Opening

The U.S. military nearly launched an interdiction and boarding operation against a Chinese cargo ship based on fake intelligence generated by AI, planes were already scrambled and on standby, and it wasn't until the last minute that officials discovered the entire report had been fabricated by a chatbot. Today, Muyan will also talk about how Anthropic plans to lock voting power firmly into the founders' hands before going public, and how 22 countries have joined forces to call on the United Nations to step in and regulate AI, though I'll tell you in a moment who's conspicuously missing from that list. Plus, there's a great story about how OpenAI's Astra and Anthropic's Claude Opus 5 teamed up to crack an Enigma cipher left over from World War II.

Today's Top Stories

1. U.S. Military Nearly Took Action Against a Chinese Cargo Ship Over Fake AI Intelligence

• Source: CNN (https://www.cnn.com/2026/09/18/politics/us-military-ai-false-intelligence-china-ship)
• Summary: An analyst at U.S. Special Operations Command fed a Chinese cargo ship's manifest into a chatbot, which generated a report claiming the vessel was carrying components for a nuclear weapons program. The military scheduled an interdiction, armed personnel prepared to board the ship, and planes were scrambled, until officials discovered at the last minute that the entire report was AI-generated misinformation.
• Most Surprising Detail: A source familiar with the matter said plainly that this nearly triggered an actual war.
• Taiwan Perspective: Given the already tense air and maritime situation around Taiwan, a case like this, where an AI hallucination turns into a military misjudgment, isn't just a curious story from afar; it's a direct warning for us. Verification mechanisms need to be solidly in place before AI is introduced into intelligence systems.
• Discussion Points:
  • How should military intelligence workflows be designed to catch AI hallucinations?
  • Will this incident make the Pentagon more cautious about adopting AI?
  • What does a similar risk mean when applied to the Taiwan Strait situation?
• Script Suggestion: I think this is the most terrifying story today. An analyst fed a cargo manifest to a chatbot, and the AI confidently fabricated intelligence about nuclear weapons components, nearly leading to an actual ship interdiction and armed confrontation. Just imagine what could happen if a wildly wrong AI report like this showed up amid tensions around the Taiwan Strait. This is absolutely not something we can take lightly.

2. Anthropic Founders Want to Lock Down Voting Power Before Going Public

• Source: TechCrunch (https://techcrunch.com/2026/09/25/anthropics-founders-seek-voting-control-ahead-of-ipo/)
• Summary: Anthropic is asking shareholders to approve a special stock structure that would give Dario Amodei and six other co-founders a combined 50.1% of voting power, meaning that even if shareholders object in the future, decision-making authority would remain firmly in the founders' hands.
• Most Surprising Detail: A company that brands itself around the idea that "AI is too dangerous, we need to hit the brakes" is now using corporate structure to ensure that only the founders can reach the brake pedal.
• Taiwan Perspective: For Taiwanese investors and startup circles who follow corporate governance, dual-class share structures aren't anything new, but applying it to a company that markets itself as "safety-first" feels especially ironic.
• Discussion Points:
  • How does concentrating voting power like this affect corporate governance transparency after going public?
  • Is there a contradiction between the "safety-first" slogan and concentrating control in the founders' hands?
  • Will other AI companies follow suit with similar pre-IPO maneuvers?
• Script Suggestion: This one is genuinely a bit awkward. Anthropic's public image has always been the company that talks the most about safety and hitting the brakes, and yet the first thing it does before going public is make sure only the founders can hit that brake. It's not that this is illegal or unreasonable, locking in voting power before an IPO is quite common for startups, but coming from this particular company, it just makes the contradiction stand out more.

3. 22 Countries Jointly Call on the UN to Regulate AI

• Source: NBC News (https://www.nbcnews.com/tech/tech-news/20-countries-call-global-ai-oversight-rcna599062)
• Summary: Initiated by Norwegian Prime Minister Støre and Finnish President Stubb, a total of 22 countries, including Germany, Canada, Australia, Singapore, the UAE, and Kenya, signed on to demand that the United Nations establish a new body responsible for pre-release model testing, independent evaluation, common standards, and reporting of major incidents.
• Most Surprising Detail: Neither the United States nor China is on the signatory list, and these are precisely the two countries actually training the most cutting-edge models.
• Taiwan Perspective: Although Taiwan isn't a UN member and can't directly participate in this kind of mechanism, if international standards do eventually emerge, Taiwan's AI industry and supply chain will sooner or later need to align with them. It's worth starting to pay attention to how these rules are being written now.
• Discussion Points:
  • How much actual enforcement power can an international AI governance body have without the U.S. and China at the table?
  • Is this essentially a strategy of "writing the rules first and waiting for you to join"?
  • For smaller AI players like Taiwan and South Korea, is this an opportunity or a path to marginalization?
• Script Suggestion: Twenty-two countries signing on to demand the UN regulate AI sounds pretty bold, but if you look closely at the list, you'll notice something awkward: the U.S. and China, the two countries actually training the most powerful models, are both absent. It's a bit like a group of people discussing how to set safety regulations for a factory, except the factory owner was never invited to the meeting. This carries more symbolic weight than actual enforcement power.

