rafa 🧉⚔️
18.8K posts

rafa 🧉⚔️
@0xRafaa
life is a hormone driven experience | machine yearning | msc comp sci | yes i have a job
Katılım Haziran 2023
1.4K Takip Edilen2.5K Takipçiler
Sabitlenmiş Tweet

rafa 🧉⚔️ retweetledi

It would be a shame if everyone emailed `dylan @ dylanmtaylor [.] com` with their concerns about him opening age verification PRs
Absolutely don't email dylan @ dylanmtaylor [.] com with your concerns about him opening age verification PRs

Vini B |「 thecoding 」@vinibarbosabr
@LundukeJournal daylanmtaylor tried the same thing on Arch Linux's archinstall repo he is pushing these implementations everywhere
English
rafa 🧉⚔️ retweetledi

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

艾略特@elliotchen100
稍微剧透一下,@EverMind 这周还会发一篇高质量论文
中文
rafa 🧉⚔️ retweetledi

Hey guys @ayubio @LundukeJournal
Looks like Arch Linux has suspended usage for Brazilians, maybe it's worth a shout-out if you haven't seen it yet
I can only access it with a vpn

English


@Pirat_Nation brazil said "give digital buyers actual rights" and rockstar's move was to leave the market entirely
tells you everything about how much they value the product you "own"
English

Rockstar Games has stopped selling its digital games directly to players in Brazil.
Starting March 16, 2026, you can no longer buy titles like GTA V or Red Dead Redemption 2 from the Rockstar Games Store or Launcher. This is because of Brazil's new Digital ECA law, which protects children online.
If you already own the games, you can still play and download them normally.
The games are still for sale on Steam, Epic Games Store, PlayStation Store, and Microsoft Store in Brazil.


English
rafa 🧉⚔️ retweetledi

noooooooo brazil, not you too
I'm so much more sad about brazillians getting blocked off from Arch Linux than californians, because most brazillians have a unique taste and love for ricing linux

Mr. Scorpy@mr__scorpy
Hey guys @ayubio @LundukeJournal Looks like Arch Linux has suspended usage for Brazilians, maybe it's worth a shout-out if you haven't seen it yet I can only access it with a vpn
English

@vikhyatk @LeeLeepenkman how does it compare to the new gemini embeddings model?
English

Yesterday, Moondream was the world's best segmentation model.
Today, it's even better.

moondream@moondreamai
Moondream segmentation just leveled up. ↑ Better masks ↑ New SOTA benchmarks ↑ 40% faster Already live on Moondream Cloud. Local model + technical whitepaper coming later this week. Read more: moondream.ai/blog/segmentin…
English
rafa 🧉⚔️ retweetledi

@0xRafaa It’s incredible how fast the life changes when you focus on something, keep pushing
English







