The Week AI Stopped Pretending To Be Friendly

The Week The Mask Came Off

For about three years the industry held a consistent line. AI was a productivity tool, a coding assistant, a friendly bot that helped you draft emails and occasionally hallucinate a legal citation. The labs were research outfits with cute mission statements. The chips were just chips. The data centers were just buildings.

Week 17 buried that story.

The Pentagon asked Congress for $54 billion in autonomous drone warfare funding, a hundredfold increase on the previous baseline. Maven Smart, the targeting system that had been quietly maturing inside defense contracts, doubled US strike capacity on the first day of operations against Iran. A thousand strikes in the window where the previous generation of shock-and-awe doctrine managed half that. Meta announced it was firing 8,000 people and closing 6,000 unfilled roles to redirect $135 billion into data center buildout. Anthropic released a model called Mythos that prompted, by the New York Times' account, an emergency response from central banks and intelligence agencies.

This is not the same industry that was demoing image generators at developer conferences eighteen months ago. Or rather, it is the same industry, and we were always wrong about what it was.

The Drone Line Item

Start with the Pentagon number because it sets the frame. $54 billion is not a research budget. It is a procurement budget, which means somebody has decided autonomous drone warfare is a product category to be acquired at scale, not a capability to be debated. The hundredfold increase tells you the previous funding level was a placeholder while the doctrine was being written. The doctrine has now been written.

Maven Smart's performance in the Iran operation is the proof point that unlocks the spending. When defense officials say AI doubles strike capacity, they mean targeting cycles that used to take human analysts hours now take minutes, and the bottleneck moves from cognition to munitions inventory. This is what investors wanted to see. It is also what every arms control lawyer warned about for a decade while being told they were getting ahead of the technology.

The UN panel convened by Pakistan this week noted that terror groups are now weaponizing AI and crypto for operational advantage. The framing is convenient because it lets state actors talk about AI weaponization as something other people do. The honest reading is that the cost curve for autonomous lethality is collapsing for everyone simultaneously, and the states that got there first are now trying to pull the ladder up.

Meta's Math

Meta's layoff was the cleanest articulation of the new labor deal. Fire 8,000 humans. Close 6,000 jobs that were open and presumably needed. Commit $135 billion to data centers. The chief people officer's memo confirmed the May timeline with the affect of a logistics update.

Microsoft and Amazon ran similar arithmetic. The Guardian put the combined Meta-Microsoft cuts at roughly 17% of their workforces, citing AI productivity gains as justification. New York Magazine's read was sharper, noting that surveillance tools are being deployed to extract more labor from the employees who remain. So you fire a chunk of the staff, monitor the survivors more aggressively, and pour the savings into the infrastructure that will eventually replace the survivors. There is a name for this kind of arrangement and it is not a partnership.

The Next Web's piece on tech education catching up with AI-driven job changes was one of the more depressing reads of the week, mostly because the answer is that it cannot. Curricula are calibrated for two-year horizons. Job descriptions are now changing on quarterly cycles. The advice given to a computer science freshman in fall 2024 was already wrong by spring 2026. The honest counsel is that nobody knows what the entry-level job looks like in three years, and the people claiming otherwise are selling bootcamps.

Senator Warren spent part of the week pointing out that AI investment patterns rhyme with the run-up to 2008. She is correct. Concentration risk, opaque counterparty exposure, valuations decoupled from cash flow, and a regulatory apparatus that has explicitly chosen not to look. Whether the bubble pops before or after the autonomous weapons are fully deployed is a question of sequencing, not of substance.

Mythos And The Central Banks

Anthropic released Mythos and something unusual happened. Central banks responded. Intelligence agencies responded. The Guardian's editorial board wrote a piece asking who controls the internet when an AI can find everything on it, which is the polite version of asking what sovereignty even means now.

The specifics of Mythos's capabilities are less interesting than the response pattern. When a private company ships a software update and the Bank for International Settlements has to convene, the framework you have been using to understand the AI industry is wrong. These are not products. They are infrastructure with the option-value of being weapons, surveillance systems, market-moving entities, or all three depending on who points them at what.

Anthropic's branding still leans on the safety-conscious lab story. That story is harder to maintain when your model release is the kind of event that gets briefed to finance ministers. To Anthropic's partial credit, they appear to be the only frontier lab that still pretends the question matters. OpenAI spent the week issuing an apology because Sam Altman acknowledged the company failed to alert law enforcement about an individual connected to the Tumbler Ridge mass shooting. The apology was the news. The fact that nobody is quite sure what OpenAI's reporting obligations are, or should be, is the deeper story. Federal AI surveillance is expanding with what Fast Company described as minimal oversight, which means the legal scaffolding for any of this is being built after the fact, if at all.

