Sponsored by
In Todayโs Issue:
๐ Nadella wants an emergency brake for AI
โก A utility gets the off switch on Amazon's data centers
๐ Call-center jobs are shrinking worldwide
๐ฆ Goldman's juniors now start as AI agent managers
โจ And more AI goodnessโฆ
โก The Signal
The people building AI are starting to argue about who holds the off switch.
Satya Nadella wants companies to treat every frontier model like an insider risk: watched, limited and stoppable mid-task by a human. The same question runs through today's issue. A utility can now flip Amazon's Mississippi data centers to backup power on its own, call-center jobs are shrinking after more than a decade of steady growth, Goldman Sachs starts its newest bankers as supervisors of AI agents, and churches are turning AI into a question of faith. Each story is a version of one fight: who stays in charge once the models do more of the work.
All the best,
Kim Isenberg
(Art by Mike Sullivan / The Information)
โก Amazon Hands a Utility the Off Switch
Entergy told Amazon earlier this year that it would have to start pushing AWS's Mississippi data centers off the grid at peak times, The Information reports. Amazon's answer is Project Falcon, a system that lets the utility switch the data centers to backup power, such as gas generators, without first calling AWS. Mississippi is no outlier: Texas passed a law last year letting its grid operator curtail large customers, Duke Energy has added similar contract terms, and PJM, the largest U.S. grid operator, has asked federal regulators for the same power.
๐ tl;dr: AI data centers now draw so much power that utilities are getting the right to cut even Amazon off when the grid runs short.
(WordCraft / ArtCraft on GitHub)
๐ An Open-Source Microsoft Word, Co-Written by Claude
WordCraft, a free, open-source rebuild of Microsoft Word written from scratch in Rust, has collected more than 7,000 GitHub stars since its repository was created on October 7. It opens and saves .docx files, brings Word's ribbon, styles, track changes and mail merge, and runs on macOS, Windows, Linux, BSD and in the browser. Two details make it an AI story: most of its recent commits list Claude Opus 5.5 as co-author, and all 422 of its commands can be driven by AI agents through MCP, the protocol that lets assistants operate software.
๐ tl;dr: The ArtCraft team is rebuilding one of the world's most-used apps with AI coding help, and designed it so AI agents can use it too.
(Pope Leo XIV. Photo: Vatican Media/AFP/Getty Images via WSJ)
โช AI Has a God Problem, From the Vatican to the Pulpit
From the Vatican to small Baptist churches, religious leaders are turning AI into a moral question, and the Wall Street Journal argues that this is becoming a political risk for OpenAI and Anthropic. Pope Leo XIV compared AI to the Tower of Babel in a May encyclical, the National Association of Evangelicals issued an AI policy statement this month, and a South Carolina pastor says AI anxiety in his congregation now outranks wars and the economy. Sam Altman recently distanced OpenAI from a New York Times report that Anthropic unsuccessfully lobbied the Vatican on whether AI could be conscious.
๐ tl;dr: Churches are starting to shape how ordinary Americans judge AI, adding organized moral pressure to the existing fights over data centers and jobs.
State of Product reveals what AI still hasnโt solved
80% of product professionals say AI helps them ship faster, but customers arenโt seeing value any sooner. Atlassianโs State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasnโt kept pace to gut instinct overriding customer evidence.
๐ฌ Watch This
The All-In hosts open their October 9 episode with a 35-minute debate on Anthropic's view that Claude might be conscious: whether that amounts to a new religion, the reported lobbying of the Pope, alignment risks and the push for model welfare. It is Silicon Valley's side of the argument the Wall Street Journal hears from the pulpit in Story 3. The segment starts at 2:53.
"The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least."
โ Satya Nadella, Chairman and CEO of Microsoft, on X
The closing line of his Saturday essay on treating AI models like insider risks. The Featured Story below explains the seven controls he wants built around them.
A leaks account claims an unreleased Anthropic model, which it calls "Claude-next," built this 3D car from a single-sentence prompt, working for two hours in Blender. Anthropic has not confirmed any such model. The same account has posted several one-shot demos since Saturday, including SVG drawings it credited to the "next Anthropic model."
Nadella Wants a Human Hand on AI's Emergency Brake
The Takeaway
๐ Microsoft's CEO says companies should treat frontier AI models, closed and open-weight, as insider risks: powerful, useful and never fully trusted.
๐ The controls over what a model can access and do must sit outside the model, where it cannot tamper with them.
๐ His seven principles include tamper-proof logs of every meaningful action, no model checking its own work, and an emergency brake that lets a person stop a model mid-task.
๐ Any company deploying agents can start on the playbook now, and Nadella wants it turned into industry standards where existing ones fall short.
