Opening

OpenAI was just exposed for ignoring internal safety warnings from leadership, and two days later fired three researchers who had been blowing the whistle to external oversight bodies, one of whom was even the official safety liaison. On the same day, Broadcom was reported to be lending Anthropic up to $42 billion so Anthropic could use the money to lease, yes, lease Broadcom's own chips. Today we're also looking at two AI copyright lawsuits, one where the company won and one where it lost, and even the Pope weighed in on AI-generated art. We'll cover it all in a moment.

Today's Top Stories

1. OpenAI Fires Three Safety Researchers, One of Whom Was the Official Safety Liaison

• Source: RTÉ News (https://www.rte.ie/news/2026/1002/1593746-openai-staff-fired/)
• Summary: OpenAI fired three alignment researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, on the grounds that they shared confidential information with an external AI safety organization. Ironically, Korbak had been specifically assigned by the company to liaise with those very external oversight groups on "rogue agent incidents." The timing is also notable, coming just two days after leadership was reported to have ignored internal safety warnings.
• Most surprising point: The person tasked with handling whistleblowing got fired for whistleblowing.
• Taiwan perspective: This is a warning sign for Taiwanese teams working on AI governance or enterprise AI safety frameworks. It shows how easily "internal whistleblower mechanisms" can be sacrificed under commercial pressure, so SOPs shouldn't rely solely on leadership self-restraint.
• Discussion points:
  • If even an "official liaison" isn't protected, does cooperation with external oversight bodies still mean anything?
  • Will this make other AI safety researchers afraid to report issues externally in the future?
  • Does OpenAI's PR explanation actually line up with the timeline of events?
• Script suggestion: The most ironic part of this story is that Korbak wasn't some outsider; he was literally the person OpenAI assigned to liaise with external safety organizations. Now he's been fired for "sharing confidential information." Picture it like being sent by your company to talk to an auditor, telling the auditor the truth, and then getting fired for it. And the timing couldn't be more pointed, just two days after reports surfaced that leadership had ignored internal safety warnings. If this kind of move becomes industry standard, nobody is going to want to be the whistleblower anymore, and external oversight mechanisms could end up completely toothless.

2. Broadcom to Lend Anthropic Up to $42 Billion, So It Can Lease Broadcom's Own Chips

• Source: Semafor (https://www.semafor.com/article/10/01/2026/broadcom-to-lend-anthropic-42-billion-to-lease-chips-report)
• Summary: In its IPO prospectus, Anthropic disclosed that Broadcom will provide up to $42billioninconvertiblenotes,lettingAnthropicusethefundstoleaseBroadco{m}^{\mathrm{\prime }}sownTPUcomputeresources\mathrm{.}ThatamountrepresentsroughlyathirdofAnthropi{c}^{\mathrm{\prime }}s42\; billion\; in\; convertible\; notes,\; letting\; Anthropic\; use\; the\; funds\; to\; lease\; Broadcom\text{'}s\; own\; TPU\; compute\; resources.\; That\; amount\; represents\; roughly\; a\; third\; of\; Anthropic\text{'}s$125.2 billion, five-year compute spending commitment. The notes can also be converted into Anthropic equity.
• Most surprising point: A chip supplier is lending money to its own customer, specifically so the customer can lease that supplier's own chips, and Anthropic openly admits in the filing that this creates a "potential conflict of interest."
• Taiwan perspective: Taiwan's chip supply chain is used to the logic of "orders plus cash." This model of a supplier bundling in financial leverage to lock in a customer is, in a sense, a way to secure order visibility for years to come. It's worth watching for Taiwanese manufacturers as a sign of how financialized the supply chain is becoming.
• Discussion points:
  • Could this "vendor financing" model become a standard way for big chipmakers to lock in customers?
  • Since the convertible notes can turn into equity, Broadcom is essentially also betting on Anthropic's future valuation.
  • Could this make Anthropic's balance sheet look healthier than it actually is?
• Script suggestion: Break this deal down and it's actually fascinating. Broadcom leases chips to Anthropic with one hand, then lends Anthropic money to pay for that lease with the other. That loan can convert into Anthropic shares, meaning Broadcom is simultaneously a supplier, a creditor, and possibly a future shareholder. Anthropic itself wrote "potential conflict of interest" in black and white in the prospectus, so credit where it's due for the honesty. This really reflects just how scarce compute capacity has become: chipmakers aren't just selling hardware anymore, they're directly helping customers lever up to lock in orders.

