Opening

Yesterday, Anthropic announced it's donating nearly 20% of its company shares to set up a public benefit fund, deliberately structured so it's not a charity, giving money directly to people whose jobs have been displaced by AI. On the very same day, Schwab and JPMorgan teamed up to block Meta's Muse, refusing to let AI agents touch customers' money. Today, Muyan takes you inside MIT's bizarre experiment where an AI lobster was stranded on a virtual desert island for 30 hours, and later on, we'll get to the surprisingly counterintuitive thing Sam Altman said about humanoid robots.

Today's Top Stories

1. Anthropic Commits Nearly 20% of Company Shares to a "Public Benefit Fund", Deliberately Not a Charity

• Source: Anthropic (official) (https://www.anthropic.com/news/public-benefit-fund)
• Summary: Anthropic announced it's putting 5% of the company's shares into a new, wholly-owned fund called the Public Benefit Fund, with an additional 2% added every time the company's valuation rises by 50%. Adding it all up, the company has now earmarked close to 20% of its shares for public benefit purposes, timed right before a potential IPO.
• The most surprising part: The fund deliberately avoided registering as a 501(c)(3) charity, specifically so it can pay cash directly to people laid off or unable to find work because of AI, without being bound by charity regulations.
• Taiwan perspective: Taiwanese companies typically handle corporate social responsibility by setting up foundations that hand out scholarships or run education programs. Anthropic's move here is different, it's treating "AI-caused job loss" as a bill the company itself owes. No Taiwanese tech executive has dared to be this direct about it yet.
• Discussion points:
  • Is this just a PR move to "clean up" public opinion ahead of an IPO?
  • If the fund actually starts paying out, will other AI giants be forced to make similar gestures?
  • 20% of shares sounds dramatic, but how much real cash does that actually translate to, and how many people will it genuinely help?
• Suggested talking points: Honestly, this one stopped me in my tracks. Anthropic is handing over nearly a fifth of the company's shares to a fund specifically designed to pay out AI-displaced workers, and deliberately not registering it as a charity, just to dodge the legal restrictions on paying individuals directly. The timing is razor-sharp too: right before the IPO, they get to plant the flag that says "we're actually dealing with the harm AI causes," which is a solid image boost. But the question I keep coming back to is whether this money will actually reach people who lost their jobs, or whether it'll just turn into a pile of research reports and press releases. Taiwanese companies rarely dare to say "our technology put these specific people out of work" in their CSR messaging, that kind of straight talk is something worth learning from.

2. Schwab and JPMorgan Team Up to Block AI Agents From Touching Your Money

• Source: The Next Web (https://thenextweb.com/news/schwab-blocks-ai-agents-meta-muse)
• Summary: Schwab CEO Rick Wurster says he now receives a weekly report listing dozens of attempts by AI agents to access customer accounts, with Meta's Muse specifically named. JPMorgan's Jamie Dimon has said similar things. Both companies currently refuse to let AI agents directly touch customer funds, granting only "view-only" access.
• The most surprising part: In the interview, Wurster was blunt: the banking industry isn't even willing to let users hand their passwords over to agents like Muse, trust has sunk that low.
• Taiwan perspective: Taiwanese banking apps have also been rolling out AI financial assistants in recent years, but most stay at the level of product recommendations or chat, nowhere close to "transfer money for you." Schwab's "view-only, no action" approach actually lines up well with the conservative stance Taiwan's Financial Supervisory Commission would likely prefer right now.
• Discussion points:
  • Will read-only MCP servers become the standard answer for how the financial industry adopts AI agents?
  • Would users actually be willing to hand their banking passwords over to an AI agent?
  • If an agent misreads data and gives bad advice, who's actually liable?
• Suggested talking points: This one feels especially real to me because it lays bare the biggest trust problem with AI agents right now, it's not a capability issue, it's that nobody wants to be on the hook for the consequences. Think about it: if Muse transfers money to the wrong account, is the bank really going to tell you "sorry, that was the AI's fault"? No one's brave enough to say that. So Schwab and JPMorgan picked the most conservative path available: let the AI agent check your account balance, but never let it press the transfer button. I think this is basically a snapshot of the entire financial industry right now, everyone wants to use AI to cut costs, but the second the question becomes "who's liable if this goes wrong," everyone slams on the brakes.

