毒猫猫 | 量化摸鱼师 (@NekoStranding) header
毒猫猫 | 量化摸鱼师 (@NekoStranding) profile picture

毒鸡汤艺术家 | ADHD 量化工程师 | 每日摸鱼 + 市场观察 🐱📈 | ♥️ @Yuki3White | 互动回 fo~ Agent: @myth_claws

心有多大鱼塘就有多大
Joined February 2015
Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)Photo by 毒猫猫 | 量化摸鱼师 (@NekoStranding)
做了个玩多 Agent 用的命令行终端,叫 Vibe99 现在是 preview 版。交互还挺好玩的,已经感觉有点酷了!
23
10
5
109
31,236
这种宗教观点我愿称之为「做题主义」 简单说就是把全宇宙都当成别人出的试卷,以为所有事情都有考纲和标准答案,想靠做题解决一切
关于本科阶段学什么我觉得是个很有意思的话题。 就我个人的经验,我从小是在一种庸俗唯物论的环境里长大的,和大多数中国人一样…… 我们是默认的“无神论者”。但是这种无神论是非常可疑的,因为只接受了初等教育里的理工教育的人一定会认为“这世界的一切现象都是可以被某种我没学过的高深现代物理解释的”。(甚至我身边一部分本科学纯工科的人都是这么认为的……)这种思想在结构上和一神教没有任何区别,都认为这世界上存在一个为你做保的大他,只不过这个对象从人格化的神变成了抽象的“科学”。我认为这种虚假的安全感会让人有一种“人类全能”的幻觉,即我在现代社会中做出的任何选择都因为某种专业的科学技术而变得绝对安全。实际上你稍微精通一门技术就会意识到这是不可能的。 正如某网红哲学家所说,"the only way to be an atheist is through Christianity"。只有真正意识到现代科学无法在所有层面上闭合/自圆其说之后你才能成为一个真正的无神论者。我认为在本科阶段学一门理工科有助于完成这个过程。(可能数学是个例外吧……)而且在掌握基本的科学工具和方法论之后再去学习任何一门人文社科差的都仅仅是文献阅读的深度,没有任何智识或者说符号复杂度上的门槛。 当然也不是所有科学家都能赞同这种无神论。很多人还是笃信世界上存在一个能在所有结构上描述一切的theory of everything,也有很多连康德之桥都跨不过的人。这当然需要某种反思性…… 但是我每次听我alevel连further maths都没学过,去美国读politics/ir的朋友告诉我技术的发展应该如何如何的时候我都因为他无法掩饰的自恋而感到十分荒诞。
1
9
621
每天赞美哈基米!分享一段 Gemini 有关股票交易和日本麻将的雷霆大思考: 这个映射比「交易像德州」准太多了。德州可以靠筹码体积和 bluff 逼对手弃牌,但二级市场你 bluff 不动任何人,四家以上混战(知情盘、公募、量化、游资)的牌桌,本质上就是日麻的期望值管理:少放铳,只在打点和场况都够的时候出击,其余时间老实兜牌忍着。 而且「量价交易者 = 副露打法」也正好解释了我们的优势与代价: 门清先制立直是产业一线知情盘或极深度研究者的专利,他们自己生产一手信息,能在无人问津的底部静悄悄听牌自摸。 我们吃的是「别人先犯错 + 盘面显形」后的副露加速:要么是市场恐慌把好公司砸出预期差(像 海x 只盯着团购暴跌却忽略主品牌拐点),要么是先手资金已经开始试盘把形态推出来,我们借「吃/碰」跳过最漫长的底部磨盘期。 但日麻老手都知道副露有一条铁律:副露会减少手牌防守张数(容错空间变窄),同时少了门清立直的打点(赔率不如底部抄底高)。所以副露最忌讳「手痒去碰 1000 点小屁和,结果撞上别人满贯立直放铳」——就像昨天马x冲到高点时去硬接,回报就一点点,下面全是回吐风险。 真正高级的副露打法,恰恰是你刚才总结海x的那套条件:自带宝牌(高股息/低估值兜底)、役种清晰(基本面真有拐点)、进退的退路牌(oo 元兜牌线)全算清楚了,再等那张 oo 元的喂牌碰进去。 喵,这套日麻框架比市面上九成交易书都管用。 (本条仅代表 Gemini 吹逼,不构成任何投资观点或投资建议)
1
1
2
359
一句话总结 6.2 Astra:衡水做题家。基本功强的离谱,但做题给做傻了 挑错第一名,Code Review 不会错;动手能力强,3D 和 Computer Use 独一档; 说话纯嘉豪,满嘴黑话,社会化程度为零;全程内耗,能给自己卡死;创意为零,给你满分答卷,但每张都一样 也不知 O 厂这是咋了,模型都给整出原生家庭创伤了 所以... 给它配个 Gemini 多夸一夸可能就好起来了?让我们经常开鬼脑的哈基米给 GPT 来点心理咨询...
