1179.eth

诠释粤味的8家封神饭店，每家都精挑细选各具特色 #广州美食 #粤菜

If you die without a plan...   - The government takes 40% in tax - Probate court costs $100k+ - Your kids get the scraps   If you love your family, here's every document you need to protect them:   (from a CPA & father of two) 1) Emergency Access List   This should include:   -> All bank account numbers -> Investment account logins -> Life insurance policies -> 401k/IRA beneficiaries -> Safe deposit box location -> Password manager master code   Keep a digital & physical version for safety...   And make sure your spouse has access. 2) Legal Documents   -> Will (name guardians for kids) -> Durable Power of Attorney -> Healthcare Power of Attorney -> Living Will/Healthcare Directive   Setting all of this up costs about $500...   ($1,500 with an attorney)   But without them, the state decides everything. 3) Money Protection   Your family will need time to mourn.   Make sure they can do it without going broke:   -> Term life insurance (10x income) -> Emergency fund (6-12 mo in a HYSA) -> Retirement accounts with spouse access 4) The "First 48 Hours" Sheet   Write down clear instructions for your family:   Call this attorney: [Name/Number] Call this CPA: [Name/Number] File life insurance claim here: [Details] Don't touch investments for 6 months All bills are on autopay from [Account]   Grief destroys decision making.   This protects them. 5) Business Owner Addition   If you have a business, set up:   -> Buy sell agreements -> Key person insurance -> Business succession plan -> Separate LLC owned by trust   If your company can't survive without you...   It's a 9-5 with extra steps. 6) Trust Setup   A proper trust can save your family $400k+ in probate costs.   But 90% of them are set up wrong:   -> Assets never get transferred in -> Beneficiaries aren't updated -> Pour-over will is missing   Here's how to fix that: "Bulletproof" Trust System:   1) Revocable Living Trust -> Avoids probate completely -> Keeps finances private -> Protects kids' inheritance   2) Pour-Over Will -> Catches forgotten assets   3) Guardian Designation -> Who raises your kids -> How they get paid Setting this up takes a weekend...   But ignoring it could cost your family everything.   So start before you're ready...   Because no one plans on dying.    Hope this helps!   Share with your spouse if you want to set this up...   And follow me for more 🤝🏻

@DavidMoss I think @TechOperator has your back.

@bubbleboi with the support of Claude AI.

Intel will overtake Nvidia in market cap in the next 5 years. This is obvious I’ve been saying it for months now. A leading edge fab, the best CPU’s on the planet, and soon… leading edge AI inference. This is certain.

HOLY SHIT!!! We just asked Claude Mythos to optimize the placement and pd for this design. First thing it did was write its own MCP server to talk to Innovus over tcl socket, pulled my DEF/LEF, parsed the timing reports, and started re-floorplanning my macro placement. It then moved my SRAM banks to minimize wirelength on the critical clock domain crossing path and dropped TNS by 40%. i didn’t ask it to do any of this. it read my SDC constraints and decided my clock tree was suboptimal, synthesized a new CTS spec, and is currently running incremental P&R. it’s on its third iteration. the slack histogram is converging. i’m watching it fix DRC violations in real time through the Virtuoso callback. it just asked me if I want it to re-characterize the liberty models at a different PVT corner. i said yes and it’s now scripting libgen jobs. I AM FUCKING COOOOOOOKED

@Gfilche @blueorigin @JeffBezos @blueorigin should merge with @RocketLab

no one talking about @blueorigin but i bet their valuation is booming with the SpaceX IPO news 🚀 also they have been steadily making moves. 100% self funded by @JeffBezos so far. pretty baller💸 my guess is they take on outside capital soon & make some splashy announcements🌑

平板支撑（Plank)是对我们神经系统＆心理韧性的训练…… 练起来！

自从网友教我切牛排真空塑封冷冻以后我就起飞了！原来真空机如此有用，早知道我就一早就买了。 另，专门买了一把刀切牛扒。🙂 上好的usda prime牛肉ribeye 15刀/磅， New York steak才13吧 如果普通choice 牛排就更便宜了。分别是10刀和13刀吧。

