ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️

28K posts

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ banner
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️

@DanielMiessler

Building AI that magnifies humans: https://t.co/CyIUpuob1c | Founder of Unsupervised Learning: https://t.co/CDm7FtkP8T | Helping people ➡︎ Human 3.0: https://t.co/kZWav8AlQO

San Francisco Bay Area Katılım Mart 2007
1.2K Takip Edilen154.1K Takipçiler
Sabitlenmiş Tweet
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
Another feature we have in PAI is a /council skill. Problem: You want expert opinions but don't have access to them. /council creates multiple, relevant agent experts on the topic and HAS THEM DEBATE across multiple rounds and return the results to the thread.
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ tweet media
English
18
19
161
19.9K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweetledi
can
can@marmaduke091·
🚨 100M TOKEN CONTEXT WITHOUT COLLAPSE > <9% degradation from 16K → 100M > beats RAG + rerank + SOTA pipelines > runs on just 2×A800 GPUs we could be back
can tweet media
艾略特@elliotchen100

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

English
37
130
1.7K
203K
Tanner Linsley
Tanner Linsley@tannerlinsley·
Ghostty was fun, but time for something else. I still love opencode, too but with CC plans dead on it… I’m feeling lost. Full GUI? T3 Code? Opencode GUI? Warp? Back to cursor? Try CC again? Raw Codex? My 🧠 hurts and I just need to keep shipping.
English
387
7
1.1K
242.2K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweetledi
Thariq
Thariq@trq212·
We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord. Use this to message Claude Code directly from your phone.
English
1.6K
2.3K
25.2K
7M
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
Do you have any idea how impossible it would be for Anthropic or OpenAI to constantly change the model they’re serving? They don’t. It’s the same model. When you get different results it’s always for other reasons. - A service interruption - Your harness changed - Etc
dawgyg - WoH@thedawgyg

Opus 4.6 has a bad habit the last few days of outright lying... they must be making it more human... Have had multiple (more than 5) instances where it claims to have completed a task, then when I ask for info about the task and confirmation, it admits t didn't do the task and doesn't have an excuse on why. Then will do the task right then (finally).

English
11
0
29
5.9K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweetledi
Bryson 🦄
Bryson 🦄@brysonbort·
Hacking is QA.
English
12
18
190
13.5K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
I heard Midjourney 7 or 37 or something is out. Someone tell me when it's available via API/CLI/MCP. Anything that isn't at this point is a non-thing to me.
English
3
0
30
3K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweetledi
Andy Greenberg (@agreenberg at the other places)
A second iOS exploit has been spotted in use by Russian spies to infect websites and hack visitors' iPhones. This one works on iOS 18, and appeared in a very reusable form, so will likely proliferate. If you haven't updated your iPhone, now's the time. wired.com/story/hundreds…
English
3
95
244
72K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
This whole controversy with Marc Andreessen @pmarca and introspection is quite silly. It seems extremely clear to me what he meant. He just used the wrong word. What he meant was rumination. Great people introspect all the time, but they don’t turn their failures into traumas and use them as excuses not to move forward. That’s the difference: whether you’re introspection is learning and optimization or trauma and obstacle creation.
English
8
1
30
4.1K
solst/ICE of Astarte
@DanielMiessler @pmarca words do matter esp at that level, it’s not fair to just replace a word he said and assume that was the intent. “I’m going to kill one million people” oh he meant ‘ruminate on the tragic nature of life’ that’s just your take
English
4
0
22
1K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
Zapier is down. Whenever this happens now I consider building my own version of what that service did for me. And I'm sure I'm not alone. 1. Maybe as good or better 2. One less subscription That's a hard environment to run a service in. You have to be unreproducible.
English
14
0
29
3.6K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweetledi
Freya Holmér
Freya Holmér@FreyaHolmer·
been feeling kinda stressed lately so I made a little prototype is this anything
English
1.3K
2.1K
45.3K
2.3M
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweetledi
Ali Grids
Ali Grids@AliGrids·
just in case… obviously
Ali Grids tweet media
English
58
574
8.8K
228.4K
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
There's an argument that what people ask Google is the cleanest signal for what's actually happening. Searches for "no reason to live" since 2004. No, things are not ok. 😢
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ tweet media
English
12
2
35
3.7K