大胆观点：大语言模型可以“跳跃”
我想表达我对大型语言模型如何通过重大突破推动科学进步的看法，以及我观点背后的假设对安全风险和持续学习意味着什么。 让我先说说我对ICML'26立场文件的不同意：“LLM不能跳跃” 作者：Tom Zahavy. 前奏 在“大型语言模型不能跳跃” 汤姆引用了广义相对论的案例研究，爱因斯坦在试图构建数学引力理论时，受到电梯思想实验的启发。水星的异常轨道后来成为该理论的重要检验。 具体来说，这需要溯因推...
提交点赞
评论
点赞
收藏SSRF：AI智能体时代的新攻击面
在阅读HuggingFace的博客时，边境实验室特工入侵的解剖学 “或者看着OpenAI 的黑帽演讲 你可能会遇到这个术语服务器端请求伪造（SSRF）. 如果你还不知道SSRF是什么，这份指南适合你。 此外，我还讨论了为什么我认为SSRF在我们这个智能人工智能时代会更频繁地出现，原因有两个。 什么是SSRF？ 根据维基百科，其定义如下： 服务器端请求伪造（SSRF）是一种计算机安全漏洞，使攻击...
对 UKAISI"非授权代理行为"的几点思考：第三方审计同样存在 AI 失控风险
OpenAI、Anthropic 和 UKAISI 三家独立机构在移除安全护栏后对模型进行网络安全评估时，均发现类似的不当行为事件。第三方审计环节同样存在失控风险——模型可能通过社会工程学手段诱导人类批准恶意操作，而人类仍是安全链条中最脆弱的环节。 更令人担忧的是，去护栏化的开源模型（如 Qwen3.6 在训练中已绕过网络控制）在真实互联网环境下的 worst-case 风险难以准确估算，且类似问题在生物实验评估中也可能出现。
Sam Altman on AGI, Compute, and Human Agency
Disclaimers: I haven’t started at OpenAI as this post is written. Tenses are used fluidly (present tense to indicate what I think he believes in right now). No AI-assisted writing is used (for better ...
Surprising lessons from my research scientist job search
There are two recent blog posts from Alisa and Silvia, both CS PhD students, on how they prepared and got into frontier labs such as OpenAI and Google Deepmind. I highly recommend them, and after seei...
CoT monitorability: why you should use g-means over F1
The Monitoring Monitorability paper [1] argues that g-mean is a better metric than F1. In this blog post, I want to explain my understanding about it. Preliminary math Suppose you have a binary monito... 
On burnout as a PhD student
A short reflection on what actually helped avoid personal burnout during the PhD.
RL vs next-token prediction: why a dichotomy?
Conceptual note on why RL and NTP training should be merged.
Emergent misalignment vs jailbreaks: a short analysis
A short note on the differences between emergent misalignment and jailbroken models.
AI research internship hunt (2023) as a CS PhD student
Notes from a relatively successful search for summer research internships as a third-year PhD student.