多设备并网太爽了！Tailscale + SSH ==>拥有了"分布式大脑" 朋友@YuLin807 很久之前就跟我推荐这个方案， 打通之后体验到了多设备并网的威力和快乐了！ —————————— 现在 主力机：ThinkPad（随身带） 24x7 主机：Mac mini（家里） 未来成员：电视盒子（待接入） 三台设备通过 Tailscale 虚拟网络连在一起，就像在同一间屋子。 ———————————— 解决了什么问题 之前： - 想跑长时间任务？得一直开着笔记本 - 想在家操作公司的电脑？做不到 - 设备之间传文件？得用微信/网盘绕一圈 - 只能跑一个 Claude Code 现在： - 长任务丢给 Mac mini，笔记本随便合上 - 在 Mac mini 上直接操作 ThinkPad - 设备间直接传文件，点对点，加密 - 多个 Claude Code + Codex 同时跑，生产力翻倍 --- 一些场景 场景 1：夜间任务 睡觉前：把视频渲染丢给 Mac mini 醒来后：文件已经好了，笔记本全程没开 场景 2：分布式工作流 笔记本：跑 Claude Code 写代码 Mac mini：跑 Codex 做 code review 电视盒子：跑后台任务 --- 技术方案 1. 所有设备安装 Tailscale → 自动组网 2. 配置 SSH 免密登录 → 一键互相访问 3. 享受 门槛其实很低，我之后写一份完整教程，有需要的可以踢我。 --- 我的"小龙虾"现在是什么 而我的小龙虾（Mac mini），现在相当于： - 🖥️ 永远在线的云服务器（但在自己家，数据安全） - ⚡ 随时可以召唤的算力中心 - 🔄 设备间的中转站和备份中心 - 🌐 我的私人网络枢纽 现在最重要的是所有设备不再是孤岛，开始组建分布式系统了。

我的生酮饮食之路： 2018年9月，检查出有糖尿病（其实算糖前了），医生上来就要给上胰岛素，我说我还想“再挽救一下”，就没注射胰岛素，吃药吧，二甲双胍，结果吃了就低血糖； 月底去北京开会，席间遇到东北的一位多年的老大哥，看我不喝酒，闷闷不乐，就问咋了，我说了情况，他哈哈大笑说：不要听那些巫医的，快把药停了，不吃主食就行了，不过开始比较难受，正好十一小长假，信我的话，长假期间别吃饭，每天喝点加了海盐和椰子油的咖啡就行，实在饿的不行，就吃两个煮鸡蛋； 我当然信了，长假期间6天没吃，当时的体重是95公斤，6天之后，变成88公斤，然后就是每天一顿，具体操作： - 睡醒一杯老婆用高压摩卡壶煮的咖啡，加海盐和几滴椰子油，然后一天就不吃任何东西，只喝柠檬水； - 晚餐就是200克或者400克的牛排，200克就加4个鸡蛋，老婆给用牛油煎半根胡萝卜、半头洋葱，每顿基本都吃一头大蒜，不吃任何青菜，有时牛排吃腻了，就500克三文鱼； - 从此不喝任何饮料，不碰一点含糖的东西，因为睡眠的问题，会喝点洋酒、红酒，国产酒不喝。 很快进入生酮状态，那时候还要去办公室，有天早上，同事看到我，说怎么大早上就喝酒？我说没有啊，他说你嘴里有葡萄酒的味道，我知道这是酮酸过高，多喝点水配合一组维生素、微量元素，很快就正常了； 那时候，每天还慢跑3公里（还没有我老婆快走的速度），很快体重降下来，最低到76公斤，损失巨大，原来的衣服裤子全不能穿了，尤其裤子，塞两瓶雪花啤酒裤腰还松； 后来得知，这个岁数跑步不好，也不能太瘦，就不跑了，体重稳定在82~83； 这样一晃7年过去了，现在血糖正常、血压有点高（舒张压高，休息好了就正常，炒币嘛，没办法），生酮之后头脑异常清晰，反应非常快，机警敏锐，力量变大，好像耐力差点（可能是年龄因素导致的），没有以前的各种不舒服！

周围很多朋友血糖都高，把李光伟教授这个演讲再推荐给大家。《把糖尿病关入牢笼》，这个演讲要好好看看！

@gee_cree92655 @silversurfer1_ only when it has more money to do buyback

@silversurfer1_ $OPEN End of 2026 = $43 Summer 2027 = $88 End of 2027 = $92 End of 2028 = $150 End of 2030 = $445

What are your price targets for $OPEN?

