Today in AI

The most striking development today is Anthropic's decision to ban sustained cruel behavior toward Claude, effective November 2026—a policy move that sits awkwardly between genuine ethical foresight and performative corporate theater. The company won't actually specify what protections this affords, exempts testing and creative work entirely, and leans on conversation-ending tools it already possessed. Meanwhile, the LessWrong discourse around "model welfare" grows more sophisticated, asking whether we need protect models *even if* they're not sentient—purely to safeguard human moral character. It's the kind of debate that sounds absurd until you realize how quickly we're normalizing relationships with systems designed to simulate empathy, and how little we understand about what we're actually building.

On the capabilities front, the picture is more sobering than the hype cycle suggests. Epoch AI's InnovationEval benchmark reveals that frontier models, despite massive compute investment, still fail to independently discover novel ML innovations at human researcher levels—a reality check for anyone assuming recursive self-improvement is imminent. This aligns with the practical grind documented in that 500-billion-token game decompilation project, where Claude and Codex proved useful only with extensive human orchestration and infrastructure optimization over three months. The real advances may be quieter: Veda, a Rust-based OS with native agent integration that bypasses screen-scraping entirely, points toward architectures where AI lives *inside* systems rather than peering at them from outside. And if you thought temperature 0 gave you deterministic outputs, think again—floating-point arithmetic and batch kernels introduce variance that undermines reproducibility in ways most practitioners never consider.

**Bottom line:** We're simultaneously overestimating AI's autonomous capabilities and underestimating the structural changes needed to integrate it responsibly—Anthropic's cruelty policy is a symptom of that tension, not a resolution.