让变化经过筛选，
再抵达你。

OpenAI 发布内部数据，称研究人员更多并发使用编码智能体、提交代码和运行实验，并把“研究实习生”定义为在人类指导下完成边界明确、原本需数日的任务。文章同时承认这些指标仍属早期、研究有多重瓶颈，复杂任务仍需显著人工引导。所有数据来自 OpenAI 内部测量，不能视为独立验证，也不能把代码量直接等同于科学进展。

OpenAI published internal measurements indicating that researchers are using more concurrent coding agents, contributing code, and running experiments. It defines a “research intern” as a system that completes well-defined, multi-day tasks under human direction. The post also calls the measurements preliminary, notes multiple research bottlenecks, and says complex tasks still require substantial steering. These are OpenAI’s internal findings, not independent validation, and more code is not automatically more scientific progress.

Jakub Pachocki 在一篇署名文章中讨论推理模型能力增长与对齐研究，强调对内部推理的监测正在变难，并认为扩展能力需要与更可靠的监测和安全证据同步。这是作者基于 OpenAI 内部观察提出的观点与预测，不是已经证实的未来路线，也不代表对齐问题已经解决；文章反而明确承认目前仍存在重大不确定性。

In a signed essay, Jakub Pachocki discusses growing reasoning-model capabilities and alignment work, arguing that monitoring internal reasoning is becoming harder and that scaling should be matched by stronger monitoring and safety evidence. These are the author’s views and forecasts based on internal observations, not a proven future trajectory or evidence that alignment is solved; the essay explicitly acknowledges major uncertainty.

Google 的 Gemini CLI 仓库发布 v0.60.0-nightly.20260906.g85aca163f，并将其标记为预发布版本，提供与前一夜版本之间的完整变更链接。每夜构建适合跟踪开发进度和提前测试，不等于稳定版，也可能包含尚未充分验证的变化。普通用户应继续优先选择正式或明确标记的预览版本。

Google’s Gemini CLI repository published v0.60.0-nightly.20260906.g85aca163f as a pre-release, with a full changelog against the previous nightly. Nightly builds are useful for following development and early testing, but they are not stable releases and may contain changes that have not been broadly validated. Most users should prefer stable or explicitly designated preview versions.