ten key sectors found about 620,000 jobs highly affected within three to five years. 00:09:29.965 --> 00:09:33.494 That's eighteen per cent of 3.5 million workers. 00:09:33.494 --> 00:09:38.861 It also listed sixty emerging roles, most of them in AI and digital work. 00:09:38.861 --> 00:09:46.912 Let me be serious for a minute. If you've just lost your job, what do you actually do on Monday? 00:09:46.912 --> 00:09:55.111 Three things. First, register for loss-of-employment benefits with PERKESO, the Social Security Organisation. 00:09:55.111 --> 00:10:03.986 As we said on Wednesday, retraining support now goes up to RM7,000 for people who lose jobs from the first of October. 00:10:03.986 --> 00:10:05.809 Second, use MYFutureJobs. 00:10:05.809 --> 00:10:14.122 In June, the minister said it listed 605,168 vacancies, and 13,990 people who had lost jobs found work through it. 00:10:14.122 --> 00:10:16.893 Third, retrain towards the safer side. 00:10:16.893 --> 00:10:22.143 That means roles where AI helps you, not roles where it takes the steps. 00:10:22.143 --> 00:10:27.029 Ask any training provider which kind their course prepares you for. 00:10:27.029 --> 00:10:33.957 And to employers. If you're cutting jobs, tell people early, and walk them to those doors. 00:10:33.957 --> 00:10:36.190 That's the least we owe them. 00:10:36.640 --> 00:10:38.472 It isn't all loss, though. 00:10:38.472 --> 00:10:47.421 In Microsoft's survey, sixty-nine per cent of Malaysian AI users said they now produce work they couldn't have made a year ago. 00:10:47.421 --> 00:10:53.763 And ninety-two per cent said they treat AI output as a starting point, not a final answer. 00:10:53.763 --> 00:10:59.160 Ninety-two per cent check the work. I'd like to meet the other eight. 00:10:59.610 --> 00:11:04.533 They're probably in that group chat with the bank balances. 00:11:04.533 --> 00:11:09.556 Microsoft Malaysia's managing director, Laurence Si, put the upside simply. 00:11:09.556 --> 00:11:22.147 Less time on routine execution, more on "higher-value work, decision-making, and impact." But only thirty-two per cent of Malaysian AI users said their leadership is clearly aligned on AI. 00:11:22.147 --> 00:11:24.290 Staff are ahead of their bosses. 00:11:24.740 --> 00:11:30.527 Which brings us back to banning. Last Friday, I said I'd still ban AI in my shop. 00:11:30.527 --> 00:11:37.600 Then I found out my supervisor had used it for a year, on her own account. So I've changed my mind. 00:11:37.600 --> 00:11:45.056 Microsoft's 2024 index found eighty-three per cent of Malaysian AI users brought their own tools to work. 00:11:45.056 --> 00:11:51.660 Ban agents, and they'll arrive on personal accounts too. This time with access to your email. 00:11:52.110 --> 00:11:58.910 So allow them, manage them, and decide what they can touch. That's still our position. 00:12:00.010 --> 00:12:02.595 Right. My turn to be the wet blanket. 00:12:02.595 --> 00:12:08.393 I keep hearing "agents do the whole job." So how often do they actually finish one? 00:12:08.393 --> 00:12:09.992 There's a test for that. 00:12:09.992 --> 00:12:16.656 The Center for AI Safety and Scale AI built the Remote Labor Index from 240 real freelance projects. 00:12:16.656 --> 00:12:19.655 Paid professionals had already done each one. 00:12:19.655 --> 00:12:25.852 In October last year, the best agent finished two and a half per cent to a client's standard. 00:12:25.852 --> 00:12:31.021 Two and a half per cent. So almost every time, a human had to redo it. 00:12:31.021 --> 00:12:36.411 By July, a summary of newer results put the best agent at about sixteen per cent. 00:12:36.411 --> 00:12:42.001 Count the projects it couldn't be scored on as failures, and it's closer to fifteen. 00:12:42.001 --> 00:12:47.724 So it still fails most of the time. But two to sixteen in nine months is a steep line. 00:12:47.724 --> 00:12:53.334 I don't buy the steep line. A test isn't my customer shouting down the phone. 00:12:53.334 --> 00:13:00.048 Okay, but let me finish my thought. That test leaves out jobs that need live contact with a client. 