4. The Pentagon Pours $30 Million Into an AI Lie Detector

• Source: MIT Technology Review (https://www.technologyreview.com/2026/09/25/1145144/pentagon-ai-lie-detector)
• Summary: The U.S. Department of Defense has budgeted $30.3 million over five years for a project code-named "Polygraph+." At its core, it uses an AI algorithm to score responses, combined with a technology called standoff sensing that can read physiological signals remotely, without attaching any sensors to the body, to determine whether someone is lying.
• Most Surprising Detail: Experts put it bluntly: this takes a lie detector that was never very accurate to begin with and wraps it in a completely unexplainable AI black box.
• Taiwan Perspective: Both Taiwan's cybersecurity and HR communities are watching technologies like AI lie detection and emotion sensing closely. Once the military sets a precedent, it won't be long before vendors try to sell similar concepts into corporate interviews or border security scenarios, so it's worth thinking about regulations ahead of time.
• Discussion Points:
  • The scientific basis of lie detectors is already controversial. Does adding AI make them more or less credible?
  • How should the privacy concerns around standoff sensing, remotely reading physiological signals, be addressed?
  • Once the military adopts this, will it spill over into civilian law enforcement or corporate settings?
• Script Suggestion: Lie detectors have already been a subject of decades-long debate in psychology over their accuracy, and now the Pentagon plans to spend $30 million tying one to an AI whose reasoning can't be explained, all while reading your physiological responses remotely without even touching you. Experts put it plainly: combining two unreliable things doesn't make them reliable, it just makes a black box that's even harder to question.

5. Google Confirms Gemini 4 Is Racing Toward Launch

• Source: 9to5Google (https://9to5google.com/2026/09/24/google-says-gemini-4-release-is-coming-as-soon-as-possible/)
• Summary: Koray Kavukcuoglu, the new DeepMind head succeeding Demis Hassabis, made his first public appearance and confirmed that Gemini 4 has entered the post-training phase, with a target of launching "well before the end of the year." Internally, it's already being used to power testing for Antigravity.
• Most Surprising Detail: The first thing this new leader talked about publicly after taking office wasn't a grand AGI vision, it was "hurry up and ship Gemini 4."
• Taiwan Perspective: For Taiwan's AI application developers and enterprise clients, the top three model providers racing to release new versions means shorter evaluation cycles. Teams that haven't yet decided which model to use may need to speed up their assessments.
• Discussion Points:
  • Is Google really falling behind in this round of the model race?
  • What internal pressures does the new leader's shift in priorities reflect?
  • What impact will Gemini 4's launch have on the product cadence of OpenAI and Anthropic?
• Script Suggestion: DeepMind built its identity around chasing AGI, and yet under new leadership, the first public statement is "hurry up and get Gemini 4 out the door," not some grand vision. Honestly, that's pretty candid. Put plainly, Google is currently the one falling behind in this three-way race, and the new leader's first move is to shore up the basics before anything else.

