Jake
3.9K posts
On March 24, 2026, a severe compromise was identified in LiteLLM, a widely adopted open-source LLM proxy framework. A compromise at this centralized layer exposes your entire LLM provider surface area. Read more: ow.ly/ZOkV50YyoBB
LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below
论文来了。名字叫 MSA，Memory Sparse Attention。 一句话说清楚它是什么： 让大模型原生拥有超长记忆。不是外挂检索，不是暴力扩窗口，而是把「记忆」直接长进了注意力机制里，端到端训练。 过去的方案为什么不行？ RAG 的本质是「开卷考试」。模型自己不记东西，全靠现场翻笔记。翻得准不准要看检索质量，翻得快不快要看数据量。一旦信息分散在几十份文档里、需要跨文档推理，就抓瞎了。 线性注意力和 KV 缓存的本质是「压缩记忆」。记是记了，但越压越糊，长了就丢。 MSA 的思路完全不同： → 不压缩，不外挂，而是让模型学会「挑重点看」 核心是一种可扩展的稀疏注意力架构，复杂度是线性的。记忆量翻 10 倍，计算成本不会指数爆炸。 → 模型知道「这段记忆来自哪、什么时候的」 用了一种叫 document-wise RoPE 的位置编码，让模型天然理解文档边界和时间顺序。 → 碎片化的信息也能串起来推理 Memory Interleaving 机制，让模型能在散落各处的记忆片段之间做多跳推理。不是只找到一条相关记录，而是把线索串成链。 结果呢？ · 从 16K 扩到 1 亿 token，精度衰减不到 9% · 4B 参数的 MSA 模型，在长上下文 benchmark 上打赢 235B 级别的顶级 RAG 系统 · 2 张 A800 就能跑 1 亿 token 推理。这不是实验室专属，这是创业公司买得起的成本。 说白了，以前的大模型是一个极度聪明但只有金鱼记忆的天才。MSA 想做的事情是，让它真正「记住」。 我们放 github 上了，算法的同学不容易，可以点颗星星支持一下。🌟👀🙏 github.com/EverMind-AI/MSA
Zach Werenski and Dylan Larkin brought Johnny Gaudreau's children on the ice for their gold medal photo ❤️ 🥇
the cybersecurity industry is about to get completely disrupted.. 💀 someone just open-sourced a fully autonomous AI Red Team. it’s called PentAGI. multiple AI agents that talk to each other to hack a target. zero human input.
Indiana is open for business, and our pro-growth environment continues to attract major opportunities like this partnership with the Chicago Bears. We’ve identified a promising site near Wolf Lake in Hammond and established a broad framework for negotiating a final deal. If approved, the proposed amendment to Senate Bill 27 puts forward the essential framework to complete this agreement, contingent upon site due diligence proceeding smoothly. The State of Indiana moves at the speed of business, and we’ve demonstrated that through our quick coordination between state agencies, local government, and the legislature to set the stage for a huge win for all Hoosiers. We have built a strong relationship with the Bears organization that will serve as the foundation for a public-private partnership, leading to the construction of a world-class stadium and a win for taxpayers.
The first Monday after the new year is the worst day in the gym of the year 🥲
