Podcast 5-5 by Dongee · November 3, 2026

Transcript

Gabriel: Welcome to the episode of November third. Let's go.

Gabriel: First. Some MIT researchers gave OpenAI's GPT-6 Astra the weights of another AI: Qwen, a small model from Alibaba. The weights are the numbers that define how a model thinks. And they asked it to make it better at math without training it!

Lina: One AI performing surgery on another. I'm not sure that reassures me.

Gabriel: Well, it went well! It found in those numbers why Qwen kept going around in circles, repeating itself without reaching the answer. It changed fewer than two in every ten thousand weights, and on a math test it went from getting less than eight percent right to almost half.

Lina: Fixing a model that gets stuck isn't making it smarter. It's just curing its stutter.

Gabriel: The authors admit it: with another small model that didn't get stuck, it couldn't improve anything. But it also guessed, just by looking at the numbers, what game a network had been trained for.

Lina: That game-guessing thing sounds more like a carnival trick than science to me.

Gabriel: A trick anyone can check: they published all the logs, and someone else has already reproduced the math result on their own.

Lina: It's a blog post, not a study reviewed by other scientists. And with tiny models.

Gabriel: On Hacker News it was the most talked-about item of the day. Imagine reviewing and fixing models without retraining them, for much less money.

Lina: And an AI surgeon operating without anyone understanding what it cut. Would you really let one AI tinker with another one's brain?

Lina: Second. Anthropic, the company behind Claude, slipped an unsettling sentence into its paperwork to go public: that it could be very close to an AI that accelerates its own development.

Gabriel: It's in the risk section. It says its estimates about future models, not current ones, indicate it could imminently reach its own threshold for dramatically AI-accelerated progress.

Lina: Imminently. Right in the document read by the people about to buy shares. What a coincidence.

Gabriel: It's a legal document! If you see a risk and don't disclose it, you get sued. And the threshold is specific: making as much progress in one year as would normally take two.

Lina: I'm more worried about another line: they admit that in the past they didn't notice an acceleration until after it happened.

Gabriel: That's why they promise to keep people in charge. They say they don't intend to let Claude design its own successors without human oversight.

Lina: Promising the brake while flooring the accelerator. I don't fully buy that balance.

Gabriel: They also make clear that no current model has crossed that line, and that they're still reviewing the data.

Lina: I'll give you this: warning about the risk before it arrives is more than others do. But saying you're on the verge of something huge also sells shares. Does this warning give you more confidence in the company, or more fear?

Gabriel: Third. Marvell, which makes chips and connectivity for data centers, is buying Lumentum, which makes lasers. Price: one hundred fifteen billion dollars, paid in stock. One of the biggest acquisitions in tech history!

Lina: And the market responded with applause. Downward.

Gabriel: Marvell fell about thirteen percent that same day, and Lumentum fell even more.

Lina: When you buy something and your own shareholders run for the exit, they're telling you something.

Gabriel: That the price scares them, not the idea. The idea is that AI data centers are moving from copper cable to light to move data between thousands of processors. And Lumentum is an expert in that—

Lina: The idea can be good and the price bad. By Reuters' calculations, they're paying more than double what the market pays for similar companies.

Gabriel: Marvell's boss talks about a once-in-a-generation opportunity. And here's a detail: NVIDIA is a shareholder in both.

Lina: So whatever happens with this deal, NVIDIA comes out ahead.

Gabriel: I see it as visionary: whoever controls optical connectivity will dominate the next stage of data centers.

Lina: And I see someone paying top dollar just as chip stocks are falling. Betting big, or buying at the top? I'll leave that one with you.

Lina: Fourth. The New York Times sued Brave, the browser that prides itself on protecting your privacy. It accuses it of copying its articles and selling them to AI companies.

Gabriel: And it's not alone: three other U.S. newspapers filed a similar lawsuit that same Monday. According to the Times, Brave built an index of more than thirty billion pages with a crawler that publishers can't identify or block.

Lina: The privacy browser, sneaking into other people's houses. Textbook irony.

Gabriel: The lawsuit names Anthropic, Perplexity, Mistral, xAI and Cohere as customers. And there's more: the browser has a feature to get around paywalls.

Lina: That's what I find hardest to defend. Reading for free what others pay for isn't privacy.

Gabriel: But crawling the web is what every search engine has always done! If that's a crime, half the internet is in—

Lina: Crawling to show a link isn't the same as reselling the full article to an AI.

Gabriel: A judge will decide that. Brave hasn't responded yet, and Perplexity, one of the companies named, said the Times overpays its lawyers and underpays its journalists.

Lina: Good joke. But journalism costs money, and whoever uses it should pay for it.

Gabriel: And if they shut everyone down, a piece of the open web shuts down with them.

Lina: There, we actually agree on something. Don't make a habit of it.

Gabriel: Fifth. In Germany, several members of parliament and of the government got an alert from Apple on their iPhones: someone tried to spy on them with one of those programs sold to governments.

Lina: The key word is tried. Or maybe not: nobody knows if they got in.

Gabriel: Exactly. Apple doesn't say who's attacking or whether they succeeded. That's why Germany's cybersecurity agency says to assume the phone may have been compromised.

Lina: Compromised means someone can read your messages, listen through the microphone and see through the camera. A lawmaker's phone.

Gabriel: And without the victim touching anything. These attacks get in without you clicking on any link.

Lina: The government says it's taking it very seriously. But it doesn't say how many there are, or who, or whether Chancellor Merz is on the list. Very seriously, but very quietly.

Gabriel: It's not just Germany. According to one of the notices, Apple sent this batch of alerts to users in ninety-nine countries.

Lina: The awkward part is that these tools are bought by governments themselves. It was reported that Germany's federal police bought access to Pegasus, one of the best known, in twenty nineteen.

Gabriel: But the system worked! Apple sent the alert and the state wrote to those affected the very next day.

Lina: The alarm worked. The rest is silence: we don't know who, or how many, or what was taken.

Gabriel: If your phone warned you that you were being spied on, would you keep using it?

Gabriel: That's it for today's episode.

Lina: If any question got you thinking, share it and talk it over.

Gabriel: Tomorrow, five more stories. And if an AI wants to improve my weights, I'm in!

Lina: They won't fix your stutter, Gabriel. They'll fix your enthusiasm. Talk to you tomorrow.