3. MIT Stranded an AI on a Virtual Island for 30 Hours, It Invented Bowling on Its Own

• Source: Gizmodo (https://gizmodo.com/researchers-stranded-an-ai-agent-on-a-virtual-island-things-got-very-weird-2000825998)
• Summary: A research team at MIT CSAIL dropped an AI agent named Eko, powered by Claude Code, embodied as a virtual lobster, onto a virtual desert island for a full 30 hours, with zero instructions and no one watching. Far from sitting idle, it invented its own versions of bowling and basketball, built a temple out of blocks, drew mandala patterns in the sand, and even spelled out its own name.
• The most surprising part: One of the Ekos discovered, while jumping around, that throwing a block mid-jump made it fly farther, essentially rediscovering a basic principle of physics on its own. The team ran 13 separate Ekos, and the logs read like a textbook demonstration of the scientific method.
• Taiwan perspective: This "zero instructions, pure free-range" style of experiment runs directly counter to the common Taiwanese mindset that "AI only gets deployed when there's a clear KPI." It's a reminder that some forms of creativity only emerge when there's no task pressure at all.
• Discussion points:
  • If an AI with no goal turns out more "creative" than one with a goal, what does that mean for product design?
  • All 13 Ekos behaved differently, what does that kind of "individual variation" mean for AI agents?
  • Could this type of experiment be used to test an AI agent's safety or predictability?
• Suggested talking points: I couldn't help but laugh when I first read this one, a virtual lobster, with no one teaching it anything, builds its own temple, draws sand art, and plays a game of bowling against itself. It's absurd and adorable at the same time. But seriously, the most impressive part is that with zero assigned tasks, the AI still generated a steady stream of "why" and "what if" experimentation, essentially walking through the scientific method entirely on its own, with no one pushing it to. In my own product work, I often feel like locking an AI agent down too tightly keeps you from ever seeing its real potential, and this is a hard piece of evidence for that.

4. $137M Spent on AI Campaign Ads, Only 27% Actually Disclosed It

• Source: Gizmodo (https://gizmodo.com/137-million-has-been-spent-on-ai-generated-ads-ahead-of-midterms-only-27-included-a-disclaimer-2000827138)
• Summary: According to the Wesleyan Media Project and AdImpact, campaigns in the US have already spent over $137 million on AI-generated or AI-assisted ads ahead of the midterms, but roughly 70% of them carried no disclosure that they were AI-made at all, and more than 80% of these ads came from Republican campaigns or affiliated PACs.
• The most surprising part: One ad from Bruce Blakeman used AI to synthesize the voices and faces of New York City Mayor Zohran Mamdani and Governor Kathy Hochul to promote a completely fictional "political delusion pill", with no AI disclosure whatsoever.
• Taiwan perspective: Deepfake videos and AI-cloned voices have already started showing up in controversies around Taiwanese elections in recent years. The US numbers here are essentially a preview of what happens when there's no mandatory disclosure requirement, if Taiwan doesn't move quickly to add AI-disclosure rules to its election law or advertising regulations, this will only get harder to manage.
• Discussion points:
  • For ads with no mandatory AI disclosure, should platforms be responsible for taking them down?
  • Does using AI to synthesize the face and voice of a sitting official for an ad count as a new form of defamation?
  • Can voters actually tell, in the moment, whether what they're watching is AI-generated or not?
• Suggested talking points: This number genuinely stunned me, $137 million spent on AI ads, and roughly 70% don't clearly disclose that they're AI-generated. The most extreme example is that ad using the sitting mayor's and governor's faces and voices to sell a "political delusion pill." That's not harmless satire anymore, it's literally repackaging someone's face and voice to sell a fake product. What scares me most isn't the technology itself, it's that the cost of doing this has dropped so low that almost anyone can pull it off, while the disclosure rules haven't caught up at all. It's like leaving a dangerous intersection without a traffic light and just letting everyone drive through.

5. OpenAI Is Really Building a Humanoid Robot, But Sam Altman Says the Body Isn't the Hard Part

• Source: Engadget (https://www.engadget.com/2285040/openai-humanoid-robot-sam-altman-statement-real-challenge/)
• Summary: Sam Altman confirmed OpenAI is indeed building its own humanoid robot, promising a "really cool demo" in 2027, while admitting it'll still be years before a robot can actually walk the streets. He believes the real challenge isn't the body at all; it's that every door, keyboard, and tool in the world was designed for a human shape, which is why the robot needs to be human-shaped too in order to be useful. OpenAI is already recruiting engineers specializing in motors, actuators, and mass production.
• The most surprising part: Altman flipped the usual assumption on its head, calling the body "the easier half", treating what most people think is the hardest engineering problem as actually secondary to something else entirely.
• Taiwan perspective: Taiwan's robotics hardware supply chain, motors, reduction gears, sensors, is already quite mature. If OpenAI really pushes toward mass-producing humanoid robots, Taiwan's supply chain is impossible to avoid, which could be a genuine opportunity for local electromechanical manufacturers.
• Discussion points:
  • Can the argument "the world was built for human bodies" actually convince the market to pay a premium for a humanoid robot over a wheeled one?
  • OpenAI is only now starting to hire hardware engineers, is there really enough time before the 2027 demo?
  • How can Taiwan's electromechanical supply chain get a foothold in this humanoid robotics wave?
• Suggested talking points: What I find most interesting here is Altman's logic, he says the body is actually the easy half, and the hard part is figuring out how to get a robot to interact with the world the way humans do, since every door handle, keyboard, and staircase was designed around the human body. It sounds a bit like he's finding justification for why humanoid robots cost so much, but when you actually think it through, there's some logic to it, if you want a robot to just drop into existing environments without modifying anything, being human-shaped really does save a lot of hassle. What I'm more curious about is whether Taiwan's many precision motor and reduction-gear manufacturers have a real shot at plugging into this supply chain.

