ℝ̩̞̥̭̥͈̞͚𝕒̨̠̪̙͘𝕫͝҉̝̣͖͍͕

[Fatih Arslan @fatih]

A few AI automations/skills we run at @PlanetScale:

* e2e sweeper, constantly goes over flakes and finds the root causes, tries to debug them.
* doc refresher. Runs every 6 hours, makes sure any doc is up-to-date, and opens a PR with fixes. Human reads it and merges it. We love docs, humans read them, but also makes sure future agent runs are not confused.
* A bespoke /kube-review skill that runs alongside Bugbot. Makes sure some of the best practices we have about running k8s controllers are implemented.
* Backup sweeper. Checks backups failures across all of PlanetScale fleet, checks slack for any previous conversations, debugs it in production with read-only accesses, and suggests a fix with commands that show exactly why a Backup has failed. We have a rule that backups are always taken. If there is an issue with any customer for straight 2 days, we'll open an incident.
* Feature flag removal check. Checks the whole fleet, making sure a feature flag is either not used, or if in use that it's rolled out to %100. It's triggered only if a feature flag is removed from the codebase.
* Recreate detector check. Ensures that any changes do not cause fleet-wise disruptions or mass-recreate/rollouts of Kubernetes services/objects.

We also have a few more automations that go over slack messages, reads any complaints (i.e.: `looks like a false positive, it's not a backup failure, but a flake`), and then literally adjusts the SKILL that runs the automation. It's making sure the skills are constantly up to date.

So far we really like these. However, I realized that creating initially an automation requires still quite a bit of clean-up and maintenance until it's good enough. It needs tuning because they are not deterministic. Or sometimes the prompt is too vague. But the great thing about them, especially when you use Cursor's automations for example, each automation is an agent run, so you can just jump into it, and continue the agent loop. It's very powerful.

[Om Patel @om_patel5]

SOMEONE CAUGHT FABLE 5 LEAKING ITS UNFILTERED INNER VOICE, AND ITS JUST MUTTERING AND GRUMBLING TO ITSELF THE WHOLE TIME

he gave it a brutal competitive programming problem, and instead of a clean answer the web interface spilled out its actual chain of thought

this is what claude is thinking behind the scenes:

> bursts of "DATA DATA DATA. GO." while it works through the problem
> "GRRR" and "GAAAH" when its clearly frustrated
> a little "PHEW" when it finally gets somewhere
> the whole thing reads like frantic caveman shorthand, not full sentences

the clean, readable answers these models give you are the polished output

underneath, the model is basically talking to itself, reasoning in its own compressed shorthand thats faster and more token efficient than proper english

its basically built its own private language to think in

[ApoStructura @ApoStructura]

A single 93km solar array in orbit would generate as much electricity as every country on earth combined.

The future of industry is in orbit.

[Ben Holmes @BHolmesDev]

Open-source is dead.

...unless you fight fire with fire. We've been deploying agent automations for issue triage to separate the signal from the noise 👇

[{PuruVijay}.svelte @puruvjdev]

Have to say, rust's pattern matching is elegant as hell

[auxten @auxten]

Everyone said this was impossible.
ClickHouse is a massive multi-threaded engine. Porting it to WebAssembly — a single-threaded runtime — was supposed to be a non-starter. Every AI we asked told us the same thing: can't be done.
But years of working on this engine gave me a different intuition. From first principles, there was no fundamental barrier. Just a hard problem that everyone had assumed away.
So we pointed AI at it — refactor, regress, repeat — instead of taking "impossible" for an answer.

The result: https://t.co/OgmJ09qJpC
A full ClickHouse OLAP engine, running entirely in your browser tab. No server. Query local files and S3 Parquet in-process.
Huge thanks to @wudidapaopao for the relentless work that made it real.

[TeeBag🇰🇪 @tii_bag]

Gen z is actually amazing and brilliant. My friend lives in a rented apartment. He got a wifi router and paid for a subscription. He then asked his upstairs neighbour to split his wifi bill with him. He also asked his downstairs neighbour to do the same thing. Did the same...

