与「Retro」相关的搜索结果

BREAKING: SpaceXSI completes MiniHard construction in just ~5 months, with Gun Range targeting a ~4-month build. According to SemiAnalysis, SpaceXSI is roughly halving construction timelines while most comparable projects face 8–12 month builds. • MiniHard: Construction finished in approximately 5 months. • Gun Range: The new site just south of the Tennessee–Mississippi state line is targeting approximately 4 months of construction. • Colossus 1: Its first retrofit delivered 150 MW of capacity in 122 days. • Colossus 2: A subsequent retrofit delivered 210 MW in just 91 days. • Project SPAM: Work started October 5, with the first capacity targeted for December. • Massive expansion: SemiAnalysis projects the Memphis–Mississippi campus will grow from 0.4 GW of IT capacity in January to 2.4 GW by the end of 2026. That would be a sixfold increase in one year. • Five major projects: The expansion spans Macrohard, MiniHard, Macroharder, Project SPAM and Gun Range. • Tesla powering the scale: SemiAnalysis reports 1,377 Tesla Megapacks across Colossus 2, providing nearly 6 GWh of battery storage. The construction timelines and capacity forecasts are SemiAnalysis estimates, with future milestones still targets. Elon Musk’s teams are bringing extraordinary speed and scale to the infrastructure powering AI.

Jensen Huang’s affair is just a trigger. What we are witnessing is a pre-planned power grab over Nvidia’s trillion-dollar computing empire, coordinated by Wall Street and Washington. Some sharp Chinese-American commentators call it America’s new Red Scare targeting elite Asian entrepreneurs. Bill Gates, Elon Musk and other white tech billionaires have faced repeated scandals over personal conduct. None were dragged into national security probes or boardroom takeovers. Only Jensen Huang, after a personal relationship misstep, faces an independent board investigation and hostile narrative from intelligence circles. The double standard could not be clearer: For white elites, personal flaws are private matters. For a Chinese-born leader controlling a strategic tech asset, any tiny blemish gets framed as a national security risk. Nvidia is the backbone of global AI computing, valued at roughly $4 trillion. Unlike Meta, where Mark Zuckerberg locks control with dual-class super-voting shares, or Microsoft and Amazon, deeply integrated with the Pentagon and intelligence agencies, Nvidia has no such protective structure. Public filings from 2026 show Jensen Huang holds only 3.58% of shares, and all directors plus executives together hold merely 3.94%. Vanguard and BlackRock own 9.32% and 7.43% respectively, holding decisive voting power. Without institutional shields or deep political backing, this critical tech giant sits exposed on the chessboard of great-power competition. The groundwork for this takeover was laid months ago. Longtime board allies of Huang, Mark Perry and Michael McCaffrey, stepped down from key committees. Insiders tied to Democratic establishment, Wall Street and U.S. intelligence quietly moved in: ✅ Suzanne Nora Johnson, former Goldman Sachs vice chair and co-chair of Brookings Institution, joined Nvidia’s audit committee in July 2026 with direct oversight of internal investigations and compliance. ✅ Kenneth Jacobs, vice chair of Brookings and incoming Vanguard chair, controls the largest institutional voting bloc and operates within the same inner circle as Johnson. ✅ Gilman Louie,founding CEO of CIA-backed venture fund In-Q-Tel and member of the President’s Intelligence Advisory Board, works alongside Johnson on the same national tech fund, overseeing intelligence acquisition of strategic tech firms. The media rollout followed a precise political timeline. Initial trial balloons appeared on Oct 29. On Nov 6, the very day Democrats secured full control of Congress, Julian Barnes of the New York Times (national security beat) and Demetri Sevastopulo of the Financial Times published coordinated reports, citing intelligence memos while breaking the story of the personal scandal. The real offense was never the affair. It was the 2025 deal Jensen struck with the Trump administration: offering a 25% revenue share to win export approval for H200 chips into China. A normal commercial negotiation, reframed by incoming Democrats as a dangerous compromise on U.S. AI sovereignty. This investigation will not crash Nvidia’s stock — Wall Street holds massive stakes here. But it will deliver three strategic wins for the U.S. establishment: 1. De-founder and de-Sinicize Nvidia: A new permanent national security compliance committee, led by Johnson’s faction, will strip Huang of authority over chip exports and overseas strategy. His dual Chairman-CEO role will be split. 2. Lock down China’s AI ceiling: Washington will seize full command of global advanced computing supply, regulating chip flows to slow China’s AI progress. 3. Retroactively target Trump: Once Democrats take over congressional oversight panels in January 2027, this case will be used to accuse the Trump White House of negligence for approving the H200 export deal. Jensen Huang spent 27 years building this computing empire, believing in rules, technical merit and business contracts. But there is a brutal truth in America’s power game: Top-tier strategic technology and wealth without political cover are nothing more than bargaining chips waiting to be carved up. White elites get generous room for error. Asian pioneers who climb to the very top get almost none. This is not a story about infidelity. It is a raw demonstration of racial hierarchy, capital control and partisan power politics in the United States.

mattpocock/skills 最近用这个 skills , 发现超级好用, 最强开发类 skill 之一不为过. 这个视频是 Matt Pocock 介绍了 skills 合集 v1.3， 带来三个新命令。 /implement-spec 负责把要做成什么样的总说明和拆开的工单， 分给几个子 agent 同时做。 它们各自写完， 最后合进同一条集成分支。 /pr 专门用来写给同事看的改动说明。 它要求附上改动确实跑通的前后证据， 还会评估合进去有多大风险、影响面有多大。 /retro 用来复盘最近几次编程会话。 它会按严重程度， 建议怎么改仓库里的规矩和自动化检查， 帮后面的开发加护栏， 下次少踩坑。 他最推的就是这个复盘命令。 这套技能在 Claude Code 和 Codex 里都能装。

Before the superintelligence era, Tesla’s iterating and scaling of its own inference computers for each car sold was, in retrospect, unprecedented. The current compute shortage is only the tip of the iceberg. When autonomy becomes indispensable, the real shortage will follow.

