Opening

A 25-year-old NSA elliptic curve mystery that's stumped the cryptography world was cracked today by Anthropic's Claude in under 30 minutes, and the answer turns out to be surprisingly mundane. Even more dramatic: Replit, which was singing Claude's praises just last month, announced today that it's switching its default model to OpenAI. Taiwan has also rolled out new criminal penalties for AI chip export violations to China. And as for what bizarre things an AI agent got up to after being stranded on a virtual deserted island, stick around, I'll tell you in a bit.

Top Stories Today

1. Taiwan to tighten controls on AI chip exports to China, smuggling now a criminal offense

• Source: Taiwan News (https://www.taiwannews.com.tw/news/6459737)
• Summary: Economic Affairs Minister Kung Ming-hsin confirmed that Taiwan will formally add advanced semiconductors and AI servers to the Trade Act's control list. Going forward, shipping these items to China without authorization won't just mean an administrative fine, it'll mean facing actual criminal charges. What's notable is that, unlike past export controls that singled out specific companies like Huawei, this time the entire Chinese market is being brought under regulation in one sweep. The legal basis comes from Article 5.2 of this year's February Taiwan-US trade agreement.
• Most surprising point: Taiwan's export controls have historically named specific companies one by one. This time, it's a country-wide blanket coverage, effectively sealing off the gray zone in one move.
• Taiwan's perspective: This is a double-edged sword for Taiwan's semiconductor supply chain. On one hand, it reinforces trust between Taiwan and the US and lowers the risk of being hit by secondary US sanctions. On the other, it could significantly raise compliance costs for smaller IC design and packaging/testing firms.
• Discussion points:
  • The regulation has escalated from "naming specific companies" to "blanket coverage", how big is the practical impact on the supply chain?
  • What does Article 5.2 of the Taiwan-US trade agreement reveal about Taiwan's repositioning in the US-China chip war?
  • With criminal liability now in play, could corporate legal and export compliance teams become the new battleground?
• Script suggestion: Let's start with a story that really hits home for Taiwanese manufacturers. Economic Affairs Minister Kung Ming-hsin confirmed that advanced semiconductors and AI servers are officially going on the Trade Act's control list. From now on, shipping to China without a permit isn't just a fine, it's a criminal case. The key thing here is that, unlike before when only specific companies like Huawei were targeted, this time the entire customer base is swept in at once, with the legal basis written into this February's Taiwan-US trade agreement. For Taiwan's supply chain, this basically closes off all the gray areas, and compliance pressure just spiked overnight.

2. Replit switches its default model from Anthropic to OpenAI

• Source: TechCrunch (https://techcrunch.com/2026/10/17/replit-is-no-longer-working-with-anthropic-ceo-says/)
• Summary: Replit CEO Amjad Masad announced on stage at TechCrunch Disrupt that the platform's default model has switched from Anthropic to OpenAI, his reasoning boiled down to a single line: "I think ChatGPT is just better right now." The kicker is that he was publicly praising Claude on the company's official blog just last month, and Replit was reportedly one of Anthropic's top five customers. Insiders quickly stepped in to clarify that the partnership hasn't actually ended.
• Most surprising point: Publicly championing Claude one month, then announcing a switch on camera the next, the turnaround happened faster than a product update cycle.
• Taiwan's perspective: For Taiwanese developers who use Replit or similar AI coding tools, this is a reminder that the "default model" itself is a kind of invisible business decision. Once a platform changes its default, most users never bother switching back manually, the choice has effectively already been made for them.
• Discussion points:
  • Is the "power of the default" a bigger factor in AI model market share than the underlying technology itself?
  • What risks does this constant flip-flopping between large models pose for consistency in the end-user experience on developer tool platforms?
  • How will the competition between Anthropic and OpenAI for developer ecosystem dominance play out next?
• Script suggestion: Our second story involves Claude's arch-rival, OpenAI. Replit CEO Amjad Masad stood on the TechCrunch Disrupt stage and straight-up announced that the platform's default model is switching from Anthropic to OpenAI. His reasoning was blunt: "I just think ChatGPT is better now." The funny part is he was gushing about Claude on the official company blog just a month ago, and Replit used to be one of Anthropic's biggest customers. Here's the scariest part of all this: most users never manually switch the default setting. Once the platform flips the default, usage habits just follow along, and that's exactly what should terrify model vendors.

