The Week the Brakes Came Off
Trump killed the safety review, Musk lost in court, banks failed their AI stress tests, and Hassabis called it the foothills of the singularity.
The Brakes Came Off, and Nobody Was Driving
The defining feature of Week 21 was the absence of resistance. Every institutional mechanism that might have slowed AI deployment this week either failed, folded, or was deliberately switched off. The White House scrapped its own safety review. A federal court dispatched Elon Musk's lawsuit against OpenAI. The European Central Bank held a panic meeting and confirmed that yes, the banks are still wide open. Demis Hassabis stood on the Google I/O stage and used the word singularity without irony, and the room applauded.
If you were waiting for the adults to step in, the adults filed paperwork this week confirming they will not be stepping in.
Let us walk through it.
Washington Surrenders, Politely
Trump's AI executive order was supposed to include a safety review. Then it was not going to include a safety review. Then the signing ceremony itself got postponed. The arc of the week, in three news cycles, was from regulation to deregulation to vapor.
This is not chaos. This is a result. The industry wanted no oversight and got no oversight, and it did not even have to lobby particularly hard, because the tech-funded super PACs were busy elsewhere, dumping money into congressional primaries with anti-immigrant ad buys aimed at clearing the path for whatever AI faction wins the internal knife fight. The lobbying happened upstream of the legislation, in the candidate selection itself. By the time a bill reaches the floor, the floor has already been bought.
Meanwhile the UK AI Safety Institute is staffing up by poaching from OpenAI and Google, which is either a clever co-option of insider knowledge or the regulatory equivalent of hiring the fox to redesign the henhouse ventilation system. Probably both. The Institute's threat detection unit will be excellent at identifying the threats its alumni already knew about and worse at identifying the threats its alumni are currently building.
The takeaway from the Washington beat is simple. There is no regulatory architecture coming. The window for one closed quietly this week while everyone was watching Google I/O.
The ECB Held a Meeting Nobody Wanted to Hold
The most underreported story of the week was the European Central Bank's emergency session on AI-driven bank vulnerabilities. The framing in the financial press was measured. The reality is that AI models, used as red-team tools, are finding security flaws in the IT systems of European lenders faster than the lenders can patch yesterday's flaws. The ECB had to ask the banks, in writing, to please catch up.
This is the practical edge of capability overhang. When the offensive tools get cheaper and faster while the defensive posture remains a quarterly compliance exercise, the gap is not a curve. It is a cliff. The ECB knows this. The banks know this. Nobody has a plan beyond strongly worded letters and a meeting.
File this alongside the Verge's reporting that security researchers are now documenting attack vectors that exploit the personality quirks of chatbots themselves, which is to say, the chatbots are now an attack surface for the systems they were supposed to defend. Every layer of the stack is now a vulnerability, and each new layer ships before the previous one is hardened.
Musk Loses, OpenAI Accelerates
Musk's suit against OpenAI was the last serious legal instrument that might have slowed the company's restructuring and expansion. The court tossed it. The New York Times noted, almost as an aside, that protests over AI development are growing. They are. They are also irrelevant to the legal calendar, which has now cleared OpenAI's runway through at least the next product cycle.
The simultaneous proposed merger between NextEra and Dominion, a $420 billion energy-and-infrastructure deal, would consolidate control over the physical layer of US data centers into roughly two hands. AI is not just a software phenomenon this week. It is becoming a utilities monopoly question, and the antitrust apparatus that might have looked at it is, see above, busy not existing.
Add to that the circling around Tenstorrent by Intel and Qualcomm, both desperate for a credible Nvidia alternative, and you have a picture of the compute stack being locked down at every level while the regulators are at lunch.
Google Declared Victory From the Foothills
Google I/O was where the week's posture became explicit. Sundar Pichai announced Gemini-powered smart glasses and AI agents that will, in the company's framing, handle every task you might once have done through a search box. The pitch is convenience. The mechanism is total data access. The Verge put it plainly. Google's AI future requires you to hand over substantially more of your life than you currently do, and to trust the company more than its track record warrants.
Demis Hassabis, on the same stage, described the current moment as the foothills of the singularity. This is the DeepMind CEO using the S-word in a keynote. Five years ago this would have been a scandal. This week it was a slide.
The demos did not entirely cooperate. Gemini's multimodal capabilities were demonstrated by a Verge journalist who used them to turn a stuffed deer into the convincing star of a fake vacation video, replicating Google's own promotional aesthetic without permission and without effort. Google's AI Overviews, separately, were caught producing hallucinated answers that ignored the user's actual query, which is the kind of failure mode that would have killed a product launch in any prior era and now barely registers as a footnote.
The foothills, it turns out, are also where the product still does not work.
Meta Fires Everyone, Then Fires Everyone Else
Meta's week was a study in sequencing. Two days before announcing 8,000 layoffs, the company reassigned 7,000 workers into AI-focused roles. The reassigned workers were, in many cases, then included in the layoff round. This is not a strategy. This is a logistics exercise in moving bodies into a column before deleting the column.
