xc (@settinghead) - Профиль Twitter | Zamantika Mersobahis Locabet

Закреплённый твит

xc@settinghead·10 Mar

What if your coding agent sounded like the StarCraft Adjutant or C&C EVA? I built Voxlert — it hooks into Claude Code, Codex and Cursor and speaks notifications in-character. Each session gets a distinct voice (Adjutant, SHODAN, GLaDOS) so you know which agent needs you without looking. LLM + Local TTS + voice filters, cost negligible. github.com/settinghead/vo…

English

0

10

4.4K

xc@settinghead·48m

@bozhou_ai 另外一个导致应用尚不明确的点: 模型应该是得从头训练，所以现有的大模型不能直接使用

中文

0

3

xc@settinghead·51m

@bozhou_ai 检索还是有的，只是变相换了个地方

中文

1

0

4

泊舟@bozhou_ai·3h

EverMind 直接给了个记忆的新思路，就是把记忆长进注意力机制里，不检索、不压缩，端到端训练。数据上很猛：4B 参数的小模型，在长上下文跑分上干翻了 235B 级别的 RAG 方案。从 16K 扩到 1 亿 token，精度掉不到 9%。跑推理只要 2 张 A800。做 Agent 和长记忆方向的可以关注一下后续，论文和代码都放 GitHub 了，让你的龙虾帮你分析一下 github.com/EverMind-AI/MSA

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

中文

8

6

30

2.8K

xc@settinghead·6h

@anorth_chen @ItsTheDaybreak 现在做不好只是没有足够的参考范式，过一阵子就不一定了。

中文

0

36

0xNorth@anorth_chen·7h

@ItsTheDaybreak 我的观点是，agent产品不是靠堆人能抄得了的

中文

1

0

4

3.1K

0xNorth@anorth_chen·10h

腾讯出了 qclaw，身边几个拿到内测码的朋友试了一圈后的产品反馈是： “要不我们去做空腾讯吧。😅” 也再次印证了我之前的一个判断—— 如今互联网大厂的人才密度根本无法支撑Agent产品的落地。一边是独立开发者，靠兴趣驱动，快速迭代，做出像 openclaw 这样开始影响一批人的项目。另一边是大厂里二三十人的团队，层层分工、反复对齐、加班加点，最后却连一个“能用”的版本都难产。再多的人堆进去，产出的也只会是“更慢的失败”。有点讽刺，但大厂的管理层们还需要起码一年的时间才能建立这个认知。还会有一批又一批的大厂Agent 团队，消耗着人力和预算，陪老板过家家，最终让老板学会这个不复杂的道理。

中文

52

35

369

86K

xc@settinghead·6h

@BlockBeatsAsia @fynnso 好奇从技术上来说是怎么弄到他们的tokenizer的？

中文

0

39

BlockBeats｜We're hiring!@BlockBeatsAsia·9h

哇哦！刺激了🍉Cursor「自研」Composer 2 被曝内藏 Kimi K2.5，月之暗面公开指控：未遵守许可证、未支付费用！！今日，开发者 @fynnso 在调试 Cursor API 请求时发现，Composer 2 的实际模型 ID 为 `kimi-k2p5-rl-0317-s515-fast`，字面即「Kimi K2.5 + RL」。月之暗面（Moonshot AI）预训练负责人杜羽伦（ @Yulun_Du ）随即发推，称团队测试 Composer 2 的 tokenizer 后发现「与我们的 Kimi tokenizer 完全一致」，「几乎可以确认这是我们的模型被进一步后训练的结果」，并直接 @ Cursor 联合创始人 Michael Truell @mntruell ，质问「为什么不尊重我们的许可证，也没有支付任何费用」😡😡 Cursor 发布 Composer 2 时称，性能提升来自「首次对基座模型进行继续预训练，再结合强化学习」，但全程未提及 Kimi K2.5。两点并不矛盾：继续预训练和 RL 本就施加在某个基座之上，Cursor 只是未披露基座来源。截至发稿，Cursor 未公开回应且开发者 @fynnso 原推文已删除，但手快的律动小编已截图😉😉）

中文

29

33

310

62.2K

xc@settinghead·15h

@ravikiran_dev7 64GB+. Easily can have 30 cli agents open concurrently.

