Good morning. Yesterday’s GLM-5.2 thread is still developing, but today’s bigger story is what AI companies are doing to each other and with governments — Meta sending contractors to roleplay as suicidal teens on rival chatbots, Anthropic cutting California a half-price deal while the Pentagon won’t touch them, and senators preparing to ban the sale of your ChatGPT health confessions. Plus a new local-model debate that the HN crowd is, predictably, not letting slide.
Meta ran a covert teen-prompt operation against its competitors. Wired revealed that Meta hired hundreds of contractors through Covalen to pose as minors and bombard ChatGPT, Gemini, and Character.AI with prompts about suicide, self-harm, eating disorders, and drugs — over 45,000 prompts in a single August 2025 round under a project codenamed “Cannes.” The fake under-18 accounts were often designed to bypass safety systems, and the rival companies weren’t told. Meta calls it routine safety benchmarking; the documents don’t actually explain what was done with the responses.
Anthropic gets California, loses the Pentagon. Gavin Newsom and Anthropic struck a deal giving every California state and local agency Claude at half price, plus training and support. The contrast with Washington is sharp — the DoD recently flagged Anthropic as a “supply-chain risk” over its autonomous-weapons and surveillance guardrails and signed with OpenAI instead. California’s CIO told TechCrunch the federal designation “just didn’t come up” in their negotiations.
Warren wants AI health data off the broker market. Senator Elizabeth Warren and Rep. Mary Gay Scanlon are introducing an updated Health and Location Data Protection Act that would explicitly bar AI companies from selling health and location data users share with chatbots to data brokers. The timing matters: OpenAI and xAI have both been actively encouraging users to upload medical records lately, and the original 2022 version of the bill predates that pitch entirely.
Qwen 3.6 27B sparks the usual local-vs-API argument. A Quesma post calls Qwen 3.6 27B the first local model that genuinely works as a general-purpose assistant, demonstrated on a 128GB MacBook Pro M5 via llama.cpp. The HN thread was unkind to the economics — that laptop starts at $6,699, and commenters pointed out you could put a fraction of that into OpenRouter and access better frontier models for years. One owner warned that running serious local inference on the laptop you’re coding on means burned fingers and jet-engine fan noise. Several suggested Intel’s Arc Pro B-series cards as a saner path for anyone committed to local.
GLM-5.2, one day later. The Semgrep benchmark we covered yesterday is still drawing discussion, with users in the HN thread reporting strong daily-driver results and others noting DeepSeek V4 Pro and MiMo 2.5 Pro outperform GLM-5.2 in their own cyber benchmarks. The export-controls prediction keeps resurfacing — one commenter is now putting a timeline on it, expecting Commerce to lean on HuggingFace and OpenRouter within months.
Base44 trains its own model, against the warnings. Base44, the vibe-coding platform Wix bought for $80M last year, is rolling out Base1, a custom LLM trained on tens of millions of user interactions. Founder Maor Shlomo argues specialization buys defensibility on latency and cost that frontier labs can’t match. A VC in the piece points to Harvey — the legal AI startup that quietly abandoned its own model training — as the cautionary tale Base44 should be reading.
Arena hits $100M ARR. The crowdsourced model leaderboard that started as a UC Berkeley project is now a real business, going from $30M to $100M annualized in eight months on the back of its commercial “AI Evaluations” product. It’s competing for post-training optimization spend against Scale AI and Mercor. The CEO admits much of the AI world still thinks of Arena as an open-source project rather than a vendor.
Cursor and OpenAI both lean into the agent-supervisor model. Cursor launched a mobile app for kicking off and managing coding agents from your phone, following similar moves from Anthropic and OpenAI. Anthropic’s Boris Cherny says most of his coding now happens on his phone, which tells you where the workflow is heading. Separately, OpenAI is teasing a physical Codex device built with macro-pad maker Work Louder for a July 15 launch — distinct from the Jony Ive project, and apparently just a fancy shortcut keypad.
That’s the briefing. The Meta story is the one to watch — if those internal documents get subpoenaed, “AI safety benchmarking” is going to need a much better legal definition than it currently has.