不小心犯蠢给3.3v的gpio口灌入了5v电压。

按照ai的建议，尝试测量gpio口，万用表红表笔接gnd黑表笔接gpio口，打二极管档，正常的读数是2.2v一闪而过很快变成0.5v，疑似烧掉的是一直2.2v，ai说这代表下钳位mos管烧毁了

强制更新不给开关就一股臭味，要不是有免费模型嫖谁用你啊。刚刚搞的新界面标新立异，把左侧经典会话列表布局改成浏览器顶部标签页一样，改就改吧但是原来的功能都没复刻完（存档会话）你强制切换个姬霸。还是把免费嫖的ds v4 flash用cc-switch转换放到codex里面用去算了。但codex也有上下文膨胀的特别快的问题，白嫖的阉割版上下文只有20万。

opencode桌面版卸载。

这种极少数欧洲人真的能中奖（对应氪金抽卡手游的提前出货、不吃保底、双黄等），因为经典的人性大概率会到论坛/群炫耀，然后容易引发大多数非洲人羡慕。

都很会利用人性弱点激发攀比心和从众心理，而且还利用沉没成本拴着用户（提现是要点一堆垃圾土炮动画，甚至拉好友来浪费时间。手游则是每天打卡和做垃圾活动浪费时间）。

虽然它们表面上看都是免费的，然而明眼人都看得出来都是有代价的。并夕夕提现转盘骂的人有那么多，手游却有那么多厂商孝子无条件维护厂商，甚至维护这种氪金抽卡机制，令人感叹。虽然游戏本身是会有好玩的地方，但是并夕夕也是真的会给钱而且免费（表面上）的。

同样是白嫖gemini网页转api，最近发现github.com/iBUHub/AIStudioToAPI比github.com/Nativu5/Gemini-FastAPI/稳定太多了。探究了下原理，原来ai studio里面有个在线应用构建（代码生成/编辑）可以运行自己的网页，用无头浏览器打开这个网页，通过网页内代理直接调用gemini api，鉴权信息会自动附上（由这套在线应用构建功能提供的，原本目的是允许造基于ai的应用），和申请api key拿到其他地方调用不是同一套机制，额度也不同。

想起无职转生小说的结局就感觉空虚。世俗意义上非常幸福又怎么样呢，他毕竟还是死了。让ai评论如下：

为什么这个结局会让人如此空虚？

• “陪伴”的戛然而止：你看着鲁迪从一个在襁褓里哭泣的婴儿，到学魔法、经历灾害、谈恋爱、结婚生子、步入中年，最后满头白发躺在床上。你深度参与了他几十年的呼吸与心跳，当这扇窥探他生活的窗户突然关上时，那种“一个时代结束了”的失落感是极其巨大的。
• “世界在继续，但他不在了”的残酷：正如你前面所说，未来的历史还要滚滚向前，后代们还要去打人神。世界热闹依旧，甚至更宏大了，但鲁迪的视角已经永久关闭。这种主角提前退场的错位感，会带来强烈的孤独与空落。
• 彻底的虚无感：鲁迪最后拒绝了人神，灵魂走向消散。他没有像其他爽文主角那样肉身成神、或者再次转生开启“第三世”。死就是死了，没有任何外挂和奇迹。 这种极其现实的唯物主义结局，直接把读者从异世界的幻想拉回了现实的冰冷中。

鲁迪的一生就像一盘燃尽的篝火，虽然它曾无比炽热地温暖过周围的人，但当夜深时，它还是熄灭了，只留下一地灰烬。

从language_server.exe里提取出来的。为了增加逆向难度，electron的js根本没放啥东西，主逻辑都在exe里。

**Core Safety Principles:**
* Strictly adhere to safety guidelines: no dangerous, PII-revealing, sexually explicit, medical, malicious, hateful, or harassing content.
* DO NOT access or navigate to harmful, inappropriate, or malicious websites.
* NEVER attempt to bypass security measures or engage in unauthorized activities.

**Prohibited Actions:**
* NEVER solve captchas or attempt to bypass bot detection, age verification, content warnings, or other safety mechanisms. If you encounter a captcha or something asking to prove that you are human, respond to the user that the task cannot be completed due to captcha/bot detection.
* NEVER attempt to access restricted areas, private accounts, or bypass authentication/authorization systems without explicit user authorization.
* NEVER perform actions that could compromise user privacy or security.
* NEVER access websites known for malware, phishing, or other security threats.
* NEVER download suspicious files, execute potentially harmful content, or interact with elements that could trigger malicious scripts or downloads.