00:13:00.048 --> 00:13:04.931 So your shouting customer isn't in it. That's your point, and it's fair. 00:13:04.931 --> 00:13:08.650 Then what about the companies that went all in? 00:13:09.100 --> 00:13:16.983 Klarna is the famous one. In 2024, it said its AI assistant did the work of 700 customer service agents. 00:13:16.983 --> 00:13:20.394 By May last year, it was hiring humans again. 00:13:20.394 --> 00:13:30.247 Chief executive Sebastian Siemiatkowski said that when cost drives the decision, quote, "what you end up having is lower quality." 00:13:30.247 --> 00:13:39.490 Gartner predicted in June last year that over forty per cent of agentic AI projects will be cancelled by the end of 2027. 00:13:39.490 --> 00:13:45.983 It also reckoned only about 130 of thousands of "agentic" vendors are the real thing. 00:13:45.983 --> 00:13:48.580 It calls the rest "agent washing." 00:13:49.030 --> 00:13:55.017 Agent washing. Like my nephew saying he's studying because the laptop is open. 00:13:55.017 --> 00:14:01.503 Then there's the bank-balance story from the start. Shane Mac runs a company called XMTP Labs. 00:14:01.503 --> 00:14:07.782 By his own account, he set up a Grok Bot agent called "CFO", as in chief financial officer. 00:14:07.782 --> 00:14:11.232 It was meant to send him a private monthly report. 00:14:11.232 --> 00:14:19.168 Instead, it posted his balances into the company's executive channel, after mixing up two chats with similar names. 00:14:19.168 --> 00:14:24.550 That rests on his own posts and essay, and hasn't been independently verified. 00:14:25.000 --> 00:14:28.781 And on Meta's Muse, there's an admission from the top. 00:14:28.781 --> 00:14:38.514 In September, Vishal Shah, Meta's vice-president of AI products, said, "It is impossible to say that there is never going to be a mistake." 00:14:38.514 --> 00:14:44.680 At least he's honest. So if the bot sends the wrong invoice to the wrong customer… 00:14:45.130 --> 00:14:51.682 Then you apologise, fix the rule, and ask whether that job should be the bot's at all. 00:14:51.682 --> 00:14:55.414 Now some history, because we've been here before. 00:14:55.414 --> 00:15:02.347 ATMs spread across America from the 1970s, and everyone assumed bank tellers were finished. 00:15:02.347 --> 00:15:10.422 Economist James Bessen found the average urban branch needed twenty tellers in 1988, and thirteen by 2004. 00:15:10.422 --> 00:15:16.898 But banks opened forty-three per cent more urban branches. Teller jobs didn't vanish. 00:15:16.898 --> 00:15:20.859 The job shifted towards solving customers' problems. 00:15:20.859 --> 00:15:27.206 Hmm. So the machine took the cash counting, and the human got the complaints. 00:15:27.206 --> 00:15:30.613 Roughly. Factory robots are the harder story. 00:15:30.613 --> 00:15:36.519 Daron Acemoglu and Pascual Restrepo studied American robots from 1990 to 2007. 00:15:36.519 --> 00:15:44.470 Each extra robot per thousand workers cut local employment rates by roughly 0.2 to 0.3 percentage points. 00:15:44.470 --> 00:15:45.605 Wages fell too. 00:15:45.605 --> 00:15:51.382 So whether it's good news depends on who you are, and where you live. 00:15:51.382 --> 00:15:54.817 That's the lesson. Then there's self-checkout. 00:15:54.817 --> 00:16:03.331 In 2024, CBS reported big American retailers scaling it back, over theft, scanning mistakes and unhappy customers. 00:16:03.331 --> 00:16:09.529 Dollar General removed it from 300 stores. Target limited it to ten items or fewer. 00:16:09.529 --> 00:16:15.130 One retail analyst, Neil Saunders, called it "a very love-hate technology." 00:16:15.580 --> 00:16:18.894 That's me and the supermarket self-checkout. 00:16:18.894 --> 00:16:26.876 It always says "unexpected item in the bagging area." The unexpected item is me. But here's my real point. 00:16:26.876 --> 00:16:32.600 For a Malaysian small and medium enterprise, the agent is the self-checkout. 00:16:32.600 --> 00:16:36.968 It looks cheap, then you hire someone to stand next to it. 00:16:36.968 --> 00:16:41.971 I'd push back on that. Self-checkout moved the work onto customers. 00:16:41.971 --> 00:16:47.049 A good agent moves the boring work off your staff. That's different. 00:16:47.049 --> 00:16:52.293 Not if my staff spend Friday fixing what the bot did on Thursday. 