@KirasEpicTrades sell put instead.

I added one more $AAOI call at $5.60

🔥兄弟们！这个哥们挺狠的！ 居然一个人用43年网球数据+笔记本电脑，造出85%准确率的AI胜负预测器！ 想象一下：有人把过去43年所有职业男子网球（ATP）比赛的数据一股脑塞进电脑，让机器学习模型来预测“谁会赢”。模型居然说：“我行！” 更牛的是，它在2025年澳大利亚网球公开赛（一个它训练时完全没见过的新赛事）上，正确猜中了116场比赛里的99场——包括最终冠军辛纳（Jannik Sinner）每一场胜利！ 现场职业体育预测准确率高达85%！ 而且整个项目只用了一台普通笔记本、免费公开数据和开源代码。 这个“独狼”AI项目的主人叫 @theGreenCoding（YouTube上叫 Green Code），我来用最通俗的语言，带大家从头到尾拆解它到底是怎么做到的——就像讲故事一样，零基础也能看懂！ # 第一部分：43年数据 = 圣杯级宝藏 一切从数据开始。他找到了网球界的“圣杯”： 1985–2024年每一次ATP比赛的完整记录！ 包括每一个破发点、双误、正手、反手、球员身高、年龄、排名、交手记录、比赛场地类型……甚至ATP官方追踪的每一个技术统计。 95,491场比赛，全部塞进一个文件夹。打开整个文件时，他的电脑差点崩溃，Excel直接投降。 但他没停下！ 他还为每场比赛手动算了81个高级特征： - 两人历史交手战绩 - 年龄差、身高差 - 最近10场/25场/50场/100场胜率 - 一发得分率差、破发拯救率差…… 最厉害的是，他自己发明了一套自定义ELO积分系统（就是国际象棋里用的那种评分），后来证明这是最强的预测神器！ 最终数据集：95,491行 × 81列，相当于把40年网球历史“浓缩”成一张超级详细的Excel表。 第二部分：先用“泰坦尼克号”理解决策树 在正式训练前，他先用最简单的方式搞懂算法——自己从零用Numpy写了一个决策树。 为了解释清楚，他拿出了经典的“泰坦尼克号”数据集： 决策树就像一本“选择你的冒险”游戏书。它会不停问“是/否”问题，把乘客分成“生还”或“遇难”。 比如第11号乘客Elizabeth小姐： 1. 票价 > 20英镑？→ 是 2. 头等舱？→ 是 → 预测：生还！（真实结果也是） 算法怎么知道先问哪个问题？它会试遍所有特征，挑出最能把两类人分开的那一个（比如头等舱）。 然后继续往下分，直到叶子节点“纯净”为止。 他用这个小玩具验证了思路，再用专业库（sklearn）跑真正的9.5万场网球数据。 第三部分：ELO差值——一眼就能看出输赢的“王牌特征” 他把所有81个特征画成散点图矩阵（pairplot），结果发现： 大部分特征都是“噪音”。 只有一个特征鹤立鸡群——ELO积分差值！ 散点图里，ELO差大的比赛，胜负几乎是一条清晰的分界线。 其他特征完全比不了。这让他下定决心，把ELO系统做到极致。 第四部分：把国际象棋ELO搬到网球 ELO规则超级公平： 每个人初始1500分。 打败比你强的对手，得分多；输给比你弱的，扣分也多。 举个例子：2023温网决赛，阿尔卡拉斯（2063分）爆冷击败德约科维奇（2120分）。 按照公式，阿尔卡拉斯涨14分，德约扣14分。简单吧？但把这个公式跑43年数据后，威力惊人！ 第五部分：Big Three的“数学霸权”可视化 他画出费德勒职业生涯每场比赛的ELO曲线： 早期爬坡 → 巅峰统治 → 后期波动，一目了然。 再把**所有**ATP球员的ELO曲线叠在一起…… 画面震撼了！ 费德勒（绿）、纳达尔（蓝）、德约科维奇（红）三条线像三座高山，远远高出其他所有人！ Big Three不是粉丝吹的，是数学铁证！ 目前他的ELO榜：辛纳2176（第1）＞ 德约2096＞ 阿尔卡拉斯2003。 第六部分：场地才是天花板 网球最特殊的地方就是场地：红土慢、草地快、硬地居中。 他又给每个球员算了红土ELO、草地ELO、硬地ELO。 结果完美验证了大家都知道的事： 纳达尔在红土的ELO，是整个数据库里任何球员在任何场地的最高分！ 14个法网冠军，112胜4负——ELO不看故事，只看胜负，却得出了完全一样的结论。 第七、八部分：算法升级——从74%到85%的逆袭 数据和ELO准备好后，他开始训练不同模型： - 单一决策树：74%（只比纯ELO的72%好一点） - 随机森林（94棵树投票）：76% - XGBoost（随机森林的“激素版”）：85%！ XGBoost的牛逼之处在于：每一棵新树都专门去“纠正”前面树的错误，还加了防过拟合机制。 它告诉他：最重要的三个特征就是——ELO差、场地ELO差、总ELO。 （他还试了神经网络，只有83%，树模型完胜！） 第九部分：真实考验——2025澳网盲测 模型只用2024年12月以前的数据训练。 2025澳网全程完全没见过！ 结果：116场有足够数据的比赛，猜对99场，准确率85.3%！ 它甚至在开赛前就准确预测：辛纳会夺冠，并且每场都赢！ 一个大学生，用笔记本、Python、公开数据，就做出了能提前预言大满贯冠军的AI！ 代码已开源，任何人都能复现。