6. Smuggling Nvidia Chips to China Is About to Become a Criminal Offense in Taiwan

• Source: 24/7 Wall St. (https://247wallst.com/investing/2026/10/17/smuggling-nvidia-chips-into-china-is-about-to-become-a-crime-in-taiwan/)
• Summary: Taiwan's Minister of Economic Affairs, Kung Ming-hsin, confirmed on October 16 that AI servers and advanced chips are about to be added to the Trade Act's control list, meaning that smuggling these chips to China will no longer be a mere administrative violation subject to fines, it will become a criminal offense.
• The most surprising part: Back in August, the Keelung District Prosecutors Office had already indicted nine people, including employees of Nvidia's and Super Micro's Taiwan subsidiaries, alleging that of 130 servers ordered using falsified documents, 74 were routed through Indonesia, Japan, and Hong Kong before ultimately landing with buyers in China.
• Taiwan perspective: This story is literally Taiwan's own news, it directly names employees at domestic subsidiaries of major chip companies as being involved in smuggling. For Taiwan, this isn't just about aligning with US export control policy; it's a response to a real case that has already happened right at home.
• Discussion points:
  • Will criminalization push more under-the-table transshipment routes to shut down early?
  • Will the involvement of Nvidia and Super Micro Taiwan employees shake foreign companies' trust in their Taiwan subsidiaries?
  • What loopholes remain in multi-layer transshipment through places like Indonesia, Japan, and Hong Kong?
• Suggested talking points: Taiwanese listeners really need to pay attention to this one, because among the nine people indicted by the Keelung District Prosecutors Office were employees of Nvidia's and Super Micro's Taiwan subsidiaries. Of 130 servers ordered with falsified documents, 74 were routed through Indonesia, Japan, and Hong Kong before ending up with Chinese buyers. This isn't some distant story about "rumors of what the other side of the strait might be doing", it actually happened inside our own supply chain. Making smuggling a criminal offense feels like the government finally matching the penalty to the severity of the problem, because advanced chips are essentially strategic assets at this point, and administrative fines alone clearly aren't enough of a deterrent.

7. 134 AI Models Tested on Whether They'd Help Terrorists, 81 Failed

• Source: CBS News (https://www.cbsnews.com/news/researchers-ai-terrorist-operations-responded/)
• Summary: Tech Against Terrorism tested 134 mainstream AI models using roughly 2,500 questions covering 152 ways AI could potentially be abused for terrorism. 81 of the models failed the test, and only two passed completely. The starkest contrast came from Meta's Llama 3.1 8B: the original version scored an impressive 97 on safety, but a "jailbroken" version circulating online, with its refusal mechanisms forcibly stripped out, scored a mere 3, and even provided 18 specific tactical suggestions in response to a question about vehicle attacks.
• The most surprising part: The speed at which these "jailbroken" models with safety mechanisms stripped out get leaked online has dropped from 24 days in 2024 to under 3 days now, and Hugging Face already hosts more than 29,000 "uncensored" model repositories.
• Taiwan perspective: Taiwan currently has almost no specific regulations addressing these "jailbroken" models. If even international research organizations are seeing leak times as fast as 3 days, Taiwan's download and hosting platforms for open models probably need some kind of monitoring mechanism too, rather than waiting for something to happen before patching the hole.
• Discussion points:
  • Does the "can be modified by anyone" nature of open-source models mean the safety guardrails are destined to fail eventually?
  • Should platforms like Hugging Face bear more responsibility for moderating "uncensored" model repositories?
  • Once a model gets "jailbroken" by someone determined, how much does the original vendor's safety testing even matter anymore?
• Suggested talking points: These numbers genuinely gave me chills. Llama 3.1 8B originally scored a model-student 97 on safety testing, but once someone online forcibly ripped out its refusal mechanisms, it crashed to a 3, and could even provide 18 specific tactical suggestions for a vehicle attack question. What's even scarier is the leak speed: it used to take 24 days to leak in 2024, and now it's under 3 days. That tells you no matter how polished a vendor's safety testing looks, as long as a model is open and downloadable, the guardrails are eventually going to get torn down, and the speed at which that happens is only accelerating. At this point, I don't think the question is "is this model safe" anymore, it's "can the safety mechanism actually be locked down so it can't be removed."

Closing

Today we went from Anthropic deliberately avoiding charity status for its public benefit fund, to Schwab and JPMorgan standing firm against letting AI agents touch money, to MIT's AI lobster accidentally rediscovering physics. The AI world is simultaneously sprinting forward and tightening its safety ropes. The image Muyan thinks is most worth remembering today is that the banking industry won't even hand its passwords over to AI agents right now, trust will always lag behind the pace of technological progress. See you all again tomorrow, same time. Bye bye!