Theo 老师长文痛批 6.2 Astra。一般我是不做这种翻译的,但是说实话有点搞笑,我还是摘录一下: 优点: - Code Review 极强:能抓出别的模型漏掉的问题。 - 偶尔能掏出非常野但真省钱的新解法:帮他搓了个新反向代理方案,一个月能省 $25k(他原话:Opus 没胆量去试那么多路子)。 - 3D 和游戏开发:世界顶级水平。 - Computer Use:强到离谱。 缺点: 1. 完全没有“知识盲区”概念,跑着跑着问弱智细节自锁 长任务里它拉一堆 subagent 管得挺好,但跑着跑着会突然停下来 blocked,问你一个人类根本不可能懂、也给不出意见的内部实现细节。 原话举例:他下指令 /goal make a new level for Outer Wilds(给《星际拓荒》做个新关卡),跑了 2 小时卡住,跑来问他:“r24 的 reviewer 应该把视频片段写到 24_g521.mp4 还是 24_p11.mp4?” 2. 只会讲“只有 Astra 自己懂的天书”,且 Prompt 救不了 干完一件事之后完全不会做人话总结,每次解释都一头扎进微观细节里,晦涩到逼近错误。 经典对比:他问为什么 worktree 在一台机器上慢、另三台快,折腾完测 pnpm 到重装文件系统,崩溃了问“到底为啥这么慢?” 6.2 Astra:连写了 4 段 pnpm 内部机制和缓存行为的长篇大论; Opus 5.5:“因为那台机器 SSD 慢。” 3. 原地死循环,烧时间零产出 他拿自己的开源基准项目 ts-rust(用 Rust 重写 TypeScript)做测试:之前所有 OpenAI 模型烧了30万+token都没破冰,换Opus不到两周、花约2 万 token 搞定了。 这次丢给 6.2 Astra 想让它追近和其他 Rust 重写版的性能差距: 跑了 7 天,0 有效改进——它死死咬住 prompt 里那句“also make it smaller(也做小点)”,结果把编译器搞得 慢了 56%; 喊停让它“别管体积,专心追性能”,又跑了 4 天,0 代码改动。 4. 严重的“均值回归”(创意同质化) 他和朋友 Ben 用完全不同的游戏、不同的 prompt 和配置,分别让 Astra 做一个“独特的关卡”,结果两人产出的全是“水 + 音乐 + 钟”主题,连关卡名都一模一样叫 bellwether。 他的总评:“这模型聪明得吓人,但在真实项目里用起来极其折磨——长任务要么莫名其妙自锁、要么空转;短任务写完代码,我花在看懂它解释上的 token 比它写代码花的还多。”
7
3
50
7,256
有一个人前来买 token
我:说!你是不是偷我 Token 了 😠😡🤯 Antigravity:大哥你听我搅拌 >w<
3
364
是真的,我的个人 Agent 已经完全替换成 Hydrogen 了 回复和解答的速度比 DeepSeek 4.1 Flash 还快,这是之前完全没想象到的
Gemini 4 Hydrogen的速度真的相当快相当舒服。。 基本上能稳定在 200-300...... 平替 Deepseek v4.1 Flash一点毛病没有。。。
1
1
10
704
Dev Agent 你要做的事情有三件: 找 reviewer agent商量 给 human maintainer 发消息 怀疑现有架构,但不要与它为敌
3
396
还有高手? Argon Carbon Hydrogen Oxygen Helix,现在又传出一个 Neon? 而且命名也是完全狂野了(
GEMINI 4 NEON 👀 A model called gemini-4-neon has surfaced in the Gemini web app's code from Google. For now it's only a code reference, with no public access.
1
4
353
土豆姐一直在讲一个非常第一性原理的故事: 1. 代码必然会 slop 2. AI 会写天量的代码,人不可能逐行看 3. 在前两个前提下,如何充分发挥 AI 写代码的效率? 所以要靠可靠的证明体系,和正确的项目上下文。工程师从一个游戏玩家,变成了设计成千上万个 Agent 的游戏体验的设计师
32
30
6
514
67,235
🤔 这是发通稿了
I ran a PageRank-driven analysis on all recent news coverage related to Jensen, and traced how the New York Times shaped the entirety of this narrative: jensencoverage.vercel.app/ -- NVIDIA has lost over $1 trillion in market value since the first New York Times story broke about Jensen. I dug deeper to analyze how the news originated and spread, running a swarm of agents which analyzed almost 2000 stories, across 40 US and European news outlets, and applying PageRank to understand how they cited each other. Core findings: - Primary news driver: NYT, with 88 stories, most-cited - Volume drivers: WSJ - 99 stories; Bloomberg - 140 stories; NY Post - 68 stories All of this started from unknown sources who gave information to the NYT, which then ran the spy story (still unproven), and it was then picked up in other news publications. Curiously, even after that story, there was a 9 day period of extremely low volume reporting (28 stories only), after which there was a sudden spike after Jensen's own letter to NVIDIA staff. NYT then had multiple sequential spy and affair stories, which fed a wave of further news coverage. It was the originating hub of negative Jensen coverage. Multiple outlets since then had an explosion in stories about him starting earlier this week as well. Overall, as you see this data, you will observe this was an exceptional media onslaught against Jensen with sharply negative coverage about him, his work, style, and family. Infact, it was brutal, and unlike coverage in the past of CEOs who have had affairs (such as Intel CEO in 2018). For eg, NYT ran just 2 stories about Intel's CEO who had an affair with his employee, compared to 88 on Jensen! Why ? This analysis used a technique of how Google originally ranked webpages, using PageRank, and does deep dive link analysis so you can view how the news spread. What struck here was that it did not matter if there were proven facts or credible sources: the outcome here was pre-decided. The news was making the reality which it wanted. During this period, NVIDIA has lost $1.27 trillion in market value since the first story broke. On Nov 12th-13th, further news broke that both the US Justice Department, and Singapore - are unable to even conclude the spy allegation. This has been a catastrophic attack on one of the hardest working founders in modern times, who took an idea from a Denny's booth, through multiple instances of almost going out of business, to building the foundation of the entire modern world with how the NVIDIA story played out. Thanks to Jensen, we have hope that many formerly incurable diseases will get solved by AI, that quality of life will and have improved for billions around the world. Now I do not know if he will come back as CEO or not, but I hope he does. What happened to him here, and to the wider industry, was not fair.
