AI 排名重新洗牌：分数、成本与真实任务，究竟该信哪一个？

Carbon's reported benchmark lead and rapid iteration are promising, but restricted access prevents most users from testing those gains in their own workflows.

Lower-than-expected capital spending can reflect physical bottlenecks rather than weak demand when cloud revenue is accelerating and investment forecasts remain intact.

The investigation described in the episode extends beyond one transaction to questions about distribution control, investment relationships, and customer pressure.

Benchmark redesign is necessary as tasks evolve, but a ranking change caused by a new measuring system is not itself a change in the model.

Private datasets and agentic evaluations can make a benchmark harder to optimize for superficially, while tests of mergeable code ask whether an output is actually usable.

Allocating one quarter of the index to science is a substantive design choice that may not match the everyday coding needs of many users.

The wider open-weight gap in v5 describes performance under the redesigned index, not a universal verdict on every open model and every task.

A model can fall in the aggregate ranking while retaining strong coding results, so the drop alone does not establish benchmark gaming.

Model selection should compare task-level cost and reasoning effort alongside capability, because a maximum-effort leaderboard may not represent normal usage.

The episode describes the US-China capability gap as a narrowing time lag of about 3.3 months on a historical progress scale, rather than predicting a future catch-up date.