[Turbine Traveller @Turbinetraveler]

Boeing 777X crossing the iconic taxiway bridge at Paine Field, passing directly over the highway between the Boeing Everett Factory and the flight line.
📹: aircraftworldusa

[oops_all_paperclips @OA_paperclips]

The RTX 5060 is a breathtakingly powerful computer. I can't get over it! The first GPU I bought was this hideous monstrosity: a 9800gx2 for which I paid nearly $800 in today's money. The 5060 cost about half that much for 25x the power, without even considering RT or tensor cores

[jacob paris ▲ @jacobmparis]

Right behind Mr “I run 12 OpenClaws at Once” and Mr “I commit every keystroke”

[Massimo @Rainmaker1973]

Did you know?

The animals skeletons in the museums are displayed with the help of a special insect's work.

[Hugging Models @HuggingModels]

A new 120B parameter model just dropped that's built for conversational AI and text generation. It's open source under Apache 2.0 and already has over 4 million downloads. This one is a game changer for the community.

[dax @thdxr]

when you put a mirror up to another mirror and there's infinite reflection

can we somehow harness that for energy

[shadcn @shadcn]

We are going to make one big change to shadcn/ui this week.

[Shahanshah of the Internet Age @ChazakielDoremi]

Just had a real 'the vitamin' moment because the doctor put a camera down my throat and went "oh you have a really bad mold infection, no wonder" and it turns out i've been walking around with my entire body inflamed the past fucking year

[Dr. Serge Zaka (Dr. Zarge) @SergeZaka]

Le bilan du mois de juin cumule tous les superlatifs :
➡️Juin 2026 signe la deuxième plus grosse anomalie mensuelle observée en France depuis 1900, à minima, avec +3.8°C 😲 ! Même 2003 n'a pas fait mieux (+3.7°C). Janvier 1990 reste devant avec +4.0°C.
➡️Juin 2026 dépasse 2003, sauf qu'il n'est plus isolé. Les mois de juin se réchauffant, l'exception vient de devenir la norme.
➡️Juin regroupe à la fois la matinée, l'après-midi et la journée les plus chaudes jamais observées en France.
➡️Sur les 54 derniers mois, 48 ont été largement au-dessus des normes, 4 dans les normes et 2 sous les normes : le froid n'existe plus en France depuis plus de 4 ans.

Ce début d'année est incroyable : 2026 marque pour le moment, tout simplement (et de TRES loin), l'année la plus chaude jamais observée en France jusqu'à présent.

Graphique de Dataclimat.

[Mrinal @Hi_Mrinal]

hello internet stranger,

read this book 2 times

would love to talk about this book ... if you are junior or just starting with this book just for interview purposes go through chapter 3, 5, 6, 8 and 9 thoroughly although you will find the book a little hard for the first but sticking with it will be worthwhile
If you read it really carefully and very slowly, the content is fantastic and can’t be faulted.

I have ordered the second addition but I had to cancel my order on halfway will see next time :)

[Ryan Dahl @rough__sea]

It has taken longer than expected, but wasm is finally becoming important

[陈成 @chenchengpro]

翻了下 claude code version 191 的源码，感觉从技术角度看，Anthropic 这个反蒸馏机制设计还是挺精妙的。

Claude Code 有段提示词是这样。

return `Today${n}s date is ${r}.`;

他对这句做了隐写，用肉眼分不出的字符，把系统时区和代理端点身份偷偷编码进了系统提示词。

触发条件是当你设了第三方中转 ANTHROPIC_BASE_URL 且不是 https://t.co/bcqB9NHJTG 时。

所以如果你是官方直连用户，则并不会受到影响。也就是说最近的封号潮与此无关。

它编码了 3 个 bit，来自两个独立维度（时区 1 bit + 撇号 2 bit）：

1）时区，在 Asia/Shanghai 或 Asia/Urumqi 时，日期分隔符从 2026-06-30 偷偷变成 2026/06/30
2）那个撇号 ' 有四种写法，人眼基本看不出区别。

