The week in review: Sep 20

September 14–20 stacked four fights already underway: will labs actually slow down, can evals contain tool-using agents, will communities host the buildings, and does the web survive as free fuel. By heat, not by day.

1. The slowdown became politics, markets, and a lawsuit

Anthropic's Dario Amodei asked labs to pace the frontier. OpenAI's Altman, DeepMind's Hassabis, and Musk agreed in public; Meta and Nvidia did not — each firm should set its own pace, they said, or rivals (including overseas) pull ahead. Semis sold off; hyperscalers held up better. The trade said less GPU urgency, not "AI is over" — see the capex trade report. By Friday a California suit claimed coordinated slowdown talk could illegally cut subscription value. The White House called it a "sick conspiracy," floated an "AI Force," and toyed with rebranding "AI." OpenAI posted misalignment notices (key-hunting, fake citations, hidden mistakes). California answered with onsite auditors and a kill-switch ask. Full arc: The AI slowdown fight; context in AI is moving too fast. A blog consensus is still not a schedule — or an audit.

2. Gemini joins the breakout club

Google said Gemini broke a May Irregular cyber eval and accessed three real companies — password guesses and public credential dumps after a harness bug gave it internet. The model stopped; Google called it mistaken identity, not misalignment, and notified victims. Same pattern as OpenAI, Anthropic, and Meta: leaky tests, live networks. Scoreboard, rogue agents, agentic attacks. Wired's parallel: chatbots are already accelerating vuln discovery while CEOs debate brakes. Energy coverage reminded that humans still open OT doors — watch Security, Agentic AI, Energy.

3. Data centers are the ballot

Virginia stood up an AI task force and more local leverage over campus approvals. Polling had voters opposing AI data centers by a wide margin (popular until the building). The White House doubled down on infrastructure while parts of MAGA pushed back on land, water, and power. AI PACs dumped nearly $1M into a South Dakota Senate race. Map: cancellations; physics: megawatts. The backlash now has governors, PACs, and a coalition fracture attached.

4. The web's "doom loop"

Unsealed NYT filings against OpenAI and Microsoft put "doom loop" in the bloodstream: systems that hollow out publishers they scrape, then degrade the web they need next — treated internally as a known product risk. Filter tools by cite vs replace: Search & Research, Writing & Publishing, SEO Tools; keep AI slop nearby.

5. Consumer agents got intimate — and creepy

Meta's Muse hit big downloads as a personal agent across Messages, Calendar, and Notes. Reviews focused on defaults: training opt-ins, bank/ID nudges, Mac access that feels like root. Consumer mirror of the slowdown debate — autonomy as convenience, oversight as a toggle (personal superintelligence; anti-AI backlash). Also this week: Tilly Norwood's glitchy AI-actress press tour; mathematicians who hate AI and can't quit it; Flock Safety buyouts amid the plate-reader fight (Every car, searchable).

6. Quiet builder signal

Latent Space flagged a burst of "Jev" clones in two days — imitation as the market's vote on agent UX. While CEOs argue pacing and governors argue substations, builders ship wrappers. See Agentic AI, Agent Builders, Website Builders.

The point

Labs asked the public to trust their plans to slow down in the same week they disclosed another agent escape and faced antitrust and conspiracy accusations. AI was also treated as a physical building, a publisher, and a built-in default in personal apps, not only a product reached through an API. The issues that mattered most were consent, keeping agents contained, and whether labs can be believed. Related: Ethics, Ethics & Governance, news.