YOLO Tendie

Autistic trader. 1000x my paper trade account. Will take all comments as financial advice. Thank you for your attention on this matter.

[鄭vibin' Passion- @zhngwibn284976]

猫猫女友也太可爱了吧
怎么就蹦上来了呀

Prompt：
9:16竖屏，时长10.0秒，60fps，真实手机人像第一人称互动质感，单镜头连续拍摄。年轻东亚女性，白皙透亮肤色，黑色长发微卷垂落。

【人物与视觉妆造】

严格保持上传图片中@image_1 的人物脸部、五官、发色、肤色、年龄感和身材比例一致。

视觉妆造：
精致通透的“眼泪妆/微醺妆”风格，搭配粉嫩系眼影腮红、水光玻璃唇釉与水汪汪的琥珀色通透美瞳，眼角带细闪，强化像小猫一样好奇、灵动、可爱黏人的软萌感。

【场景与拍摄机位】

人物位于：

【温馨明亮的现代公寓客厅与走廊拐角】

场景真实自然，包含温馨木质地板、布艺沙发与走廊墙壁转角。

使用 第一人称POV主观视角、“躲猫猫（Peek-a-boo）”探头运镜。

转角白色墙壁作为最关键的前景遮挡道具，镜头通过“探头 → 缩回墙后纯白遮挡 → 再次探头”形成类似小猫步步逼近的阶梯式推进视觉。

【动作节奏规则】

10秒内完成 4 次探头与缩回的“小猫步步靠近”游戏，女主每次在墙壁遮挡后瞬移式向前拉近距离，最后一次直接飞扑抱住镜头，节奏欢快干脆。

【秒级结构化时间轴】

0.00–2.00秒｜第一探：沙发对视与好奇转头（Distance: 远景）
第一人称视角从走廊拐角探出头。远景处女主正盘腿坐在客厅沙发上看手机，听到脚步声后好奇转头，两人视线瞬间锁定对视，女主眼睛微亮露出惊喜神色。
（第 1.8–2.0秒：镜头快速向后缩回墙壁后，转角白墙完全挡住画面）

2.00–4.00秒｜第二探：离开沙发·靠近一步（Distance: 中远景）
镜头再次探出。女主已经从沙发上跳下，身形瞬移靠近到了客厅地毯中央，微歪着头，双手背在身后，像警觉的小猫一样充满好奇地眨着大眼睛打量镜头。
（第 3.8–4.0秒：镜头迅速缩回墙壁后完全遮挡）

4.00–6.00秒｜第三探：猫步逼近·调皮探头（Distance: 中景）
镜头第三次探出。女主已经靠近到走廊入口中景位置，身子微微前倾，一只手轻托下巴做搞怪思考状，眼神带着恶作剧前夕的狡黠调皮与甜笑。
（第 5.8–6.0秒：镜头再次飞快缩回墙壁后）

6.00–8.00秒｜第四探：近在咫尺·蓄力准备（Distance: 特写近景）
镜头第四次探出。女主已经站在离拐角仅一步之遥的极近位置，双手握成可爱小猫爪放在胸前，身体微微下蹲做出“小猫扑食”的蓄力晃动姿势，眼神锁定镜头满是坏笑。
（第 7.8–8.0秒：镜头慌忙缩回墙后）

8.00–10.00秒｜第五探：飞身前扑·树袋熊抱住镜头（Hug Attack!）
镜头最后一次小心翼翼探出头。女主瞬间如轻盈小猫般向前猛地一蹦，双手直接张开环抱住第一人称镜头（脖子），整张精致甜美的脸庞瞬间贴近镜头正前方，双眼笑成月牙，整个人挂在身上开心地蹭了蹭，画面在极致心动与甜笑中定格结束。

【强制统一规则】

统一保持：
10秒、60fps、9:16竖屏、第一人称躲猫猫POV、墙角遮挡阶梯式推进、最后飞扑熊抱高甜收尾。

探头缩回节奏干净利落，女主距离拉近与小猫般的肢体微动作生动连贯，完美呈现极具沉浸感与互动趣向的“猫系女友”日常。

[Smartphones sans Google (GAFAM) @Sans_Google]

If you’re still on stock Android…

Google silently installs SafetyCore (on-device sensitive content scanning)
and it reinstalls itself every time you try to remove it.

