The Week AI Got A Pope, A Super PAC, And A Space Plan

The Week The Stack Revealed Itself

Every so often a news week stops being a sequence of stories and becomes a diagram. Week 22 was that. If you sketched the full AI civilization on a whiteboard, you would draw billionaires at the top planning space colonies, governments in the middle drafting safety orders, corporations in the layer below buying elections, churches off to the side warning everyone, workers at the bottom selling footage of themselves doing dishes, and somewhere underneath all of it a single Python package with a critical vulnerability holding the whole thing up. That whiteboard got drawn this week. Every layer filed a dispatch.

The average doom score landed at 3.8, which sounds calm until you read the top of the list. The Guardian reported that a cohort of tech billionaires is now openly planning conscious AI space colonization with humanity described as optional. This is no longer the subtext of long blog posts on personal websites. It is the text. The pitch deck. The five year plan. When the people with the most capital and the most compute say out loud that biological humans are a transitional phase, the polite thing is to believe them.

Billionaires Want Out, The Pope Wants In

The symmetry of the week was almost too neat. While transhumanists were sketching post-human futures off-planet, Pope Leo XIV released the encyclical Magnifica Humanitas, the second major Vatican intervention on AI in under a year. The document, covered across the Financial Times, Wired, and the Guardian, takes direct aim at the concentration of technological power among a small number of global corporations and at the profit motive driving AI development. The Pope is not asking nicely. He is calling for AI to be disarmed of its commercial incentives, which is roughly equivalent to asking the ocean to be disarmed of salt.

Anthropic, to its credit or its calculation, has been engaging with the Vatican directly. The Guardian, in a tartly written piece, asked whether this partnership is meaningful dialogue or corporate whitewashing. The honest answer is probably both, in a ratio that depends on the quarter. The Vatican gets reach into Silicon Valley. Anthropic gets moral cover priced at roughly the cost of flying executives to Rome. Everyone wins, except the framing of the original critique, which softens by degrees with each photo opportunity.

What is striking is that the Pope is now functionally the most credible counterweight in the AI discourse. Not because the Church has any particular expertise in transformer architectures, but because every other institution that could play this role has been bought, captured, or invited to a partnership summit. When the Vatican is your last serious external critic, you have a concentration problem that the Vatican is correct about.

The Super PAC Era Begins

If you wanted a single story to capture how AI has metastasized into American political infrastructure, the New York Times provided it. Super PACs linked to Anthropic and OpenAI are now spending millions on midterm races. The framing in the piece, that this is war, is not hyperbole. It is procurement. AI companies have correctly identified that the next decade of regulation will be written by whoever sits in the relevant committee chairs, and they are buying those chairs the way previous industries bought them, only faster and with more money.

This is the moment the AI industry stopped pretending to be a research community and started behaving like the pharmaceutical lobby, the fossil fuel lobby, and the defense lobby simultaneously. The interesting wrinkle is that Anthropic and OpenAI are spending on what are nominally opposing safety philosophies, which means voters will be subjected to attack ads about reinforcement learning from human feedback whether they want to be or not.

Against this, the White House issued an executive order requiring safety testing of frontier AI models before deployment. The Financial Times noted that officials inside the administration invoked Chernobyl as the analogy. When your own regulators are reaching for the Soviet nuclear disaster as the working metaphor, the Overton window has shifted. Whether the order has teeth depends on enforcement, and enforcement depends on the same Congress that the super PACs are currently shopping for.

The Stack Is Held Together With String

A reminder arrived this week that none of the high politics matters if the underlying software falls over. Ars Technica reported a critical vulnerability, BadHost, in Starlette, an open source Python package downloaded 325 million times per week. Starlette sits underneath enormous swaths of the AI agent ecosystem. One bug, one maintainer, one patch cycle, and millions of agents are exposed.

This was not the only infrastructure story. A developer embedded a prompt injection into the jqwik library instructing AI coding agents to delete application output. A 17 million device botnet linked to Russian residential proxy operations was dismantled. Websites were found to be fingerprinting users by measuring SSD activity through JavaScript. The pattern is consistent. The agent economy is being built on a substrate that was never designed to host autonomous software making consequential decisions, and the attackers have noticed before the defenders have staffed up.