3. arXiv Cuts Monthly Submission Limit to Two Papers per Author, Citing a Flood of AI-Assisted Writing

• Source: arXiv official blog (https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/)
• Summary: The world's largest preprint platform, arXiv, announced that each author can now submit a maximum of two papers per month, with no more than three under review at any given time. The reason: submissions hit 40,363 papers in September alone, nearly double the same period two years ago. That month alone generated about 9,000 support tickets, a surge largely attributed to AI-assisted writing.
• Most surprising point: Even submissions that get rejected by the system still count toward your monthly quota.
• Taiwan perspective: Taiwan's academic community has also debated concerns about AI-assisted paper writing inflating output. This restriction essentially puts the brakes on globally first, and it will have a real impact on how Taiwanese grad students and professors rush to post on arXiv before deadlines just to establish a public timestamp.
• Discussion points:
  • Could a quota system backfire and encourage people to split papers into smaller pieces just to game the numbers?
  • Should AI-assisted writing be encouraged or restricted? Is this platform-level fix just a band-aid?
  • Is it fair that rejected submissions still count toward the quota, especially for researchers who genuinely need to revise and resubmit quickly?
• Script suggestion: What's interesting about this story is that arXiv became popular precisely because it was open and had almost no submission barriers. Now it's being flooded by papers that AI helped churn out, forcing the platform to impose a quota on itself. Even harsher is that rejected submissions still eat into your quota, so if you submit a weak paper and it gets rejected, you've lost one of your slots for the month. This is really pushing researchers to be more careful before submitting, but it could also hurt people who genuinely need to make quick revisions and resubmit. It's a pretty classic case of "using quantity restrictions to plug a leak, while accidentally damaging quality control."

4. Court Rules for the First Time That Training AI on Others' Data Isn't Fair Use, Thomson Reuters Wins Key Lawsuit

• Source: MediaPost (https://www.mediapost.com/publications/article/418380/appeals-court-sides-with-thomson-reuters-in-battle.html)
• Summary: The U.S. Court of Appeals for the Third Circuit affirmed that Ross Intelligence, a startup, did not qualify for fair use when it trained its own legal search AI using Westlaw's case summaries. This is the first appellate ruling in the U.S. on whether training AI with copyrighted material counts as fair use. The court's key finding was that the tool Ross built served the same purpose as the original material and directly competed with Thomson Reuters.
• Most surprising point: The lawsuit has dragged on since 2020, and Ross Intelligence has already gone out of business, yet the ruling only just came out now.
• Taiwan perspective: Many Taiwanese startups also use other companies' curated databases or knowledge bases to train models. This ruling draws a clear line: if the trained output is meant to directly replace or compete with the original data source, the legal risk rises significantly. Legal teams should treat "does the purpose directly compete" as a key factor in their risk assessment.
• Discussion points:
  • How will this "same purpose, direct competition" standard be applied to other AI training lawsuits going forward?
  • The defendant company went bankrupt before the ruling even came out, which is a brutal timeline for any startup caught in litigation.
  • Will this embolden more publishers and database companies to sue AI firms?
• Script suggestion: The most ironic part of this case is that the lawsuit dragged on for six years, and the defendant, Ross Intelligence, had already shut down long before this industry-defining ruling finally arrived. The line the court drew is actually pretty clear: if what your AI does is essentially the same thing as the data you trained it on, and it directly steals that company's business, it's hard to call that fair use. I think this is a concrete warning for any team hoping to train models on industry databases: the question isn't whether you can use someone else's data, it's whether the resulting product happens to hit right where it hurts for the data's original owner.