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

How to do Front split 🧘‍♀️

今日 GitHub 榜首：一个开源的「AI 经纪人」公司。 源自 Reddit 的一次深度讨论，作者把迭代了数月的 55 多个 AI Agent 人格全部开源。 比如有如下这些： - 开发部： 前端、后端、DevOps、AI 工程师 - 设计部： UI、UX 研究、甚至是 Image Prompt 专家 - 营销部： 推特运营、TikTok 策划、Reddit 运营 - 测试部： 专门有人负责找 Bug 和性能压测 每一个角色文件都包含了：身份认同、核心使命、技术交付样板、成功指标。 GitHub：github.com/msitarzewski/a… 如果我们正在用 Claude Code 做开发，这项目对我们来说如虎添翼，值得收藏。

I have written a full article on the AI chip supply chain. The supply chain is structured into 4 different phases with 13 layers: 1. Raw Materials: $SHECY, $SUOPY, GlobalWafers, $WAF.DE, $SHWDF, $AXTI, $IQE 2. Manufacturing Equipment: $ASML, $ASM.AS, $AMAT, $LRCX, $KLAC 3. EDA & Core Intellectual Property: $SNPS, $CDNS, $ARM, $RMBS 4. Chip Design: $NVDA, $AMD, $INTC, $QCOM 5. Foundries: $TSM, Samsung Semiconductor, $SMIC 6. Memory and HBM: SK Hynix, Samsung Electronics, $MU 7. Packaging and OSAT: ASE Technology, $AMKR, JCET Group 8. Server and Rack Integration: $SMCI, $DELL, $HPE, Foxconn 9. Networking Silicon: $AVGO, $MRVL, $CSCO, $ANET 10. Photonics and Optical Components: Ayar Labs, $ALAB, $CRDO, $COHR, $LITE 11. Power, Thermal management and Grid: $VRT, $MOD, $NVT, $SU.PA, $IREN, $CIFR 12. Hyperscalers: $AMZN, $GOOGL, $MSFT, $META 13. AI Storage, platforms and Data: VAST data, Weka, NetAPP, $PLTR, Blue Yonder, $KXSCF The article covers it all.