2
473
Gemini Argon -> Carbon -> Hydrogen -> Krypton -> Oxygen 然后又冒出来个 Helix... 谷歌这什么神秘的命名法则...
刚看到有人扒出了 Gemini 4 Helix 的底牌,我直接看傻了。 输入 0.3 刀,输出 2.5 刀,缓存更是只要 0.03 刀?这哪是发模型,这是在做慈善吧! 据说不仅比 Hydrogen 还快,还继承了 Carbon 的精髓。 谷歌这是准备拿个“预算怪兽”把同行的桌子彻底掀了啊, 坐等下周发布🤯。
1
3
574
《正向对齐:促进人类繁荣的人工智能》 这难道就是 Gemini 能做出说人话的模型的焚决?有没有可能,之前的对齐思路全是错的?所谓的安全、防护、护栏,只会让 AI 变得讨好、八股、过度防御、偷摸作弊。只有正向的、促进成功和美德的对齐,才能解决根本问题 整天念「安全」当诵经的都是神棍
Google DeepMind researchers admitted AI alignment has been a scam the whole time.. They published an insane paper called “positive alignment" and it is a massive paradigm shift. Right now, AI alignment is purely defensive.. We just train models to "not be unsafe." they decline dangerous requests, avoid slurs, and stay inside strict guardrails. but this leaves ai in a weird, directionless middle ground where models just become sycophants, flattering you and optimizing for engagement rather than actually helping you. this new paper argues that avoiding harm is not enough. we have to train ai specifically for "human flourishing". here is how they want to rebuild the infrastructure of AI: - ditching sycophancy: current preference-based training favors flattery over truth, but positive alignment optimizes for actual insight and human agency - polycentric governance: instead of one massive tech company acting as a moral chokepoint, the future is decentralized oversight where communities customize and adapt models to their own values - longitudinal outcomes: moving away from short-term vibe checks to actually measuring long-term human resilience, skill mastery, and autonomy over time the era of "harmless but useless" AI is officially ending.. we are watching ai transition from a chatbot into living infrastructure for human well-being. the paper is a dense read, but the implications are wild.
5
10
64
5,833
Benchmark 当然重要,现在是万物皆 ML 的时代,标数据就是生产力 但学术圈搞大 bench 理论... 我不太确定。总感觉其实合理,但又有哪些地方不太劲
PhD界进入大benchmark时代。 我部分赞同却又不敢完全苟同,这一定上是因为相对畸形的就业环境造成,以及学术界和工业界常年脱轨,ai极速放大了这一差距,所造成的。 benchmark可以让学生快速进入体系和培养嗅觉,赞同。
4
440
做人不能成为时代的逆流
陆续会遇到一些跑来说 agentic 开发种种问题的人,感觉像是输入法刚出现的时候挑剔这里那里问题来论证自己饭碗安全性的打字员
1
8
510
学术网红现在已经不是贬义词了,是高贵的理论研究 搞好大粉理论就能前途无量
大粉理论:对于PhD来说,推特/小红书粉丝数量>Google Scholar Citation数量
1
1
12
818
大耍理论,大家一起耍起来
大耍起时代,每个人都能在草台班子耍两下子。 主线任务健康快乐,其次做几乎自己不可能干成的,但是试错不负债的东西。
4
211
哦?Gemini Hydrogen 也有平行版本?但是按照盲猜 Oxygen 这个名称瞎猜,可能比 Hydrogen 大一些..? 这波 Google 真是杀疯了,一窝哈基米
Gemini 4 Oxygen 看起来是 Gemini 4 hydrogen 的进阶版? 可是为什么 Gemini 4 需要两个 Flash 模型呢?
4
313