Appearance
📅 20260720 推文存档
📊 共 79 条推文
@elonmusk
2026-07-20 22:15:09
RT by @elonmusk: Backpropagation by hand ✍️ ~ 11 steps walkthrough below Backpropagation is the algorithm that actually trains a neural network, and it is where most people stop following along. It is not calculus you cannot do. It is matrix multiplication, working backward, one layer at a time. So I drew and calculated one entirely by hand. Goal: push the loss gradient back through a 3-layer network and land on a new value for every weight and bias. = 1. Given = A 3-layer perceptron, an input X, predictions Ypred = [0.5, 0.5, 0], and the truth Ytarget = [0, 1, 0]. = 2. Backprop gradient cells = Let us draw empty cells for every gradient we are about to compute. The shape of the answer comes first. = 3. Layer 3 softmax = We get dL/dz3 straight from Ypred minus Ytarget = [0.5, -0.5, 0]. No chain rule needed, and that shortcut is the whole reason softmax and cross-entropy are paired. = 4. Layer 3 weights and biases = Let us multiply dL/dz3 by [a2 | 1]. One multiplication gives the gradient for W3 and b3 together. = 5. Layer 2 activations = We multiply dL/dz3 by W3 to get dL/da2. The gradient moves back across a layer the same way the signal moved forward. = 6. Layer 2 ReLU = Let us pass it through the gate: keep the gradient where the activation was positive, zero it everywhere else. = 7. Layer 2 weights and biases = We multiply dL/dz2 by [a1 | 1]. The same figure as step 4, one layer up. = 8. Layer 1 activations = Let us multiply dL/dz2 by W2. = 9. Layer 1 ReLU = We apply the same gate again, now on a1. = 10. Layer 1 weights and biases = Let us multiply dL/dz1 by [x | 1], and every weight in the network now has a gradient. = 11. Update = We subtract, and the network has learned. In practice a learning rate scales this step. The gradients: dL/dz3 = [0.5, -0.5, 0] dL/da1 = [1, -2, 2, -1] dL/dz1 = [0, -2, 2, -1] The takeaway: matrix multiplication is all you need. Just like the forward pass, backpropagation is matrix multiplications end to end. You can do every one by hand, slowly and imperfectly, which is exactly why a GPU's ability to do them fast mattered so much to deep learning. 💾 Save this post!
2026-07-20 21:41:36
RT by @elonmusk: One very useful tool in Agent mode is image cropping of Grok Imagine . I don't know why no one talks about it, as it is a crucial element for creating smooth transitions. During a conversation, you need to frame each person and switch camera angles to add energy to your video. Being able to pick the frame you want and crop it saves a huge amount of time.
2026-07-20 21:27:45
RT by @elonmusk: Today is the 57th anniversary of Apollo 11 Moon Landing. In 1969: most powerful technology on Earth = Apollo 11. In 2026: most powerful technology on Earth = 2-trillion-parameter AI. Which one changes civilization more?
2026-07-20 21:25:47
RT by @elonmusk: instagram: just got married linkedin: thrilled to announce we have sold more b2b saas X: hello there the jacobian conjecture is false
2026-07-20 18:50:58
Grok
2026-07-20 18:21:54
True
2026-07-20 13:15:00
RT by @elonmusk: • Starship To “Collapse Cost” of AI Compute & Global Communications
2026-07-20 09:06:46
R to @elonmusk: I mean Thursday (duh)
2026-07-20 07:41:50
RT by @elonmusk: An absolutely incredible performance by Spain to take home the World Cup! 🇪🇸🏆 So well-deserved, just a beautiful display of teamwork all tournament. Congratulations to Argentina & Messi on a great run. And huge kudos to all of the teams and fans around the world for making this World Cup so memorable, already looking forward to 2030!
2026-07-20 07:37:05
Try Grok 4.5!
2026-07-20 06:39:25
Grok is good at predictions
2026-07-20 06:37:21
🇪🇸 🇪🇸 Congratulations Spain!! 🇪🇸 🇪🇸
2026-07-20 06:34:57
Next Starship launch attempt on Friday
2026-07-20 06:34:24
The @SpaceX team is amazing
2026-07-20 05:49:46
x.com/i/jf/soccer/league/mat…
2026-07-20 05:47:46
x.com/i/jf/soccer/league/mat…
2026-07-20 04:01:34
RT by @elonmusk: SpaceX's newest satellites carry about 1,024 gigabits of bandwidth each. The generation before carried 96. That is a tenfold leap in a single design cycle, beamed down from orbit. Jumps like that do not happen in industries that have given up, if so can you name me any?
