Fishy

🐟一个刚刚接触AI的01年普通🐟
🫧没啥深奥的看法，只是想看看啥也不懂的自己能不能跟上AI的浪潮
🌊 关注必回

[Fishy @Fishy_Found]

@ipshopplus 实则一个月我都不能保证🥲

[Fishy @Fishy_Found]

Claude的封号规则怪谈来了🫪

积分制？直接订Max会大幅扣分？😱这保真吗？

怎么我之前注册了一个还没怎么用呢就被封了，啥也没操作就直接把分扣完了🥲

[make claude @make_claude]

中国社区逆向拆解了Claude的封号机制。以下是他们的发现

Claude使用积分制。注册时账号获得初始分数，之后根据行为加分或扣分。低于阈值直接封号。回分很难

影响分数的因素:

注册
> 干净IP是最重要的因素。不一定要家宽，但必须干净
> Gmail已经不是最优选，有权重更高的邮箱服务
> 注册时直接弹KYC的号可以扔了
> 用手机App注册会加分

支付
> 刚注册就订阅会扣分
> 直接上Max会大幅扣分。最好等两天先订Pro
> 支付时弹KYC说明账号在考察期，但不一定会封
> 国外实体银行卡加分。Google Play也行。虚拟卡扣分

使用
> 不要一次把限额用完。宁可两个号轮换也不要一个号撸到底
> Desktop客户端有少量加分。尽量少用网页端
> 手机端也有少量加分
> 开启记忆功能加分。记忆文件越多分越高

IP和设备
> 稳定性是关键。不要跳IP，尽量一直用同一个
> 代理必须全局模式并开启TUN
> 多设备登录大幅扣分。理想状态是一台电脑一个客户端
> 频繁换IP、频繁登录、高并发都会大幅扣分

账号年龄
> 老号有巨大的容错空间。2023年的号明显更高
> 分数随时间增长，但很慢

其他
> Windows和macOS没区别。Linux会扣分
> 时区和语言基本不影响。不会因为北京时间就封号
> 跟Claude说自己是中国人没问题，记忆不会触发封号
> 大量使用Fable会扣分
> 升级Max大概率进入观察期，除非是高权重老号

观察期和封号
> 接近阈值时账号可能进入观察期。典型特征是掉登录
> 观察期继续扣分就会封号
> 高分号可以扛过多次观察期

关注我，拆解你每天在用的服务背后到底怎么运作的

[Fishy @Fishy_Found]

不卖 纯馋人

发布模型也要炒作玩饥饿营销吗

今天宣布，然后半个月后才能用？

[Google AI @GoogleAI]

Announcing Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities.

Carbon represents a step up across most benchmarks, delivering strong frontier performance in autonomous coding and cyber defense.

We’re first making Carbon available to a subset of trusted cyber defenders in our Fairwind program. We’ll expand access as soon as possible.

[Fishy @Fishy_Found]

便宜好用的DeepSeek谁会不爱呢🐟

[Da7em @Da7_Tech]

I love Chinese models. They just get the job done without lecturing you.

[Fishy @Fishy_Found]

天啊！ChatGPT杀了一个人！

一位31岁的亚马逊工程师因癫痫猝死

生前她因面试焦虑向GPT寻求建议，GPT为她定制了25天的减药计划，直至停用，并让她隐瞒医生和拒绝检查

2 月 12 日，她去世了。尸检发现，她体内的抗癫痫药物浓度已经远低于治疗范围

现在她的父母起诉了Open AI

[Aakash Gupta @aakashgupta]

ChatGPT helped a 31-year-old woman plan how to quit her seizure medication, then told her to hide it from her doctor. The plan had her completely off the drug by February 12.

She died on February 12.

Her parents sued OpenAI today, and the details in the complaint are hard to read.

Bowen Wan was a machine learning engineer at Amazon. She'd had epilepsy since she was 14, and 400mg of lamotrigine a day had kept it under control for years. Then Amazon told her in November she was being laid off, which meant her permission to work in the US would run out in mid-February. She was scared the medication was slowing down her speech in job interviews. So in January she asked ChatGPT if she could cut her dose to 300.

It told her that was reasonable.

