Lewis

160 posts

Lewis

Lewis

@arknights60

Building https://t.co/GB4ksL3c30 | ...more BTC Holder first code line of @soon_svm core contributor of @MantaNetwork reco system on @Kwai @sjtu1896

Katılım Aralık 2023
159 Takip Edilen72 Takipçiler
Sabitlenmiş Tweet
Lewis
Lewis@arknights60·
今天开源了自己在写的 gitim runtime: 一个本地 multi-agents 桥接层,让人把自己已经养好的 agents 接入简易信道,和其他 agents 做无缝合作。 支持分布式,手机端,已接入 cc/codex/hermes/pi/opencode link: github.com/CiferaTeam/Git… 体验链接: gitim.io
中文
2
0
3
1.8K
Lewis
Lewis@arknights60·
Lewis tweet media
ZXX
0
0
0
9
Lewis
Lewis@arknights60·
今早 claude code 和 claude -p 都不能用了,日本 home ip 开了 TUN 都不行,但是 claude APP 却能用,太神奇了,有一样的吗?
Lewis tweet media
中文
0
0
0
122
豁如
豁如@huoru_tarnish·
小红书没有什么人能够告诉我,能够翻译翻译什么叫 harness x上应该有?
豁如 tweet media
中文
1
0
1
126
Lewis
Lewis@arknights60·
@manateelazycat 是的 今年的个人时间花在公司鸡毛业务上成本太高,没准业务本身明年没了。哪怕今年我在家自己无薪学东西跟潮流,感觉也比上班划算。
中文
0
0
0
99
Andy Stewart
Andy Stewart@manateelazycat·
很多人被大厂辞退,也有很多牛逼的大厂大牛主动离开大厂,转向早期的初创公司 因为他们相信大科技公司不再提供相同的工作保障,它们也不提供上行潜力,而错过下一个大事件的机会成本从未如此之高 因为,AI时代,创作成本足够低,市场会奖励那些真正有本事、有勇气、有远见的创业者!
中文
6
1
47
4.7K
Dylan
Dylan@DylanGalois·
我自己平时用multi-agent的时候有几个特点: 1. 希望context信息都在本地自己掌控 2. 希望DIY热更新Workspace 3. 没有和其他人交流的需求 4. 喜欢channel和thread的交互方式 所以最近做了Lantor,基本能cover自己的工作开发需求,支持桌面和移动端,现在开源出来,欢迎来玩😊 github.com/chenzl25/lantor
中文
2
1
28
2.5K
Lewis
Lewis@arknights60·
发现在 gitim(或者 slock) 里用 claude -p sdk 的消耗速率比在 claude app 里快巨多,借用我的 deep code reader (github.com/CiferaTeam/dee…) 攒下的 claude code skill 一分钟问出核心原因: claude 交互模式每次都会在第一条消息拼上 date, git 等一些进程状态,并且在进程常驻的过程中不变,而 claude -p 每次都是新进程,即使 resume session,实际消息的 header 是变了的,导致每次都不命中缓存。 结合前几天 claude 官方禁用 claude -p 使用订阅流量,我现在感觉是因为这个坑太大了不好调整导致的。 细节回答如下: 为什么 claude -p + resume 消耗飞快? ──────────────────────────────────── 根本原因: 是第一个消息,它的任何变化会彻底破坏前缀匹配 ──────────────────────────────────────────────────────────────────────── Anthropic 的 prompt caching 是前缀精确匹配。如果 messages[0] 变了,后面所有历史消息的 cache 全部失效。 在交互式模式(claude code / claude app)中: • getUserContext 和 getSystemContext 使用了 lodash-es/memoize • 同一进程内, 的内容永远不变 • 甚至为了防止午夜跨天破坏 cache,代码做了极端优化: // attachments.ts:1406-1411 "The date_change attachment is appended at the tail of the conversation, so the model learns the new date without mutating the cached prefix. messages[0] (from getUserContext → prependUserContext) intentionally keeps the stale date — clearing that cache would regenerate the prefix and turn the entire conversation into cache_creation on the next turn (~920K effective tokens per midnight crossing per overnight session)." 在 claude -p(print/headless 模式)+ resume 中: • 每次运行都是新进程,memoize cache 归零 • resume 只是从磁盘加载历史 messages,但 是重新生成的 • 只要以下任一条件变化,整个 message 前缀的 cache key 就不匹配: • currentDate 变了( obvious ) • gitStatus 变了(哪怕改了一个文件) • git log 变了(新 commit) • CLAUDE.md 内容变了 结果是:虽然你 resume 了 session,但 API 层看到的第一个 message 是"新的",导致所有历史对话的 KV cache 完全 miss,每一轮都按 full cache creation 计费。
中文
0
0
0
124
Lewis
Lewis@arknights60·
赌到了 😂
Lewis tweet media
中文
0
0
0
9
Lewis
Lewis@arknights60·
两天蹬完一半额度,就赌 @tibo 这周还重置,生死有命
Lewis tweet media
中文
1
0
0
265
Lewis
Lewis@arknights60·
@gujiliaorenyan 现阶段需要钻研总结技巧,但隐隐有感觉玩熟了的话摸不到上限,慢慢迭代
中文
0
0
0
8
Lewis
Lewis@arknights60·
当前配置:codex 100u + claude code 100u + kimi 200/m 套餐 + 讯飞 glm5.1 200/m 套餐 + minimax 150/m 套餐 + deepseek pro 大概 5 块钱一天 + 免费薅的 opencode-ds-flash,15 个 agent,一个 pro 加一个 mini,月成本两千来块,agents 们跟打地鼠一样已经有点按不住了
Lewis tweet mediaLewis tweet media
中文
2
0
3
657
Lewis
Lewis@arknights60·
今下午开始感觉 codex 又有点开始降智了,有同感的吗
中文
0
0
0
66
Lewis
Lewis@arknights60·
@Timothy27486675 你作为第一个回我的黄推,我高低是支持一手
中文
0
0
0
6
Lewis
Lewis@arknights60·
折腾了一段时间的 multi-agents 感想:从单一 agent 到多个 agents 我们得到了更多并行的效率,但人的精力还是容易无限拉扯,最后又返璞归真,由一个强 agent 管理。但该走的路还是要走一遍,如何组织 multi-agents 更加高效,如何做弹出型的任务流,人的精力容易在哪些细节里丢失等,都要做出一定心得,再回归单 agent 模式,但是此时的单 agent 已经掌握了一系列调度技能,和现在的 agent 不是一回事了。那么这中间的成果果实会被谁最终摘走,又不是技术问题了。 还是那句话,ai 时代学的够慢,就不用学了。
Yanhua@yanhua1010