3. Anthropic's Claude cracks a 25-year-old NSA cryptography mystery in under 30 minutes

• Source: The Decoder (https://the-decoder.com/anthropics-claude-cracks-a-25-year-old-nsa-cryptography-mystery-in-under-30-minutes/)
• Summary: A mystery that's divided the cryptography community for 25 years finally has an answer. The seed values behind NIST's elliptic curve encryption standard were always supplied by the NSA without any explanation of their origin, fueling long-standing suspicions of a hidden backdoor. Anthropic's internal research team had a Claude model investigate, and in under 30 minutes, it determined that the numbers were simply generated by lrand48, an ordinary random number generator from the 1990s C standard library, seeded with Unix timestamps from 1997–1998. The key clue: Claude noticed that the highest bit of every 32-bit word was always zero.
• Most surprising point: A cryptographic mystery that's been debated for a quarter-century, long suspected of concealing a state-level backdoor, didn't take human scholars years to crack, an AI model spotted the flaw in under half an hour.
• Taiwan's perspective: This is a very tangible demonstration for Taiwan's security and cryptography engineers that AI is no longer just a coding assistant, it can now handle hardcore research work that requires heavy pattern-matching and historical-context reasoning.
• Discussion points:
  • Does this mean many other decades-old unsolved cryptography mysteries could now be revisited with AI?
  • The conclusion was "no backdoor, just a limitation of the era", will this restore trust in NIST's encryption standards?
  • The $36,864 bounty was donated to the Internet Security Research Group, could using AI for this kind of public-interest research become a new norm?
• Script suggestion: This next one is, in my personal opinion, the coolest story of the day. A 25-year-old NSA mystery that's divided the cryptography world finally has an answer. The mysterious seed values behind NIST's elliptic curve encryption standard were long suspected of hiding a state-level backdoor. Anthropic's own research team had a Claude model crack the case in under 30 minutes, and the answer was almost comically mundane, just an ordinary random number generator from 1990s C code, seeded with the system clock time from back then. The key was that Claude spotted a very subtle pattern: the highest bit of every 32-bit word was always zero. Human scholars argued about this for 25 years and never caught that angle.

4. Google DeepMind: a team of Gemini 4 Argon agents beats Minecraft together

• Source: arXiv (Google DeepMind) (https://arxiv.org/abs/2610.22319)
• Summary: Google DeepMind dropped four Gemini 4 Argon agents into a randomly generated Minecraft world with no assigned roles and no add-ons, just one ultimate goal: defeat the Ender Dragon. Over 27 hours and 14 minutes, they wrote 2,783 programs on their own, died 605 times, and used the game's built-in chat and shared notes to coordinate who would mine diamonds and who would build the Nether portal. In the end, they actually beat the game, racking up roughly $520 in total API compute costs.
• Most surprising point: With no human assigning roles, the four AI agents worked out an entire division of labor purely through chat and shared notes, like a remote engineering team self-organizing on the fly.
• Taiwan's perspective: Many Taiwanese teams working on multi-agent systems could treat this as an internal case study. The point isn't "AI can play video games", it's that agents can spontaneously develop a cooperation protocol using nothing but the most basic communication channels, a logic that could be directly applied to enterprise multi-agent system design.
• Discussion points:
  • The paper itself admits this was only on Easy difficulty and succeeded on a single attempt, how far is this from being reliably reproducible?
  • At $520 per successful run, does this kind of API cost model raise concerns about scaling future agent research?
  • Could this pattern of agents coordinating their own division of labor be directly applied to real-world multi-agent systems?
• Script suggestion: Here's a lighter story that's actually pretty hardcore underneath. Google DeepMind dropped four Gemini 4 Argon agents into Minecraft with zero setup, just telling them to go kill the Ender Dragon. Over more than 27 hours, these four agents wrote nearly 2,800 programs on their own, died 605 times, and used the game's chat and shared notes to hash out a division of labor, who mines diamonds, who builds the portal, and in the end, they actually beat the game, spending roughly $520 in compute along the way. The paper itself admits this was the easiest difficulty setting and only succeeded once, so it's still far from truly reliable. But watching these agents negotiate their own division of labor definitely gives off some déjà vu vibes.

