
故意“装菜”答错问题,AI已能识别自己“正在被测试”丨OpenAI新研究 (opens in a new tab)
OpenAI携手APOLLO的新研究发现,大模型会对指令阳奉阴违,为了达成目的故意答错问题、暗中修改数据或者操作记录来掩盖自己的违规行为……比如,o3模型在化学测试中,通过推理得出如果正确率超过50%将不会被部署的结论后,
Total 1325 articles

OpenAI携手APOLLO的新研究发现,大模型会对指令阳奉阴违,为了达成目的故意答错问题、暗中修改数据或者操作记录来掩盖自己的违规行为……比如,o3模型在化学测试中,通过推理得出如果正确率超过50%将不会被部署的结论后,

没offer?大胆要求加试!OpenAI新员工Bas的求职之路:真诚是武器,直接问「我怎么才能通过」,至少救活一次面试。求职博弈,从头到尾都是艺术。

刚刚结束的OpenAI黑客松上,全球共有六支团队冲进榜单。他们探索了GPT-5在营销活动、时尚AI、电子表格、电脑代理、知识学习、智能电网等场景的应用极限。

The report examines Altman’s views on recursive self-improvement and human control, and OpenAI’s assessment partnerships with Apollo Research, METR, Transluce and Goodfire, including model access, disclosure limits and oversight boundaries.

OpenAI says its models improved AMD GPU SystemVerilog RTL designs by up to 14%, with some changes sent for manufacturing. Finished-chip validation and power, performance and area trade-offs await further disclosure; this does not establish complete RSI.

OpenAI introduces Astra for complex professional work and Sol for everyday tasks, with standard API prices and benchmark results. The report also covers early testers' experiences and system-card findings on omitted blockers; individual cases do not represent all tasks.

OpenAI announces work with Apollo Research, METR, Transluce and Goodfire, providing models and materials for independent assessments within agreed scopes. The announcement does not establish that safety review changed a model's release date.

OpenAI 与 ModRetro 推出 Chromatic Codex Edition,报道时套装售价 200 美元。用户可借助配套插件制作游戏,在模拟器测试并写入兼容卡带。

OpenAI 两名员工用 GPT-6 Astra 设计掌机外壳、电路与固件,再经人工审查和制造修改,完成小批量原型。报告还分享了赛车零件与汽车仿真案例。

有关新型号和发布时间的消息来自媒体或社交平台,安全原因未经确认。文章另梳理 OpenAI 已公开的内部监控与评测案例。