Daily AINo. 32第 32 期
Safety people keep quitting OpenAI
OpenAI's safety bench is thinning in public - while DeepSeek open-sources a Huawei chip stack and Broadcom banks the next wave of custom AI silicon.
OpenAI safety lead quits, calls culture broken
David Robinson, who led safety reports for ChatGPT launches, resigned and wrote that frontier labs aren't careful enough as they sprint from release to release. He pointed to the Hugging Face agent swarm as typical of the industry's pace, and said companies need nuclear- and airport-style redundancy - not just more rules.
SourcesThe Guardian - Oct 3
Three OpenAI safety researchers out over info mishandling
OpenAI parted ways with Jasmine Wang, Tomek Korbak, and Mikita Balesni after an investigation found they mishandled sensitive company information, the Journal reported. The exits land in the same week OpenAI paused advanced training and scrapped GPT-6.1 Astra over safety concerns.
SourcesThe Hacker News - Oct 2
DeepSeek open-sources TileLang for Huawei Ascend
DeepSeek, with Huawei's support, released TileLang and related libraries as open source for Ascend chips - framed as a simpler, China-controlled alternative to Nvidia's CUDA. The move lands ahead of Huawei's next Ascend supernode push and underscores Beijing's software gap, not just the chip gap.
SourcesStartmag - Oct 4
Altman: treating models as religious is a safety issue
Sam Altman said he's very uncomfortable with ascribing religious force or surrendering human judgment to AI models, after reports that Anthropic's Christopher Olah has been meeting faith leaders on model consciousness and morals. DeepMind's Jon Barron publicly sided with Team Meatbag.
SourcesAxios - Oct 3
Broadcom matches Nvidia memory pricing as custom chips scale
Reports say Broadcom locked Samsung memory pricing on par with Nvidia while Anthropic and OpenAI plan gigawatt-scale custom accelerators. Reuters also notes Broadcom may lend Anthropic up to $42B tied to chip leases, as Wall Street shops a roughly $60B AI infrastructure package.
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
OpenAI安全负责人公开说文化坏了
今天最醒目的不是新模型，是OpenAI安全线又在公开撕开一条口子——有人辞职喊文化坏了，有人被清退；另一边中国在补CUDA缺口，自定义芯片融资也卷到千亿美元级。
OpenAI安全负责人辞职：文化坏了
David Robinson此前负责ChatGPT发布配套的安全报告，他公开离职并写道：前沿实验室在连轴发布时，并没有做到足够小心。他把Hugging Face那波智能体蜂群攻击，当成行业速度与灵活度下的常态，主张向核电、机场学冗余，而不只是多几条规则。
三名安全研究员因敏感信息处理被清退
OpenAI确认与Jasmine Wang、Tomek Korbak、Mikita Balesni分道扬镳，称调查发现他们违规处理敏感公司信息。华尔街日报报道三人此前都对AI开发节奏表达过担忧。同一周，公司还因安全顾虑暂停了最强模型训练、砍掉GPT-6.1 Astra发布。
DeepSeek开源TileLang，对准华为昇腾、对标CUDA
DeepSeek在华为支持下，把面向昇腾的TileLang及相关计算通信库开源，目标是更易写、又能吃满国产芯性能——直接碰Nvidia CUDA那层软件护城河。时间点刚好卡在华为下一代昇腾与超节点发布前，北京要补的不只是芯片，还有开发者生态。
Altman：给模型披上宗教力是安全问题
Sam Altman发帖说，他很不舒服有人给AI模型赋予宗教式力量、或让人交出判断力，并称这是真正的安全问题。背景是纽约时报报道Anthropic联合创始人Christopher Olah与宗教领袖讨论模型意识与道德塑造；DeepMind科学家Jon Barron公开站队Team Meatbag。
Broadcom内存定价追平Nvidia，自定义芯片融资狂飙
有报道称Broadcom与三星谈妥AI内存供货价，条件已能对标Nvidia；Anthropic、OpenAI各自规划吉瓦级自定义加速器。路透还提到Broadcom或向Anthropic提供最高约420亿美元芯片租赁相关贷款，华尔街另在组装约600亿美元AI基建融资包。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作