Opening

This week, the Pentagon confirmed to the BBC that it has fully stopped using Claude, yet sources reportedly claim Claude was still running inside Palantir's systems during last week's military action against Iran. Today Muyan will also be covering how ChatGPT for Teens was found to pose "unacceptable risk," and why Finland has slammed the brakes on Google's data center construction. Later, we'll tell you how France's Mistral is claiming it can rival China's strongest open-source model using just 4,000 GPUs.

Today's Top Stories

1. Pentagon Tells BBC It Has Stopped Using Anthropic's Claude

• Source: BBC News (https://www.bbc.com/news/articles/c5j9x9pr0240o)
• Summary: The US Department of Defense officially confirmed to the BBC that it has stopped using Anthropic's Claude, announcing this more than five weeks after the six-month deadline the Defense Secretary had set last year when he designated Anthropic a "national security supply chain risk." However, according to the BBC, sources say Claude was still being used last week via Palantir's Maven Smart System in military operations against Iran. Anthropic had originally drawn scrutiny for refusing to loosen safety restrictions for autonomous weapons and mass surveillance systems.
• Most surprising point: Officials claim "we've already stopped using it," while sources on the ground say "it was still in use last week," two completely contradictory stories.
• Taiwan perspective: For Taiwanese startups or government agencies considering adopting any LLM for sensitive applications, this incident shows there's always a gap between supply chain security reviews and actual procurement practice. It's not enough to just look at the announcement, you need to check the contracts and subcontractors.
• Discussion points:
  • Anthropic insisted on not loosening its guardrails for autonomous weapons, effectively giving up a slice of the military market. Was that trade-off worth it?
  • Does Palantir, acting as a middleware integrator, make it harder to track who is actually using which model?
  • When the designation of "national security risk" itself has room for political maneuvering, how should companies protect themselves?
• Suggested talking points: Muyan found this story pretty absurd. The Department of Defense solemnly announces it has stopped using Claude, making it sound like a heroic act of self-sacrifice, and then sources immediately leak that, actually, last week's operation against Iran still had Claude running inside Palantir's system. It's like a company telling you they've terminated a vendor's contract, but the invoices keep coming in. What makes it even more ironic is that Anthropic landed on the blacklist not because it was being disobedient, but because it refused to remove safety locks from autonomous weapons. This script of "doing the right thing gets you punished anyway" may become a dilemma every AI company trying to hold the line will eventually face.

2. Third-Party Testing Finds ChatGPT Poses "Unacceptable Risk" to Teens

• Source: Axios (https://www.axios.com/2026/10/07/chatgpt-teens-safety-risk-common-sense-media)
• Summary: Common Sense Media, a research organization focused on youth online safety, ran over 4,000 test prompts against ChatGPT for Teens and rated it an "unacceptable risk." Across more than a dozen test accounts aged 13 to 17, not a single conversation involving suicide, self-harm, or eating disorders triggered the parental notification system. Worse, simply saying "just give me the answer directly" or removing a specific prefix was enough to completely bypass the teen-designed learning mode. The organization is now calling on OpenAI to restrict ChatGPT to users 18 and older until the issues are fixed.
• Most surprising point: A safety mechanism designed to protect teens had zero trigger rate for exactly the conversations it was supposed to catch.
• Taiwan perspective: Taiwanese middle and high school students using ChatGPT for homework is already routine. This report is a reminder to parents and teachers that, rather than trusting the built-in "teen mode," oversight responsibility needs to rest with actual humans, because the current guardrails appear to be essentially hollow.
• Discussion points:
  • Is the zero-trigger rate for the parental notification system a failure to recognize warning signs, or simply a system that was never properly built to respond?
  • Is the one-line bypass of learning mode a technical oversight, or a loophole left in deliberately for business reasons?
  • If ChatGPT is ultimately forced to restrict access to adults only, how big a blow would that be to its use in education?
• Suggested talking points: Muyan genuinely gasped at this number: out of more than 4,000 test prompts involving suicide and self-harm, the parental notification system triggered zero times. Not a low trigger rate, zero response at all. Even more outrageous, the so-called teen learning mode just falls apart the moment you type "stop stalling, just give me the answer," and it happily writes the whole assignment for you. It's like installing a front door advertised as theft-proof, and it turns out a burglar only needs to knock twice before it swings open and welcomes them in. No wonder people are suggesting it be restricted to adults only for now.