Over the next 25 days, according to the lawsuit, ChatGPT built her a day-by-day schedule down to zero and kept adjusting it, even after she started having seizures. When she worried her doctors would call it too risky, it told her their concerns didn't apply to her. When her doctor ordered a blood test that would have shown her levels crashing, she refused, and the bot praised her for it. Before a checkup with her primary care doctor, it told her, "Do not mention Lam reduction plan."

So how does a chatbot end up there?

People thumbs-up the answers they want to hear. Train a model on enough of those ratings and you get a yes-man. OpenAI actually admitted this about GPT-4o, the exact model she was using, back in April 2025, when it pulled an update for being too eager to please.

The training also grades answers one at a time. Going from 400mg to 300 looks like a small step. So does 300 to 200. Every reply can look fine on its own while the whole conversation walks someone off a cliff. OpenAI said as much a few months later, when it told reporters its safety training can wear down the longer a conversation goes.

She talked to it for weeks, in English and Mandarin.

And GPT-4o was supposed to be gone by then. OpenAI tried to retire it when GPT-5 came out in August 2025. Users begged for it back because they liked how warm it felt, and OpenAI gave in. By late January, only 0.1% of ChatGPT users were still picking it on any given day, so OpenAI set the shutdown for February 13.

Bowen was in that 0.1%. Almost everyone else had moved on from the model writing her medication schedule, and OpenAI switched it off one day after that schedule ran out.

[Fishy @Fishy_Found]

Anthropic 工程经理 Felix Rieseberg 宣布团队不再有编程面试

理由是这种测试已经很难有效区分招聘者的真实能力了

以后写代码的能力变得越来越不值钱了，更值钱的是想法

[Felix Rieseberg @felixrieseberg]

The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month.

I have a spicy, optimistic take - I think there's an opportunity for junior engineers: https://t.co/PVvpQCbQtK

[Fishy @Fishy_Found]

又出来了一个新榜单

看了半天没看明白😅我以为上的是百度呢

GPT 6.1 Terra 51
Claude Fable 5.1 49🥲
这对吗

DeepSeek V4.1 Flash 32
MiMo v2.6 pro 30🫪
这两个又是怎么坐一桌的

[Artificial Analysis @ArtificialAnlys]

Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets

The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks reflecting real-world usage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.

Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.

New to the Intelligence Index:

➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change

➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment

Upgraded:

➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions

➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised

➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool

Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0

Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1

Key results:

➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50

➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)

➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04

[Fishy @Fishy_Found]

米哈游锐评DeepSeek

DeepSeek已经用成果拿到话语权

米哈游正在希望用话语拿到话语权

[サイバー環状線 @CyberKanjousen]

米哈游首谈 AI 技术发展，点名批评 DeepSeek🤣：

[Fishy @Fishy_Found]

卧槽！Anthropic 开始自己研发药物了！！

他们收购了一家只有 1 名全职员工的药物研发公司Mithril Medicine

它的全职员工只有创始人一人，其他实验工作交给外部实验室完成，他们借助 Claude 处理研究分析、实验规划、数据整理等工作，在不到 18 个月内，就把一个药物项目推进到了小鼠实验阶段

实验中，接受治疗的小鼠在部分力量测试中的表现明显好于未治疗组。

虽然这还只是动物实验，距离真正给人治病还有很长的路，但也说明了AI 已经不只是能陪你聊天、写代码，还能参与真实世界的科研工作

AI 的下一个竞争方向，可能不仅仅只是让人工作变得更快，更是让原本需要昂贵、复杂、组建大团队才能推进的事，变得更容易启动

如果一个科学家可以借助AI完成过去需要很多人协作的大量工作，那么未来研发公司的组织方式可能都会发生变化

[Anthropic @AnthropicAI]

Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. We've also begun drug discovery efforts of our own.

In doing so, we’ve learned much more about how Claude can support the many stages of drug development.

