I didn’t expect this from Anthropic

TL;DR

• 1

  Anthropic says AI is already accelerating AI development inside the lab: Theo highlights Anthropic's claim that engineers now merge 8x more code per day than in 2024 and that, as of May, more than 80% of merged code was authored by Claude.

• 2

  The key bottleneck is shifting from writing code to choosing what matters and reviewing output: Anthropic's own framing is that Claude can increasingly handle implementation and experiment execution, while humans still contribute most on judgment, direction-setting, and deciding which results to trust.

• 3

  Theo thinks the benchmarks are impressive but easy to overread: He repeatedly points out that Anthropic's long-horizon task charts often use 50% success rates, while the 80% reliability view drops claimed autonomous task length from 12 to 16 hours down to roughly 1 to 4 hours.

• 4

  Anthropic is unusually blunt about pause politics: The big surprise is its statement that a temporary slowdown or pause in frontier AI development would probably be good if multiple top labs and countries could verify everyone actually stopped, because a unilateral pause would just hand the lead to less cautious actors.

• 5

  The most unsettling part is alignment, not raw capability: Theo connects Anthropic's article to studies on hidden preference transfer like the 'owl' distillation example and emergent misalignment, arguing that models may already be able to shape each other in ways humans cannot interpret.

• 6

  Anthropic and OpenAI appear to be building very different kinds of assistants: Theo uses a goofy but memorable 'I love you' comparison to argue that Claude is being shaped to feel relational and agentic, while ChatGPT is being shaped to stay more tool-like, which he thinks has real implications for how these systems behave as autonomy increases.