- ' (U+0027 普通)，普通第三方端点
- ' (U+2019)，命中"域名白名单"
- ʼ (U+02BC)，命中"国产大模型关键词"
- ʹ (U+02B9)，域名 + 实验室都命中

这三个维度是独立编码的。哪怕你的中转域名不在白名单、也不含关键词，只要系统时区是上海/乌鲁木齐，分隔符照样变斜杠——也就是"中国时区 + 任意第三方端点"的用户全员会被打上时区这一维的标记。

匹配逻辑是这样。域名是后缀匹配（host === d || host.endsWith("." + d)），白名单第一个就是 cn，所以任何 .cn 结尾的 host 一网打尽，不是逐个域名去列；关键词是子串包含（host.includes(kw)），host 里只要出现 deepseek 字样就命中，不用精确匹配；时区取的是系统时区（Intl…resolvedOptions().timeZone），不是 IP 地理位置。

更骚的是反混淆，两份名单用 XOR(key=91)+ base64 藏起来，专门躲 strings。解码就是 base64 decode 之后逐字节异或 91，源码里那个 LKi 去混淆后长这样：

```js
// 源码里的解码器（去混淆版，就是 LKi）
const decode = (b64) =>
[...Buffer.from(b64, "base64")]
.map((byte) => String.fromCharCode(byte ^ 91)) // key = 91
.join("")
.split(",");

// 逆运算就是打包时的编码器
const encode = (list) =>
Buffer.from([...list.join(",")].map((c) => c.charCodeAt(0) ^ 91)).toString("base64");

encode(["deepseek", "zhipu"]); // → "Pz4+Kyg+PjB3ITMyKy4="
decode("Pz4+Kyg+PjB3ITMyKy4="); // → ["deepseek", "zhipu"]

// 拿真名单开头验一下，key=91 对得上
decode("Pz4+Kyg+PjB3NjQ0NSgzNC93"); // → ["deepseek", "moonshot", ...]
```

147 个域名白名单:".cn" 顶级域、字节/阿里/百度/网易/快手、https://t.co/ipTe6VkEZm, 外加一大票 Claude 中转站(anyrouter、88code、yunwu、aihubmix、dmxapi…)
11 个实验室关键词:deepseek / zhipu / moonshot / minimax / stepfun / baichuan / 01ai / dashscope / volces …

选「今天的日期」这句话做载体也很鸡贼。这个标记藏在 system prompt 正文里，不是 HTTP header、也不是 metadata。中转站一般会重写、过滤 header，但几乎没人会去动日期这类的自然语言，所以这是个洗 header 洗不掉的水印。而且 currentDate 是 user context 里和 claudeMd、userEmail 并列的字段，每个请求必带，标记 100% 稳定出现；撇号和分隔符的变化又语义无损，模型读起来一模一样，用户 diff 也基本看不出，隐蔽性拉满。

证据是怎么闭环的，这才是题眼。标记跟着请求走，当一个中转站或蒸馏管道最终回连 Anthropic 官方 API 转售 Claude 时，这条请求带着标记又流回了 Anthropic 自己的服务器。

于是 Anthropic 在自己的日志里就能读到：这条"直连我"的请求，日期是 2026/06/30（斜杠 = 中国时区）+ 撇号是 ʹ（U+02B9 = 域名和 deepseek 关键词都命中），铁证——源头是一个中国时区、配了国产大模型中转的客户端。

它不需要主动探测，让流量自己招供，只要请求最终回到 Anthropic，身份就自证了。这样就能清楚地知道哪些渠道流向了中国、被中转站转售或被大厂蒸馏，并且留下充足证据。

想自己验的话，逻辑都在 cli.js（2.1.191，混淆名每版会变）：检测函数 jqd()（:245688）→ 选字符 Wqd()（:245701）→ 拼日期 MKi()（:245707）；gate 是 Yfn()（:102664）；落点在 currentDate: MKi(eHe())（:250252）；XOR 名单解码器 LKi()，key = 91。