AIMarketsNewsTech

The Sunday brief: Nadella’s AI emergency brake, DistroKid’s human takedowns, and the chips under AI debt

Microsoft’s Satya Nadella says every AI model should be treated as compromised from the start, with a person always able to stop it mid-task. DistroKid is pulling songs over Universal Music’s AI-slop lawsuit, and some of them were made by people.

A law firm’s memo on GPU-backed data-center loans made the rounds on Wall Street, Prime Video bought an AI-drawn series the day Nicolas Cage said he refused its AI waiver, and a Walmart patent covers the pricing Walmart says it will never use.

In this Brief

  • Nadella wants an “emergency brake” on AI models and a harness kept separate from the model it runs
  • DistroKid confirms it is removing tracks over UMG’s claims while no judge has ruled
  • Quinn Emanuel’s memo on the debt behind AI data centers, and the chips it rests on
  • Prime Video and Netflix keep adding AI to production as actors fight over waivers
  • Walmart’s 2023 patent claims cart-based price changes on digital shelf tags

Nadella says to assume every AI model is compromised

Setup: In a long post on X Saturday, Microsoft CEO Satya Nadella wrote, “We must assume a model is compromised and contain it from the start. Think of it like an emergency brake,” TechCrunch reports. He called for “separating the model from the harness that orchestrates its work,” “externalizing controls and safeguards,” “tamper-proof human readable evidence” of every meaningful model action, and “an authorized person” who can always “pause or shut down a model mid-task.”

The Verge notes most of his list matches what others in the industry have proposed, including incident disclosure, independent audits and containment, and that he goes slightly further on containment.

The harness is the software around a model that hands it tasks, tools and permissions, and that is the part Microsoft builds and sells with other companies’ models plugged in. Its new decision model, Decision-1, runs on Alibaba’s Qwen, and Microsoft says it will rebase it on its own and OpenAI’s models, The Register reported Friday. A brake is only as good as whoever checks it, and the passages quoted so far name “an authorized person” without saying who does the authorizing.

DistroKid’s AI cleanup is taking human artists’ songs

Setup: DistroKid confirmed to The Verge that it is removing tracks in response to claims from Universal Music Group, which sued it in federal court in Delaware on Sept. 15, accusing it of running an “AI-slop pipeline.” The musician McGwire told The Verge he lost six songs, including a Stevie Wonder cover he says he licensed. DistroKid’s vice president of artist services, Amanda Ferri, said “a very small number of recordings” had been removed.

UMG’s 52-page complaint points to one account, “Lofi Chill,” that released 4,562 tracks in 12 months, and asks a judge to make DistroKid “promptly remove from all Digital Services all infringing tracks.”

No judge has ruled on any of it. A distributor accused of leaving infringing tracks up after being warned has every reason to pull first, and the mistakes land on solo artists with no label to call. Five days before it sued, UMG announced a licensed AI music platform with ElevenLabs.

See full breakdown at @Aaron_Harme on X.

Longer cut: DistroKid’s UMG takedowns are pulling human artists’ songs along with AI tracks.

The chips under the AI data-center loans

Setup: A 20-page memo from the law firm Quinn Emanuel on how the debt behind AI data centers could end up in court has been circulating on Wall Street this week, Semafor reports. GPUs are the collateral on many of those loans, and rental rates for Nvidia’s H100 have fallen roughly 70% to 90% since 2023, the memo says. CoreWeave writes its GPUs down over six years. Engineers put their useful life at three to four.

The memo also walks through Meta’s Louisiana campus, where a company 20% owned by Meta raised $30 billion, about $27 billion of it in loans that never appear as a Meta liability. Meta’s promise to cover those investors, up to $28 billion, sits in the footnotes of its annual report.

Quinn Emanuel is a litigation firm and would get the court work if these deals go bad, Semafor notes, so the memo doubles as a sales pitch. The deal terms it cites are still the companies’ own. Fortune, reporting on the CBO’s debt math, counts heavy AI hyperscaler borrowing among the reasons long-term yields sit at a 24-year high.

See full breakdown at @Aaron_Harme on X.

Longer cut: Quinn Emanuel’s AI data-center debt memo hangs on how long a GPU lasts.

Streamers add AI to production while actors fight the waiver

Setup: “I didn’t sign the waiver to let them use AI on Spider-Noir,” Nicolas Cage told a New York Comic Con panel Saturday, The Hollywood Reporter reports, adding that he’s “probably not going to be working with that studio again.” That studio is Prime Video. The same day, Variety reported Prime Video had taken global rights to “Çukur: Vartolu,” a spin-off of the Turkish hit “The Pit” in which, per producer Ay Yapim, “every frame” was generated with AI over real actors’ performances.

At the Busan International Film Festival on Saturday, Netflix’s Sung Q Lee said the company has used AI on about 300 titles, in previsualization, production design, set-dressing concepts, editing and VFX, THR reports. “The human touch is critical and necessary,” he said.

A waiver is the permission that lets a studio run AI over a performance, and “Çukur: Vartolu” shows what that permission buys: the actors performed every scene and AI drew the characters on top. Netflix’s 300 titles say the tools are already routine on the other side of the camera. Cage can afford to refuse in public and lose a studio over it. The waiver still lands in front of actors who can’t.

See full breakdown at @Aaron_Harme on X.

Walmart’s patent covers what Walmart says it won’t do

Setup: “Putting tuna in a customer’s cart does not cause Walmart to charge that customer more for mayonnaise. We do not do that, and we will not,” Walmart corporate affairs chief Dan Bartlett wrote this week, Gizmodo reports. He was answering consumer advocate Lindsay Owens, who has been citing Walmart’s patents. A patent “describes an invention and potential capabilities,” Bartlett wrote.

The tuna and mayo come from the patent’s own example, granted to Walmart in 2023. Its first claim, the part that defines what the patent legally covers, takes the list of products in a shopper’s cart, changes another product’s price based on one of them, and sends the new price to the digital tag on the shelf.

Bartlett’s letter commits Walmart for as long as Walmart wants it to. The patent is active, and Google Patents estimates it runs to about 2039, which also lets Walmart stop any other retailer from pricing that way.

See full breakdown at @Aaron_Harme on X.

Also noted

  • The Hollywood Reporter: Sony put Fede Goldenberg, YouTube’s head of film and TV partnerships, in charge of its Screen Gems genre label, which will now develop theatrical movies from online creators.
  • Fortune: CBO director Phillip Swagel said growth alone probably can’t stabilize U.S. debt. His rough math needs 5% to 6% real growth a year, against 2.2% last quarter.
  • MacRumors: Apple disclosed to the European Commission a deal to hire staff from AI podcast startup Huxe and license its technology, 12 days after Huxe shut its app down.
  • TNW, citing Bloomberg: Ramp raised about $1.85 billion at a roughly $60 billion valuation before the new money, up from $44 billion in June.

Sources


Discover more from Bitter Fool

Subscribe to get the latest posts sent to your email.

Leave a Reply

Your email address will not be published. Required fields are marked *