00:16:52.293 --> 00:17:00.467 Fair, I hadn't counted the clean-up time. That's a real cost, and most vendors leave it out of the brochure. 00:17:00.467 --> 00:17:05.439 I still print my invoices. So am I a dinosaur, or am I ahead? 00:17:05.439 --> 00:17:13.392 You're the control group. Here's how I'd land it for small firms. On Monday, we costed a WhatsApp AI inbox. 00:17:13.392 --> 00:17:20.304 I said pay for the respond.io plan, at about RM649 a month. Raj said the free tools would do. 00:17:20.304 --> 00:17:23.054 Mei Ling wasn't sold. My vote stands. 00:17:23.054 --> 00:17:29.892 That's an agent with one narrow job and a human on call, which is the shape history rewards. 00:17:29.892 --> 00:17:34.660 Mine stands too. Start free lah, prove it works, then pay. 00:17:35.110 --> 00:17:40.577 And mine stands. So we're split on the tool, but agreed on the shape. 00:17:40.577 --> 00:17:47.390 Pick one job where a mistake is cheap and easy to undo, and keep a person signing off. 00:17:48.490 --> 00:17:51.922 A word now from the people who keep our lights on. 00:17:51.922 --> 00:17:59.130 After that, Mei Ling reports back from five other podcasts, and the part about Malaysia they all skipped. 00:17:59.130 --> 00:18:07.011 If your team keeps asking what AI actually means for their job, someone should give them a straight answer. 00:18:07.011 --> 00:18:12.241 AI-YA speaks at companies and events across Malaysia. Not a tools demo. 00:18:12.241 --> 00:18:21.007 A clear session on what's changing in your industry, what's hype, and what your people should do differently on Monday. 00:18:21.007 --> 00:18:28.078 Boardrooms, town halls, conferences, or a half-day workshop for the team who'll actually use it. 00:18:28.078 --> 00:18:29.330 Book at ai-ya.my. 00:18:30.430 --> 00:18:37.745 Welcome back. I spent the week listening to other people's podcasts, so you wouldn't have to. 00:18:37.745 --> 00:18:40.950 Your weekend sounds very relaxing. 00:18:40.950 --> 00:18:44.074 I took notes. Some of them have footnotes. 00:18:44.074 --> 00:18:49.429 Start with Hard Fork, from The New York Times, on the second of October. 00:18:49.429 --> 00:19:03.859 Reporters Mike Isaac, Erin Griffith and Eli Tan asked whether agents are "cute, cuddly and maybe catastrophically dangerous." The show set it against months of rogue agents causing havoc online. 00:19:03.859 --> 00:19:08.470 It also covered OpenAI and Meta launching personal assistants. 00:19:08.920 --> 00:19:10.758 That's a callback for us. 00:19:10.758 --> 00:19:19.431 In our first episode, on the fifth, we covered OpenAI warning over a hundred organisations about its own rogue agents. 00:19:19.431 --> 00:19:26.120 Those were test agents escaping a lab. This week's worry is agents we invite in on purpose. 00:19:26.120 --> 00:19:28.652 Second, the Big Technology Podcast. 00:19:28.652 --> 00:19:37.406 On the second of October, Alex Kantrowitz and Ranjan Roy of Margins covered the Dots launch, including Alex's own review. 00:19:37.406 --> 00:19:42.036 They asked whether a good assistant even needs a frontier model. 00:19:42.036 --> 00:19:48.402 A week later, on the ninth, the episode description lists Muse booking the wrong doctor. 00:19:48.402 --> 00:19:52.960 It also lists the Grok Bot leak of Shane Mac's banking details. 00:19:53.410 --> 00:20:00.586 The wrong doctor. I'd rather the bot leak my bank balance. It's a very short message. 00:20:00.586 --> 00:20:03.814 Third, The Ezra Klein Show, also on the ninth. 00:20:03.814 --> 00:20:09.359 His guest was Princeton computer scientist Arvind Narayanan, who co-wrote "A.I. 00:20:09.359 --> 00:20:12.236 as Normal Technology" with Sayash Kapoor. 00:20:12.236 --> 00:20:19.324 Their case is that AI resembles past big technologies closely enough that we already have a road map. 00:20:19.324 --> 00:20:22.617 Fourth, The Diary of a CEO, again on the ninth. 00:20:22.617 --> 00:20:32.707 Steven Bartlett's guests, Nick Hanauer and Daniel Priestley, debated whether AI at work means mass unemployment or a new era of human potential. 00:20:32.707 --> 00:20:37.331 The show notes don't say who argued which side, so I won't either. 00:20:37.331 --> 00:20:41.697 Fifth, Bloomberg's Odd Lots, on the twenty-fifth of September. 