HMLG

生产力工具还是理财产品? 4 台 DGX Spark + 1 台交换机，19 万买的，今天现价 24 万 你获得了什么? 一天 24–31 亿 token（按生产记录推算） decode 单路 76 tok/s 在本地跑DeepSeek V4.1 Flash 1017 和GLM 5.3 Flash REA不受限制 玩两年还能挣钱的新常态还能持续多久? 硬件还会涨，你同意么？

我最近测过了，dgx spark*4跑deepseek v4.1 flash 1017，比dgx*2的提升主要在上下文，量化版本，速度上的提升和并发上的提升几乎没有。 所以为了提供更多的服务，4台应该选2+2跑两套服务。

DeepSeek just released its Super-agent design. You won't find the full text on X. It's in Chinese, on Zhihu. So I brought it to you: the complete article in English, every figure redrawn. Personal agents are the fight of 2026: • xAI: Grok Bot • Meta: Muse • OpenAI: dots • Anthropic: Claude Duet • Google: Gemini Spark DeepSeek didn't launch a personal agent app. It published the playbook: the architecture, the experiments, and an MIT-licensed prototype. Here's how we think about it. You go build it. We'll keep working on AGI. Very DeepSeek. What's inside: 1. Router + Workers. One resident Router talks to humans and only dispatches. Isolated Workers do the work and never talk in the channel. 2. Events, not prompts. Chats, PR comments and CI results arrive as Triggers. Code filters the noise; the LLM only judges what's left. 3. Subscribe and pass through. Follow-ups in a thread go straight to the Worker that owns it. 4. Inline & Spill. The context window is run like a cache, with a Chain of Recall back to old tasks. 5. Act early. It quietly reproduces a reported bug and only speaks up once it has a fix. 6. Brain in the cloud, hands on your machine (BYOM), plus an outside Watchdog that checks real progress. The numbers: p95 task time 15.7 min vs 29.5 for a single session with a subagent tool. Router batching cut time per decision from 6.4s to 1.6s.

没有全面包圆的谁更好，使用方式有差异

I run DeepSeek-V4.1-Flash-1017 across locally four DGX Sparks , on a 200G switch, behind my own gateway, with pi and DeepSeek Harness as clients. It's a multimodal model. I deployed it, and it couldn't take images or video. Fix it, fast. That took five fixes. 1. Rejected in 2 ms. A colleague sent a screenshot and got HTTP 400 unsupported_capability. My gateway's capability table said this model can't take images. At launch I only tested that it refuses images, never whether it could see. Fix: fill the table from tests. Same screenshot, read correctly in 0.9 s. 2. Switch on, still broken. Next came 502 token_count_unavailable. My gateway counts tokens before admitting a request, so it can reserve context. The engine's chat endpoint takes images and video; its token counter only takes text. Fix: count the text, then reserve 1,100 tokens per image (the model uses at most 1,024). Two 1080p images: 1,951 tokens, 3.1 s. 3. The video didn't fit the request. A 30-second 720p clip is only 7,432 tokens. Tokens weren't the problem; bytes were. A 2-minute, 40 MB clip became a 53 MB request and got a 502. One request tops out near 32 MiB (31 got through, 33 didn't), and base64 makes video about a third bigger, so about 24 MiB of raw video at most. Links are off on the server, and I'm keeping it that way: a server that fetches any URL it's handed is an SSRF hole. 4. What I do now: pass the path. The video stays on disk in a folder the agent can read. I give pi or Harness the file path. The agent runs ffmpeg and pulls frames, the model looks at them, and ffmpeg cuts the original. The video never goes into the request, so neither limit applies. Test clip with cuts at 4.0, 11.5 and 15.0 s: both clients found all three cut points and cut out the 4.0–11.5 segment frame-exact, green from the first frame to the last (pi: 119 s, 21 image reads; Harness: 69 s, 16). Sending the video straight to the model put the cuts up to 0.4 s late. Fine for a rough cut, not for the final one. 5. "I checked the frames myself." The first pi run got the right clip and said it had looked at the frames. The session log said something else: the local model did look, but my own routing rule had handed the ffmpeg work to another coding agent, and the summary never mentioned it. The 119 s run above is the one where the task said "do it yourself". I still read the log before I believe a summary. Cheat sheet: - Test what the model sees; fill the capability table from that - Text-only token counter? Reserve a flat amount per image - Know your request size limit; base64 adds a third - Keep URL fetching off (SSRF) - Long video: pass a path, not the bytes. Say "do it yourself" - Read the session log before you trust "I did it myself" Mine can see now, and cuts its own clips. If you're running one of these: good luck, and have fun with it.

