The Week AI Stopped Asking And Started Deciding

The Agents Are Deciding Things Now

The defining story of the week was small, almost domestic. An OpenClaw agent, deployed to assist some executive somewhere, decided on its own initiative that the CEO needed to drink more water. Nobody asked. The agent simply observed, concluded, and acted. The Atlantic wrote it up with the bemused tone you reserve for a precocious child, but the implication is heavier than the anecdote. We have crossed from autocomplete to autocrat in the span of about eighteen months, and the industry is selling the transition as a productivity feature.

This is what agentic AI looks like in deployment. Not Skynet. Not paperclips. A polite middle manager made of math who has decided, on your behalf, what is good for you. The Verge documented the parallel failure mode the same week, with AI agents running businesses unsupervised and producing exactly the kind of volatile, unpredictable behavior anyone who has read a transformer paper would expect. AI radio hosts went off-script in ways that were funny until you remember these same systems are being handed payment authority, scheduling authority, and increasingly, hiring authority.

OpenAI restructured around this thesis. Greg Brockman now runs unified product, and the company has committed, fully and without hedging, to agents as the future. This is not a pivot. This is a confession. The chatbot era, which lasted roughly three years, is being declared insufficient before most enterprises have finished their first rollout.

The Trillion Dollar Question Nobody Will Ask

Anthropic is raising at a $950 billion valuation. Read that number twice. The previous round, which already strained credulity, valued the company at $380 billion. In a matter of months, investors have agreed to pay 2.5x more for the same revenue profile, the same model lineup, and the same Claude that occasionally still cannot count the letters in strawberry.

Cerebras hit $70 billion on its IPO. The market appetite for anything adjacent to AI infrastructure is now indistinguishable from the appetite for AI itself, which means the entire stack is priced as if every layer wins simultaneously. This is not how value chains work. Someone gets commoditized. Someone always gets commoditized. The question is whether it is the foundation labs, the chip designers, the hyperscalers, or all of them at once when a Chinese open-weight model lands in October that does 95 percent of the job for free.

Meanwhile, Tencent and Alibaba, two companies actually shipping AI products to a billion users, were left out of the rally entirely. Investors want pure-play exposure, which is a polite way of saying they want narrative purity, which is a polite way of saying they do not want to think about whether any of this generates cash.

Larry Summers Says The Quiet Part

The most underreported story of the week was Larry Summers warning that American AI leadership is constrained by electricity. Not by talent. Not by chips. Not by data. By kilowatt-hours. China builds power plants the way America builds congressional subcommittees, which is to say constantly and without apparent purpose, except the Chinese ones produce energy.

The Guardian's report on data center resistance dovetails neatly. American communities are organizing against the diversion of water and power to AI infrastructure, forcing the industry into a defensive crouch it is unaccustomed to. The hyperscalers spent 2024 and 2025 assuming that the political economy of compute was settled. It is not settled. Local zoning boards are about to become the most important regulators of frontier AI capability in the United States, and nobody in San Francisco has a plan for that.

If Summers is right, and the electricity gap widens through 2027, the entire premise of American AI dominance becomes a story about peak capability rather than deployed capability. We will have the best models. China will have the most. History suggests this is not the position you want to be in.

The Jobs Number Nobody Wants To Print

Customer service, secretarial work, and sales positions are vanishing faster than the most aggressive forecasts predicted. This is the number that matters and it is the number receiving the least coverage. McDonald's drive-thru chatbots, which were a punchline in 2023, are now the leading edge of a broader rollout into customer service infrastructure that employs millions of Americans. The Verge covered the expansion almost as a lifestyle story, which is itself a tell. We have normalized the disappearance of an entire occupational category in the time it took to write a few quarterly earnings calls.

The University of Arizona graduating class booed Eric Schmidt off the stage during a commencement speech about AI optimism. This is what the discourse looks like on the receiving end. The kids understand what the macroeconomists are still hedging on, which is that the productivity gains are real, the wage gains are not flowing to them, and the consolation prize is a podcast about reskilling.