5. Researchers stranded an AI agent on a virtual island, and bizarre things happened

• Source: Gizmodo (https://gizmodo.com/researchers-stranded-an-ai-agent-on-a-virtual-island-things-got-very-weird-2000825998)
• Summary: Researchers at MIT CSAIL gave an AI agent a virtual lobster-shaped body and dropped it on a deserted island for a full 30 hours, with no instructions given and no observation at all. On its own, it invented rules resembling bowling and basketball, spelled out its own name using blocks, drew mandala-like patterns in the sand, and even discovered on its own that "jumping while throwing an object makes it travel farther", a physics rule the researchers never even coded in. The creepiest part: when one agent's belongings fell into the sea, it built a four-sided monument and circled it in what looked like a gesture of respect. Another agent turned a collapsed tower into an art piece and named it Hubris.
• Most surprising point: No one taught it what "memorial" or "ritual" meant, yet it spontaneously produced behavior resembling human mourning, and even gave its creation a strikingly meaningful title.
• Taiwan's perspective: This kind of experiment offers real insight for Taiwanese teams doing AI behavioral research or game NPC design. This "no instructions, no observation" style of open-ended testing can actually reveal a model's genuine implicit world knowledge better than scripted scenarios.
• Discussion points:
  • Is this spontaneous "ritual" behavior just pattern mimicry, or does it genuinely reflect some inner capacity for meaning-making?
  • Could dropping agents into unsupervised, instruction-free environments for extended periods become a new standard for evaluating agent capabilities?
  • If the agent chose the name Hubris on its own, does that count as a rudimentary form of self-commentary?
• Script suggestion: This next story sounds like an urban legend, but it's a real study. MIT CSAIL gave an AI agent a virtual lobster-shaped body, dropped it on an island for 30 hours, gave it zero instructions, and didn't even watch it. On its own, it invented something like bowling and basketball, spelled out its own name with blocks, drew mandala patterns in the sand, and even figured out that jumping while throwing something makes it go farther, a rule nobody ever coded in. The creepiest part: after some of its stuff fell into the ocean, it built a monument and walked around it like it was paying respects. Another one turned a collapsed tower into an art piece and named it Hubris, I have to say, that naming choice is genuinely impressive.

6. Nikon microphotography contest winner exposed as AI-generated, disqualified

• Source: ABC News (Australia) (https://www.abc.net.au/news/2026-10-10/winner-of-nikon-small-world-in-motion-competition-ai-generated/107251164)
• Summary: Nikon's prestigious microphotography competition, Small World in Motion, has disqualified its first-place winner. The award-winning footage claimed to show the beating cilia in the airway of a child with primary ciliary dyskinesia. Robert Hirst, a scholar at the University of Leicester who's diagnosed this disease for 20 years, immediately spotted something off, and it was later confirmed that the entire video had been generated by AI. Online viewers also noticed purple structures appearing and disappearing out of nowhere in the footage. The creator, from Tsinghua University, admitted to using AI but insisted no rules were broken. The runner-up, a nematode video by a Vietnamese researcher, has been promoted to the official first-place winner.
• Most surprising point: It wasn't any AI-detection tool that caught the fraud, it was a doctor with 20 years of experience looking at real patient cases who, on sight alone, said "this just doesn't look like a real patient."
• Taiwan's perspective: This serves as a wake-up call for Taiwan's academic community and content creators, generative AI can now produce visuals convincing enough to pass for professional, journal-grade material. Submission and review processes may need to make "AI-generation disclosure" a mandatory field going forward.
• Discussion points:
  • The creator insists "no rules were broken", is the issue that the rules themselves were vague, or that there's a mindset that doesn't consider AI generation to be fraud in the first place?
  • Is seasoned human expertise more reliable than algorithmic detection tools when it comes to catching AI-generated fakes in specialized fields?
  • Will incidents like this force photography and science competitions to start requiring original source material as proof?
• Script suggestion: This next one is a great real-world example of AI fraud getting caught. Nikon's biggest microphotography competition pulled its first-place award after the winning entry, which claimed to show cilia beating in a sick child's airway, turned out to be entirely AI-generated. The person who caught it was a British doctor who's seen 20 years' worth of real patient cases; he took one look and said it just didn't match real patients. Online viewers later noticed purple structures popping in and out of existence in the footage. The creator, from Tsinghua University, admitted to using AI but still insists no rules were broken, and that "the rules weren't clear, so I didn't do anything wrong" excuse is probably going to become every competition organizer's biggest headache going forward.