6. Astra and Claude Opus 5 Team Up to Crack a WWII Enigma Cipher

• Source: TechCrunch (https://techcrunch.com/2026/09/25/astra-and-opus-just-passed-turings-other-test/)
• Summary: Two cryptanalysis hobbyists each handed a puzzle to OpenAI's Astra and Anthropic's Claude Opus 5, giving them only a single instruction: "go find an unsolved message in the database and crack it." The models went on to research on their own, dig up historical clues, and even write an Enigma simulator, ultimately cracking a WWII-era cipher that had gone unsolved for decades. What took Turing's entire team months to accomplish back then, the AI replicated in just a few days.
• Most Surprising Detail: The developers never walked the models through the steps at all; the models figured out how to build their own decryption tools on their own.
• Taiwan Perspective: This ability to hand an AI a vague task and have it break it down and execute on its own is both a warning and an opportunity for Taiwan's forensic security and legacy system maintenance engineers. Down the line, a lot of grunt work might genuinely be outsourced to models.
• Discussion Points:
  • Does this signal that AI now possesses a genuine form of "research capability"?
  • Will the cryptography community re-examine the security of legacy encryption systems as a result?
  • What would it look like if this autonomous research-and-tool-building ability were applied to cybersecurity offense and defense?
• Script Suggestion: This story genuinely gave me chills, not because of the cipher being cracked itself, but because of the process. The engineer just typed one sentence, and the AI went off, dug through the database, traced the historical context, and even wrote a simulator from scratch to run it. What Turing needed an entire team to accomplish back then just got replicated by one person plus one model in a matter of days. Honestly, that's more unsettling than the cipher-cracking itself.

7. Tesla's Optimus Mass Production Is Stuck on the "Hands"

• Source: Gizmodo (https://gizmodo.com/tesla-wants-to-build-20000-optimus-robots-a-week-first-it-has-to-figure-out-hands-2000817695)
• Summary: Tesla has ramped up Optimus production from dozens of units per week to several hundred per week last month, but that's still three orders of magnitude short of the goal of 20,000 units per week. What's actually holding back progress isn't the AI brain, it's the hands. The V3 hand and forearm contain over 100 parts, still have to be assembled by hand, and the tactile sensors frequently malfunction.
• Most Surprising Detail: In order for the robot to eventually replace human hands doing work, it currently takes a large number of human hands to assemble the robot's hands, one at a time.
• Taiwan Perspective: This is actually an opportunity for Taiwan's precision mechanical components, sensors, and automated assembly supply chain. The tiny components and assembly yield control needed for robotic dexterous hands happen to be exactly what Taiwan's manufacturing sector excels at.
• Discussion Points:
  • Is the bottleneck in humanoid robot mass production more about mechanics and manufacturing than AI, more than people realize?
  • What does this gap at Tesla mean for other humanoid robot makers?
  • Does Taiwan's supply chain have an opportunity to break into the dexterous-hand segment?
• Script Suggestion: Everyone's been talking about how smart humanoid robots' AI brains are, and yet Tesla's bottleneck turns out to be the most basic thing: the hand. One hand has over 100 parts, still needs to be assembled by hand one at a time, and the tactile sensors keep breaking down. It's honestly a bit ironic, building a robot meant to replace human hands, and right now the scarcest resource is exactly that: human hands. I actually think this represents a solid opportunity for Taiwan's precision manufacturing supply chain.

8. AI Scans 4.6 Million Compounds in Hours, Boosting Drug Design

• Source: Phys.org (https://phys.org/news/2026-09-ai-scans-million-compounds-hours.html)
• Summary: A research team used AI to scan 4.6 million compounds within hours, predicting the positions of hydrogen atoms in drug molecules, a notoriously difficult problem in crystallography and drug design that traditionally required expensive experiments to answer.
• Most Surprising Detail: AI's most practical contribution here isn't discovering new drugs on its own, it's compressing a screening process that would normally take years into a single afternoon.
• Taiwan Perspective: Taiwan has quite a few biotech startups and research institutions working on drug development. If tools like this that accelerate screening become widespread, it would be a huge boost for smaller teams with limited resources, since they wouldn't have to burn expensive lab equipment at every single step.
• Discussion Points:
  • Once AI accelerates screening, will experimental validation become the new bottleneck?
  • Will tools like this eventually be made available to academia or startups?
  • Compared to those other sensational AI headlines, is this kind of quiet, no-fuss application actually more worth paying attention to?
• Script Suggestion: Compared to the earlier stories about a near-war and a fight over voting power, this one is much quieter, but I think it's actually the most practical of the bunch. AI compressed a compound screening process that would normally take years into a single afternoon. It doesn't sound flashy, but this is exactly where drug development actually gets faster. Of course, validation still has to happen back in the lab, AI isn't magic, but the time saved is very real.

Closing Remarks

Today we went from a military crisis nearly triggered by fake AI intelligence, to Anthropic's power play before going public, to Gemini 4 jockeying for position and Optimus getting stuck on its hands, all kinds of AI stories from every angle, and honestly, every single one is worth thinking about a bit more. Muyan believes that the faster AI progresses, the less we can afford to overlook the details of governance and real-world implementation. See you next time at the same hour, bye for now.