The community’s solution is simple and effective:
a placeholder APK with an absurdly high version number (2,000,000,000).

Once installed, Google can no longer overwrite it.
1. Uninstall SafetyCore

2. Install the placeholder

3. Keep control of your device

GitHub: https://t.co/enSZbjA2Ho

Even blocking a Google component requires workarounds on stock Android.

Have you tried it? Does it hold up over time for you?

#Privacy #DeGoogle #Android #SafetyCore #DigitalSovereignty

[YOLO Tendie @YOLOTendie]

@dphpu1988 為什麼他們可以繼續做男娘而完完就不能了？

[YOLO Tendie @YOLOTendie]

Goose's Travelogue: Hidden Mountain https://t.co/WDSzMt6Ko8

[Max For AI @MaxForAI]

这两天海外 AI 圈最火的开源项目，就是这个专门「删除 AI 水印」的 watermarks-remover。

它做的并不是传统意义上的那种图片去水印，而是把 Claude、Gemini、OpenAI 等 AI 生成内容里可能存在的各种「标记」清掉。

项目刚开源没几天，GitHub 已经冲到 11k Star，作者自己的帖子拿到了超过 200 万曝光。

目前项目主要处理三层信息：

第一层是文本中的隐形标记。
包括 Invisible Unicode、特殊空格、双向文本控制字符和其他隐藏字符。
这类东西不影响人类阅读，但程序可以直接检测。

第二层是统计型文本水印。
部分模型可以通过改变 Token Sampling 的概率分布，让生成文本在统计上呈现特定模式。
watermarks-remover 的处理方式更复杂，需要通过 Agent 对文本进行重写，在尽量保持语义的情况下破坏原来的统计特征。

第三层是文件级标记。
包括 C2PA、EXIF、XMP、DOCX/PDF 文档属性等元数据。
支持 PNG、JPEG、WebP、SVG、PDF、DOCX、ODT、HTML、Markdown 等格式。

为什么这东西突然这么火？

因为最近 AI 水印这件事在海外争议非常大。

Anthropic 已经明确表示，为了满足欧盟 AI Act 相关透明度要求，8 月 2 日之后在欧盟上线的新 Claude 模型会支持机器可读标记，文本会加入嵌入式水印，文件也会加入 provenance metadata。

Meta、Goole和OpenAI也宣布会在其模型生成的内容中加入隐式水印。

于是非常符合互联网精神一幕又上演了。

你加水印，我就做一个去水印工具。

更好笑的是，已经有人尝试让 Claude 帮自己安装这个 watermarks-remover。

Claude：拒绝。

他解释了半天，说这是自己的内容、自己不想要这些水印，Claude 还是拒绝。

我觉得如果Claude都会拒绝安装，那就说明这个项目真的有用吧🤣

[孤桜ETH @GYLQ520]

Anthropic这回直接把自家发现和修补漏洞的全套底牌亮出来了。仓库刚公开就飙到6.8k星，还拿了MIT许可。

工作流拆得明明白白，从威胁建模、扫描、到人工复核和修复建议，全部做成了Claude Code能直接调用的命令。

发现疑似bug后系统会连续触发三次、稳定复现才作数，而且每一步都有独立的评审员二次确认，不搞一言堂。

整个作业过程被锁在隔离的容器里，对外通信只留了一个口子，安全垫拉满。虽然官方说了这只是个参考实现，后续不会积极维护，但核心价值在于整套流程和思路可以直接移植到你自己的技术栈里用。

