本來是解掉奇數的 case，後續則是由 gpt-5.4-pro 解掉了偶數的 case：

The even case, which Claude couldn’t finish, was then cracked by Dr. Ho Boon Suan using GPT-5.4 Pro to produce a 14-page proof for all even m≥8, with computational checks up to m=2000.

然後是 GPT + Claude 再給出另外的 construction 同時可以用在一般性的 case (同時用在奇數與偶數)：

Soon after, Dr. Keston Aquino-Michaels used GPT + Claude together to find simpler constructions for both odd and even m, by using the multi-agent workflow.

整個問題在 LLM 大幅參與幫助的情況下解掉了。

其中 1-bit Bonsai 8B 這個 model 的介紹是這樣：

The first commercially viable model with 1-bit weights. Requiring only 1.15GB of memory, 1-bit Bonsai 8B was engineered for robotics, real-time agents, and edge computing. It has a 14× smaller footprint than a full-precision 8B model, runs 8× faster, and is 5× more energy efficient, while matching leading 8B models on benchmarks. This results in over 10× the intelligence density of full-precision 8B models.

原版的 llama.cpp 無法使用，會需要裝 fork 出來的版本：

gguf_init_from_file_ptr: tensor 'output.weight' has invalid ggml type 41. should be in [0, 41)
gguf_init_from_file_ptr: failed to read tensor info

另外 Simon Willison 有提到可以在 iPhone 上面用「Locally AI - Local AI Chat App」跑，我試著抓了下來，看起來 iPhone SE3 + iOS 18 沒辦法跑，不過其他模型可以跑，不確定目前是什麼情況。

最近冒出好多小模型可以玩...

Every Mac with Apple Silicon has a built-in LLM - Apple's on-device foundation model, shipped as part of Apple Intelligence. Apple provides the FoundationModels framework (macOS 26+) to access it, but only exposes it through Siri and system features. apfel wraps it in a CLI and an HTTP server - so you can actually use it. All inference runs on-device, no network calls.

Hacker News 上的討論有提到「Updates to Apple’s On-Device and Server Foundation Language Models」這個頁面，裡面就有提到在 device 端的模型參數量大約 3B：

Our latest foundation models are optimized to run efficiently on Apple silicon, and include a compact, approximately 3-billion-parameter model, alongside a mixture-of-experts server-based model with a novel architecture tailored for Private Cloud Compute. These two foundation models are part of a larger family of generative models created by Apple to support our users.

跟 id=47628820 這邊提到的感覺差不多：

AFAIK the current model is on par with with Qwen-3-4B, which is from a year ago [0]. There's a big leap going from last year Qwen-3-4B to Qwen-3.5-4B or to Gemma 4.

另外可以看到 context 不大，主要是拿來處理簡單的任務。

新聞稿裡面有提到三個壟斷行為 (中文是 gpt-5.4-mini 翻譯的)：

When Amazon and a competitor are price-matching one another, resulting in a downward pressure on price, either Amazon or the competitor agrees, through their common vendor, to increase the retail price or make the product temporarily unavailable, so that the other retailer can match the increased market price, increasing the price for consumers.

當亞馬遜（Amazon）與競爭對手彼此進行價格匹配，導致價格承受下行壓力時，亞馬遜或競爭對手其中一方會透過雙方共同的供應商同意提高零售價格，或讓產品暫時無法供貨，如此一來，另一家零售商就能跟進較高的市場價格，進而提高消費者的購買價格。

A competitor offering a discounted price on a product will increase its retail price at Amazon’s request (a request made through the vendor), so that Amazon can then match the increased retail price, thereby increasing the price for consumers.

一個競爭對手若以折扣價提供某項產品，亞馬遜會透過供應商要求其提高零售價格，讓亞馬遜接著能夠跟進調高後的零售價格，進而使消費者支付更高的價格。

The vendor removes a product from a competing retailer who is offering a lower price than Amazon, so that the lower price is no longer available in the market and Amazon then raises its retail price, resulting in a higher price for consumers.

賣家將一項商品從一家競爭零售商下架，而該零售商當時提供的價格比 Amazon 更低，讓市場上不再有這個較低價格，接著 Amazon 提高其零售價格，導致消費者得付更高的價格。

因為這些行為對消費者的傷害是持續的，所以加州政府 (檢察總長) 申請了初步禁制令：

California Attorney General Rob Bonta last night filed a request for a preliminary injunction in California’s existing case against Amazon for price fixing.

不過過了一個月，目前有翻到的是「California judge questions timing of price-fixing injunction in Amazon suit」與「Legal standard in Amazon antitrust case revisited with Calif. bid for injunction」這兩篇，但因為 paywall 的關係不確定後續的細節，目前媒體好像沒看初步禁制令生效的報導...

TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.