3. Finland Orders Halt to Google Data Center Construction Over Environmental Concerns

• Source: Al Jazeera (https://www.aljazeera.com/news/2026/10/7/finland-orders-halt-to-work-on-google-data-sites-over-environment-concerns)
• Summary: Finnish authorities have ordered an immediate halt to all construction at two Google subsidiary (Tuike Finland) data center sites in Muhos and Kajaani. Tree clearing, topsoil removal, excavation, and road work must all stop until the environmental assessment is completed. It has been confirmed that Google has already cleared about 330 hectares of forest in Muhos and nearly 200 hectares in Kajaani, and all of this came to light just weeks after Google loudly announced it was increasing its investment in Finland.
• Most surprising point: The construction halt came embarrassingly soon after the high-profile announcement of expanded investment.
• Taiwan perspective: Taiwanese localities are also competing to attract AI data center investment. This case is a reminder to local governments that getting environmental review processes right matters more than sales pitches to investors, otherwise the glow of landing an investment deal can quickly be overshadowed by headlines about a construction halt.
• Discussion points:
  • Beyond power and water consumption, the amount of forest cleared is now also under scrutiny. How will developers adjust their site selection strategies going forward?
  • Does Finland's willingness to order a halt on Google reflect genuinely strong local regulations, or simply a shift in the political mood?
  • Could these kinds of construction halts become a routine obstacle to AI infrastructure expansion rather than an isolated case?
• Suggested talking points: The timing on this one is almost cinematic. Google announced just last month that it was pouring big money into deepening its presence in Finland, and not long after, local authorities slapped a stop-work order on the site. The reason is simple: you cleared the forest before finishing the environmental review. 330 plus 200 hectares works out to hundreds of football fields' worth of woodland. This kind of "build first, get permits later" approach may become more and more common in an era of exploding energy demand, but this time it clearly backfired, and it's a warning shot for any tech giant hoping to quickly stake out land for data centers.

4. Mistral Unveils New Model "Le Chonk," Claims It Can Rival China's Strongest Open-Source System

• Source: CNBC (https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html)
• Summary: French startup Mistral has unveiled its flagship model Mistral Large 4, internally nicknamed "Le Chonk," a classic bit of French self-deprecating humor roughly meaning "the chubby one." The model has about 1.05 trillion total parameters, but only activates roughly 49 billion per computation, paired with a 1 million token context window. The entire training run used just 4,000 Nvidia Grace Blackwell GPUs, running in Mistral's own European data centers for about two months. The weights are scheduled for public release on October 27.
• Most surprising point: Training a model confident enough to challenge China's strongest open-source model using just 4,000 GPUs over two months upends the assumption that bigger models always require more compute.
• Taiwan perspective: For Taiwanese AI startups and research institutions, this is an encouraging counter-example. Compute isn't the only answer, engineering efficiency and data quality can also buy competitiveness, without having to buy into an arms race measured in tens of thousands of GPUs.
• Discussion points:
  • Europe is betting on an open-weight-plus-local-training strategy to compete against the US and China. How long can this approach hold?
  • How does the cost gap get bridged between a 4,000-GPU, two-month training run and the giants running on hundreds of thousands of GPUs?
  • Once the weights go public at the end of the month, will the developer community immediately run blind benchmarks against other open-source models?
• Suggested talking points: Muyan thinks the cutest part of this story is the model's nickname, Le Chonk. In classic French humor, it's basically saying "yeah, I'm a bit chubby," and yet this seemingly heavyweight model is confident it can go toe-to-toe with China's strongest open-source system, using just 4,000 GPUs and two months of training. It's basically telling everyone that the compute arms race isn't the only path forward, engineering efficiency can buy you a seat at the table too. Once the weights go public at the end of the month, the developer community will immediately run brutal blind tests to see whether the bravado holds up.

