而且價位其實不貴，英國版是 £55.00，日幣賣 ¥8,800 (CASIO Collection STANDARD F-B100W-1AJF)。

不是充電的，但使用的鈕扣電池 CR2016 大約可以用兩年：

Power supply and battery life
Approx. battery life: 2 years on CR2016

可以透過 app 同步時間，如果不同步的話，精確度大約是每個月約差 15 秒內：

Accuracy
Accuracy: ±15 seconds per month (with no mobile link function)

以目前資料看起來應該就只有步數，沒有其他的資訊，主要是賣這個設計... 超懷念 (?)

用 coding agent (codex) 找台灣的情況，在「電信設備審驗合格清單」這邊可以查到目前通過的清單，台灣在 2026/08/18 已經過型式認證了 (CCAQ24LP0144T0)，申請人是台灣卡西歐股份有限公司，驗證機構是優力國際安全認證有限公司：

LP,24,CCAQ24LP0144T0,CCAQ24LP0144T0,Watch,CASIO,F-B100W,台灣卡西歐股份有限公司,,優力國際安全認證有限公司,,N,N,,,20260818,,,,Y,

看起來應該會是會有公司貨...

這次是新任 CTO Vlad Fedorov 貼的：「The August 17 outage, and the work ahead (via)」。

裡面幾張圖就蠻能解釋問題了，首先是各種指標，包括 merged PR 數量、commit 數量、新的 repo 數量，都可以看到 AI 時期的爆炸性成長：

然後提到美國中部機房沒撐住導致 authentication failure：

Our investigation found that the outage began when traffic reached a new peak, and a critical infrastructure component in our Central US data center failed to scale with it. The resulting capacity pressure spread through our systems, causing authentication failures and disrupting multiple GitHub services.

他說明這些成長壓力不是炸掉的直接主因 (上面提到的 authentication 才是，但其實也跟量有關...)：

Since April, monthly commits have grown from 1.4 billion to 2.9 billion. That growth explains the pressure on our systems, but it does not excuse these outages.

然後打了母公司 Azure 的廣告，提到大多數的擴充都是透過 Azure 撐起來：

Today, Azure serves roughly 58% of GitHub’s platform load and half of all Git operations, up from 12% of platform load in May.

但看起來還是有關鍵的基礎設施沒進到 Azure 擴充...

話說很少在領跑地位還看到這種等級的成長，算是受益者還是受害者...？

實務上 KV cache 以及其他的東西加起來差不多與 model 佔同樣大小，所以 32GB VRAM 大概跑 30B 左右的 4-bit quantization model 還 OK，最近幾天很紅的就是 Qwen3.8-27B 了，可以看到各家都有出來轉檔。

一般的 quantization 是直接對每個參數轉換，而 Unsloth 針對 GGUF 格式支援動態取 quantization 的技術，挑出「比較重要的參數」給予更多的 bit，用少量額外的空間來換取更好的精確度，而這次看到 3.0 版的技術，看起來又再比之前 2.0 提升了一些：「Unsloth Dynamic 3.0 GGUFs」。

官方說 unsloth/Qwen3.8-27B-GGUF 有放上 3.0 版，剛剛試著重跑 UD-Q4_K_XL 版本就發現重新下載了，應該是直接用就可以了...

當時有些人在猜是不是因為現在 AI 的金流太大，而 OpenAI 剛跑掉所以直接綁樁：(id=49324302)

I wonder if this deal is primarily just to buy payment volume.

OpenAI just announced earlier this week that Ayden would become their payment provider (when it was previously Stripe).

And OpenRouter has a large percentage of overall AI payment volume for all the major labs.

Both OpenAI and OpenRouter represent ~$100B in payment volume, whereas Stripe in total doing ~$2T. Two customer doing ~5% of your total volume who didn’t even exist a few years ago, must be kind of scary for Stripe.

但不是很確定，畢竟這種併購都是要談蠻久的，時間線未必對的上... 從新聞稿也看不出什麼東西。

Hacker News 上有些人提到中國的 model 競爭性，不過中國的 model 已經出來一段時間了，我覺得時間點上真的有給出壓力的應該是 Grok 4.6 吧，美國籍的 model，但 $2/$6 對上 GPT-5.6 Sol 與 Opus 5 的價錢實在太殺？

我是跑去買 subscription plan，這一個禮拜用起來發現 CP 值不錯，方案的層級是 $10/mo、$30/mo、$100/mo 然後是 $300/mo，升級的步調大約都是三倍左右，不像其他兩家 $20/mo 後就直接跳 $100/mo 了。

不確定這個 discount 會維持多久，但目前看起來有愈來愈多 model 跟上來了...