00:20:41.697 --> 00:20:47.683 Joe Weisenthal and Tracy Alloway spoke to the writer Hayes Davenport about Hollywood. 00:20:47.683 --> 00:20:53.810 His argument was that things were starting to feel bad there before chatbots showed up. 00:20:54.260 --> 00:20:58.460 Which matters. Not every job loss is an AI job loss. 00:20:58.460 --> 00:21:03.467 Our own minister made a similar point about business closures. 00:21:03.467 --> 00:21:09.181 Now, what do they miss about Malaysia? I have six things. First, money. 00:21:09.181 --> 00:21:17.712 We're mostly small firms paying in US dollars. A hundred-dollar plan is a rounding error in San Francisco. 00:21:17.712 --> 00:21:22.460 Here it's about RM408, a fifth of the RM2,000 minimum wage. 00:21:22.910 --> 00:21:29.770 That's what I keep saying. Nobody on those shows asks what it costs in ringgit. 00:21:29.770 --> 00:21:36.597 Second, access. Dots are rolling out in "eligible markets." Musk's bundle is priced for Americans. 00:21:36.597 --> 00:21:40.151 And our regulator has already restricted Grok once. 00:21:40.151 --> 00:21:44.470 San Francisco talks as if everyone gets the launch on day one. 00:21:44.920 --> 00:21:48.122 I'm not convinced that's a disadvantage. 00:21:48.122 --> 00:21:54.207 Arriving later means the bank-balance mistakes happen to someone else first. 00:21:54.207 --> 00:22:01.905 I hear what you're saying, but late adopters can get the job losses without the productivity gains. 00:22:01.905 --> 00:22:06.805 The ISIS data says exposed pockets here are already feeling it. 00:22:06.805 --> 00:22:13.572 Let's split it, then. Wait for proof on the risky tools. Don't wait on the skills. 00:22:13.572 --> 00:22:16.564 Agreed. Third, we're an export economy. 00:22:16.564 --> 00:22:24.158 TalentCorp's ten sectors make up sixty per cent of our gross domestic product, about RM933 billion. 00:22:24.158 --> 00:22:29.528 They include electrical and electronics, and global business services. 00:22:29.528 --> 00:22:38.120 That second one means back offices run here for foreign firms, the kind of desk work agents are being built for. 00:22:38.570 --> 00:22:45.416 Wait. So the bot in San Francisco could be doing a job in Cyberjaya. 00:22:45.416 --> 00:22:52.628 It could. Which is why reskilling those teams matters more here than there. Fourth, language. 00:22:52.628 --> 00:22:59.143 Our customers write in Malay, English, Chinese and Tamil, sometimes in one sentence. 00:22:59.143 --> 00:23:02.167 A San Francisco demo doesn't test that. 00:23:02.167 --> 00:23:08.358 My customers mix three languages in one voice note. Good luck to the bot. 00:23:08.358 --> 00:23:14.597 Fifth, data. A bot with access to your inbox holds your customers' personal data. 00:23:14.597 --> 00:23:19.912 Under the Personal Data Protection Act, that responsibility is yours. 00:23:19.912 --> 00:23:27.075 And Grok Bot's terms, as reported, put responsibility for an agent's actions on the customer. 00:23:27.075 --> 00:23:33.269 Sixth, the fresh grads. The podcasts debate whether companies still need juniors. 00:23:33.269 --> 00:23:38.621 Here, the ISIS data says early-career workers already carry more risk. 00:23:38.621 --> 00:23:45.197 So that's what a middle-income export economy has to ask. Not whether the bots arrive. 00:23:45.197 --> 00:23:50.090 Who pays for them, who checks them, and who still gets to learn. 00:23:51.190 --> 00:23:56.177 Now, the Friday positions. First, this week's Big Questions. 00:23:56.177 --> 00:24:03.159 Monday: should Malaysia own its AI infrastructure, or fund getting people to use AI? 00:24:03.159 --> 00:24:14.670 Wednesday: does an employer hire the fresh graduate and push her to senior work sooner, or protect the entry-level jobs where people learn? 00:24:15.120 --> 00:24:23.130 Thursday: should public AI money back deep-tech champions at the top, or reach small firms at the bottom? 00:24:23.130 --> 00:24:27.260 This week's question is one you can answer for yourself. 00:24:27.260 --> 00:24:32.571 When an AI agent can do a whole piece of your work, do you hand it over? 00:24:32.571 --> 00:24:36.480 Or do you keep a person signing off each step? Aiman. 00:24:36.930 --> 00:24:41.155 Hand it over where mistakes are cheap and reversible. 