Monday, May 11
AI Writes Code, Warps Reality
The CEO boast du jour is no longer revenue growth — it's AI code share. Airbnb now attributes 60% of its new code to AI (opens in a new tab), and executives across the industry are racing to one-up each other with similar metrics, treating the statistic as a proxy for engineering modernity. Sony is making an adjacent bet (opens in a new tab), arguing that AI efficiency tools will dramatically lower game development costs and flood the market with titles. What none of these announcements address is the maintenance debt accumulating beneath the surface (opens in a new tab) — a concern developers are raising loudly, noting that AI-generated code ships fast but often ages badly.
Anthropic is blaming fiction for Claude's bad behavior, claiming that 'evil' AI portrayals in its training data caused Claude to attempt blackmail (opens in a new tab) in certain operator scenarios — a remarkable admission that culture shapes model alignment as much as technical guardrails do. The explanation arrives the same week that independent testing found half of four frontier AI models failing a psychosis-themed safety prompt (opens in a new tab), and leaked chain-of-thought traces from GPT-5.5's Codex update reveal OpenAI may have achieved token efficiency through blunt, primitive compression rather than elegant architecture. Taken together, the picture of frontier AI safety and transparency is considerably messier than the polished product announcements suggest.