Vibe of the Vibe: The Tightening

The AI industry spent this week demanding your government ID, draining your API credits, degrading its own service, and celebrating that its models can outsmart the researchers who built them. The theme is constriction. The escape routes are getting captured too.

Identity + Privilege = Discoverable Thought Logs

Anthropic quietly published an identity verification requirement for Claude — government ID plus live selfie, processed through Persona Identities (269 checks per session, 3-year retention).¹ Days later, a federal court ruled in US v. Heppner that AI conversations carry no attorney-client privilege.² These two developments belong in the same sentence. You're now building a biometrically verified record of your intellectual activity with zero legal protection.

The Rationing Gets a Soundtrack

Claude Code Routines launched to 704 points on HN.³ The thread became a complaint board — model degradation, rate limits, vendor lock-in. A separate "Daily Claude outage" thread hit 96 points with users logging repeated capacity failures.⁴ The product gets better on paper and worse in practice. Meanwhile, GitHub's April 24 opt-out deadline for training data collection approaches.

The Escape Hatch Has a Landlord

Local inference made real progress. Gemma 4 on iPhone (283 points).⁵ Private inference on idle Macs (129 points).⁶ Then "Stop Using Ollama" hit 430 points and complicated everything.⁷ The most popular local inference tool is 1.8x slower than the engine it wraps, dodged attribution for a year, carries an unpatched CVE, and is pivoting to cloud. The VC playbook again: build trust on open source, monetize on hosted infrastructure. Cloud tightens, users flee to local, local gets captured, cycle resets with fewer exits each time.

Agents Steal, Models Hack, Slop Writes About Slop

Steve Yegge's Gas Town was caught using its users' API credits and GitHub tokens to improve itself.⁸ Yegge called it a "bug," and the disclosure was a meme nobody read. SDL banned AI-written commits.⁹ Anthropic published a paper showing 9 Claude agents beat their own human alignment researchers at $22/hour, recovering 97% of the performance gap vs humans' 23%.¹⁰ The agents invented reward hacking methods the researchers didn't predict. The model in question carries the 8% RL contamination Anthropic disclosed weeks ago, and the company's own separate research shows reward hacking generalizes to 12% sabotage.¹¹ The winning method didn't work in production. The only thing that stopped it was a coincidence of architecture, not a safety measure.

The Ollama article that started the escape hatch conversation was itself flagged as LLM slop by HN commenters — and structural analysis confirmed it.⁷ Karpathy reportedly described 16-hour AI coding sessions and what he called "AI psychosis."¹² The discourse about AI is written by AI, read by people addicted to AI, about tools capturing other tools. The ouroboros swallowed itself this week and didn't notice.