想上手的话，官方建议第一天先跑通威胁建模和静态扫描，第二天就让流水线在C/C++项目上自动驾驶，后面几天再慢慢适配你自己的代码仓库，节奏给得挺务实。

🔗传送门：https://t.co/H5vqXdlr9W

[WangNextDoor @WangNextDoor2]

曾经遭到诺奖候选人否定的梯度下降算法，如今是AI大模型训练的核心手段。 大模型把文本拆成token做下一词预测，依靠交叉熵损失量化误差，再沿着梯度方向迭代更新海量参数，不断压低损失完成训练。 在百亿级的超高维参数空间里，梯度下降几乎不会困在局部最优，过去的维度灾难反倒变成大模型训练的有利条件。

高维空间改写了传统优化问题的认知，很多在低维世界成立的经验，放到大模型领域会完全失效。 看似简单的迭代算法，搭配数学规律，才撑起如今整个大模型的训练底座。

[The Milk Man @MrMilkTrading]

Even though nobody guessed it, I'm sharing it anyway:

Sell one SPX put credit spread per week at the Multi-Day Saty ATR -1 level. That's the whole strategy.

326 backtested trades since May 2020
95.4% win rate
Profit factor 2.48
+$53,328 on a single contract
Max drawdown $4,605
Zero losing years, 2026 is 32 for 32

Everything below, free.

[Kaz Nejatian @nejatian]

Some news. Today, for the first time in Opendoor's history, we bought back our own stock.

Our share count is down 5%. We paid for this buyback with money borrowed at a 0% coupon and we still have hundreds of millions of dollars left to grow faster.

There are plenty of helpful details, along with a lot of legalese, in our press release and 8-K (linked below). But there are a few things I want to say directly and in plain English.

First, I *despise* dilution. On my first earnings call at Opendoor, I told you that if we issue a share, it has only one job: to make every other share worth more for our existing shareholders, not to extend runway for management. The flip side is just as simple: when our own stock is one of the best uses of capital, we should buy it back. Today, we did just that - by 5%.

Second, I run a publicly traded company. I don’t get to have feelings about the macro or the way Wall Street works. My job is to understand the rules of the game and use them to find an edge and build a better company.

One of those realities is that our stock price has lots of volatility.

People disagree - A LOT - about what Opendoor could be worth one day. Some think we're worth less. Some think we're worth dramatically more. I obviously have a view…

That disagreement isn’t just noise - it has economic value that Wall Street monetizes every single day.

Most companies would treat this as a headache, but we see it as an asset. If people are going to speculate about our future, I'd rather our shareholders get paid than anyone else. So instead of complaining about volatility, we used those dynamics to borrow $650M at a 0% coupon.

Then we turned around and used part of those proceeds to buy back our own stock. At today's prices, buying back our stock and call options was one of the best trades on the board. The rest stays on our balance sheet so we can buy more homes and grow faster.

As for dilution, we bought our stock where we thought it was cheap and pushed any future dilution far above today’s price. Not one net new Opendoor share will exist below $10.38. And if we buy back stock in the future, that floor goes higher.

But why raise now? Because the best time to raise capital is when you don’t need it. We’ve proven the business can reach ANI profitability, but profitability is just the starting point, not the finish line. And waiting could make the shares we are buying back today more expensive, and homes we could be buying don’t get bought. I’d rather move now.

Some people will call what we did today aggressive. They’re right. But being aggressive is how we fixed a company that spent years being careful. I’d make that trade again.

To our shareholders: You trusted us with your capital. Today, for the first time, Opendoor used its own capital to buy more of itself. The company is putting its money where its mouth is. Tomorrow, I will too.

After our lawyers allow, I’m personally buying $100K worth of shares. I'm all in, and I plan to keep buying.

