That statistic comes courtesy of Pew, which analyzed our so-called link rot situation earlier this year. One in five pages was inaccessible just two years later, according to its findings.

Sobering stuff, and not just because I’m in the publishing business. In a world where data is (again) the new oil, that should be concerning to anyone training AI models on the open web. —Andrew Nusca

P.S. Fortune colleague Jason Del Rey has published a fantastic new profile about Matt Garman, the CEO of Amazon’s cloud computing business. Give it a read.

Want to send thoughts or suggestions to Data Sheet? Drop a line here.

Sony buys 10% of ‘Elden Ring’ owner Kadokawa

Japan’s Sony said yesterday that it had paid about $320 million for a 10% stake in Kadokawa, the company behind the bestselling video game Elden Ring.

Kadokawa does quite a bit more than just games; the Japanese company is known for its anime, manga, and other publishing endeavors. 

But its ownership of Tokyo’s FromSoftware, the creator of the George R.R. Martin-connected fantasy game Elden Ring, is considered its most coveted asset. The title has sold more than 28 million units since its debut almost three years ago; 12 million of those were sold in the first two weeks.

The deal makes Sony the largest shareholder in Kadokawa. The electronics giant has held a stake in the smaller company since 2021.

In a joint statement, the companies said the arrangement would lead to joint investments, joint promotions, and “mixes of both companies’ IP.” Kadokawa fans can expect to see the company’s intellectual property make its way to live-action films and television dramas and other conglomerate-spanning projects.

Apple: Meta interoperability requests could harm user privacy

Apple and Meta have gone to war in the EU.

Thanks to the EU’s new pro-competition Digital Markets Act, Apple is going to have to open up a lot of iOS functionality. The European Commission launched a consultation yesterday, laying out each function that may need to be exposed to third-party developers, from AirDrop to iOS notifications.

Apple responded by publishing a paper about how dangerous interoperability supposedly is, and accusing Meta of requesting access to iOS functions “in a way that raises concerns about the privacy and security of users.”

Meta’s response? “Here’s what Apple is actually saying: they don’t believe in interoperability. In fact, every time Apple is called out for anticompetitive behavior, they defend themselves on privacy grounds that have no basis in reality.”

The consultation closes on Jan. 9th and Apple would have to implement the changes by the end of 2025. —David Meyer

Google reveals its own reasoning AI model

Google on Thursday released its own version of an AI reasoning model that pits it against OpenAI’s o1, released in September.

Called Gemini 2.0 Flash Thinking—just rolls off the tongue, doesn’t it?—the experimental model promises to apply “deeper thinking" to queries posed to it. 

It shows “promising results when we increase inference time computation,” said Jeff Dean, the chief scientist for Google DeepMind and Google Research.

Reasoning models are different from other AI models because they incorporate feedback loops that allow them to “continuously learn” to arrive at a better answer. 

That AI “thinking” uses extra computing power and requires more time, but may offer a better pathway to solve particularly complex problems. (Think: complex logical puzzles, multi-step problem-solving situations, and anything that requires one to draw inferences, rather than only recognize patterns.) Though it isn’t without its flaws.

Gemini 2.0 Flash Thinking is available now through the Gemini API in Google AI Studio and Vertex AI.

YouTube plans crackdown on ‘egregious clickbait’

In other Alphabet news, YouTube said this week that it would strengthen its efforts to address “egregious clickbait” from creators who use deception to stoke engagement and ultimately ad revenue.

Starting in India—the company’s largest market by far; with an estimated 46 million users, it’s double the size of the U.S. market—YouTube will increase enforcement against “videos where the title or thumbnail promises viewers something that the video doesn't deliver,” most commonly found in news clips.

“This can leave viewers feeling tricked, frustrated, or even misled—particularly in moments when they come to YouTube in search for important or timely information,” the company says. 

YouTube says it will first remove content without “issuing a strike;” that will change over time as creators get used to the new standard.

Users have been complaining about YouTube’s clickbait content for years as the company has tweaked its product; perhaps the problem has now grown to the point where it’s beginning to affect engagement—a perpetual concern for the world’s second-largest social media platform.

More data

—Next year’s job interview question of choice: “Tell me how you used AI.”

—Apple reportedly in talks with Tencent, Bytedance for AI features in China.

—Oura valued at $5.2 billion. The ringmaker’s latest fundraise is among the largest for a non-AI Euro tech company.

—Hackers stole $2.2 billion in crypto this year, up from $1.8 billion.

—Congress allocates $3 billion to remove Chinese telecom gear from U.S. phone networks.

—Netflix places a huge bet on Squid Game. Prepare for real-life “Experiences.”

Endstop triggered