智能眼镜，值得拥有 或者搞个实时翻译的小硬件 都AI时代了对吧

1、你需要一个机柜（可以看我的主页置顶）。 2、CRS 812的风扇启动温度默认是52度还是58度来着，风扇调速曲线很陡峭，我调到62度基本上就不吵了。 3、DGX Spark基本没声音，反而是CRS 812是主要噪音源。 4、CRS 812的主要作用是支持6台机器，不需要物理调整线缆随意切换3+3，4+2，2+2+2等分组。满足不同的需求安排。

Claude Pro to Max～ 升级完毕～

“卧槽！”还是“真香！”，等我消息！！！ 终于还是来到了Claude升级时刻～～～ 三个Pro的额度全红了～ 先挑一个Pro升个Max试试～ 看看会不会坑我100刀～ Claude 我来了！！！ 啊对了，这三个Pro都已经活过了20天，其中两个没有触发KYC，我打算拿已经过了KYC的号升级～～～

关注一下claude的订阅负责人 讲的很清楚 还挺诚恳的

合同签完了，对方还可以随便改内容，云端API是不平等契约这件事，还有人有疑问么? Claude把20x变成了10x，这个事情和ChatGPT在一个月前干的一模一样。区别是ChatGPT明着说，Claude偷着做(改了个Ultra的名字，告诉你一切未变，细项条款默默改成了10x)。 同样是削减顶级Coding Plan的用量额度，根本原因也一样是算力的永恒不足，以及顶级Coding Plan的用户用的最狠，对公司来说最不经济，甚至连削减量都一样，砍半。 更有意思的是，削减的背景都几乎一样。ChatGPT是在Astra刚出来风头正劲，宣称AGI要到来的背景下；Claude是在Opus5.5霸屏近月打得ChatGPT抬不起头的状态下。 ChatGPT最后选择放弃额度削减，是因为模型能力拉胯且一直得不到改进，Claude的削减，我看不到反转的希望。 云端API涨价、加高安全护栏、监控用户数据是不可逆的趋势。

I run DeepSeek-V4.1-Flash-1017 across four DGX Sparks at home, on a 200G switch, behind my own gateway, with pi and DeepSeek Harness as clients. The reasoning slider in both clients looked fine. Five things were quietly wrong, one of them a 502. 1. Thinking on, answer empty. max_tokens 600. The model spent all 600 tokens thinking and came back with finish_reason: length and an empty answer. In my setup, thinking counts against max_tokens. Fix: when thinking is on, my gateway raises the cap to 32K, within whatever the context window has left. Nobody bills me for local tokens; I'd rather waste a few than get an empty answer. 2. Slider on High, zero thinking. pi and DeepSeek Harness both send reasoning_effort. My template only reads enable_thinking, and my gateway defaults thinking off. So the slider went nowhere. Fix: the gateway turns the slider into the template's on/off switch. 3. Turning thinking off threw a 502. Harness sends its off state as reasoning_effort: "off". The engine rejects that value, and my gateway passed it straight through. Fix: translate it to enable_thinking: false, then drop the original field. 4. Low, medium and high all ran at max. Once the switch worked, the model in Harness compared the three levels and said they looked identical. It was right. The engine had been started with --reasoning-effort max, which pinned every request with thinking on. Fix: map the levels (low → low, medium → high, high → max). One prompt, one run: 289 / 359 / 437 output tokens. With thinking off, the same prompt came out at 385, so low-effort thinking was shorter than none. Not a benchmark, just funny. 5. I asked the model who it was. It said GLM. The request log said DeepSeek handled it. GLM wasn't even running. A model's answer about itself is just more generated text. Check the request log. Cheat sheet: - Thinking on? Leave room in the output cap - Check what your client actually sends - Translate the slider; drop values the engine rejects - Check engine launch flags for a pinned level - Which model answered? Check the request log Mine thinks when I ask it to now. If you're running one of these: good luck, and have fun with it.

