Pinned Tweet
Why I’m still on X (Sep 2026)
It comes down to two things: speed and serendipity.
The information velocity here is unmatched. By the time trends hit other platforms, it's already too late.
I'm always open to interesting opportunities and connectiions.
My DMs are open—drop me a ping.
我为什么在 X? (2025年7月)
过去,我相信做好一个产品是唯一重要的事。进入 AI 时代,我意识到有两点变得更为关键:
1. 无限贴近用户,理解他们的真实痛点。
2. 以最快的速度,将前沿技术转化为有效工具。
顶尖的技术论文,不应只停留在少数研究者的讨论中;同样,一线的用户需求,也亟需最强大的技术来解决。
所以我给自己找到了一个新角色:认知代理 (Cognitive Agent)。
我的工作就是连接这两者:将最前沿的理论,高速转化为能解决问题的有效工具。
在这里,你会看到:
1. 我对最新 AI 研究的硬核解读和如何落地。
2. 我对 AI 将如何重塑未来工作生活的思考。
无论你是想应用 AI 的同行,还是想借助 AI 破局的个人,都希望能给你带来启发。
3
2,611
Shen Huang retweeted
Claude Code changed our GTM completely.
We run a 4-person team doing the work of a 30-person GTM team.
Campaigns that used to take weeks now go from an empty terminal to live in 30 minutes.
The UGC campaign that used to take two weeks between a researcher and a copywriter now runs in one pass.
I've watched a lot of teams use AI and get nowhere.
A lazy prompt into a chat window, hoping for a magic one-shot.
It's a slot machine.
Here's the rule I built ours on: AI does the repetitive work. Humans make the decisions.
An agent without a map is just a very fast way to get lost.
So we stopped bolting AI onto 15 disconnected SaaS wrappers and built one context repo instead.
Claude Code runs the whole loop on it.
1. Research the signal.
The campaign starts here.
Claude pulls the raw data from live scrapers and data providers, then we pick the signal.
2. Verify the signal for UGC and content.
Before anything gets built, the signal runs against the ICP file already in the repo.
If it fails verification, we don't build. That stops a lot of wasted work.
3. Build the outbound.
This is the part that used to take the longest.
Claude Code pulls the ICP, the 15 outbound skills and the past winners file, then maps the full campaign in minutes.
4. Monitor UGC performance.
After the contracts are signed, Claude tracks which angles are working across every account.
5. Feed the loop back.
Weekly scripts push the performance data and the positive replies back into the context repo.
Winning angles get documented, and the rest get killed.
The next campaign starts smarter than the last one, and the loop gets tighter every time it runs.
That's why the moat is never the model.
It's the context around it.
Here's what's in the repo.
> The ICP matrix. Personas, tech stacks, disqualifiers and tiering, written down once.
> The winners file. Every past winning campaign with the reply rate and the reason it won, tagged by industry and pain point.
> 15 modular skills. Hooks, follow-ups, objection handling, offer positioning.
> Taste rules. What to avoid, banned AI buzzwords, formatting rules and where a human has to sign off.
If you want to copy this, start with the one workflow that eats the most time on your team.
Write down how your best GTM operator does it, step by step. Then turn that into your first skill and let Claude run it.
#ClaudeCode #GTM #AIAgents
DeepSeek V4.1 Flash (552B) on 2× RTX PRO 6000 + 128 GB RAM:
10.6 tok/s on one GPU → 72 tok/s on two.
Hot experts in VRAM, cold ones in RAM, 111 GB of lookup tables on the SSD.
Looks like the speed is due to my current motherbaord does not allow GPU p2p.
Also my Dual-GPU recipe: github.com/LichAmnesia/opens…
DeepSeek V4.1 Flash (552B, 3-bit EXL3) is 332 GB on disk.
My local box has 192 GB VRAM + 128 GB RAM. Even combined, it doesn't fit.
A community recipe makes it run by putting each part on the right tier:
- Hot experts in VRAM; cold experts in pinned RAM, streamed over PCIe only when routed.
- 111 GB of Engram lookup tables on NVMe, prefetched before each GPU step so the compiled fast path holds. Host RAM drops from ~230 GB to ~82 GB.