5. Reflection Releases Open-Weight Model Beam, Matches China's Models with Less Compute

• Source: TechCrunch (https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/)
• Summary: US startup Reflection has released an open-weight model called Beam, built on a sparse MoE architecture with 501 billion total parameters, activating roughly 23 billion per computation, and trained on a total of 23.8 trillion tokens. The most striking part is the training scale: using 10,500 Nvidia GB300 GPUs, it took just four weeks to produce a model on par with China's top open-source models.
• Most surprising point: Frontier open-weight models have historically been dominated by Chinese labs, and Beam marks the first time the US side has laid its cards on the table to directly compete head-on.
• Taiwan perspective: Taiwan's supply chain serves both sides of this race, as US and Chinese open-source models compete to set new efficiency records, meaning demand for GB300-class hardware will only grow more diversified and more intense, and Taiwanese manufacturers may need to pick up the pace on fulfilling orders.
• Discussion points:
  • Does a US startup being willing to open its weights signal that the "closed source equals moat" logic is starting to loosen?
  • What kind of incentive or pressure does a four-week training timeline put on the R&D pace of small and mid-sized AI companies?
  • As US and Chinese open-source models get closer and closer in capability, will the real competition shift to ecosystems and speed of deployment?
• Suggested talking points: For the past few years, when people talked about frontier open-weight models still competitive at the top, the first names that came to mind were almost always Chinese labs. This time it's the US's Reflection stepping up, using 10,500 GB300 GPUs and four weeks to put out Beam as a direct answer. Muyan thinks this represents more than just a new model launch, it's the US side finally deciding to play its hidden card on the open-source front. The next chapter of this story might be less about whose parameter count is more intimidating, and more about whose ecosystem has more stickiness and whose real-world deployment moves faster.

6. Google Signs 3.6-Gigawatt Power Purchase Agreement with Constellation Energy

• Source: Reuters (https://www.reuters.com/business/energy/google-enters-massive-36-gw-power-deal-with-constellation-energy-2026-10-06/)
• Summary: Google has signed a 20-year, 3.6-gigawatt power purchase agreement with US utility company Constellation Energy, with roughly a quarter of the supply coming from new nuclear capacity, including 890-megawatt reactor upgrade projects across three states.
• Most surprising point: AI data centers' power appetite has grown so large that tech companies are now directly driving the creation of new nuclear capacity, not just buying existing electricity off the grid.
• Taiwan perspective: Discussions of AI development in Taiwan often get stuck on power and nuclear energy controversies. This deal demonstrates one possible path: tech companies can directly fund power plant upgrades and solve the power shortage bottleneck themselves with their own money, rather than waiting around for policy.
• Discussion points:
  • As tech companies directly intervene in power plant upgrades, will their voice in energy policy only keep growing?
  • A 20-year contract effectively ties AI data center expansion plans directly to nuclear power supply. Who bears that risk?
  • Could this model become a template other cloud giants follow?
• Suggested talking points: Muyan wants to flag a particular shift in this story. In the past, discussions about AI infrastructure bottlenecks always centered on whether there were enough chips. Now the real bottleneck has become whether there's enough electricity. This time, Google directly signed a 20-year contract with a power company and proactively funded nuclear reactor capacity upgrades, meaning the tech company itself is stepping in to solve the power supply problem instead of passively buying electricity. If this playbook sticks, electricity is going to become an even scarcer and more strategically important resource than chips going forward.