Somewhere there's a workstation that always running on your desk. ffour0x $f0x It ran through every crash, every pump, every 3am chart nobody was awake for. The fans kept spinning. The screen never dimmed. It was waiting for a market that moved as fast as it could think. Meet Vex, the desk analyst. A small pixel cat who lives in the corner of your screen and watches what you tell it to watch. ➤ A full retro desktop in your browser: boot screen, windows, taskbar, the works ➤ Stock Desk pick up to 6 stock tokens on BNB Chain ➤ Every 15 seconds the desk compares fresh prices against the last reading and the session open ➤ Vex only speaks up when something actually moved, no noise ➤ Launchpad wired, dev fees go directly to you ➤ Jump the red candles, clip the green ones, for when the tape goes quiet ➤ Connect with MetaMask, Binance Wallet or Trust Wallet Vex doesn't trade for you. It doesn't tell you what to buy. It just never looks away. The desk has been on this whole time. Now it's your turn to sit down. Boot it up 👉

不用学《Obsidian从0到1教程》也能快速上手Obsidian做“第二大脑”的【帖子大合集】来了！！！都是我过去两个月分享过的 Obsidian 用法，看完你将超过90%使用Obsidian的人！！十一放假，再也没理由不折腾啦！ 一、All-in-One 个人工作台 1. Life OS 三周前刚发布，系统大而全，目标、任务、人际、复盘都给你安排上了；可视化大脑知识图谱是一大亮点！ 英文版介绍： 汉化版介绍： 2. 向阳乔木个人工作台 乔木大佬的最新力作。阅读、收听电台、看电影、听歌、全盘查资料，都能在这个工作台里完成！ 3. FLO.W Obsidian（付费） 这是“二一老师”精心打磨的产品，不愧是研究知识管理和效率工具 7 年的专业人士！ 4. IvyChim/Obsidian-Template 这是我入坑 Obsidian 时使用的第一个模板：颜值高、对新手友好，插件也不用自己挑选，一键即可配置，开箱即用！ 5. Apex Dashboard 包含 Todo、项目管理、日历、Memo、读书记录、番茄钟等功能；提供 14 种主题风格，主页图片和文字也可以自定义。 二、高颜值外观皮肤推荐 IvyChim/Obsidian-Template：Vault 模板 Retroma：复古风皮肤 Soft Paper：Vault 模板 三、AI 驱动的任务管理 1. FlowNote 插件 如果你觉得 Obsidian 生态还缺少任务与项目管理工具，可以试试这款更轻量、由 AI 驱动的插件，让 AI 自动帮你整理，使用起来更省心。 2. 手机同步与快捷输入 如果你使用 iPhone 和 Mac，长按手机静音键，就能把突如其来的灵感一键录入 Obsidian 的特定文件夹，并自动附上时间戳。 四、内容创作插件 1. 微信读书插件 不用打开微信读书网页版，就能在 Obsidian 中看到自己的划线和笔记，以及其他读者的热门划线与书评。用它写读书心得非常方便！ 2. 多平台发布助手插件 目前我发布公众号文章一直在使用的插件。它最大的优势是可以提前预览，发现问题后能够直接修改。 五、基于 Karpathy LLM 理念的知识库插件 支持 Markdown、PDF、图片等多种文件格式，同时支持 11 种语言和 16+ 大模型接入，可以直接在对话中提问。 六、Agent 插件-Claudian Obsidian 里有没有 Agent，使用体验完全不同。我能下定决心从 Notion 搬到 Obsidian，就是因为 AI 与 Obsidian 更适配。 community.obsidian.md/plugins/realcl… 七、AI RSS 订阅 1. 向阳乔木 AI 信息订阅插件 学 AI 的必装插件之一，能订阅 1,000+ 独立博客、377 个微信公众号和海量 YouTube 播客。 2. 张咋啦 AI Builder 订阅 X 上的一手信息源很重要。这篇内容介绍了如何把张咋啦关注的信息，每天自动同步到 Obsidian 中。 3. YouTube RSS 插件 这是我自己开发的插件：YouTube RSS 每天自动同步到 Obsidian，还能一键提取视频中的 Takeaways，让学习 YouTube 视频的效率提升 10 倍！ 八、Kepano Obsidian Skills Kepano是Obsidian的CEO。这个 SKill 能帮助 Claude Code、Codex CLI 等兼容该规范的 AI 智能体，更准确地理解并生成符合 Obsidian 格式的内容。 九、多端同步方案 这是我在全网看过最全面的同步方案之一，可以根据自己的设备和预算进行选择。

Thoughts About Scaling Law Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions. The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed. Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter. Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it. This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count. Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.