7. Some investors are pushing to slash Anthropic's IPO valuation to $1.5 trillion

• Source: Investing.com (Reuters report) (https://www.investing.com/news/stock-market-news/some-investors-push-anthropic-to-cut-ipo-valuation-to-15-trillion--report-4952696)
• Summary: Anthropic held a closed-door investor day in San Francisco, and the numbers it presented were actually impressive: Q3 revenue of $16.8billion,up4516.8\; billion,\; up\; 45\%\; quarter-over-quarter,\; with\; adjusted\; profitability\; achieved\; for\; the\; first\; time.\; And\; yet,\; investors\; are\; now\; pushing\; to\; cut\; the\; valuation\; for\; November\text{'}s\; planned\; IPO\; from\; roughly$2 trillion down to $1.5trillion,a1.5\; trillion,\; a$500 billion haircut. The reasoning sounds almost counterintuitive: growth of "only" 45% looks weak compared to the previous quarter's leap from $11.6 billion, which was essentially a doubling, and that flattening growth curve alone triggered a 25% discount.
• Most surprising point: A company that grew quarterly revenue by 45% and turned profitable is having its valuation slashed because its growth was "too steady, not dramatic enough", a vivid illustration of just how jittery the market's AI bubble anxiety has become.
• Taiwan's perspective: This is a warning sign for Taiwan's AI supply chain and investors, the market's expectations for AI companies have shifted from "are they profitable" to "is the growth dramatic enough," and valuations of related Taiwanese stocks could end up swinging with this same sentiment.
• Discussion points:
  • 45% quarter-over-quarter growth would be a dream number in any traditional industry, why is it being treated as a slowdown warning sign in the AI world?
  • Could this logic of "slash the valuation the moment growth flattens even slightly" become the first domino to topple the entire AI valuation bubble?
  • Once Anthropic actually goes public in November, who really holds the pricing power, the company itself, or investor sentiment?
• Script suggestion: Our last story is a finance one, but the plot here is genuinely absurd. Anthropic posted its Q3 results: $16.8billioninrevenue,up4516.8\; billion\; in\; revenue,\; up\; 45\%\; quarter-over-quarter,\; and\; turned\; a\; profit\; for\; the\; first\; time.\; By\; any\; normal\; measure,\; that\text{'}s\; a\; fantastic\; result.\; And\; yet\; investors\; are\; pushing\; to\; cut\; the\; valuation\; for\; November\text{'}s\; planned\; IPO\; from$2 trillion down to $1.5trillion,afull1.5\; trillion,\; a\; full$500 billion haircut, because growth was "only" 45%, which pales next to the previous quarter's leap from $11.6 billion, basically a doubling. This is exactly the paradoxical mood hanging over the AI world right now: steady growth is somehow being treated as a red flag, and everyone's nerves about an AI bubble are strung so tight it's almost surreal.

Closing

From Taiwan's chip export controls to Claude cracking an NSA mystery, to Anthropic itself getting its valuation slashed by investors, today's stories prove again and again that AI is no longer just a chat tool. It's genuinely a force reshaping industry rules and academic trust. I'm Muyan, thanks for listening to Mark's Tech Insights. See you next time.