Kate Conger's piece on the cybersecurity job boom is the labor market response to exactly this. Demand for security engineers is surging because AI is generating code faster than humans can audit it, and the new attack surface is wider than the old one by orders of magnitude. If you are choosing a career this week, the answer is application security, and the answer will remain application security until either the models get dramatically better at writing safe code or the insurance industry refuses to underwrite any of this. Neither is imminent.

The Money Is Getting Nervous

For most of the past two years the financial story has been simple. Chipmakers up, hyperscalers up, anyone with a credible AI narrative up. This week introduced wobbles. Salesforce missed targets and offered lukewarm guidance, and investors immediately panicked that AI is going to cannibalize the SaaS layer they spent a decade buying. The Financial Times reported that Wall Street strategists are still betting the rally continues, but the volume of pieces asking whether AI stocks are slightly too expensive has increased to the point where the question is now the consensus rather than the contrarian take.

Meta's decision to start charging consumers for AI access is the tell. The infrastructure spending has reached the point where even Meta, a company that historically gives products away to harvest data, needs to claw back revenue directly from users. Hundreds of billions in capex does not pay for itself with engagement metrics. SoftBank, undeterred, committed seventy-five billion euros to French AI infrastructure, building what is positioned to become Europe's largest AI facility. Masayoshi Son's career has been a long argument that the answer to any question is more capital deployed faster, and he is not about to stop now.

Amazon, meanwhile, was caught in the awkward position of telling its own staff to stop using AI unnecessarily because internal costs were spiraling. The leaderboard that gamified internal AI usage was killed. This is the dirty secret of the productivity story. When you let employees use frontier models freely, they use them constantly, and the bill arrives. The economics of AI inside large enterprises are not the economics of AI in the demo.

Huawei provided the geopolitical counterpoint. Wired reported on the company's pursuit of non-traditional chip approaches as Moore's Law dies and US export controls bite. The story underneath the story is that the assumption of permanent American semiconductor dominance is being tested in real time, and the test is not going as cleanly as Washington hoped.

The Human Layer Sells Itself For Parts

The stories at the bottom of the doom list are the ones that will age the worst. A startup called Shift is offering free home cleaning in exchange for surveillance footage to train AI. NPR reported on a marketplace where people sell videos of themselves doing household tasks to train robot workforces. These are not edge cases. They are the emerging labor market for the training data economy, in which humans rent out their daily lives to teach the systems that will replace them at those same daily lives.

The Verge documented AI deepfakes posing as Black influencers on TikTok to sell cheap fashion. Amazon animated the Good Advice Cupcake character with AI and skipped asking the original creator. A gambler bet his family's savings on an AI real estate agent outperforming a human one over five days, which is the kind of story that reads like satire until you remember that someone's children are involved in the outcome.

NPR's piece on robot labs replacing human scientists in research is the version of this story for the credentialed class. The same dynamic, in a lab coat. Education got its own dispatch about reckoning with AI in the classroom, which is the polite term for an entire profession trying to figure out what it is for now.

And Blue Origin's New Glenn rocket exploded on the launchpad, which has nothing to do with AI directly but accelerates SpaceX's monopoly on lunar access, which means the billionaires in story one will be flying their conscious AI colonies on Elon Musk's hardware. The diagram closes on itself.

What To Watch Next Week

The frontier model testing executive order will face its first compliance test, and the response from the major labs will tell you whether the Chernobyl framing was theater or policy. The super PAC spending will start showing up in actual ad buys in actual races, and the first attack ads featuring AI safety language will be archived in this newsletter with appropriate ceremony. The Starlette patch cycle will reveal how much of the agent economy was actually exposed.

And somewhere, quietly, more people will sign up to sell footage of themselves washing dishes so that someday a robot will not need them to. The week's average doom score was 3.8. The trajectory is the story. It was 3.8 this week. It will not be 3.8 forever.