Martin

Studying AI, loves sports science, cycling and running. Occasional DJ

[Martin @martinflo_]

The math that turns chaos into clarity🚀

[Martin @martinflo_]

@lvwerra @huggingface Brilliant

[Martin @martinflo_]

@sh_reya DM

[Martin @martinflo_]

@cerebras @bridgemindai

[Martin @martinflo_]

@CoachDanGo Partner on the left cheating or what? Shared bed 134 days while the partner on the right only 128?

[Martin @martinflo_]

@Brady_H Wasn’t this quite clear? But the picture changes when you do the same test within group of «great» or «good»

[Martin @martinflo_]

@Brady_H Durian rider was right all along

[Martin @martinflo_]

@ProfTimNoakes @Cdenejnjendjen @PaulBLaursen I partly agree with the theory, but in my experience feeling after race day is also much better and you don’t feel as «dead» as you otherwise would. Why is that?

[Martin @martinflo_]

@claudeai Next step: your agent view can talk to other people’s agent view

[Martin @martinflo_]

@godofprompt Reminds me of the prompt repition paper: https://t.co/XtjrUpcAAs

[Martin @martinflo_]

@milesdeutscher For my projects, I like to create a /knowledge/plan folder. There, I iterate on the implementation plan. This makes it easy to regularly clear my chat. Then I do a couple of passes with Opus on the final plan, maybe even send it to another LLM, before implementation.

[Chris @everestchris6]

this OpenClaw bot finds restaurants with ugly menus, rebuilds them as live web menus, and mails the owner a postcard...on autopilot.

here's how agencies can land recurring contracts with this system:

- scrapes every restaurant in a city in real time
- filters by review count + rating + last menu update + photo quality
- pulls the real menu items from the official site, PDF, or Google reviews
- samples the brand palette from the restaurant's own visual identity
- renders a 9:16 brand-matched menu, hosted live at a QR-accessible URL
- writes a personalized postcard referencing a real reviewer and a real dish
- mails it to the registered office addressed to the owner by first name

every step from discovery to brand-matching to outreach is automated.

reply "MENU" + RT and i'll send you a free guide so you can build this too

[Martin @martinflo_]

@everestchris6 MENU

[Martin @martinflo_]

@Tu7uruu @huggingface Specialize it in a specific language

[Martin @martinflo_]

@DJKilmartin @ProfTimNoakes 3:20 pase for him is probably comfortable within z2. Maybe even z1. Also just 1h run

[Martin @martinflo_]

Today marked the day I first heard the word "credit rollover" (talking about LLM subscriptions).

I hope you don’t roll over your tokens

[Martin @martinflo_]

@Cyberhonk26 @akshay_pachaar Curious, Why would local voice be better?

[Martin @martinflo_]

@karpathy Soon obsidian will build extensions for the LLM

[Martin @martinflo_]

@lawrencecchen Brilliant

[Martin @martinflo_]

@seikixtc Would be nice to see if you can add TurboQuant to achieve more space https://t.co/mHrNpVvesD

[Martin @martinflo_]

We’re so back

[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。🌟👀🙏

https://t.co/Or3PVUv67q