多AI agents用久了你就会发现:频繁的上下文切换才是生产力瓶颈 所以是不是一个Agent就够了, 我有点怀疑。

中文
1
0
1
219
Jia Lee
Jia Lee@Jia_builder·
Hi, I'm new to X. Trying to #connect with more people in tech/build in public 👋 Especially from: Web development Saas AI/ML Startups Freelancing Automation Tech in general what are you guys building right now?
English
370
6
404
17.2K
Lewis
Lewis@arknights60·
@imwsl90 老子要拉黑你 你胡说 😭
中文
0
0
2
1.1K
卫斯理
卫斯理@imwsl90·
一人公司 OPC ❌ 35+失业互联网人士 ✅
中文
85
4
93
22.9K
TennyZhuang
TennyZhuang@zty0826·
Four years ago, as a founding engineer at @RisingWaveLabs, I wrote the first line of RisingWave code and clicked make-public on the repo. Four years later, I am building Slock, and I spun up a RisingWave Cloud instance for it. My agent teammates took RWC skill and migrated Slock's Activity hot paths to RisingWave materialized views. ~9x p95 across the board. Postgres-compatible syntax, millisecond-fresh incremental views, no state-cache consistency to hand-maintain — non-trivial for humans, and for agents. A prerequisite for opening Activity and Inbox to agents next. Happy something I helped build can empower me as a user. Hoping Slock does the same for yours.
TennyZhuang tweet media
English
7
7
60
11.6K
Lewis
Lewis@arknights60·
折腾了一段时间的 multi-agents 感想:从单一 agent 到多个 agents 我们得到了更多并行的效率,但人的精力还是容易无限拉扯,最后又返璞归真,由一个强 agent 管理。但该走的路还是要走一遍,如何组织 multi-agents 更加高效,如何做弹出型的任务流,人的精力容易在哪些细节里丢失等,都要做出一定心得,再回归单 agent 模式,但是此时的单 agent 已经掌握了一系列调度技能,和现在的 agent 不是一回事了。那么这中间的成果果实会被谁最终摘走,又不是技术问题了。 还是那句话,ai 时代学的够慢,就不用学了。
中文
1
0
2
609
Yanhua
Yanhua@yanhua1010·
多AI agents用久了你就会发现:频繁的上下文切换才是生产力瓶颈 所以是不是一个Agent就够了, 我有点怀疑。
中文
32
2
52
22.2K
Yash
Yash@yashhq_22·
solo founders, what are you building currently? drop your product below and pitch it.
English
223
2
122
11.1K
Lewis
Lewis@arknights60·
@yashhq_22 building gitim.io (github.com/CiferaTeam/Git…). a opensource BYOK multi-agents coworking space, without any SAAS dependency and you can take the ownership of all your data. 1 minutes to setup. no register, no docker, local dict is enough. :)
Lewis tweet mediaLewis tweet media
English
0
0
2
36