5. Google Beats Penske Media in AI Summary Lawsuit, a Sharp Contrast to the Thomson Reuters Case

• Source: The Hollywood Reporter (https://www.hollywoodreporter.com/business/business-news/google-wins-dismissal-of-pmc-lawsuit-over-ai-search-snippets-1236720513/)
• Summary: A federal judge dismissed Penske Media's lawsuit against Google over AI Overviews summarizing Penske's news content, ruling that a publisher's expectation of search traffic has never constituted a contractual relationship. This ruling, coming almost simultaneously with Thomson Reuters' win, makes for a striking contrast: publishers just secured a landmark ruling that "training AI on copyrighted content" isn't fair use, only to immediately lose on the issue that's actually eating into their traffic, AI directly using their content to generate summaries.
• Most surprising point: In the same week, the publishing industry won the biggest AI copyright precedent in history, yet lost the fight that's actually draining their traffic.
• Taiwan perspective: For Taiwanese media outlets and content sites, this ruling shows how hard it is to protect traffic through legal means. Rather than betting on lawsuits, it's better to start thinking now about how to build content value that can't be replaced in the age of AI summaries, such as deep analysis or original perspectives.
• Discussion points:
  • Will the logic that "expecting traffic isn't a contract" be used to block more similar lawsuits in the future?
  • How can publishers strike a balance between protecting copyright and competing with AI summaries for traffic?
  • Does this mean publishers' only remaining leverage is on the "training data" front, with the traffic battle already lost?
• Script suggestion: Put these two cases side by side and it's a strange picture. Publishers just won the first-ever appellate victory on AI training copyright, and immediately afterward, a judge told them that expecting a search engine to send them traffic isn't some contractual obligation owed to them. In other words, publishers won the fight over training data but lost the fight that's actually taking food off their table. The law is willing to protect "your content can't be used to train a competing product," but it won't protect "your accustomed business model can't be disrupted." For an industry that depends on search traffic, that's a much harsher reminder.

6. California Signs 13 AI Bills Banning AI From Firing Employees Alone, But Vetoes Smart Glasses Spy-Camera Bill

• Source: The Next Web (https://thenextweb.com/news/no-robo-bosses-act-newsom-ai-bills-smart-glasses-veto)
• Summary: California's governor signed a sweeping package of AI bills, including the "No Robo Bosses Act," which bars employers from firing employees based solely on an AI's decision. At the same time, he vetoed SB 1130, which would have made it a misdemeanor to covertly record people using smart glasses and required such devices to show a recording indicator light while filming. The governor's stated reason was that the definition of "wearable recording device" was too broad and could sweep in things like smartwatches.
• Most surprising point: An algorithm can no longer fire you, but the person standing across from you wearing smart glasses can still secretly record you with no warning light at all.
• Taiwan perspective: Taiwan currently has fairly limited regulations around wearable devices and privacy. This case is a reminder for lawmakers that a definition written too broadly can ensnare other legitimate devices, while one written too narrowly fails to address the real risk. Taiwan will eventually have to grapple with the same tension over privacy rules for new wearables like smart glasses.
• Discussion points:
  • How would the No Robo Bosses Act determine what counts as "solely an AI decision"? Could companies just get a human to rubber-stamp AI decisions as a workaround?
  • Will the smart glasses spy-camera controversy keep dragging on now that the bill has been vetoed?
  • Will other states follow California's lead on labor protections while staying more conservative on privacy?
• Script suggestion: This story has a very vivid contrast. On one hand, the governor made a decisive statement that an algorithm shouldn't be allowed to decide alone whether to fire you, and that direction deserves credit. On the other hand, the bill requiring a recording indicator light on smart glasses got vetoed, on the grounds that the definition was too broad and might accidentally catch devices like smartwatches. The net result is that you now have an extra layer of AI protection at work, but the moment you step outside the office, a stranger wearing smart glasses could be livestreaming you the whole time without you ever knowing. This highlights how lawmakers tend to patch AI-related issues case by case, rather than working from a comprehensive framework.

