Hosts: James & Maya
In this episode:
• Welcome to the Pivot 5 daily briefing. Hi, I'm James. Today we start with a court filing that puts a human voice on the fight over how AI gets trained...
• Hi, I'm Maya. Unsealed documents in the New York Times copy
Pivot5 | 5 Headlines & Unprompted
James: Welcome to the Pivot 5 daily briefing. Hi, I'm James. Today we start with a court filing that puts a human voice on the fight over how AI gets trained.
Maya: Hi, I'm Maya. Unsealed documents in the New York Times copyright case show OpenAI and Microsoft warning of a web 'doom loop,' and one Microsoft research director calling training data collection the largest theft of labor in human history.
James: That's a stark phrase from inside the company. The real story isn't the headline, though. It's that internal doubts are now public evidence in a case that could reshape how models are built.
Maya: Worth noting the caveats here. Microsoft says the line was read out of context. But the consequence is real: this language now sits in the record, and lawyers will use it. From filings to code, Maya.
Maya: Jarred Sumner says the Bun runtime was rebuilt in Rust from 535,000 lines of Zig, with AI help. He estimates four months instead of a year, and about 165,000 dollars in tokens.
James: Here's why this changes everything for engineering budgets. A full-language port that once meant a year of salaries now reads like a line item in cloud spending. That math will get every CTO's attention.
Maya: The data tells a different story on quality, though. Zig's creator, Andrew Kelley, publicly criticized shipping the result. The consequence: speed is proven, but trust in the rewritten code is still an open question. Next, a harder one.
James: Six former DraftKings staff told the New York Times the company built a model to score customers by how much they're likely to lose per promotion. Four others said work to predict problem gambling stalled or was killed.
Maya: Let's look at what actually happened. One model got resources; the protective one didn't. That's a choice about priorities, not a technical limit.
James: And the consequence lands on regulators. If a sportsbook can rank likely losers but shelves harm detection, expect gambling authorities to start asking which models actually shipped. Maya, take us to something lighter.
Maya: PrismML released Bonsai 2 27B on Thursday. Using ternary compression, they squeezed a 56-gigabyte multimodal model down to 5.9 gigabytes, under an Apache 2.0 license.
James: That size matters. It's small enough to run on a single consumer GPU or an Apple M5 Max chip, with no further quantization needed.
Maya: The consequence: capable models move off rented servers and onto hardware a company already owns. Worth noting the caveats, though. Independent testing of its accuracy is still thin. Now, a serious near-miss.
James: CNN's sources say a faulty AI intelligence report nearly triggered a US boarding of a Chinese ship. Armed personnel were preparing to board, and aircraft were already airborne.
Maya: The interdiction was only called off after someone re-examined the underlying intelligence. A person caught the error, not the system.
James: The consequence is sobering. When AI feeds military decisions, a bad report can put people minutes from a confrontation. Human review isn't a formality here; it's the last line. Let's move to the rapid-fire.
Maya: Five quick updates now, each one worth a moment of your attention.
James: Chewy expects AI to cut about 50 million dollars from costs in fiscal 2027. CEO Sumit Singh says the savings come from internal tools in customer service, pharmacy and vet clinics, and shouldn't be added to the existing 2026 margin path.
Maya: Nscale, a two-year-old London AI cloud provider, is seeking 3 billion dollars in a New York IPO. CryptoBriefing reports its filing shows a 56.4 billion dollar contracted backlog and more than 8 billion dollars in debt.
James: Ukrainian drone makers Frontline, Ark and Twist Robotics say soldiers want onboard autonomy so one operator flies many drones through Russian GPS jamming. Ukraine's defense ministry insists a human still makes the final call to strike.
Maya: Disney named Character.AI chief Karandeep Anand as its first CTO. He starts October 2 with authority over enterprise technology, data and AI platforms, a year after Disney sent his previous startup a cease-and-desist letter.
James: And BleepingComputer found the Tilly Norwood call hotline face-scans callers for an age check, run by Spanish provider Didit and added to its terms this month. It applies worldwide and expires when the hotline shuts down September 27.
Maya: The thread today: whether it's warships or sportsbooks, the story keeps coming back to who checks the model before it acts.
James: Well said. That's your briefing. Thanks for listening to the Pivot 5 daily briefing.