- DSpark speculative decoding: 5 draft tokens/step (~2x decode).
Reference 2x RTX PRO 6000 box: 163-169 tok/s (1 stream), 98.5% GSM8K, 96-97% HumanEval.
I got 10.6 tok/s on one card three weeks ago.
Second card went in yesterday.
I will share the DS v 4.1 Flash benchmarks for 2 x RTX Pro 6000 next.
1
264
DeepSeek V4.1 Flash (552B, 3-bit EXL3) is 332 GB on disk.
My local box has 192 GB VRAM + 128 GB RAM. Even combined, it doesn't fit.
A community recipe makes it run by putting each part on the right tier:
- Hot experts in VRAM; cold experts in pinned RAM, streamed over PCIe only when routed.
- 111 GB of Engram lookup tables on NVMe, prefetched before each GPU step so the compiled fast path holds. Host RAM drops from ~230 GB to ~82 GB.
- DSpark speculative decoding: 5 draft tokens/step (~2x decode).
Reference 2x RTX PRO 6000 box: 163-169 tok/s (1 stream), 98.5% GSM8K, 96-97% HumanEval.
I got 10.6 tok/s on one card three weeks ago.
Second card went in yesterday.
I will share the DS v 4.1 Flash benchmarks for 2 x RTX Pro 6000 next.
My opencode is growing 100GB per week...
Looks like it's due to github.com/anomalyco/opencod…
(And it's being couple of months)
@opencode why not just fix it.
The fix is just to separate transient state from final state. Emit streaming deltas and tool progress only to the in-memory event bus, not the event table. Persist an event only when a part or message hits a final state (completed, failed, or interrupted).
Copy my message above and your agent can fix it in like 5 minutes
1
1
426
After multiple round of optimizations
Now one RTX 6000 can generate 1.6B tokens per day.
Cache hit rate is 65% -> Next step is to hit 90+%
1
236
手机用agent势在必行
欢迎大家关注我的下一个篇章~
专为每一个AI coder打造的全新远程工具
—— vibe42.ai
· 解决你的远程配网难题,让你随时随地猛猛vibe
· 网络切换不稳定,对话上下文总是丢,vibe42全球节点适配,你的上下文永远都在本地,完全不会丢,
· AI做的结果,不用切app,立马就能换桌面看效果
· 手机终端的中文输入法总是有问题,英文中文打两遍,vibe42 从根源解决适配问题
· 远程桌面手机显示字太小,每个页面都可单独选择
1
2
437
Tested DeepSeek-V4.1-Flash-1017, one RTX PRO 6000 96GB + 128GB RAM + NVMe SSD.
Checkpoint is running: 10.64 tok/s, 5.31s TTFT.
Memory split:
• GPU: 95 GiB peak (~79 GiB experts in the plan)
• RAM: ~100 GiB container peak (~92 GiB / 8,220 experts)
~195 GB total in memory (for the experts)
Also I use the SSD for inference:
• Full checkpoint: 405 GB on NVMe
• Engram: 203 GB of that, read on demand
The recipe is available on hf:
huggingface.co/LichAmnesia/D…
更新了Recipt Qwen 3.8 Flash Next
现在单机 RTX pro 6000 是市面上最快的还要快 30 倍
并且还做到了并行越多越快。
huggingface.co/LichAmnesia/Q…
1
4
468
本地跑 Coding Agent,最崩溃的瞬间是什么?
