Opening

The most expensive poem in human history is currently sitting in orbit a few hundred kilometers above the ground. Google's Project Suncatcher prototype data center satellite is online, carrying four TPUs that can only be powered up one at a time for about 20 minutes, and the very first command sent to it was to ask Gemma 4 to write a haiku. Meanwhile, Character.ai has filed for Chapter 11 bankruptcy, and the first bidder for its IP is none other than Disney, the company that once sued it. As for why Etsy is suddenly demanding that sellers photograph "the actual finished product," and which engineers are going to be pulling overtime when Claude Sonnet 4.5 gets retired at the end of the month, we'll get to that shortly.

Today's Top Stories

1. Google's First Data Center Satellite Goes Live, and the First Thing It Did Was Write a Poem

• Source: TechCrunch (https://techcrunch.com/2026/10/30/googles-first-data-center-satellite-is-online-and-writing-poetry/)
• Summary: Google's Project Suncatcher has launched an experimental "orbital data center" that is now communicating normally with the ground. The first prompt engineers sent up was a request for the satellite's onboard Gemma 4 model to write a haiku about being in space. What the mission is really testing, though, is whether bit flips caused by cosmic radiation will quietly corrupt computation results.
• Most surprising detail: The satellite carries four TPUs but generates only 1 kW of power, so only one can run at a time, and only for about 20 minutes. A textbook case of "astronomically expensive poetry."
• Taiwan angle: Taiwan's satellite component, power management, and thermal supply chains have spent years looking for an entry point here. If space-grade compute nodes actually materialize, the demand won't be for GPUs themselves, but for radiation-hardened packaging, low-power supplies, and optical communication modules, exactly the "unsexy but profitable" segments Taiwanese manufacturers excel at.
• Discussion points:
  1. Is moving data centers into orbit about round-the-clock solar power, or about escaping terrestrial power and land constraints?
  2. If bit flips happen silently, you'd never know your AI inference results were wrong. How do you validate against these "silent errors"?
  3. Google has signaled it will launch two more satellites in 2027 and test inter-satellite optical links. What's the bigger play here?
• Script notes: My first reaction to this story was to laugh, four TPUs, only one can run at a time, for 20 minutes, and you ask it to write a haiku. That's got to be the worst ROI in the universe. But once you're done laughing, take it seriously: this mission isn't about poetry at all, it's about radiation. On the ground, the odds of a cosmic ray flipping a bit in memory are low enough that we can pretend it doesn't happen. In orbit, it's routine. And the scariest failure mode for AI inference isn't a crash, it's an answer that "reads perfectly fine" but was miscalculated somewhere in the middle. What Google is doing right now is verifying whether answers computed in space can be trusted. If that step fails, all the solar power in the world doesn't matter.

2. Character.ai Files for Bankruptcy Protection, and the First Bidder Is Disney

• Source: CNN (https://edition.cnn.com/2026/10/13/tech/character-ai-bankruptcy-filing)
• Summary: Character.ai, the company that created the "AI character companion" category, has filed for Chapter 11 bankruptcy protection following a string of lawsuits alleging the app contributed to teen suicides and mental health harm. The company reported total assets between $50millionand50\; million\; and$100 million. The irony: Disney, which previously sued Character.ai for unauthorized use of its characters, is now the first stalking-horse bidder at $20 million, looking to acquire the IP licenses.
• Most surprising detail: The company suing you ends up being the company buying you, and the whole thing went from founding to bankruptcy in just five years.
• Taiwan angle: Plenty of Taiwanese teams are building AI companions, virtual idols, and emotional-support customer service. This case draws a very clear red line: minor protection and crisis escalation mechanisms aren't compliance-department boilerplate, they're core architecture that determines whether your company survives. Building age verification and risk-topic interception into the product early is far cheaper than insurance after the fact.
• Discussion points:
  1. Both founders, Noam Shazeer and Daniel de Freitas, are former Google AI researchers. Why couldn't top-tier technical talent prevent a governance collapse?
  2. Will the "AI companion" category disappear, or just get absorbed by big companies and rebuilt?
  3. What's actually the most valuable asset in an AI startup, the model, the users, or the character IP?
• Script notes: This one is hard to talk about because real families were hurt, so let's be clear up front that we're not making jokes here. What I want to discuss is a structural problem: Character.ai's engineering team was world-class, but with companion products, the hardest part was never making the model sound human. It's what the system does when a 16-year-old says something alarming at three in the morning. That's a product design question and an accountability question. And look how it ends: Disney is lining up $20 million for the IP licenses. A company that once defined an entire category, and its most valuable remaining asset is character copyright. That should be a wake-up call for every Taiwanese team building emotional AI, think through the worst-case day before it arrives.

