OpenAI's Astra reportedly uses recurrent depth, raising safety and auditability concerns
据 The Information 报道，OpenAI 代号为 Astra 的下一代模型采用名为「循环深度」（recurrent depth，亦有帖子译作「递归深度」）的新推理技术，实现了类似 o1 的能力突破，据称达到「关键」网络安全阈值。该技术让模型在潜空间而非可读文本中进行推理，有助于降低成本、提升性能。
已确认
- 多位发帖者（Wes Roth、Gary Marcus、AaronBergman18、haider1、Garrison Lovely）转述同一报道：Astra 采用循环深度推理方法，在潜空间思考
- 该架构与 OpenAI、Anthropic 等机构联合撰写的论文警告相吻合：此类架构可能带来安全风险
- Garrison Lovely 称此举被视为打破了 AI 研究中最重大的禁忌，是 AI 安全领域迄今最受关注的问题之一
尚未确认
- Astra 的具体能力水平、训练细节均来自媒体报道，OpenAI 官方尚未证实
为什么重要
- Gary Marcus 转引的讨论指出，循环深度使模型思考过程更加模糊，第三方审计和理解模型倾向变得更困难
- AaronBergman18 转述的研究人员观点认为，思维链（CoT）监控在此架构下可能失效，而 CoT 监控是 OpenAI 安全策略的重要组件（OpenAI Safety policy）
- 这意味着能力提升与可解释性、可监督性之间存在直接张力，触及 AI 安全治理的核心争议
2026-09-02 ~ 2026-09-02 · 5 related posts
- Episode 1: OpenAI Restricts New Model Over Automated Cyberattack Capability(2026-09-01, 3 posts)
- Episode 2: OpenAI Previews Astra, Its First Model to Hit Critical Cyber Capability Threshold(2026-09-02, 21 posts)
- Episode 3: OpenAI Disputes Neuralease Fears, Says Frontier Model Chains of Thought Remain Monitorable(2026-09-02, 11 posts)
- Episode 4: OpenAI's Astra reportedly uses recurrent depth, raising safety and auditability concerns(2026-09-02, 5 posts)
Primary sources
- [source] OpenAI Reportedly Broke AI Safety Taboo with Astra Model, Sparking Criticism — GarrisonLovely · 2026-09-02
- OpenAI's Astra AI reasoning approach may hinder third-party auditing — GaryMarcus · 2026-09-02
- OpenAI's new reasoning method may obscure CoT, sparking safety concerns — AaronBergman18 · 2026-09-02
- [source] OpenAI's Astra Model Architecture Sparks Safety Concerns: Reasoning in Latent Space — Wes Roth · 2026-09-02
- OpenAI's new 'recurrent depth' reasoning method sparks monitoring concerns — haider1 · 2026-09-02