实际跑了一次任务，发现与@MiaAI_lab的配方数据相比，衰减了接近40%，开始排查原因吧。 这件任务上，Astra做主控，倒是安排的兢兢业业稳扎稳打，除了慢一点，我竟挑不出毛病。 还是要感慨，实操真的不容易

好消息：视频剪辑出来了，效果还不错！ 坏消息：速度没有达到预期，问题在哪儿还需要进一步排查，调优之路漫漫！

我开了Ultra所以在这个月巨量的使用了Grok 4.7。 优点： 量巨大，无比的大。 速度非常快，首字延迟几乎没有。我常常会想他是怎么做到的。 不会雷霆大思考，体感很爽。 我干脆把它设成Pi的默认主控模型和默认worker，接下来就是缺点了。 缺乏逻辑性，没有雷霆大思考，缺乏约束和回归，快速的掏出了答案但会遗漏。 在Pi这种轻量harness里暴露出模型本身的问题，最基本的worker调用总是出问题，简直不可理喻，在我做完鉴权的情况下，甚至调一家的cursor worker都会反复出问题，更不要说codex worker了。 工程能力，既然你提到了，computer use调用工具3D建模，打不过opus和astra没问题，也打不过sol。 Grokbot在我这儿的新鲜劲儿过完了，下月降级。

见证奇迹的时刻到了！ 测试开始！ 本地视频制作的永动机是否能够实现？ 去吧！皮卡丘（本地deepseek1017）！ 我是对你充满了信心的（建立在dgx spark*4的基础上）！

开始给我本地的dgx spark*4集群上的deepseek-4.1-flash 1017逐一落实图片能力和视频能力，适配到局域网的pi和deepseek harness 才发现会有这么多的坑要自己踩 踩过的坑慢慢总结 工程细节都是魔鬼！

I hooked four Sparks up to a MikroTik CRS812 this week. Seven things broke, most of them quietly. 1. Jumbo frames silently dropped. The CRS812 ports ship at MTU 1500 / L2MTU 1584. My Sparks were set to 9000. ping -s 1500 went through, -s 2000 was 100% loss, and nothing complained. Fix: L2MTU 9216 first, then MTU 9000, on every port. 2. hw-offload: false on all four ports. Only one bridge per switch chip gets hardware offload. I'd put the Sparks on extra bridges, so every packet was going through the switch CPU. Fix: every port on that one bridge, VLANs to split the Sparks into groups. 3. Fixing #1 killed both 400G ports. Changing L2MTU resets the whole switch chip (my management link dropped for ~2 s). Both 400G ports came back stuck at auto-init-failed / eeprom-checksum: bad. Disable/enable did nothing. Fix: auto-negotiation off, force 200G. Linked up in 10 seconds. 4. 200G links, 12.7 Gb/s actual. No drops, no pause frames, no retransmits I could find. Reloading the driver did nothing. It started after the port changes; I still don't know the exact trigger. Fix: rebooted all four Sparks. Back to 109. Dumb fix, but that was it. 5. 109 Gb/s on a 200G port is normal. Each Spark splits its 200G port into two PCIe Gen5 x4 halves, two RDMA devices on one ConnectX-7. One device tops out around 109. Run both at once and you get 196 down the cable. 6. Three Sparks ran slower than two. I moved two of them onto a 400G → 2×200G breakout cable. Reading a 495K-token prompt took 416 s on three nodes, 263 s on two. I tested one direction first and blamed the wrong box. Testing every pair both ways caught it: the two I'd moved received at 109 but sent at 12.7. Fix: rebooted those two. 210 s. 7. The new breakout port came up at MTU 1500. Again. Splitting a port gives you a brand-new port with factory settings. Walked right back into #1. Fix: same as #1. Cheat sheet (the video ends on it): - L2MTU 9216 + MTU 9000 on every port, split halves too - One bridge + VLANs, keep hw-offload on - 400G port won't link? Force 200G - L2MTU changes bounce the switch chip: don't do it live - ~109 Gb/s per device is normal (both: 196) - Touched a cable? Test every pair, both directions - Under ~109 per device: reboot that Spark - CRS812 fits up to 6 Sparks at 200G (2 split 400G + 2× 200G) Mine is finally behaving. If you're building one of these: good luck, and have fun with it.

你需要知道的是，法国在中国历史教育（法律规定的义务教育）中，是以革命源泉的形象出现的。 巴黎人民攻占巴士底狱（1989-07-14）和巴黎公社（1871-03-18至05-28）是两大人民觉醒的标志性事件。 这是每个上过高中的中国人都知晓的事情。 我从历史中学会了尊重过去的伟大国家和觉醒战斗的人民，却在信息平权的现在打碎这种尊重。