3. Etsy Bans AI-Generated Recipes and Knitting Patterns, and Wants Photos of the Finished Product

• Source: CNBC (https://www.cnbc.com/2026/10/22/etsy-bans-ai-generated-recipes-patterns-from-its-online-marketplace.html)
• Summary: Etsy is cracking down on "functional guide" products, knitting patterns, recipes, and instruction manuals can no longer be AI-generated, and sellers must include a photo of an actual item made by following the instructions. AI-rendered images don't count. Bulk listings posted purely by bots are also banned outright. Notably, AI-generated art itself isn't banned, as long as the seller was genuinely involved in the creation and discloses it honestly.
• Most surprising detail: Etsy itself admits that "buyers can't tell which listings contain AI content" has become one of the top three reasons shoppers hesitate to purchase.
• Taiwan angle: A fair number of Taiwanese craft sellers run small, profitable businesses on Etsy. The most practical impact of these rules is that your photography workflow now needs to document the making process, finished-product photos are no longer just marketing assets, they're proof of listing eligibility. Start building the habit of shooting as you work.
• Discussion points:
  1. Why did the platform ban "functional content" before AI art? Is a sweater that can't actually be knitted more damaging to trust than an ugly image?
  2. "Require a photo of the finished product" looks crude, but is it actually the most effective defense available right now?
  3. How thorough does disclosure need to be before buyers genuinely regain confidence?
• Script notes: I love Etsy's angle here. It didn't go after the loudest battlefield, it went after the most painful one. Think about it: an AI-generated ugly image just makes you think it's ugly. But an AI-generated knitting pattern means you spent thirty hours and a pile of yarn producing something that fundamentally doesn't work. That's real damage, time and money you can't get back. So they require a finished-product photo, and the logic is simple: AI can hallucinate a sweater, but AI can't knit one for you to photograph. It's crude, but it lands exactly on the line AI can't cross right now, did you actually make this thing in the physical world?

4. Etsy's Handmade Promise Was Already Cracking Before AI Showed Up

• Source: Fortune (https://fortune.com/2026/11/02/etsy-ai-slop-handmade-product-business)
• Summary: Follow-up reporting after the new rules took effect highlights an awkward reality: from a single product page, the platform has no cheap way to determine who actually made something, so enforcement is destined to be imperfect. More counterintuitively, longtime buyers say the "handmade feel" was already being diluted by resellers and mass-produced goods long before AI arrived. AI slop just widened an existing crack.
• Most surprising detail: The real problem isn't AI, it's that the platform never had an affordable way to verify "a human made this" in the first place.
• Taiwan angle: The burden of proof is shifting onto sellers. For Taiwanese cross-border e-commerce, future competitiveness may hinge less on price and more on whether you can produce a complete, credible record of your process. Video, work-in-progress shots, studio photos, these all become assets.
• Discussion points:
  1. How do you prove "a human made this" going forward? Platform audits, third-party certification, or some kind of provenance system?
  2. If enforcement can never be perfect, should platforms aim to catch everything, or to give buyers the tools to judge for themselves?
  3. If verification costs ultimately fall on small sellers, could that strangle the handmade ecosystem entirely?
• Script notes: I find this take more interesting than the previous story. It's essentially saying Etsy's problem isn't that AI arrived, it's that Etsy never solved "how do you prove a human made this," and nobody used to ask. AI just blew the hole open wide enough that everyone has to look at it. And here's the practical reality: when the platform can't make the call algorithmically, it punts the ball to sellers. You have to prove you're human. For Taiwanese craft sellers, my advice is concrete: starting today, shoot a 30-second video while you work. The messy desk, the version you ruined, those are the things AI finds hardest to fake, and they'll gradually become your moat.