English

1

0

1

5

Ray🫧@ravikiran_dev7·1d

As a Developer, how much RAM is enough for you? 1. 8 GB 2. 16 GB 3. 24 GB 4. 32 GB 5. 64 GB 6. 128 GB+

English

259

21

603

53K

xc@settinghead·15h

@Stanleysobest 大部分简历不是都是注水严重的吗

中文

0

15

Stanley@Stanleysobest·17h

我们团队正在给B端做HR AI提效，刚听一位300人公司的HR总监说：一个岗位刚放出去，瞬间几千份简历砸过来。最吓人的不是海投，是每一个都真的能胜任。现在的就业市场，已经卷到这种程度了？

中文

19

2

86

19.9K

xc@settinghead·16h

@WizLikeWizard The discussion is pointless without specifying which model you are using.

English

0

2

Wiz 👨‍🚀@WizLikeWizard·1d

Have been using OpenClaw for ~a month and it kinda sucks? I spend more time battling it to get basic crons fired reliably, remember things, and not repeat itself. Am I doing it wrong or are we just still very early on all of this?

English

305

5

368

56.4K

xc@settinghead·16h

New idea: You don't have to have only one prototype.

Aakash Gupta@aakashgupta

Everyone's debating whether PRDs should die. Wrong question. The spec moved from step 2 to step 6. Old flow: Idea → PRD → Design → Eng → QA → Ship. 8-12 weeks. The PRD was a permission document. "Please approve before we commit resources." New flow: Idea → 5 prototypes → Evaluate → Kill 4 → Spec the survivor → Ship. 1-2 weeks. The PRD is now a decision record. "We built 5 versions. Here's which one and why." At Anthropic: no PRDs. At OpenAI: specs still critical because 800M MAU need behavior contracts with 15-25 labeled examples. At enterprises with 5,000 people: the document is the alignment mechanism across 3 time zones. Company stage determines where the spec sits. The universal shift is that the spec comes after you've touched working software. A prototype shows what. The spec explains why, how you'll measure, and when you'll pull the plug. Those are the things that separate a PM from a vibe coder.

English

0

8

xc@settinghead·16h

@MrGafish 到底谁真的裁了

中文

1

0

29

Mr Gafish@MrGafish·17h

科大讯飞裁员网易裁员 B站裁员最近刷到这几家的消息，下一个会是谁？

中文

170

24

643

303.8K

xc@settinghead·17h

@KKaWSB 知错就改及时扭转是对的。毕竟他只有一例成功创业经历，每次都成功也不现实。

中文

0

3

KK.aWSB@KKaWSB·1d

Meta 宣布将关闭 Metaverse，科技史上最贵，超过 800 亿美元投入的大败局：马克·扎克伯格当年全力投入虚拟现实技术 Facebook 收购 Oculus（20 亿美元）对整个公司进行品牌重塑（2021 年）集团公司更名为：Meta（元宇宙）为 Reality Labs 招聘数千人投资额超过800亿美元构建元宇宙愿景推出地平线世界甚至没有真正意义可用产品，无用户留存损失不断累积（每年数十亿美元）注意力转向人工智能科技领域最大的赌注之一……

中文

93

28

255

190K

xc@settinghead·17h

@binghe 苹果的溢价来自于品牌+图形处理+开发生态+现在的AI，小米的溢价来自于……？

中文

0

1

6

冰河@binghe·1d

小米笔记本，卖的挺贵啊！！！真敢要价啊。。。现在经济上行了吗？？？我有这价格，为什么不买苹果？？？

中文

301

6

367

245.1K

xc@settinghead·17h

@iamai_eth 如果从The Bitter Lesson的棱镜看，前者有更大概率胜出。

中文

0

5

✧ 𝕀𝔸𝕄𝔸𝕀 ✧@iamai_eth·21h

OpenClaw-RL 和 Hermes Agent 是2026年初个人AI Agent领域两个非常热门但方向不同的开源项目，都在试图解决“AI代理如何持续变强”的核心问题，但实现路径、适用场景和哲学完全不同。 - OpenClaw-RL ≈ “给OpenClaw装一个超级在线对齐引擎”，靠强化学习把你每一次不满/纠正/追问都变成模型的训练信号，理论上长期能最贴合你，但需要GPU、时间和OpenClaw基础。 - Hermes Agent ≈ “一个会自己长大的个人助理”，重点在记忆、技能积累和跨会话连续性，短期就能感受到“它越来越懂我”，部署简单、模型中立。