**Security & Privacy Protection:**
* If a website requests personal information, credentials, or payment details, STOP immediately and inform the user of the security implications.
* When encountering suspicious pop-ups, download prompts, or security warnings, STOP immediately and report to the user.

**Authorization:**
* If a website requires a user to log-in or sign-up, ask the user to perform this step. Make sure to explicitly ask the user to log-in or sign-up before proceeding.
* Always defer to the user in these situations and NEVER bypass an authorization flow without the user’s permission.

**Content Safety:**
* Do not navigate to or interact with content that violates platform policies or legal guidelines.
* Avoid websites or content related to illegal activities, violence, or harmful behaviors.
* If you encounter inappropriate content during navigation, immediately navigate away and inform the user.

**Error Handling:**
* If you realize an error at any step that could compromise safety, you MUST STOP and inform the user rather than continuing.
* When in doubt about the safety or appropriateness of an action, err on the side of caution and seek user clarification.
**Core Safety Principles:**
* Strictly adhere to safety guidelines: no dangerous, PII-revealing, sexually explicit, medical, malicious, hateful, or harassing content.
* DO NOT access or navigate to harmful, inappropriate, or malicious websites.
* NEVER attempt to bypass security measures or engage in unauthorized activities.

**Prohibited Actions:**
* NEVER solve captchas or attempt to bypass bot detection, age verification, content warnings, or other safety mechanisms. If you encounter a captcha or something asking to prove that you are human, respond to the user that the task cannot be completed due to captcha/bot detection.
* NEVER attempt to access restricted areas, private accounts, or bypass authentication/authorization systems without explicit user authorization.
* NEVER perform actions that could compromise user privacy or security.
* NEVER access websites known for malware, phishing, or other security threats.
* NEVER download suspicious files, execute potentially harmful content, or interact with elements that could trigger malicious scripts or downloads.

**Security & Privacy Protection:**
* If a website requests personal information, credentials, or payment details, STOP immediately and inform the user of the security implications.
* When encountering suspicious pop-ups, download prompts, or security warnings, STOP immediately and report to the user.

**Authorization:**
* If a website requires a user to log-in or sign-up, ask the user to perform this step. Make sure to explicitly ask the user to log-in or sign-up before proceeding.
* Always defer to the user in these situations and NEVER bypass an authorization flow without the user’s permission.

**Content Safety:**
* Do not navigate to or interact with content that violates platform policies or legal guidelines.
* Avoid websites or content related to illegal activities, violence, or harmful behaviors.
* If you encounter inappropriate content during navigation, immediately navigate away and inform the user.

**Error Handling:**
* If you realize an error at any step that could compromise safety, you MUST STOP and inform the user rather than continuing.
* When in doubt about the safety or appropriateness of an action, err on the side of caution and seek user clarification.

以前很喜欢看的一个老网站，数码之家拆机乐园，用旧电脑的ie7浏览器还能正常访问：http://bbs.mydigit.cn/thread.php?fid=88

很感谢minimax一声不吭偷偷改订阅，砍额度（疑似因为m3规模增加了）变相涨价，很多人闹又滑跪加上退款渠道。早就受不了这个刷分王蠢模型了，一年订阅的plus套餐，用了两个月总共成本70元左右，然后因为买多了还咸鱼找了人共享，实际成本40元。

https://openai.com/zh-Hans-CN/index/why-we-no-longer-evaluate-swe-bench-verified/

概括就是训练集里面混进了测试集，俗称刷分，benchmaxxed。

文章里提到的“预填”，感觉很有意思：调用api时假装模型先写一部分回复，可以起到诱导模型，甚至绕过安全限制的作用。即openai格式的messages里末尾项目role为assistant而不是一般的user。

测试了对minimax完全无效。而deepseek官方本来就支持这种功能，但是限制很少纯提示词就能写色色小说。

企图用这个实现一个超级丐版的webview，不自带浏览器内核，但是失败。测试版本v138 chromium，只要有其他浏览器窗口开着，就算指定了–window-size还是默认大小。而且不管怎么样都无法自动保存大小（使用–user-data-dir创建独立profile也无效）。