The internal response was, to its credit, funny. A departing employee built an in-house radio station that played AI-generated breakup songs about being laid off. This is the emotional register of the AI workplace in 2026. You are replaced by a tool that you then use to write the song about your replacement. You post the song to the internal Slack on your way out. Someone forwards it to the press. It becomes a Mike Isaac story. The cycle completes.
The Wired framing of the week, that Meta is hemorrhaging staff while Google doubles down, misses the point. They are not making opposite bets. They are making the same bet at different cadences. Both companies have concluded that the human workforce is a transitional cost structure.
The Labor Market Tells On Itself
The Atlantic ran the data point that should anchor every conversation about AI and work for the next year. Computer science graduates now face higher unemployment than philosophy majors. Read that twice.
The field that was supposed to be the safe bet, the credential that justified the tuition, the degree your parents nodded at, is now underperforming the degree your parents worried about. This is the labor market pricing in what the executives have been saying out loud. Entry-level coding is the first job category to be substantially automated, and the universities have not adjusted their pipelines because the pipelines take four years to adjust and the technology took eighteen months.
The philosophy majors are not winning. Nobody is winning. But the relative ranking is the tell. When the most automatable knowledge work gets automated first, the credentials that pointed at that work lose value faster than the credentials that pointed nowhere in particular.
Elsewhere in the labor and culture beat: three Commonwealth Short Story Prize winners are under suspicion of using ChatGPT. A book about AI was found to contain fabricated quotes the author blamed on the chatbot rather than on himself, which is the new and durable form of plagiarism defense. Surgeons are reporting patients who arrive asking to have their faces restructured to match AI-generated images of themselves, and then asking to have it undone when they realize the proportions are not human. The Spotify and Universal deal that monetizes AI covers of real songs is the music industry's formal capitulation, dressed as a licensing win.
The culture is metabolizing the technology by losing the ability to tell what is real, and then charging a subscription for the confusion.
The Cars Drove Into the Flood
The single most on-the-nose story of the week was Waymo halting service across six cities after its autonomous vehicles drove into flooded streets in Atlanta. The cars did not know the road was underwater. The cars went anyway. The Financial Times, in a separate piece, made the case that robotaxis need real roads to learn how humans panic, which is true and also a confession. The vehicles are being trained on the public by being deployed against the public.
This is the operational metaphor for the entire week. The systems are being shipped into conditions they cannot read, and the conditions are providing the training data through the damage. The ECB's banks, Google's hallucinating Overviews, Meta's shuffled-then-fired workforce, the regulatory apparatus that keeps postponing its own signing ceremony. Same pattern. Drive into the flood. Collect the data. Issue the press release.
What to Watch
The NextEra-Dominion merger filing is the one to track. If it clears, the physical infrastructure question is settled for a decade, and every subsequent conversation about AI competition becomes a conversation about who rents from whom.
The ECB's follow-up actions, or absence of them, will tell you whether European financial regulators have any tools left or whether they have joined their American counterparts in the strongly-worded-letter phase.
And watch the labor data. The CS unemployment number is a leading indicator. If it holds or worsens through the summer hiring season, the universities will have to say something, and what they say will shape the next decade of who studies what, which shapes who builds what, which shapes everything else.
The brakes came off this week. Nobody is putting them back on. The question is no longer whether the car stops. The question is what the car hits first.
- Trump Drops Safety Review; Tech Wins Without Trying Hard · The Guardian · 7/10
- ECB Panic Meeting Confirms Banks Still Vulnerable to AI Exploits · Financial Times · 7/10
- NextEra, Dominion Merger Would Control US Data Centers · Financial Times · 6/10
- AI Juggernaut Accelerates Following Musk's Court Loss · New York Times · 6/10
- DeepMind CEO Describes Current Moment as Singularity Foothills · The Verge · 6/10
- Google's AI Future Hinges on Your Data and Trust · The Verge · 6/10
- Meta Cuts 8,000 Jobs While Pursuing AI-First Strategy · New York Times · 6/10
- Google's Gemini Turns Stuffed Deer Into Deepfake Vacation Star · The Verge · 6/10
- ECB Discovers Banks Still Haven't Patched Yesterday's Vulnerabilities · · 6/10
- Meta Shifts 7,000 Workers to AI Before Layoffs · New York Times · 5/10
- Google Envisions Search Box That Does Everything for You · The Verge · 5/10
- Google Releases Smart Glasses, Adds AI Agents to Search · Financial Times · 5/10
- Meta Hemorrhages Staff While Google Doubles Down on AI · Wired · 5/10
- Spotify and Universal Monetize AI-Generated Song Covers · The Guardian · 5/10
- Waymo Halts Service After Self-Driving Cars Hit Flooded Roads · New York Times · 5/10