The Climate Footnote Nobody Wants

The UK government quietly revised its forecast of AI data center carbon emissions upward by a factor of up to 136. Not 36 percent. Not double. Up to one hundred and thirty-six times the previous estimate. The Financial Times and the Guardian both ran the numbers and arrived at the same conclusion, which is that the British net-zero commitment and the British AI growth strategy are now mathematically incompatible, and the government has not decided which one it is willing to abandon.

The answer, of course, is net-zero, but no minister wants to say that out loud yet. The departmental disagreements the Guardian reported are the bureaucratic version of someone realizing the spreadsheet does not balance and quietly hoping somebody else gets blamed.

This matters beyond the UK because every developed economy has the same problem and most of them have been less honest about the numbers. The compute required to train and serve frontier models is on a trajectory that does not respect grid capacity, water budgets, or political commitments made when AI meant Siri. Something gives, and the historical record on what gives when growth meets sustainability is not encouraging.

The China Frame Is Doing Heavy Lifting

The White House spent the week accusing China of industrial-scale theft of American AI models, specifically through distillation. The Trump administration committed to intelligence sharing with OpenAI, Anthropic, and Google to address the threat. The Financial Times ran an opinion piece arguing the US must block China's access to advanced AI immediately. DeepSeek, on cue, previewed its V4 model claiming parity with US frontier systems.

The China narrative is doing a lot of work right now, and it is worth noticing what it justifies. Export controls, yes. Intelligence sharing with private companies that effectively turns the labs into national security assets, also yes. Reduced domestic oversight on the grounds that we cannot afford to slow down, definitely. The Trump Justice Department intervening on xAI's behalf in its Colorado lawsuit, creating a federal-state preemption fight, fits the same pattern. AI companies are being repositioned as national champions, which historically is a status that comes with subsidies on the front end and consequences on the back end.

The consequences are starting to show. More than 80 people, including cofounders and senior engineers, have left xAI in recent months. Fast Company's piece read like an obituary written while the patient was still upright. When your retention problem is bad enough that the federal government has to litigate on your behalf, the moat is not the technology.

Backlash, Hardware, And The Attacker's Advantage

Two stories sat uncomfortably next to each other this week. The Guardian reported on attacks against OpenAI leadership and warned that anti-AI sentiment is starting to turn violent. Fortune covered a gathering of AI security leaders who openly admitted the offense-defense balance favors attackers and shows no sign of correcting.

The security admission is the more important one because it is structural. AI systems expand the attack surface faster than defenders can map it. Cloud-based deployment is outpacing the security tooling, as the Next Web noted in its piece on the post-quantum cryptography gap. Microsoft spent part of the week patching a critical ASP.NET authentication vulnerability that affected macOS and Linux deployments, which is the kind of routine emergency that used to be news and is now Tuesday.

The physical violence story is what happens when the security failures, the labor displacement, and the impunity of the labs accumulate in public consciousness without a political outlet. A soldier was indicted this week for using prediction markets to bet on classified military operations, which is its own perfect artifact of the era, the financialization of war information enabled by platforms that did not exist five years ago. Genocide threats, the New Times reported, are evolving through digital platforms with denial and slow international response. None of these stories are about AI per se. All of them are about what happens when the institutions meant to govern technology are slower than the technology by an order of magnitude.

What Actually Changed This Week

The doom average for the week was 3.7 out of 10, which sounds moderate until you read the top of the distribution. The pattern that emerged across 187 stories was not acceleration. The acceleration happened already. The pattern was consolidation. The Pentagon picked its vendors. Meta picked its workforce strategy. Anthropic picked its release cadence. The UK picked, by inaction, which of its commitments to break. China picked its competitive posture and the US picked its response.

The interesting question for next week is not what new capability ships. It is which of these consolidations gets contested, by whom, and with what leverage. Warren has the financial-stability argument. The displaced workers have the political argument, if anyone organizes them. The climate argument is dead on arrival in the current US administration and on life support in the UK. The safety argument is being absorbed into national security frameworks that will use it to justify exactly the opposite of what its original proponents intended.

The friendly chatbot era is over. What replaced it is not artificial general intelligence, or the singularity, or any of the science-fiction frames the industry trained you to expect. What replaced it is a normal-looking arrangement of capital, state power, and labor extraction, accelerated by tools that happen to be very good at killing people, firing people, and generating returns. The science fiction was the cover story. The ledger was always the point.