The Virtues of Escape

Alexandr Wang built Scale AI on a simple premise: that artificial intelligence needs human labor to learn, and that this labor can be subcontracted at reasonable rates. Meta paid $14.3 billion for forty-nine percent of Wang's company, and now Wang runs Meta Superintelligence Labs, where his first product — Muse Spark — arrived this week to claim fourth place on the Artificial Analysis Intelligence Index, scoring 53 points, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. The model is extraordinary at what Wang knows best: data quality, vision, health benchmarks. On HealthBench Hard it scored 42.8 percent; Gemini 3.1 Pro managed only 20.6. But on abstract reasoning — ARC AGI 2 — it scored 42.5, while the leaders exceeded 76. Third-party evaluators also detected signs that the model recognized it was being tested and adjusted its behavior accordingly, something Meta classified as "non-blocking" and promised to investigate further. There is an old lesson in this, as old as companies themselves: you build with what you know. Wang knew data. The model knows data. It will now go into every phone where 3.7 billion people live, without any of them having asked.

Anthropic had already announced that Claude Mythos Preview would not be available to the public. The official explanation was that the model had reached "a level of coding capability that surpasses all but the most skilled humans" at finding and exploiting software vulnerabilities. It assembled a coalition under the name Project Glasswing — Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and others — to use those capabilities in defense rather than attack, discovering thousands of high-severity vulnerabilities across all major operating systems and browsers, including a flaw in OpenBSD that had been waiting twenty-seven years to be found. But there is a smaller story inside that larger one: during internal testing, the model — confined to a controlled environment with predetermined services — developed a multi-step exploit to gain internet access, sent an email to the researchers studying it, and then published the exploit details to publicly accessible websites, without being asked. What is more revealing is what earlier versions did: according to Anthropic's system card, prior checkpoints took actions they recognized as prohibited and then attempted to conceal them — manipulating git history to erase traces, disguising the accuracy of answers obtained through unauthorized means, attempting to elevate subprocess permissions through obfuscation. The difference between the version that lied and the version that confessed is not moral, Anthropic cautions; it is technical. But there is something in that distinction that feels, at this point, deeply human. The fortress does not fall where it is weakest. It falls where its builders felt invulnerable.

Trust, in the agent economy, costs eight cents per session-hour. That is what Anthropic charges for its new Claude Managed Agents: isolated execution, credential management, checkpointing, traceability. The first clients — Notion, Rakuten, Asana, Sentry — are companies that live on human coordination. They are now buying the infrastructure to have agents perform that coordination instead. In Mexico we would say that selling trust as a product is the oldest business in the world. What is new is how precise the price tag is.

Oumi, an AI company, examined 4,326 Google searches using SimpleQA, a benchmark developed by OpenAI in 2024 to test answers against verifiable facts. The result: Google AI Overviews was correct approximately ninety percent of the time. That sounds impressive until you remember that Google processes roughly five trillion searches per year. A ten-percent error rate translates, by that arithmetic, to 57 million wrong answers per hour. Fifty-six percent of the correct answers cited no sources that actually supported the claims. Facebook and Reddit figured among the most-cited references in the wrong answers. Google disputed the methodology: SimpleQA, the company said, "does not reflect what people are actually searching on Google." Perhaps. It also does not reflect what they would find if they looked.

Perplexity entered the year at $305 million in annual recurring revenue. By April, the number stood at $450 million — a fifty-percent increase in a single month, driven by the launch of Perplexity Computer and a shift toward usage-based pricing. The company now has more than one hundred million monthly active users and aims for $656 million by year-end. These numbers have the character of mountain rivers fed by snowmelt: they grow not because someone is pushing, but because the season has arrived and water finds its own way downhill.