因為設計上 system prompt 放在 client 端，所以先前有人用 MITM proxy 抓過，不算什麼秘密... 但比較好奇的是，這種東西應該是放在 server 端才對？當初設計放在 client 是什麼原因？為了通用 API 的設計嗎？

回到 Hacker News 上面，有人找到不少有趣的內容，像是在 Opus 5 與 Fable 5 的 system prompt 補上被美國政府禁止提供服務的知識：

Claude Fable 5 and Claude Mythos 5 were first released on June 9, 2026. On June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls; the Department lifted those controls on June 30, 2026, and Anthropic restored access on July 1, 2026 (Anthropic's statement: https://www.anthropic.com/news/fable-mythos-access). These events are after Claude's training-data cutoff, so Claude knows about them only from this notice. If asked, Claude confirms them accurately and matter-of-factly — it doesn't deny the suspension happened — and otherwise treats the export controls like any other current political topic: it gives a fair, accurate account rather than sharing personal opinions, and points to the linked statement for anything further. Things may have developed since this notice, so Claude checks for newer information when it can search, and otherwise suggests checking Anthropic's site.

還有 Opus 4.6 關於 Trump 選上總統的部分：

There was a US Presidential Election in November 2024. Donald Trump won the presidency over Kamala Harris. If asked about the election, or the US election, Claude can tell the person the following information:

Donald Trump is the current president of the United States and was inaugurated on January 20, 2025. Donald Trump defeated Kamala Harris in the 2024 elections. Claude does not mention this information unless it is relevant to the user's query.

會選擇加這些知識進 system prompts 大概是 PR team 的想法吧？一般 engineering team 應該對這個比較沒興趣...

找資料的時候發現維基百科的「Elevator algorithm」這個條目，反而提到了硬碟中磁頭的演算法：

The elevator algorithm, or SCAN, is a disk-scheduling algorithm to determine the motion of the disk's arm and head in servicing read and write requests.

主要是因為兩者的行為很像？磁頭在跑的過程中會有其他的 i/o 需求進來，就像是電梯的人潮：

This algorithm is named after the behavior of a building elevator, where the elevator continues to travel in its current direction (up or down) until empty, stopping only to let individuals off or to pick up new individuals heading in the same direction.

From an implementation perspective, the drive maintains a buffer of pending read/write requests, along with the associated cylinder number of the request, in which lower cylinder numbers generally indicate that the cylinder is closer to the spindle, and higher numbers indicate the cylinder is farther away.

不過也提到現代的 OS 對硬碟的操作已經跟這些演算法無關了，SSD 的存取已經是固定時間，而且就算是傳統機械硬碟，存取架構已經從傳統指定物理位置的 CHS 變成邏輯位置 LBA 了：

The algorithm is largely obsolete for data storage. With the current generation of magnetic disks it is not possible to know the location of specific data on the disk and solid state memory devices have a constant seek time independent of location.

但回到電梯的情境，這些演算法還是存在需求的，從維基百科上面可以看到 SCAN 類的演算法與 LOOK 類的演算法，以及多台時的變化。

當作有趣的知識來看...

Hacker News 上面給了某些角度的分析，考慮到新墨西哥州大約 212 萬人口，未成年人口大約 44 萬人，累積的罰款到 9.42 億美金可以看出來有多高：

A lot of people on this story are making the usual snarky remarks that, oh, this is just a slap on the wrist—some trivial proportion of Meta’s global revenue. And that is true, but unlike the EU, for example, New Mexico is a tiny jurisdiction. It only has a little bit more than 2 million people.

So actually, a fine—or a judgment—of $942 million that applies just to New Mexico is enormous. If you look at Meta’s revenue from the U.S. and Canada and assign it proportionally by population, you find that Meta only earned about $1.5 billion to $2 billion total from New Mexico over the past five years. The exact figure depends on whether you allocate purely by population or account for purchasing power, which makes the estimate a little lower because New Mexico is a poorer state than the U.S. average.

不愧是電子毒品平台...

先前有提到「在桌機上擋 Facebook 各種廣告與演算法推薦文章的 userscript」，不過作者後來停止更新了，在「Is this userscript still maintained #85」這邊有提到需要幾個月的時間恢復：

I will be updating in a few months. I am currently incapacity and will take a few months to recover.

但後續沒看到回來，目前是有在 Artificial-Sweetener/facebook-clean-my-feeds 這邊看到 fork 出來的版本...

在這些網拍上面會看到不少便宜的電視棒，其實是把木馬放回家裡了，這些電視棒會變成廣告惡意點擊的跳板，或是直接變成 residential proxy 的跳板：

In addition to enlisting the user’s TV box in ad fraud networks, these off-brand streaming devices almost universally come with residential proxy software pre-installed.

這邊研究室透過註冊過期 C2 server 的網域後去觀察的：

Pedro Falé is a threat researcher with the security firm Bitsight. Falé told KrebsOnSecurity he was able to peer inside a vast and complex ad fraud network by registering an expired domain name that was used to coordinate fake ad clicks across a particularly popular brand of these streaming devices known as H96.

然後光是這個過期的 domain 就可以追蹤到大約 38k 個 device，在這些灰色 & 黑色產業上面就可以產生 US$50k/day 左右的利潤：

Bitsight said it tracked approximately 38,000 TV boxes globally phoning home to the expired Fengwo Group domain, and based on that number the report estimates this ad fraud network brings in revenues of close to $50,000 a day (not counting substantial revenue from the residential proxy side of the business). However, Falé emphasized that these estimates are highly conservative and based on telemetry from just one of the Fengwo Group’s core (but older) domains.

雖然知道問題，但要擊破其實很難，像是 multi-domain backup + signed commands + onion hidden services 之類的...