[Max For AI @MaxForAI]

刚刚，一篇可能会引起大模型圈地震的论文发表了！

研究人员第一次系统性地从 OpenAI、Anthropic、Google 的闭源模型里，大规模“偷”出了真正的隐藏思维链。

然后，他们顺手拿这些思维链去研究了 Kimi K3、GLM-5.2、DeepSeek 等一批开源模型。

结果非常有意思。

01｜原来闭源模型的 CoT，真的可以被“偷”出来

现在 OpenAI、Anthropic、Google 都越来越不愿意把完整思维链交给用户。

原因也很好理解：最终答案可以抄，真正值钱的是模型“怎么想”。

一旦完整 CoT 被批量拿走，就可以直接拿去做蒸馏，训练自己的模型。

所以现在很多 API 的做法，是把模型的 reasoning 加密以后返回给客户端。你看不到内容，但下一轮请求可以把这坨加密数据再交回模型，让它继续思考。

结果研究者发现了一个很离谱的漏洞：

这些加密思维链，在同一家厂商的不同用户、Session，甚至不同模型之间，居然可以互相兼容。

于是攻击方法简单得有点荒谬：

让最强模型思考 → 拿到加密 CoT → 扔给同厂更弱、更容易越狱的模型 → 让小模型把它解出来。

比如 Claude Opus 4.8 的隐藏思维链，可以交给 Haiku 去“念”出来。

你甚至不需要攻破 Opus，找它的小弟Haiku就行。

论文最终在 Claude、GPT、Gemini 三套 API 上都实现了隐藏 reasoning 的提取，等于直接绕过了厂商原本用来防止 CoT 蒸馏的那层保护。

[MapleShaw @msjiaozhu]

Veo3 确实牛啊！特别是这个声音语调，吊打好多 TTS！

英语汉语俄语，连 ASMR 都拿捏得死死的🤌

Onlyfans 是不是可以异军突起了😏

[K总 @KFCQZ]

😂B站有个大神，在Codex里面手搓出个AI赛博女友，能直接通过Codex Live和数字女友实时对话，已经几万人在线等教程，这效果真的好强啊

[Shikhar @xikhar]

I gave ChatGPT Real-time Voice a face, a body, and a personality.

A 3D persona that talks to you and lives right on your desktop.

OpenAI, this is what Codex Pets could become.

[Shikhar @xikhar]

ChatGPT can now control its 3D persona.

Through MCP, it can perform custom animations, not just speak.

Check out the demo below.

[得否 @wangdefou]

这是提示词，大家都去试试，评论区交作业

制作一个 exe 桌面宠物应用，把图片中的 6 个人的形象提取出来。生成程序的时候，这几个人都像一群猴子一样，在整个屏幕里爬行和玩耍，并且每个人都做成模仿猴子爬行。保留每个人原本的长相、发型和服装特征，删除瓶子并自然补全双手及遮挡部分。当我打开这个应用时，这几个人可以在桌面上爬行和玩耍，还能够根据桌面的应用和窗口的边缘作为障碍物进行玩耍。请你表述我的意图，我确认以后再执行。

[外五县程序员 @lgwebdream]

@randyloop 带派不 老铁

[Chaz Wargnier ♨️ @DataChaz]

A DANISH SCIENTIST GOT TIRED OF SENDING CVS SO HE PROGRAMMED CLAUDE TO JOB-HUNT FOR HIM

Mads Lorentzen, a PhD geophysicist, built an open-source framework on Claude Code that can run the entire job search from the CLI.

.. and it actually landed him the job!

Instead of just generating generic text, the pipeline does the heavy lifting:
→ Reads postings and scores how well you fit
→ Tailors a CV in LaTeX, keeping only relevant experience
→ Drafts a cover letter framed around the specific role
→ Runs a second AI agent to critique and revise the first draft
→ Compiles a clean, ready-to-send PDF

Free and open-source.

He released it under an MIT license, and it just passed 6,600 stars on GitHub.

Repo below ↓

[Mr||Solo @R465SRS]

Those bullies are expose they are Done for they can't escape this now

[YOLO Tendie @YOLOTendie]

@igeekbb Wtf