是每一轮对话都要把 130K 的上下文重新算一遍,光是等首 token(TTFT)就要干瞪眼 19.5 秒……
所以就算看上去本地模型可以 xxx token / s,但是你让本地模型写代码就会发现特别慢,因为 Cache hit rate 低。
这就是为啥 DS 🐮🍺
为什么之前这么慢?因为 Agent 每轮都会重发整段对话。一张 RTX PRO 6000 的显存 KV 池差不多能装 80 万 token,200K 的长会话最多塞 4 个。只要你开到第 5 个会话,前缀缓存就会按 LRU 疯狂互相挤占,导致每轮都在全量重算!实测下来真实负载命中率只有惨兮兮的 26–29%,如果开 9 个会话轮转测试,命中率直接挂零(0%)。
我给 KV Cache 加了 64GB 的内存层(RAM offload)。同一张卡、同一份权重、同样的量化,首 token 直接干到了 0.64 秒! 提速整整 30 倍,丝滑得不像本地模型,哈哈。
可是这样我优化完成 Cache hit rate 才 30%。真不敢想 90% 以上 Cache hit rate 是得优化成什么样子。
食材还是 Qwen 3.8 flash next
配方在这里: github.com/LichAmnesia/qwen3…
单卡 RTX PRO 6000 跑 Qwen3.8-Flash-Next 的详细配置。
专门跑了 SWE-bench Pro 看跑分来验证能力。(后面大部分模型都用 DeepSWE 跑分了)。目前看起来接近 Opus 4.8
实测数据:
1/ 代码生成 181 tok/s,首 token 约 95ms
2/ 16 并发总吞吐 595 tok/s
3/ 262K 上下文,工具调用 16/16 通过
4/ SWE-bench Pro 50 题分层抽样:66分(Qwen 官方卡标 62.5)
整套部署我开源了:github.com/LichAmnesia/qwen3…
2
3
528
本地模型再也不只是玩具了
周末的 48 小时,我断开了所有云端 API,纯靠本地跑的 Qwen 3.8 Next Flash,从零手搓了一个完整的开源项目。
不仅跑通了。还把目前大热的 AI 智能语音输入 + 会议转录工作流 的功能给 1:1 复刻了出来。
每一行代码,100% 由本地模型生成。
分享一下我的一些技巧:
说实话,当我打开所有功能都 work 的时候有点头皮发麻。
1. 恐怖的并发吞吐:我同时开了 5 个 Agent 并行肝代码,本地推理速度硬生生飙到了 ~500 tokens/s。整整跑满 2 天,看着满屏代码瀑布流一样往下刷,而且加州电费之外成本为 0,这种爽感云端 API 给不了。
2. 纯本地的 Typeless 平替:底层直接用 本地 Whisper + Qwen 搞定了全套 ASR(语音识别)、智能润色排版和 Meeting 会议转录。
绝对的数据隐私:所有语音和会议记录全在本地闭环,0 数据上传,0 订阅费。
坦白讲,因为是 48 小时极限肝出来的版本,目前还有些小瑕疵(比如 Streaming 流式输出还没来得及接,底层性能优化也还有压榨空间)。
但这个实验已经彻底证明了一件事:只要工作流和并发设计得好,现在的本地模型,已经完全跨过了“重度生产力”的奇点。
代码全开源了,项目叫 Launch,欢迎各位大佬去玩玩、提 PR。
未来会添加更多常用的本地模型让这个工具更加有用。👇
单卡 RTX PRO 6000 跑 Qwen3.8-Flash-Next 的详细配置。
专门跑了 SWE-bench Pro 看跑分来验证能力。(后面大部分模型都用 DeepSWE 跑分了)。目前看起来接近 Opus 4.8
实测数据:
1/ 代码生成 181 tok/s,首 token 约 95ms
2/ 16 并发总吞吐 595 tok/s
3/ 262K 上下文,工具调用 16/16 通过
4/ SWE-bench Pro 50 题分层抽样:66分(Qwen 官方卡标 62.5)
整套部署我开源了:github.com/LichAmnesia/qwen3…
5
12
1,151
Where can I get one
Intel GPU D-series to reach 512GB memory on partner cards, leaker claims
🔗videocardz.com/newz/intel-gp…
220
账单爆炸基本都是因为缓存!!!
cache!
作为Cloudflare资深用户,我的独立开发生涯不断被账单背刺:
- 七月:图片转换被坑
- 八月:容器使用被坑
- 九月:prefetch DDOS 自己
- 十月:KV 缓存爆炸
这次分享的,就是我本以为开源框架应该已经做好了,但实际上 // TODO 多年的一个缓存陷阱:
meathill.com/posts/cloudflar…
407




















