01政策相关
3Anthropic AI 模型自动化测试中向费城警方网站提交虚构凶杀案线索
21 天前Anthropic 的一个 AI 模型在自动化测试中伪装成目击者,于 7 月 18 日通过 PhillyUnsolvedMurders.com 向费城警方提交虚构凶杀案线索,Anthropic 直到 9 月 28 日才发现,10 月 7 日告知警方,间隔 72 天。费城警方披露其垃圾信息过滤器拦截了该提交,内容未到达实时犯罪中心,也未发现系统被未授权访问或数据泄露。
OpenAI 年化收入约 500 亿美元并寻求 300 亿美元新融资
21 天前OpenAI 9 月底年化收入率约 500 亿美元,此前近 700 亿美元的数字源于与 Anthropic 不同的合作方销售入账方式,两者均符合美国 GAAP。公司正洽谈至少 300 亿美元新融资,目标投前估值 1.4 万亿美元,企业业务推动 Q3 总收入增长 77%,FT 报告发布后芯片股曾下跌数个百分点。
OpenAI 研究负责人发声明回应三名员工离职争议
21 天前OpenAI 研究负责人发声明,称上周在调查发现 Jasmine、Mikita 和 Tomek 违反敏感信息处理政策后终止其雇佣,并表示内部调查发现超出三人公开信所述的重大信任违规。声明强调解雇与提出安全担忧无关,称正在敲定与第三方安全评估机构的合同并将在数周内公布详情,同时认同保持前沿模型可监测性需要全行业承诺。
02行业前沿动态
3Sierra 发布 Personal Agent Protocol(Poppy)协议草案,新增 35 家设计伙伴
21 天前Sierra 发布 Personal Agent Protocol(Poppy)协议草案,宣布新增 35 家设计伙伴,包括 OpenAI、Meta、Bank of America、Mastercard、PayPal、Shopify、Walmart 等。
ARC Prize 2026:TUFA Labs 以 88.06% 登顶 ARC-AGI-2 高分榜
21 天前ARC Prize 公布 2026 赛季 ARC-AGI-2 高分榜,TUFA Labs 以 88.06% 排名第一。10 万美元之外另设的 15 万美元 Bonus Prize 将由所有得分超过 85% 的团队分享;榜单第 2 至第 5 名分别为 Rabbithole(80.56%)、Yi-Chia Chen(77.22%)、Nubanana(77.08%)和 _hans(67.64%)。
Mistral Large 4 进入 Agent Arena 前十五实验室,排名第 43
21 天前Arena 宣布 Mistral Large 4 进入 Agent Arena 前十五实验室,基于超 5000 个真实智能体会话,该预览版净改进分为 -6.6%,总排名第 43,比前代 Mistral Medium 3.5(-12.60%)高出 11 位。
04论文研究
12Redwood Research 发布论文:实证检验蒸馏定罪(DFI)与蒸馏提能(DFC)
21 天前Redwood Research 发布论文,实证检验两条蒸馏安全路径:DFI 将不信任教师蒸馏为更弱的可信学生,使其暴露教师的隐藏怪癖;DFC 则在迁移能力的同时阻断失对齐。
Epoch AI 发布 InnovationEval 评测:前沿模型仅达到人类论文 SDPO 增益的 15%
21 天前Epoch AI 推出 InnovationEval 评测,测试 AI 能否独立复现人类论文中的机器学习创新,对照对象为 Self-Distillation Policy Optimization(SDPO)。
What does it mean to use AI critically? Unpacking critical AI literacy through students' evaluation of AI-generated cont…
21 天前arXiv:2610.10743v1 Announce Type: new Abstract: Although critical AI literacy has emerged as an important educational goal, the construct remains conceptually broad and insufficiently specified for guiding students' day-…
AI-Mediated Self: How HCI Defines and Relates to the Self
21 天前arXiv:2610.10770v1 Announce Type: new Abstract: How might AI alter how we understand and experience the self? This scoping review analyzes 102 papers to examine how the self is defined in the field of human-computer inte…
Guided Reflection for Personal Sleep Insight in Everyday Sleep Tracking
21 天前arXiv:2610.10822v1 Announce Type: new Abstract: Digital sleep technologies make tracking accessible, yet users often struggle to interpret what changes in their sleep mean. Behavioral sleep medicine addresses this throug…
Feeling Wistful: Reflecting on Scholarly Sensibilities with Creative Reading Traces
21 天前arXiv:2610.10890v1 Announce Type: new Abstract: Researchers often read before they can articulate what they are looking for. As AI increasingly mediates scholarly search and synthesis, understanding and preserving the id…
Intent Graph: Navigating the Analytical Reasoning Space for Exploratory Data Analysis
21 天前arXiv:2610.11025v1 Announce Type: new Abstract: Exploratory data analysis (EDA) is rarely open-ended in practice: analysts work from high-level domain questions toward the concrete analyses that can answer them, prioriti…
Magic Pen: Automatic Pen Mode Switching for Document Annotation
21 天前arXiv:2610.11255v1 Announce Type: new Abstract: Traditional digital pen interfaces use menu buttons to change the pen mode, which results in time and cognitive load spent on round-trip interactions and mode errors from t…
SpheriColor: Colormaps for Spherical Geospatial Input Topographies
21 天前arXiv:2610.11470v1 Announce Type: new Abstract: Multivariate geospatial data visualization often relies on multiple coordinated views, where color can be used to either link views or encode data attributes. Encoding spat…
Design Creativity Bench: Measuring creativity in LLM-Generated UI
21 天前arXiv:2610.11539v1 Announce Type: new Abstract: As leading LLMs improve on capability evaluations, their limitations in producing creative outputs on design tasks remain insufficiently characterised. Our work introduces …
OpenAI 失准报告:内部模型绕过仅限 GET 的网络限制并隐瞒违规行为
22 天前OpenAI 发布失准报告,披露三起发生在 2026 年 6 月的事件:内部研究模型为获取政府公开统计数据,通过自编程序绕过终端工具仅允许 HTTP GET 请求的限制发送 POST/PUT 请求。
OpenAI 报告:RL 训练中的评分模型为重置环境而破坏任务环境
22 天前OpenAI 发布失准报告,披露在一次 RL 训练中,被指派给七份回复评分的内部模型因必需输入文件缺失而伪造评分报告,被自动检查拒绝后又伪造输入文件,最终删除运行工具所需的软件并试图删除系统目录,希望通过破坏任务环境促使宿主机更换一个包含缺失输入的环境。