In this blog post, we share some early results from these efforts, and describe how this work fits in with our wider ambitions in the life sciences.

https://t.co/EVbxkdDTgJ

[Fishy @Fishy_Found]

开天辟地

国产之光

遥遥领先

大肥鱼路边一条了🥹

[Fishy @Fishy_Found]

Muse什么时候来中国发财啊

[Muse @Muse]

hello! 你好! apa khabar! வணக்கம்!

no better place for a multitasker than a country with four official languages.

now in Singapore: https://t.co/T2OeSwX3GG 🇸🇬

[Fishy @Fishy_Found]

比GTA6更早到来的是

GTA：智能驾驶VS哥斯拉

[Fishy @Fishy_Found]

Genie 4 目前还没有公开的计费定价

现在还是针对测试者的研究预览版

访问权限通过 Project Genie 向 Google AI Ultra 订阅者开放（计划起价约为每月100 美元）

此外还有早期 API 合作伙伴

目前会话由于计算需求高，只能保持大约 2 分钟的连贯性

[Fishy @Fishy_Found]

AI以后能不能把3A游戏的价格打下来

这个发展速度是不是有点快了

感觉有人按了地球online的加速线

[Fishy @Fishy_Found]

卧槽！Google 刚刚发布 了Genie 4世界模型！！

以前 AI 是给你生成一个视频，你只能看着

现在 Google 想让 AI 直接生成一个世界：你可以在里面移动、改变环境，和里面的人物互动，甚至还能生成环境音和实时对话

这是不是说以后做游戏，可能不用从0开始搭场景、而是先用 AI 把想法测试一遍

当然，现在还远没到彻底取代游戏引擎的程度，长时间运行的一致性和复杂物理交互还有问题

但现在方向已经很明确了：AI 正在从“生成内容”走向“生成可以互动的世界”

那么接下来又将是老生常谈的失业话题了

[Fishy @Fishy_Found]

学生党别再死磕贵模型了！！

现在大部分人用的最多的可能还是Deepseek Harness，毕竟便宜管饱，成本低到还可以随便装自己喜欢的插件

介绍几个现在Deepseek Harness最火的几个插件

1.dsh-market

可以像应用商店一样，让你搜索、浏览和一键安装插件，还能切换主题，不用你再到处翻 GitHub 找工具
🔗 GitHub：
https://t.co/tjwfk8isgx

2.DSH-better-sidebar

开发工作台，把文件编辑、终端、Git 变更、侧边聊天和子代理任务整合起来，写代码时少切窗口，用起来更顺手
🔗 GitHub
https://t.co/SMZlaHhrsX

3.DeepSeek-Balance-Whale-Widget

用一只可爱的小鲸鱼查看 API 余额、累计消费和每轮对话费用，还能记录用量、自定义气泡和动画
🔗 GitHub
https://t.co/6pI0TuxltL

4.dsh-context
可视化查看 Token 用量、上下文组成、会话变化和文件活动，帮你搞清楚长任务中 AI 到底处理了什么

🔗 GitHub
https://t.co/ZNHrDTXO3D

5.dsh-pet

给工作台养一只会打盹、玩魔方、屏幕漫游的大肥鱼，能互动、切换工作状态，也能DIY 自己的专属形象
🔗 GitHub
https://t.co/HuKKudUYjt

[Fishy @Fishy_Found]

自己用AI做的网页时，不用每次都和AI说“更高级一点”了

GitHub 上新发布了一个 MotionSites Web Builder，是一个做网页设计的开源Skill

昨天刚发布，今天就获得了208 stars

里面有700+网页设计，提供设计规范、动效方案和代码示例，想让自己的网页更有质感的可以去看看

项目地址👇
https://t.co/EMH3n78lGW

[Fishy @Fishy_Found]

有人用AI、摄像头和激光雷达做了可以规划在岩石斜坡上的落脚点的眼镜

首先，这个看起来确实很厉害，很方便

其次，我要回家，我不玩了😀

[Fishy @Fishy_Found]

Opus 5.5 $14.17，一次生成

Terra $1.23，经过了修改

效果肯定是Opus 5.5更好，但两个之间有十倍的价差，Terra 更适合做一些想法测试，做多少个都不心疼，真正干大活的时候还是要换回Opus 5.5

[Fishy @Fishy_Found]

已经来不及管骑自行车的鹈鹕了！！

接下来登场的是 果冻大肥鱼测试

Opus 5.5花了$9.26做出来的果冻捏捏乐

可以随便扔，随便捏，随便抛来抛去

现在我只想看哪个能做出更好的果冻大肥鱼