查了一下 Amazon 搞的 Chronos 系列也算是有力競爭對手，不過 TimesFM 走 decoder-only 架構，Chronos 走 encoder-only 架構，兩邊看起來效果都不差，是可以站上同個擂台的。

目前手邊的應用中想到的是 anomaly detection 的部分，如果基本的 z-score 不夠用的時候可以拿這個來跑看看...

Copilot is for entertainment purposes only. It can make mistakes, and it may not work as intended. Don’t rely on Copilot for important advice. Use Copilot at your own risk.

看了一下網址，okay，是微軟官網沒錯，網址看起來是 Terms of Use 沒錯...

其他家的免責大多就只講無法保障 AI 產生出來的東西的正確性，微軟寫的這麼... 直接？還是這份 Terms of Use 也是 Copilot 產生的...？

路透社的報導裡面提到 Clarifai 是人臉辨識公司，不過在其他文件沒看到有提到 Clarifai：

The ​FTC said OkCupid users were never ⁠told their information - including nearly 3 ​million photos, demographic information and location ​data - would be shared in 2014 with Clarifai, a facial recognition technology company, contrary to ​OkCupid's privacy policies.

本來以為 Clarifai 是丟圖進去直接尋找出哪個人 (肉搜？)，但好像跟想像中的不太一樣，應該是買大量資料來 train model 的公司？像是在「Facial Recognition」這邊提到的？

不過這樣也已經夠糟了...

private 的重點在這：

This program does not use:

* Content from your issues, discussions, or private repositories at rest. We use the phrase “at rest” deliberately because Copilot does process code from private repositories when you are actively using Copilot. This interaction data is required to run the service and could be used for model training unless you opt out.

也就是說，即使 GitHub 知道這是 private repository，而且也有能力 opt-out 這些資料，GitHub 在你沒有 opt-out 的時候會以 interactiion data 的名義將這些資料拿去 training。

依照說明，關掉的方式是到 https://github.com/settings/copilot/features 這邊：

Cory Doctorow 有提過 Enshittification 這個概念：

Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die. I call this enshittification, and it is a seemingly inevitable consequence arising from the combination of the ease of changing how a platform allocates value, combined with the nature of a "two-sided market", where a platform sits between buyers and sellers, hold each hostage to the other, raking off an ever-larger share of the value that passes between them.

gpt-5.4-mini 的翻譯：

平台就是這樣走向死亡的：一開始，它們對使用者很好；接著，為了讓商業客戶（business customers）獲得更好的服務，它們開始壓榨使用者；最後，連那些商業客戶也被壓榨，平台把所有價值都重新攫回到自己手上。然後，它們就死了。我把這種現象稱為「爛化」（enshittification）。這看起來幾乎是不可避免的結果，源自於兩個因素的結合：其一是平台很容易改變自己如何分配價值；其二是「雙邊市場」（two-sided market）的特性——平台夾在買方與賣方之間，把雙方彼此綁住當人質，同時從他們之間流動的價值中抽取越來越大的一部分。

長遠來看還是要想辦法從 GitHub 拆出來自己 host...

透過 dependency + postinstall 執行 malicious code：

The malicious versions inject a new dependency, plain-crypto-js@4.2.1, which is never imported anywhere in the axios source code. Its sole purpose is to execute a postinstall script that acts as a cross platform remote access trojan (RAT) dropper, targeting macOS, Windows, and Linux.

Hacker News 上的 id=47582632 提到了最近的 mitigation，設定只有在 release 超過 n 天才會被計算進來的方式，避免第一時間中獎，裡面提到包括 Python ecosystem 的 uv 與 JavaScript ecosystem 的工具都有支援了：

  ~/.config/uv/uv.toml
  exclude-newer = "7 days"

  ~/.npmrc
  min-release-age=7 # days
  ignore-scripts=true
  
  ~/Library/Preferences/pnpm/rc
  minimum-release-age=10080 # minutes
  
  ~/.bunfig.toml
  [install]
  minimumReleaseAge = 604800 # seconds

七天有點多，我覺得一天可能是比較合理的數字，極限的話三天？另外這個 comment 的作者也直接抱怨 JavaScript 這堆工具在搞事，剛好用了三個單位 XDDD

(Side note, it's wild that npm, bun, and pnpm have all decided to use different time units for this configuration.)

留言裡面 pixza 有提到 Google Maps 可以用 https://www.google.com/maps?force=canvas 這樣的網址強制跳到 canvas 版本。

目前把 gfx.canvas.accelerated 設為預設值 true 後看起來沒問題了，這樣少了一個要設定的...

話說 NoScript 造成的問題太多，常常在發現網站有奇怪的行為後 debug 發現都是 NoScript 造成的，在 Google Maps 可以透過 force=canvas 切換的情況下，看起來是可以拿掉 NoScript 了...