03.如何量化OpenAI o1的骗钱能力?被科技媒体集体忽视的官方白皮书给了答案 episode artwork

EPISODE · Sep 28, 2024 · 59 MIN

03.如何量化OpenAI o1的骗钱能力?被科技媒体集体忽视的官方白皮书给了答案

from 技术不无聊

OpenAI o1推出两周有余,市面上大多数讨论聚焦于其推理能力的优越性,也有一小部分声音质疑模型能力越来越强是否意味着更大的威胁?大多数质疑停留在无限推理之后的科幻化灾难场景,然而事实上,OpenAI早在模型发布之时,也同步释放了针对模型各类潜在威胁做出的测试白皮书“o1 System Card”,这期播客我们为大家深度解读这份白皮书,在缓和担忧的同时,也从客观的角度理解模型的能力的上限与不足。这不是一期鼓吹OpenAI如何做的好的内容,我们尽可能将未被关注到但十分有价值的分析呈现出来,试图打破大模型领域信息的藩篱。 时间线1:37 System Card白皮书具体是什么,它的受众是谁?4:19 推理能力的加入使得模型更聪明,更好的服务意味着更好的欺骗6:17 围绕结果的评估维度:不允许内容、越狱、幻觉、偏见13:17 思维链过程性评估18:19 外部红队攻击:评估AI抵抗越狱的能力21:54 外部机构Apollo research专注“AI是否会欺骗人类”研究28:01 外部机构METR对模型自动化能力的测试30:19 OpenAI推出Preparedness Framework,测试模型的信息安全、CBRN、劝说能力Persuasion和模型失控Model Autonomy53:04 MMLU(Massive Multitask Language Understanding)在多语言中测试大模型的准确性 提及COT(Chain of Thought)| Ilya Sutskever Super Allignment | 夺旗比赛Catch the Flag | Theory of Mind | CBRN(Chemical, Biology, Radiology, Nuclear) | 模型自制Model Autonomy | BBQ Bias Benchmark for QA | European Nucleotide Archive | MMLU(Massive Multitask Language Understanding) 相关引用assets.ctfassets.net 制作剪辑大奎在小宇宙查看该单集文稿

Episode metadata supplied by the publisher feed · Published Sep 28, 2024

Embed this episode

Ready to play

03.如何量化OpenAI o1的骗钱能力?被科技媒体集体忽视的官方白皮书给了答案

0:00 59:34

No transcript for this episode yet

We transcribe on demand. Request one and we'll notify you when it's ready — usually under 10 minutes.

No similar episodes found.

Frequently Asked Questions

How long is this episode of 技术不无聊?

This episode is 59 minutes long.

When was this 技术不无聊 episode published?

This episode was published on September 28, 2024.

Can I download this 技术不无聊 episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.
URL copied to clipboard!