It was one of those mornings when the rain seemed to seep into every pore, a cold damp shroud over the city. My desk lamp threw a pale glow over a cluttered office that I never quite finished cleaning, and the phone, when it rang, shrieked like a verdict.

On March 24, a New Mexico jury found Meta liable for misleading consumers about the safety of Facebook, Instagram, and WhatsApp and ordered it to pay $375 million in civil penalties — $5,000 per violation, the maximum the state's Unfair Practices Act allows. The verdict, the first against a major tech company in a state-led child safety trial, capped a nearly seven-week proceeding brought by Attorney General Raúl Torrez, who in 2023 had sued the company over allegations that lax security let predators reach minors. The next phase will be a bench trial seeking court-mandated platform changes. The case is not, strictly, about AI — it is about social platforms — but it sits at the front of a longer queue in which AI-mediated harms to children are the through-line, and the queue is getting noticeably busier.

On March 25, OpenAI made three announcements that read together as one decision. The company is shutting down Sora, its consumer video-generation app, with the web and mobile experiences ending April 26 and the API on September 24. It is also removing the Instant Checkout feature from ChatGPT — partners like Walmart had been reporting conversion rates roughly three times lower through ChatGPT than via their own websites — and indefinitely shelving the planned adult-mode ChatGPT, which had been promised for late 2025, postponed to Q1 2026, and ultimately blocked by internal safety advisers concerned about child protection and mental health. The Sora shutdown also collapsed a reported $1 billion licensing deal with Disney that would have allowed Sora users to generate videos featuring Disney characters. Read together, the three cuts look like a company tidying its desk before going public: trimming the experiments that lost money, drew lawsuits, or both, and concentrating on what the IPO prospectus is going to need to defend.

On March 26, U.S. District Judge Rita Lin, of the Northern District of California, issued a sweeping preliminary injunction in Anthropic's favor against the Department of Defense and seventeen other federal agencies, blocking enforcement of the supply-chain-risk designation that had been imposed in early March. Her 43-page ruling was unsparing: "Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government." She wrote that the record "strongly suggests that the reasons given for designating Anthropic a supply chain risk were pretextual and that [the government's] real motive was unlawful retaliation" — "classic illegal First Amendment retaliation," in her phrase. The injunction was paused for seven days to allow appeal. Pentagon CTO Doug Matty said the ban still stands. The lawsuit continues; the principle the case is testing — whether the government can punish a contractor for declining to remove its own ethical guardrails — has not yet been resolved, only postponed.

The same day's Anthropic news was quieter and arguably more consequential. On March 24, the company shipped Auto Mode for Claude Code, a research preview in which a safety classifier evaluates each proposed tool call — file write, shell command, edit — and decides in real time whether the action needs human approval. The default permission model in Claude Code had been deliberately conservative: every file write, every bash command, a prompt. Anthropic's own data showed that 93% of those prompts were being approved by users anyway, which is a polite way of saying the approval ritual had become exactly that — a ritual. Auto Mode replaces the ritual with a classifier, which is to say it replaces the human checkpoint with a model checkpoint. The risk is that the checkpoint is now itself a model. The compensating virtue is that long agentic tasks become possible without 137 clicks of "approve."

On March 24, Databricks announced Lakewatch, an agentic SIEM — Security Information and Event Management platform — designed to help enterprises defend against AI-driven attackers and to unify security, IT, and business data under a single governed environment. The product is powered by Anthropic's Claude, which Databricks says provides the reasoning layer that correlates signals across the three data domains. It's now in private preview, with Adobe and National Australia Bank as early customers. The pricing model is notable: rather than charging on data volume (the standard SIEM model that forces customers to discard 75% of their telemetry on cost grounds), Lakewatch charges on compute. The framing is competitive — Splunk and Microsoft Sentinel by name — but the underlying claim is more interesting: that agentic attackers require agentic defenders, and that "agentic" requires having all the data, which requires not pricing the data out of the building.

On March 26, Snowflake Cortex Code went generally available in Snowsight, alongside three other updates: native Windows support in the CLI, a new Agent Teams capability that breaks large assignments into coordinated parallel work, and standardized agent skills for building on data. Snowflake has had over 4,400 new Cortex Code users since November; the GA, the Windows CLI, and Agent Teams together push the product from "AI coding agent for Snowflake users" toward a control plane for moving agents into production across the data engineering workflow.

On March 26, Google launched Gemini memory and chat import tools, allowing users to transfer their preferences, context, and full conversation history from ChatGPT and Claude into Gemini. The mechanism is two-part: a copy-paste prompt that asks the source assistant to summarize what it knows about the user, and a ZIP-file upload of exported chats. Anthropic shipped the equivalent feature three weeks earlier, and OpenAI's data export covers the same ground; the convergence is partly competitive (lowering switching costs to recruit users) and partly regulatory (GDPR Article 20). The feature is not available in the EEA, UK, or Switzerland — itself a small commentary on which regulatory regimes data portability turns out to be available in.

The man across my desk shifted in his chair and started telling me about Apple. The company had just hired Mahesh Balakrishnan, formerly Google's VP for Gemini product marketing, as head of product marketing for AI and machine learning, reporting to marketing chief Greg Joswiak. Balakrishnan started at Apple in March 2026 after roughly a decade at Google overseeing Gemini, Assistant, Search, and Pixel marketing. The hire follows a sequence of Apple AI personnel moves — John Giannandrea's retirement, Amar Subramanya stepping in as VP of AI — that together suggest the company has concluded its Apple Intelligence problem is not only a model problem but a story problem. WWDC will run June 8–12; the keynote, by which point Apple will need a narrative, is set for June 8.

A couple of other items, briefly: on March 24 Epic Games announced it would lay off more than 1,000 employees, roughly 20% of its workforce, citing a decline in Fortnite engagement that began in 2025 and a market environment CEO Tim Sweeney described as "the most extreme we've seen since" the company's early days. Sweeney was explicit that the cuts were not AI-related, which is itself worth recording in a year when nearly every layoff memo finds a way to blame the technology. And on March 24 NASA Administrator Jared Isaacman announced the cancellation of the Lunar Gateway orbital station, redirecting roughly $20 billion over seven years toward a permanent moon base — a major reshaping of the Artemis program, an implicit acknowledgment of competition from China's 2030 lunar goals, and not, strictly, an AI story, though the agency's increasingly compressed timelines depend on AI-assisted engineering at every step. (Three other items from the original Week 13 brief — Roku's Howdy expansion to Prime Video, Pokémon Champions' April 8 release, and the Puig–Estée Lauder merger talks — have been left out. All real, all newsworthy, none AI.)

The man left. I sat for a while listening to the rain. The cases this week were mostly the same case told different ways: a court asking a government whether punishment may follow from disagreement; a company canceling its more interesting experiments to look more profitable; a coding agent learning to decide for itself when a human is required and when not; a verdict that a platform may be liable for what it knew. The pieces don't quite fit together, but they don't quite fall apart either. The city outside waited, as cities do, indifferent and full of light.