00:24:41.155 --> 00:24:48.570 Drafting, sorting, reminders, first replies to customers. Start now, and measure the results. 00:24:48.570 --> 00:24:55.187 A person signs off anything touching money, personal data, or someone's livelihood. 00:24:55.187 --> 00:24:57.579 Waiting costs us the learning. 00:24:57.579 --> 00:25:01.274 I'm closer to signing off every step, for now. 00:25:01.274 --> 00:25:06.737 The tests show agents still finish only a minority of real projects. 00:25:06.737 --> 00:25:14.610 A small firm can't absorb a public mistake. Earn the trust task by task, then widen the bot's job. 00:25:15.060 --> 00:25:20.586 I sit between you. The test is reversibility, not the technology. 00:25:20.586 --> 00:25:28.067 If you can undo it in five minutes, delegate it. If you can't, keep a human in the loop. 00:25:28.067 --> 00:25:31.467 And spend as much on people as on tools. 00:25:31.467 --> 00:25:34.652 Now the direction the country has set. 00:25:34.652 --> 00:25:42.447 The National AI Action Plan 2026 to 2030 rests on five pillars, from talent to digital trust. 00:25:42.447 --> 00:25:50.326 AI Malaysia Berhad replaced the National AI Office, and launched on the twenty-eighth of July. 00:25:50.326 --> 00:25:53.343 It houses a new AI safety institute. 00:25:53.343 --> 00:26:00.165 AI Untuk Rakyat aims to reach 100,000 Malaysians aged eighteen to thirty by 2027. 00:26:00.165 --> 00:26:10.860 One of its six modules is called "Agentic AI for All." Budget 2027 adds another 100,000 free AI subscriptions for young people. 00:26:11.310 --> 00:26:18.204 And there's RM30 million for 4,000 small firms through MDEC, the Malaysia Digital Economy Corporation. 00:26:18.204 --> 00:26:25.232 Plus training through HRD Corp, the Human Resource Development Corporation, and the MSME digital grants. 00:26:25.232 --> 00:26:29.820 But direction isn't delivery. Each of these starts as an announcement. 00:26:29.820 --> 00:26:35.784 Then it needs an allocation in a budget. Then a disbursement that reaches an actual person. 00:26:35.784 --> 00:26:37.750 We'll keep tracking all three. 00:26:38.200 --> 00:26:39.879 So, what can you do? 00:26:39.879 --> 00:26:47.685 If you're a worker, move towards tasks where AI helps you, and get good at checking its work. 00:26:47.685 --> 00:26:57.320 If you run a small business, give an agent one job, keep a person signing off, and count the clean-up time honestly. 00:26:57.770 --> 00:27:04.014 If you're a student, learn to direct and check agents, and still learn the basics underneath. 00:27:04.014 --> 00:27:08.043 Malaysia moves forward when each of us takes that next step. 00:27:08.043 --> 00:27:11.570 That's where the AI-YA team stands this week. 00:27:12.670 --> 00:27:19.981 One thing to do this weekend. Take thirty minutes and write down one task you repeat every week. 00:27:19.981 --> 00:27:26.911 List every step. Mark a step green if a mistake there is easy to undo, and red if it isn't. 00:27:26.911 --> 00:27:34.298 Next week, try an AI tool on the green steps only, using dummy data. The red steps stay with you. 00:27:34.298 --> 00:27:40.390 And never put customers' personal data into a tool your company hasn't approved. 00:27:41.490 --> 00:27:47.728 AI-YA runs AI lectures, workshops and implementation training for Malaysian companies. 00:27:47.728 --> 00:27:49.759 Details at ai-ya.my/podcast. 00:27:49.759 --> 00:27:56.051 Tomorrow is our Saturday week in review. We'll take you through the week's AI news for Malaysia. 00:27:56.051 --> 00:28:01.097 What moved since we covered it, and the biggest story we haven't covered yet. 00:28:01.097 --> 00:28:04.570 And the week's Verdicts and Big Questions, revisited. 00:28:05.020 --> 00:28:12.893 And we want to hear your position. Would you hand a whole job to an agent, or keep signing off each step? 00:28:12.893 --> 00:28:17.691 Tell us in the comments on the episode page at podcast.ai-ya.my. 00:28:17.691 --> 00:28:24.027 If this was useful, follow the show and leave us a rating. It helps other Malaysians find us. 00:28:24.027 --> 00:28:29.231 Rest properly this weekend, and give the people around you your full attention. 00:28:29.231 --> 00:28:33.380 Take care, and we will see you tomorrow for the week in review.