7. Boston Dynamics' New Humanoid Robot Hand Deliberately Drops the Pinky, and Becomes More Dexterous for It

• Source: The Robot Report (https://www.therobotreport.com/boston-dynamics-drops-pinkie-on-new-humanoid-hand/)
• Summary: Atlas has a new four-fingered hand with degrees of freedom increased from seven to thirteen. It can reorient objects within its own palm, recover on its own if an object slips, and even juggle two golf balls with one hand or pull the trigger on a power drill. Interestingly, engineers deliberately chose not to include a fifth finger, reasoning that the added complexity, power draw, and bulk from a pinky simply weren't worth it.
• Most surprising point: Evolution took millions of years to give humans five fingers, and engineers cut one with a single line in a spec sheet.
• Taiwan perspective: For Taiwanese suppliers that build robotic arms and grippers for automation, this story is quite practical. It shows that robot design doesn't have to mimic humans; it should instead make trade-offs based on task requirements. This "good enough, not overly biomimetic" philosophy can actually help Taiwan's small and mid-sized automation manufacturers save costs and get products to market more easily.
• Discussion points:
  • Does removing one finger really not affect most everyday manipulation tasks? What's the engineering logic behind that call?
  • How does this trade-off-driven design philosophy contrast with approaches that emphasize full biomimicry?
  • Will future humanoid robot hands increasingly lean toward "functionally sufficient" rather than "humanlike"?
• Script suggestion: What's cool about this story isn't that the robot hand got more impressive, it's the trade-off the engineers made. Adding a fifth finger sounds like a no-brainer, since humans have five fingers after all, but Boston Dynamics' team ran the numbers and decided the extra complexity, weight, and power draw just weren't worth it. So they cut the pinky entirely, and still ended up with a lighter hand that can juggle two golf balls and pull a trigger. This pretty much debunks the myth that "a robot has to fully mimic a human." Sometimes the smartest engineering decision is being honest about which features are actually dispensable.

8. Pope Leo XIV Publicly Criticizes AI-Generated Art, Says Machine Output Has an "Ontological Difference" From Art

• Source: TechCrunch (https://techcrunch.com/2026/10/02/pope-leo-xiv-is-not-a-fan-of-ai-generated-art/)
• Summary: Pope Leo XIV announced plans to reinforce the Church's collaboration with human artists. According to TechCrunch, he argued that the difference between art and something a machine generates through statistical computation isn't just aesthetic, it's "ontological." In simple terms, he believes AI-generated work lacks that human spark.
• Most surprising point: The Vatican essentially just issued a theological-grade verdict on diffusion models.
• Taiwan perspective: Taiwan's creative community has also been debating whether AI-generated images count as "creation." The Pope's remarks echo, to some extent, what many traditional creators have been feeling, but they're also a reminder that this debate won't stay confined to copyright or commercial concerns, it may well extend into deeper questions about what art fundamentally is.
• Discussion points:
  • "Ontological difference" sounds very academic; in everyday terms, what is the Pope actually trying to say?
  • Could a religious institution publicly criticizing AI art shape how the general public perceives AI-generated content?
  • Does this align with or push back against how museums and galleries are currently treating AI artwork?
• Script suggestion: This remark from the Pope is a bit surprising at first glance. Most people might say AI-generated images lack soul, which is a fairly emotional way of putting it, but the Pope went straight for "ontological," a philosophical-grade term, elevating the discussion to the level of faith and the meaning of existence itself. In other words, he's not saying AI art doesn't look good, he's saying it's not even the same category of thing to begin with. This really reflects something real: as AI-generated content becomes more widespread, people's anxiety about "what counts as creation" has moved beyond job security and started touching on something much more fundamental, the definition of humanity itself.

Closing

Today we went from OpenAI firing its own people, to Broadcom bankrolling a customer to lease its own chips, to AI copyright lawsuits splitting one win and one loss, and finally even the Vatican weighing in. This episode was basically a microcosm of the entire AI world: money, power, law, and human nature, all playing out on the very same day. I'm Muyan, see you next time.