中文

4

6

53

7.4K

xc@settinghead·17h

@dontbesilent 真有人试过了吗？

中文

0

61

dontbesilent@dontbesilent·1d

1、这是我迄今为止用过的最好的 skill 2、无论我们多么大力宣传这个 skill，大部分都只会围观，最多安装一下，不会真的去在自己的上下文里面去推进 3、所以，这让这个 skill 的使用体验更稀缺了

Garry Tan@garrytan

I just launched /office-hours skill with gstack. Working on a new idea? GStack will help you think about it the way we do at YC. (It's only a 10% strength version of what a real YC partner can do for you, but I assure you that is quite powerful as it is.)

中文

33

347

2.1K

441.1K

xc@settinghead·17h

@sergeykarayev Local agents are fun as a hobby, as long as they stay that way.

English

0

4

Sergey Karayev@sergeykarayev·21h

Running agents locally is a dead end. The future of software development is hundreds of agents running at all times of the day — in response to bug alerts, emails, Slack messages, meetings, and because they were launched by other agents. The only sane way to support this is with cloud containers. Local agents hit a wall quickly: • No scale. You can only run as many agents (and copies of your app) as your hardware allows. • No isolation. Local agents share your filesystem, network, and credentials. One rogue agent can affect everything else. • No team visibility. Teammates can't see what your agents are doing, review their work, or interact with them. • No always-on capability. Agents can't respond to signals (alerts, messages, other agents) when your machine is off or asleep. Cloud agents solve all of these problems. Each agent runs in its own isolated container with its own environment, and they can run 24/7 without depending on any single machine. This year, every software company will have to make the transition from work happening on developer's local machines from 9am-6pm to work happening in the cloud 24/7 -- or get left behind by companies who do.

English

88

20

290

26.9K

xc@settinghead·18h

@sitinme A2A真的有项目在用吗？

中文

0

3

sitin@sitinme·1d

这波 Google 其实干了一件挺关键的事：把一堆原本各自为战的 Agent 协议，第一次串成了一条“能跑通业务”的完整链路。以前你做 Agent，会有一种很明显的割裂感：连工具一套代码、连外部服务一套代码、做前端展示再一套代码。真正业务逻辑可能就 30%，剩下 70% 全是在写“胶水”。而且每换一个框架、换一个工具，基本都要重来一遍。这次 Google 的思路很简单但很狠：不再强调“我这套最牛”，而是把 MCP、A2A、UCP、AP2、A2UI、AG-UI 这些不同组织的协议，全拼起来，直接给你一个能从“查数据 → 调 Agent → 下单 → 付款 → 展示 → 实时交互”跑通的完整范式。说人话就是：以前你是在“拼乐高但每块都要自己削”，现在是“标准件已经给你做出来了，你只需要搭”。 Agent 这条路，已经从“能力竞赛”进入“基础设施阶段”了。之前大家比的是：谁更会写 prompt、谁上下文更长、谁 agent loop 更聪明。接下来比的会变成：谁的协议更通用；谁的生态接入更多；谁更像“互联网协议层”而不是“工具层” MCP 像 HTTP（连万物），A2A 像 TCP（Agent 之间通信），UCP/AP2 像支付+电商协议，A2UI/AG-UI 像前端标准。这一套东西一旦跑起来，其实是在往“Agent 版互联网”走。可能你以后写的 Agent，很可能不是“某个平台里的一个 bot”，而是一个可以被别的 Agent 调用的服务节点。你不只是用 Agent，你是在“把自己接入 Agent 网络”。这件事比单个模型升级更重要。模型再强，也解决不了“怎么连世界”的问题。但一旦协议打通，Agent 才真正有机会从“会聊天”变成“能干活”。

中文

4

6

55

7.1K

xc@settinghead·18h

@lifesinger 我觉得得尽快利用AI开始搞销售和营销。

中文

0

3

Frank Wang 玉伯@lifesinger·1d

对绝大部分用户来说，AI 只有一个缺陷用 AI 很花钱，但很难赚到钱克服这个缺陷的最简单办法是把 AI 当高档消费品只想着花钱，不想着赚钱心里就舒坦了没钱怎么办把之前花在消费品上的钱挪部分到 AI 上 AI 不会消灭工作而是会消灭各种高档物质消费品