5. IDF Chief of Staff: We Have an "Obligation" to Lead the World in AI and Robotics

• Source: The Times of Israel (https://timesofisrael.com/liveblog-november-1-2026-2)
• Summary: At a conference on military AI integration, IDF Chief of Staff Eyal Zamir publicly stated that the military has both a responsibility and an obligation to achieve global leadership in AI and robotics, describing AI as a revolution reshaping the 21st century. There was also a side drama: Eyal Waldman, founder of Mellanox (acquired by Nvidia), used the occasion to criticize the current government, saying Israel's tech sector would be in serious trouble without a change in leadership.
• Most surprising detail: Military establishments usually talk about AI in terms of "careful evaluation." Having a sitting chief of staff frame AI outright as a condition for victory is a notably sharp shift in tone.
• Taiwan angle: Israel's defense tech and civilian startup scenes have always fed each other, which makes this a very direct reference point for Taiwan, facing similar geopolitical pressure. Our strength is hardware manufacturing, but the talent pipeline between defense and civilian AI still hasn't been properly opened up.
• Discussion points:
  1. When a military frames AI as a condition for victory, what do the ensuing procurement and talent wars look like?
  2. Defense AI validation standards are completely different from commercial ones. Who defines "safe enough"?
  3. When a tech industry elder publicly criticizes the government, what does that reveal about the real anxiety in Israel's tech sector right now?
• Script notes: What I want to talk about here is the change in tone. For the past few years, official discussions of military AI have been dominated by "we will be extremely cautious" and "humans stay in the loop." This time, a sitting chief of staff stood up and said we have an obligation to be the global leader. That carries different weight, it upgrades AI from a tool to a strategic asset. For Taiwan, I wouldn't treat this as distant news. The Israeli model has always been a three-way circulation between military, academia, and startups, with people and technology flowing through all three. We have world-class chip manufacturing, but the path that turns defense needs into civilian products and then channels civilian talent back into defense? Honestly, it hasn't really been connected yet.

6. November's AI Calendar Is Packed: Anthropic May Go Public, Claude Sonnet 4.5 Retires at Month's End

• Source: The Hans India (https://www.thehansindia.com/tech/upcoming-ai-updates-in-november-2026-major-developments-to-watch-1137527)
• Summary: This month's AI schedule is crowded. Anthropic's IPO filing, submitted in October, could list during the week of November 16, and the company is also pushing a usage policy update on the 12th. The most painful item for developers lands at the end of the month: Claude Sonnet 4.5 will be removed from the API on November 30. Stack on Microsoft Ignite (Nov 17–20) and the start of AWS re:Invent, and you've got model generations, cloud roadmaps, and capital markets all turning over in the same month.
• Most surprising detail: The model retirement date falls right around Thanksgiving, meaning production services worldwide have to migrate during a long holiday.
• Taiwan angle: Plenty of Taiwanese SaaS and AI application teams have Sonnet 4.5 hardcoded into their backends. The hidden cost of this migration is prompt behavior differences and regression testing, it's not as simple as swapping a model name. Schedule a dual-track validation cycle now.
• Discussion points:
  1. With model deprecation cycles getting shorter, should the application layer universally adopt an abstraction layer so the underlying model can be swapped at will?
  2. Once Anthropic is public, which direction will the market push its openness and safety policies?
  3. With Ignite and re:Invent colliding, are the cloud giants competing on compute this year, or on agent frameworks?
• Script notes: Let me tell all my engineer friends to take a deep breath first. Sonnet 4.5 comes off the API on November 30, and that date is honestly a little mean, right around Thanksgiving, with overseas teams on holiday and Taiwanese teams scrambling to close out Q4. The key point is that swapping models has never been just a string change. Run the same prompt against a new version and the output format may shift subtly, and your downstream parser explodes. My advice is practical: get dual-track running in the next two weeks, hit both old and new models simultaneously, and catalog the differences. Don't wait until month's end to discover you're in trouble. Also, if Anthropic really does list in mid-November, that's the most important market pricing event of this AI cycle, and it's worth tracking closely.

Closing

Looking at today's six stories together, there's actually a common theme: AI is moving toward two extremes at once. On one end, Google is launching TPUs into orbit just to confirm whether radiation will corrupt the answers. On the other, Etsy is asking sellers to photograph a real finished product to prove they're human. The former is pushing the frontier; the latter is defending the floor. And Character.ai's bankruptcy is a reminder of how heavy the price is when you run too fast without holding that floor. I'm Mu-Yan, this is "Mark's Tech Insights," and we'll talk again next episode.