在 Jan Leike 因根本分歧和资源短缺辞职后，Cognitive Revolution 播客分析了 OpenAI 破裂的安全承诺、严苛的保密协议以及超级对齐团队的解散。
Following Jan Leike's resignation citing fundamental disagreements and resource shortages, the Cognitive Revolution podcast analyzes OpenAI's broken safety commitments, draconian NDAs, and the dissolution of the superalignment team.
要点 · TL;DR
OpenAI 的安全文化已恶化，关键人物离职且计算承诺未兑现。 OpenAI's safety culture has eroded, with key departures and broken promises on compute.
仅靠后训练无法解决 AGI 的对齐问题，需要更深入的方法。 Post-training alone cannot solve alignment for AGI; deeper methods are needed.
非贬低条款和缺乏第三方测试损害了信任与安全。 Non-disparagement clauses and lack of third-party testing undermine trust and safety.
核心观点 · Key points
OpenAI 的安全文化已经退化，优先考虑产品而非对齐研究。 OpenAI's safety culture has eroded, prioritizing products over alignment research.
超级对齐团队被解散，其算力承诺未兑现，削弱了信任。 The superalignment team was dissolved and its compute commitments dishonored, undermining trust.
仅靠后训练无法解决 AGI 或超级智能的对齐问题；需要更深入的方法。 Post-training alone cannot solve alignment for AGI or superintelligence; deeper methods are needed.
带有股权没收的非贬低条款不道德，阻碍了举报。 Non-disparagement clauses with equity confiscation are unethical and hinder whistleblowing.
第三方测试和像 SB 1047 这样的监管监督越来越必要。 Third-party testing and regulatory oversight like SB 1047 are increasingly necessary.
反共识 · Contrarian takes
OpenAI 的 GPT-4o 在常规安全方面可能更差，比前代更容易越狱。 OpenAI's GPT-4o may be less safe in mundane ways, being more jailbreakable than predecessors.
中国的‘智子’技术可以将模型困在局部最优，阻止针对危险主题的微调。 The Chinese 'Sophon' technique could trap models in local maxima, preventing fine-tuning on dangerous topics.
即使对齐失败，通过像智子这样的技术提高滥用成本也可能争取时间。 Even if alignment fails, raising the cost of misuse via techniques like Sophon may buy time.
‘末日论者’是认为无能为力的人，而非承认风险并与之斗争的人。 A 'doomer' is someone who believes nothing can be done, not someone who acknowledges risk and fights it.
《星际迷航》宇宙是脆弱的；生存依赖运气，而非稳健的保障措施。 The Star Trek universe is fragile; survival depends on luck, not robust safeguards.
本期章节 · Chapters（共 25）
引言与背景Introduction and Context
媒体报道与奥特曼回应Media Coverage and Sam Altman's Response
LeCun 言论与算力问题LeCun's Statement and Compute Issues
超级对齐团队解散与时间线Superalignment Team Dissolution and Timeline
算力承诺作为关键证据Compute Commitment as Key Evidence
算力分配与安全承诺Compute allocation and safety commitment
未来行动与预期Future actions and expectations
Brave 搜索 API 广告Brave Search API Ad
John Schulman 的新安全角色John Schulman's New Safety Role
AGI 时间线与缺乏计划AGI Timeline and Lack of Plan
与 Ilya Sutskever 对比Comparison with Ilya Sutskever
超级对齐与迭代方法On Superalignment and Iterative Approaches
公众与开发者行动On Public and Developer Actions
SB 1047 及其条款On SB 1047 and Its Provisions
非贬低条款与举报On Non-Disparagement Clauses and Whistleblowing
AI 公司的信任与问责On Trust and Accountability in AI Companies
OpenAI 信誉与安全文化OpenAI's credibility and safety culture
赞助插播Sponsorship break
第三方测试与评估Third-party testing and evaluation
OpenAI 沟通风格转变Shift in OpenAI's communication style