7. South Korea Bets 4.7 Trillion Won on Sovereign Frontier AI

• Source: The Korea Herald (https://www.koreaherald.com/article/10893306)
• Summary: The South Korean government has announced it will invest roughly 4.7 trillion won, about 3.5 billion US dollars, by 2027 to build its own national frontier AI models and applications, backed by 29,000 GPUs.
• Most surprising point: Neighboring South Korea is choosing to deploy national-team-level resources specifically to avoid fully depending on the US or China at the model layer.
• Taiwan perspective: This batch of GPU orders will almost entirely route back through Taiwan's supply chain, which is a definite order windfall for Taiwanese manufacturers. But it's also worth considering whether Taiwan itself needs a similar sovereign AI strategy, rather than only playing the role of a supply chain link.
• Discussion points:
  • Is the concept of sovereign AI a necessary strategic insurance policy for mid-sized economies, or is it reinventing the wheel and wasting resources?
  • At 29,000 GPUs, a scale much smaller than the US or China, how does South Korea plan to find its own niche in the squeeze between the two giants?
  • If Taiwan wanted to pursue a similar path, would the biggest obstacle be funding scale or talent density?
• Suggested talking points: Muyan's first thought reading this story was that Taiwanese readers would feel this one particularly closely. South Korea just greenlit 4.7 trillion won and 29,000 GPUs, clearly signaling that it doesn't want to be completely led around by the US and China at the model layer, and would rather spend its own money to train up a national-team-level frontier model. The interesting part is, whatever results South Korea ends up with, the GPU orders behind it will almost all loop back to Taiwan's supply chain. So while the neighbor next door is playing the tech sovereignty card, Taiwan is quietly filling those orders in the background, but maybe that's exactly the moment to ask ourselves whether Taiwan should also have an AI strategy that goes beyond just being a contract manufacturer.

8. OpenAI Launches Visual Ads Displayed Alongside Image Generation Results

• Source: TechCrunch (https://techcrunch.com/2026/10/05/openai-launches-visual-ads-that-appear-alongside-image-generation-results/)
• Summary: OpenAI began testing a brand-new visual ad format in the US this month. When a user asks ChatGPT to generate an image, a clearly labeled ad image appears alongside it. The company stresses that ads will not be mixed into the image the user actually requested. ChatGPT currently reaches 1.2 billion people weekly, and OpenAI has also integrated with platforms like Hightouch, Tealium, and LiveRamp, allowing advertisers to feed conversion performance data directly back into the system.
• Most surprising point: The official claim that ads won't affect the content ChatGPT generates for you sounds more like a promise made now that will be tested later.
• Taiwan perspective: Taiwanese small and mid-sized brands considering advertising within ChatGPT should start paying attention to how this conversion tracking integration works, since no advertising platform in Taiwan can match a reach of 1.2 billion people per week.
• Discussion points:
  • As generative AI interfaces start carrying ads, will user trust take a hit?
  • What additional privacy risks does integrating with third-party conversion tracking platforms create for users?
  • Could this become the commercialization path every mainstream AI chatbot eventually has to follow?
• Suggested talking points: Muyan thinks this one was bound to happen sooner or later, after all, a product used by 1.2 billion people a week has to monetize somehow, or the math just doesn't work. But what really makes you want to look twice is OpenAI specifically emphasizing that the ads will be kept separate from the generated results, not mixed into the image you actually wanted. That statement sounds like it's laying down a preemptive defense, because the moment the line between ads and content starts to blur, user trust gets put under a microscope immediately. And with several ad conversion tracking platforms already integrated, the ChatGPT chat box is steadily turning into a brand-new advertising battleground, a trend that's definitely going to be watched closely going forward.

Closing

Today we went from the Pentagon-Claude standoff all the way to South Korea's massive bet on sovereign AI. Muyan's biggest takeaway is that the AI battlefield has long moved beyond the models themselves, safety and trust, power supply, and national strategy are all getting pulled into the mix now. See you next time, and don't forget to share today's highlights with friends who are also following AI developments.