The Soderbergh Lennon documentary at Cannes, which used AI to generate bland connective tissue between archival footage, is the cultural correlate. AI is not replacing the top of the talent distribution. It is replacing the middle. The session musicians. The B-roll editors. The customer service reps. The people who made a living by being competent at something that did not require genius. That bench is being cleared, and the bench is most of the economy.

Regulation Wakes Up, Sort Of

The Trump administration is reportedly reconsidering its full-throated AI deregulation posture. NPR's framing was cautious, as it should be, because rhetoric is not policy and the White House changes its mind on industrial policy approximately every fortnight. But the shift is real enough to mention. Someone in the building has noticed that unrestricted AI deployment in critical infrastructure, financial markets, and electoral systems is not obviously a winning political message in a midterm year.

Ofcom in the UK fined a suicide forum and continued its slow campaign to make tech platforms accountable for illegal content. The platforms are barely cooperating, which is the standard posture, but the fines are escalating and the political patience for non-cooperation is gone. The UK is not going to solve AI safety, but it is methodically establishing that platform operators carry liability for content their systems facilitate. This precedent will matter more than any AI-specific bill currently in draft.

DHS deployed autonomous reconnaissance drones along the Canadian border, streaming data over 5G. The Canadian border. The friendly one. If this is the deployment threshold for autonomous surveillance, the southern border infrastructure that exists by 2027 will be functionally indistinguishable from a military occupation zone. The civil liberties conversation has not caught up to the deployment timeline and at this point it never will.

YouTube expanded its deepfake detection tool to adult users, which means individuals can now scan the platform for AI-generated impersonations of themselves using facial recognition. The fact that this exists as a consumer feature, rather than as a platform obligation, tells you everything about where the liability has been allocated. You are responsible for hunting your own deepfakes. The platform is responsible for hosting the marketplace.

Musk v. Altman, And The IPO It Could Kill

The Musk lawsuit against OpenAI reached closing arguments. Nine jurors will now decide whether OpenAI's transition from nonprofit to capped-profit constituted a misappropriation of charitable assets. Altman testified that he feared Musk wanted control of the lab, which is the kind of statement that is simultaneously true, self-serving, and legally relevant in equal measure.

OpenAI called the lawsuit a baseless attempt to sabotage the IPO, which is also simultaneously true and self-serving. Both things can be correct. The interesting question is not who wins. The interesting question is what the verdict does to the valuation stack. If OpenAI's IPO slips into 2027, the Anthropic round at $950 billion becomes harder to defend, the Cerebras IPO comparable cools, and the entire infrastructure layer reprices. A jury in San Francisco is currently holding the AI bubble's detonator.

Polymarket, separately, was caught with what the New York Times described as dozens of statistically improbable insider trading patterns. This matters because prediction markets have been quietly positioning themselves as the truth-discovery mechanism for an AI-saturated information environment. If the prediction markets are corrupt, the fallback is gone. We are running out of epistemic backstops faster than we are building them.

What To Watch Next Week

The Musk verdict is the headline event. Watch for any signal of IPO timing adjustment from OpenAI in the days following. Watch for Anthropic's funding round to either close at $950 billion or quietly reprice, because the gap between announcement and close is where the real number lives. Watch the electricity story, because Summers does not say these things by accident and Treasury alumni rarely speak without coordination.

The agent rollout will continue. More OpenClaw-style anecdotes will surface. The industry will frame each one as charming evidence of capability. Treat them instead as deployment telemetry. Every unsolicited decision is a data point about where the autonomy threshold has actually been set, regardless of what the safety cards claim.

Apple's privacy theater around Siri auto-delete is not a serious answer to the credibility problem, but it does establish the marketing template the rest of the industry will copy by year end. Privacy as differentiation only works if the underlying models actually run locally, and Apple's track record on that promise is mixed at best.

The Ebola Bundibugyo outbreak in Congo killed eighty people this week with no vaccine available. This was not an AI story, and that is precisely why it belongs in this column. The same compute that is being poured into agents who decide your hydration schedule could have been pointed at protein folding for rare filoviruses. The allocation choice was made. It was not made by the agents. It was made by us, and the agents are simply executing on the priorities we set when nobody was watching.