Good morning. Last week on ARQIV we asked who checks an AI agent's work. This week Microsoft's CEO gave his answer: assume the model could be compromised, and build the walls around it.
Also inside: the debt pile behind the AI building boom, a spacecraft on its way to a Mars moon, and a new kind of math proof.
In today’s Dispatch
- 🤖 AI: Microsoft's Nadella: treat frontier AI models like insider threats
- 💻 Tech: CoreWeave's debt is turning into the AI industry's problem
- 🚀 Space: Japan launches MMX, a mission to bring back dust from a Mars moon
- ⚛️ Quantum: Quantinuum claims a record for error-corrected qubits
- 🎲 Surprise Science: Mathematics: An AI-written proof of the quasi-Riemann hypothesis gets human scrutiny
- 🔭 Futurism: What's Coming: Apple's "Pro and tell" event
🤖 AI
Microsoft's Nadella: treat frontier AI models like insider threats
In an essay posted on X on Oct 10, Satya Nadella argued that companies should treat powerful AI models the way they treat employees with deep access: verified identity, the least privilege needed, logged activity, and limits on the damage one mistake can do. He points out that nobody can trace a model's behavior back to specific training data or weights, so he wants controls outside the model, tamper-proof records of what agents did, independent audits, and a way for a person to pause a model mid-task. He does not claim the models are malicious.
Why it matters: Companies are giving AI agents real access to real systems. The people selling the models are now saying the safety work has to happen around them, not just inside them.
Source: TechCrunch
Quick hits
- Hackers hit the AI coding tool OpenCode, running up charges on users' AI accounts, TechCrunch reports. (TechCrunch)
- Amazon released TorchNeuron, a developer tool aimed at making PyTorch, the free toolkit most AI apps are built on, run well on chips other than Nvidia's, starting with Amazon's own. (TechCrunch)
From ARQIV
The gear behind the show
The cameras, mics and lights we use to make ARQIV, all in one list.
💻 TECH
CoreWeave's debt is turning into the AI industry's problem
CoreWeave rents out AI computing power and has borrowed heavily to build it, reportedly around $50B of debt. Last week its bonds slipped into what traders call distressed territory, meaning investors demanded at least 10 percentage points more yield than on government bonds. The cost of insuring its debt against default also hit a record, per Reuters. Its CEO told staff the backlog of signed contracts more than covers the debt.
Why it matters: A lot of the AI buildout runs on borrowed money. When one big borrower wobbles, lenders start asking harder questions of everyone else.
Source: The Verge
Quick hits
- Gas-fired power is booming again. Global Energy Monitor data cited by The Verge shows gas capacity in development rose 31 percent in 2025, with the US leading and more than a third of its growth expected to power data centers. (The Verge)
- The IMF is again warning that an AI-investment bubble could hurt the wider economy if it bursts. (Futurism)
🎬 New on ARQIV
AI Agents Are Making Decisions. Who's Checking Their Work?
Nadella's essay is exactly this question. Steve Ryan sits down with Matt Davis of Critical Propulsion to work out what agents should decide alone, where a human approval gate belongs, and why rolling out agents is a change-management problem.
▶ Watch on YouTube or listen on Spotify · Apple Podcasts
🔬 SCIENCE
🚀 SPACE
Japan launches MMX, a mission to bring back dust from a Mars moon
Japan's space agency JAXA sent the Martian Moons eXploration probe up on an H3 rocket from Tanegashima on Oct 20 local time. It should reach Mars in about a year, then try to collect at least 10 grams from the surface of Phobos, helped by a French-German rover called IDEFIX. Space.com reports it would be the first spacecraft to return to Earth from Mars orbit.
Why it matters: Nobody agrees on where Phobos and Deimos came from: captured asteroids, or debris from a giant impact on Mars. Real samples could settle it.
Source: Space.com
⚛️ QUANTUM
Quantinuum claims a record for error-corrected qubits
Quantinuum says its Helios machine reached a logical fidelity of 99.9995%, a tenfold drop in the logical error rate. A logical qubit is a group of physical qubits working together to catch and fix errors. This comes from the company's own press release, and we haven't seen outside confirmation yet.
Why it matters: Quantum computers only become useful once errors are fixed faster than they pile up, so this is the number researchers watch.
Source: The Quantum Insider
🎲 SURPRISE SCIENCE: MATHEMATICS
An AI-written proof of the quasi-Riemann hypothesis gets human scrutiny
OpenAI recently released hundreds of AI-generated math manuscripts from an unreleased model, including a proof of the quasi-Riemann hypothesis. That is a weaker cousin of the famous Riemann hypothesis, which remains unsolved. Quanta reports that, unlike other recent AI proofs, this one has been unpacked by human experts, and one mathematician called the implications "staggering". Other results in the batch are still unchecked, and many mathematicians are urging careful review before anyone celebrates.
Why it matters: If AI can produce proofs that people can actually read and verify, the question shifts from whether it can do math to how we check it.
Source: Quanta Magazine
Quick hits
- Researchers report that honey beehives act as "living immune systems", with layered antimicrobial defenses that protect the colony, its young and its food stores. (Phys.org)
- Scientists are growing human brain tissue in living mice. A new commentary asks how anyone can judge the animals' welfare when the organ that expresses their mental states has been changed. (Phys.org)
🔭 FUTURISM: WHAT'S COMING
Apple's "Pro and tell" event
Oct 28: Apple has announced a special event for Oct 28 under the tagline "Pro and tell". We'll wait for the event before saying what it covers.
Source: 9to5Mac
MMX reaches Mars
Aug 2027: ESA's launch notice puts MMX's arrival at Mars about a year after liftoff. Landing and sampling on Phobos follow around 2030, and the capsule is due back on Earth in 2031.
Source: ESA
Anthropic's huge West Virginia compute lease starts service
Late 2027: Reports say Anthropic signed a roughly $45B, six-year lease for about 460 MW of Nscale computing in West Virginia, with service expected to begin in late 2027. Neither company has formally confirmed the terms.
Source: ITPro
📼 From the ARQIV vault
The Infinite Glitch: Why the Universe Refuses to Be Pixelated — a look at how we treat infinity and numbers like π, a good companion to this week's math story.
From ARQIV: Your one-line ad could be here. Book this spot →
If an AI agent had access to your email, what is the one thing you would want it to ask you before acting? Hit reply and tell me. I read every one.