Microsoft's CEO wants companies to stop trusting their own AI models. In an essay posted on X on Saturday, Satya Nadella argued that businesses should treat frontier models, closed and open-weight alike, as insider risks: capable actors with access to sensitive systems that can make mistakes or be compromised. With ordinary software, engineers can trace a behavior back to a specific code path. With today's models nobody can, yet companies are handing agents their most sensitive data and the power to act on their behalf. "We simply can't outsource responsibility for what intelligence does on our behalf," he wrote.
(Satya Nadella at a Microsoft event in San Francisco on October 7. Photo: David Paul Morris/Bloomberg via Getty Images)
His fix comes from corporate security. Companies have spent decades learning to handle powerful insiders: establish identity, limit privileges, log activity, draw containment boundaries. Nadella wants the same rules for models, with the controls placed outside the model so it cannot tamper with them. His seven principles run from model diversity, so that no model checks its own work, to tamper-proof, human-readable logs of every meaningful action and incident disclosure when something fails. The centerpiece is containment: "Think of it like an emergency brake. An authorized person should always be able to pause or shut down a model mid-task." He also calls readable reasoning traces (chain of thought) non-negotiable, while conceding they are not enough on their own, because nobody yet knows how to make a model's stated reasoning reliably faithful.
(Graphic: Superintelligence, based on Nadella's essay)
The proposal carries weight because of who is making it. Microsoft sells frontier models through its Foundry platform, announced its own MAI-Code-1.1 Flash coding model last week, and runs the Windows and Copilot software where many companies will deploy agents. The essay joins a wave of calls from tech leaders for stronger AI safeguards, at a moment when President Trump opposes slowing U.S. AI development. Nadella's plan sidesteps that fight: it asks nobody to build slower, only to build so that a human can always reach the brake.
Why it matters: Agents are moving into company data, code and transactions faster than anyone can explain how they reason. Nadella's answer is to assume a model can fail and to build the brake before it does.
The chart: An a16z chart built on Revelio Labs workforce data tracks the year-over-year change in global call-center employment. For most of the 2010s the industry grew by roughly 4% a year, dipped to about 2.5% in 2020 and rebounded to nearly 5% in 2021. Growth then slid, turned negative in early 2025 and reached about minus 4% in the latest 2026 reading.
The lesson: Call centers were long expected to be among the first jobs AI would take over: scripted conversations, measurable results, enormous volumes. This chart shows that expectation turning into lost jobs worldwide. a16z notes that even lower-middle-income countries, which held up through 2024, slipped into decline by the fourth quarter of 2025.
The caveat: The slowdown started when interest rates rose, slightly before ChatGPT launched, so AI is not the only cause. a16z's argument is that AI turned a slowdown into a steep decline.
๐ฆ Goldman's Newest Bankers Start as AI Agent Managers
โก Bottom line: Goldman's Asia-Pacific president says the bank's junior hires now start their careers supervising AI agents instead of doing grunt work.
๐ก Why it matters: Years of entry-level work were how Wall Street trained its bankers, and middle managers ran that system.
๐ What it means: With juniors directing agents, the coordinating layer shrinks and the classic banking pyramid gets flatter.
At Goldman Sachs, the newest recruits now begin their careers managing AI agents instead of spending years on entry-level grunt work. "When our young folks now start work, they're managing agents," Kevin Sneader, Goldman's president of Asia Pacific outside Japan, said on Thursday at the Milken Institute Asia Summit in Singapore. "They have to make the most of this virtual army that they've now got."
(Kevin Sneader, president of Asia Pacific ex-Japan at Goldman Sachs. Photo: Milken Institute)
For generations, that grunt work was how bankers learned the trade: juniors built models, checked numbers and drafted pitch books, and middle managers organized and reviewed what they produced. In the setup Sneader describes, the coordinating job moves to the juniors who direct the agents. "That management task is no longer with middle management," he said, according to Bloomberg.
That puts pressure on the middle of the pyramid. Banks are traditionally built with a wide base of juniors, fewer experienced bankers above them and a small group of senior leaders at the top. When the base supervises software directly, the layer whose job was coordination has less to do, which is why Bloomberg's headline says the shift leaves middle managers "in limbo."
(Photo: Bloomberg)
Today's Graph of the Day shows a harsher version of the same shift lower down the white-collar ladder, where call-center jobs are disappearing outright. At Goldman, the newest hires arrive as managers instead: each gets a team of agents and, from day one, a responsibility that used to take years to earn.
Unify Your Teams and Tech Stack With HubSpot
Connect your customer data, teams, and tools without the hassle of complex integrations or lengthy setups. One easy platform gives marketing, sales, and service teams a unified customer view and the tools to turn it into growth.
Why HubSpot and what's new
โข Generate leads and automate marketing with Marketing Hub
โข Build your pipeline and close more deals with Sales Hub
โข Scale customer support and drive retention with Service Hub
โข Keep customer data clean, connected, and actionable with one, unified platform
Join 306,000+ in over 135 countries using HubSpot to grow their businesses.
See what a more connected approach to growth can do for you and your team. Get setup quickly and start checking off your hardest tasks.
