Four throughlines run through today’s issue.
The Bill Comes Due: Courts and Regulators Close In
Anthropic and Google are both facing new accountability mechanisms, one from a lawsuit over internal chats and one from an EU transparency channel built for exactly this kind of audit.
- Sony, Warner Chappell Sue Anthropic Over Staff Piracy Chats. Sony and Warner Chappell are suing Anthropic, citing internal staff chats that allegedly show its CEO endorsing torrent downloads of pirated libraries containing copyrighted song catalogs.
- EU Classifies ChatGPT as a Very Large Online Platform. Brussels applied its toughest online safety law to OpenAI’s ChatGPT, plus Reddit and Roblox, after each service topped 45 million monthly users in the EU.
- DSA data access lets researchers audit Google’s election AI Overviews. AlgorithmWatch used the EU’s new DSA data access channel to run 4,480 queries against Google’s election AI Overviews and found the answers uneven, thin on sources, and inconsistently worded.
ChatGPT’s Institutional Reach: The Pentagon and the Ad Business
ChatGPT is showing up in more institutions at once, from the Pentagon workforce to OpenAI’s own ad business.
- Department of War Puts ChatGPT in Front of 3 Million Staff. OpenAI’s ChatGPT Mil went live on the Pentagon’s GenAI.mil platform, cleared for controlled unclassified work across the Department of War’s entire 3 million person workforce.
- ChatGPT Ads hits $1B run rate, OpenAI says, in under 200 days. OpenAI says ChatGPT Ads reached a $1 billion run rate in under 200 days, an unaudited internal figure, as self-service buying opens in India, Europe, the Middle East and North Africa.
The Ghost in the Machine: Coding Agents Multiply
Three labs shipped coding agent updates within days of each other, each staking its claim on a different strength.
- Meta bets its coding agent on receipts, not benchmark bragging. Meta’s Muse Code pairs its Muse Spark 1.2 model with replayable JSONL logs and per-agent git worktrees, competing on auditability rather than raw benchmark claims.
- OpenClaw 2.0 turns an installer fix into a 16,000-PR rewrite. A planned cleanup of setup and the browser app grew into OpenClaw’s largest release yet, touching memory, skills, models, and security across 16,000 pull requests.
- Z.ai Ships ZCode, a Desktop Coding Agent Built on GLM-5.3. Developer Flavio Copes ran Z.ai’s new desktop coding agent, built on GLM-5.3, against a 1,700-post codebase and called it capable but missing rivals’ subagents and terminal control.
Models That Delete a Step: Rendering Without Code, Forecasting Without Fine-Tuning
Two releases today take a stage engineers used to build by hand and absorb it into the model. One removes the code beneath an interface, the other removes task-specific training.
- Runway Ships Solaris, an AI That Renders Apps Instead of Code. Runway calls Solaris an Interface World Model: it draws each frame of a working interface as the user clicks and drags, with no code generated in between. Runway says testers preferred it to a site built by Claude Opus 5, which is the company’s own claim rather than an independent result.
- Google’s TimesFM-3 forecasts several metrics at once, no fine-tuning. Google’s 330 million parameter TimesFM-3 forecasts multiple related metrics in a single pass and reads known future events, with no task-specific fine-tuning required.
Builder’s Toolkit: Infrastructure for the Agent Era
Two tools built by individual engineers, aimed at the parts of agent work that only hurt once you have shipped something.
- One engineer’s fix for agent memory: plain Markdown files. Engineer Cal Paterson argues most agent memory systems are overbuilt and proposes storing memories as portable Markdown pages backed by a SQLite vector index instead.
- A terminal tool that watches what your coding agent does to a database. diffium-db streams a live diff of schema and row changes while a coding agent runs, catching a bad migration before it gets committed rather than after.
Quick Hits
- Mollick: design where AI agents must stop and ask a human. Wharton’s Ethan Mollick argues operators should decide in advance when an autonomous agent must stop and ask a human, not improvise the boundary after it has already acted alone.
- Frontier AI Is Splitting Into Closed Camps. Investor Tom Tunguz argues Salesforce, Anthropic, OpenAI, and Nvidia are drawing tighter lines around who can run their strongest models, leaving buyers with less negotiating leverage than before.
- OpenAI is quietly testing pay-only-if-it-works pricing. OpenAI is reportedly testing pay only if it works pricing with a handful of large accounts, billing for completed work instead of tokens, though the company has not confirmed it.
- Google Prototypes “Rooms” Collaboration Spaces for Gemini Enterprise. TestingCatalog spotted an unreleased Gemini Enterprise feature called Rooms that would let teams and Gemini share a single, goal-driven collaboration workspace inside the enterprise product.
- A Fintech Investor Says This Agent Finally Feels Like a Product. Fintech investor Simon Taylor argues the Instinct agent, which works from a text message instead of an app, is the first consumer AI agent that feels like a real product.