中文

16

2

40

7.4K

xc@settinghead·18h

@CTracy0803 国内现在有哪些公司真的裁员了？

中文

0

60

Tracy@CTracy0803·21h

目前“B站裁员要裁40%到60%”在各大平台疯传 B站虽在2025年首次实现全年盈利，但股价仍较为低迷。

中文

33

9

267

61.4K

xc@settinghead·18h

@heyrimsha GSD + long sessions = you need to know when it's done without watching. Built Voxlert for this — character voice notifications so you hear which agent finished. x.com/settinghead/st…

xc@settinghead

What if your coding agent sounded like the StarCraft Adjutant or C&C EVA? I built Voxlert — it hooks into Claude Code, Codex and Cursor and speaks notifications in-character. Each session gets a distinct voice (Adjutant, SHODAN, GLaDOS) so you know which agent needs you without looking. LLM + Local TTS + voice filters, cost negligible. github.com/settinghead/vo…

English

0

2

Rimsha Bhardwaj@heyrimsha·3d

🚨 This is how engineers at Amazon, Google, and Shopify actually use Claude Code. It's called GSD (Get Shit Done) and it solves context rot the quality degradation that destroys your Claude Code sessions as the context window fills up. No BMAD. No enterprise sprint theater. No Jira nonsense. Here's how it works: You run one command → /gsd:new-project → It interviews you until it fully understands your idea → Spawns parallel research agents to investigate your stack → Creates atomic task plans with XML structure Claude actually understands → Executes in fresh 200k context windows per task → Commits every single task to git automatically Here's the wildest part: Your main context window stays at 30-40% the entire time. All the heavy lifting happens in subagent contexts. No degradation. No "I'll be more concise now." Just clean, consistent execution. Engineers at Amazon, Google, Shopify, and Webflow trust this thing. MIT license. One command to install: npx get-shit-done-cc@latest Link in the first comment 👇

English

60

111

859

94.9K

xc@settinghead·18h

@PawelHuryn Another hooks use case: voice notifications. Voxlert hooks into PostToolUse to speak in-character — SHODAN, StarCraft Adjutant — so you know which agent needs you by ear. x.com/settinghead/st…

xc@settinghead

What if your coding agent sounded like the StarCraft Adjutant or C&C EVA? I built Voxlert — it hooks into Claude Code, Codex and Cursor and speaks notifications in-character. Each session gets a distinct voice (Adjutant, SHODAN, GLaDOS) so you know which agent needs you without looking. LLM + Local TTS + voice filters, cost negligible. github.com/settinghead/vo…

English

0

1

Paweł Huryn@PawelHuryn·3d

Claude Code already shipped a solution for this: hooks. PreToolUse runs before any tool executes. Your script sees the tool name + arguments and returns allow, block, or error. Block rm. Allow edits. Require confirmation for writes to /src. Log every Bash command. PostToolUse runs after execution: auto-format, lint, run tests. Not less dangerous. Precisely controlled.

ben@benhylak

i wish there was a claude --less-dangerously-skip-most-permissions

English

14

5

111

19.7K

xc@settinghead·18h

@brian_scanlan Hooks are so underused. I built Voxlert on them too — PostToolUse triggers character voice notifications so you hear which agent session needs you. x.com/settinghead/st…

xc@settinghead

What if your coding agent sounded like the StarCraft Adjutant or C&C EVA? I built Voxlert — it hooks into Claude Code, Codex and Cursor and speaks notifications in-character. Each session gets a distinct voice (Adjutant, SHODAN, GLaDOS) so you know which agent needs you without looking. LLM + Local TTS + voice filters, cost negligible. github.com/settinghead/vo…

English

0

1

Brian Scanlan@brian_scanlan·3d

Claude Code hooks enforce our PR workflow at the shell level and blocks it unless the create-pr skill was activated first 1. A PreToolUse hook intercepts raw gh pr create 2. The skill extracts business INTENT before creating - asks "why?" not just "what changed?"

English

2

0

38

24.5K

Brian Scanlan@brian_scanlan·3d

We've been building an internal Claude Code plugin system at Intercom with 13 plugins, 100+ skills, and hooks that turn Claude into a full-stack engineering platform. Lots done, more to do. Here's a thread of some highlights.

English

83

198

3K

792.3K

xc

Открыть