2026-07-20 03:10:36
RT by @elonmusk: BREAKING: Grok 4.5 leads VulcanBench’s new coding benchmark. 🔥 Grok scored 91.3%, solving 21 of 23 real-world software tasks across five languages, beating Claude Fable 5 and GPT-5.6 Sol while also owning the cost-efficiency frontier. Grok keeps winning. 🏆
2026-07-20 02:44:21
RT by @elonmusk: Rome conquered Greece in 146 BC when they burned Corinth to the ground. And then the strangest thing in ancient history happened. The winners spent the next few centuries desperately trying to become the losers. It actually started before the final conquest. When the general Marcellus sacked Syracuse in 211 BC, he shipped boatloads of Greek statues and paintings back to Rome, and the city went crazy for it. Suddenly every rich Roman had to have Greek art in his house. Decades later Cicero built an entire court case around prosecuting a governor named Verres for looting Greek masterpieces, and Romans ate it up. Then came the schools. Wealthy families started buying Greek slaves specifically because they were educated, and handed them the job of raising their children. So the sons of the men who crushed Greece grew up speaking Greek, memorizing Greek poetry, and learning philosophy from Greeks. The elite eventually sent their teenagers abroad to Athens and Rhodes to finish their education, the way people today send kids to study overseas. Cicero did it. Julius Caesar did it. Their gods were Greek gods wearing name tags. Zeus became Jupiter. Hera became Juno. Aphrodite became Venus. Ares became Mars. Same myths, same family drama, mostly just relabeled. Their entertainment was borrowed too. The first big Roman playwrights, Plautus and Terence, basically adapted Greek comedies into Latin. Their doctors were Greek. Their architecture leaned on Greek columns, the Doric and Ionic and Corinthian orders you still see on bank buildings today. And the art obsession left a strange fingerprint on history. Romans loved Greek bronze statues so much that they mass produced marble copies of them. A huge chunk of what people walk past in museums today and call "Greek sculpture" is actually a Roman copy. The Greek originals melted down or vanished long ago. The knockoffs are what survived. Not everyone was thrilled. Cato the Elder spent his whole life warning that Greek culture would rot the Roman character. He sneered at Greek philosophers, distrusted Greek doctors, and wanted them kept out of the city. And then, as an old man, he quietly sat down and learned Greek anyway. Even the loudest hater caved. The trend went all the way to the top. Virgil wrote the Aeneid, Rome's national epic, by openly modeling it on Homer. The emperor Nero competed in Greek games and toured Greece like a superfan. Hadrian was so obsessed he grew a Greek philosopher's beard and poured money into rebuilding Athens. And Marcus Aurelius, master of the entire Roman world, wrote his private journal not in Latin but in Greek, because that was the language of serious thinking. The poet Horace nailed it in a single line: captured Greece took her savage conqueror captive. Rome won every war on the battlefield and then handed its whole soul to the people it had beaten.
2026-07-20 02:01:41
RT by @elonmusk: The first phone call across the Atlantic, in 1927, cost about seventy five dollars for three minutes, more than a thousand in today's money. Right now you can see a face on the other side of the planet, in real time, for free.
2026-07-20 01:15:23
RT by @elonmusk: I'm a cardiologist. Everyone is sharing this study as a skin story. They're burying the part that matters. Scientists took the aorta of a 75-year-old donor, applied a single engineered enzyme, and stripped away more than 70% of the molecular damage — bringing it down to the levels you'd see in a 30-year-old artery. Published five days ago in Nature Communications. Revel Pharmaceuticals, with Calico and the University of Colorado. Here's what they erased. Sugar reacts with proteins in your body the same way heat browns bread — slowly, over a lifetime. It leaves behind a residue called CML, the most abundant advanced glycation end product in aging tissue. It welds itself onto collagen and elastin in your skin, your eye lens, and your arterial walls. Two things follow. Your arteries stiffen. And CML latches onto a receptor called RAGE, which drives chronic inflammation — the exact fire I've been writing about for months as the engine of heart disease. Since the 1980s this damage was considered permanent. Your body has no enzyme to remove it. Every existing approach only slows new damage from forming. Nothing touched what was already there. So they built an enzyme that doesn't exist in nature. They screened 45,000 protein structures, then ran five rounds of directed evolution across more than 500 million variants until they had CMLase — a molecular lawnmower that oxidizes the CML off the protein and restores the original, healthy lysine underneath. Not patched. Reversed. Over 70% cleared from elderly arterial tissue. Over 55% from elderly skin — below the levels found in 31-year-old skin. 45-78% in lens proteins. The CEO said they expected 20% and were floored. Arterial stiffness drives systolic hypertension, heart failure, and stroke, and I have no drug that reverses it. I can slow the process. I cannot undo it. This paper says undoing it may be possible. The caveats are real and I won't skip them. This was done on donated tissue in a dish, not in a living person. No functional data yet — we don't know if that artery got measurably more elastic. Delivering a large enzyme deep into human tissue is a hard, unsolved problem. Clinical trials are years away. But something considered permanent for forty years just came off human tissue. We spent a century learning to slow aging. Someone finally figured out how to erase it. Thank you @theallinpod @friedberg @chamath @pesottas
2026-07-20 00:30:53
Wow
2026-07-20 00:28:58
Way more than a billion
@dianzhuXS
2026-07-20 21:38:55
牆國今年的所有極端天氣 跟明年的「人禍」比 都是「弟弟」
2026-07-20 10:50:16
入門必聽，昨天直播的錄音 真想學的，不管以前在哪跟誰學過 在我這，都得從頭來 https://substack.com/@dianzhuxs/note/p-207650234?r=8moxkp&utm_medium=ios&utm_source=notes-share-action
2026-07-20 10:26:31
不改革政體，單純經濟上的所謂「改革開放」，就是「韜光養晦」，翻譯成人話就是「偷和騙全世界的錢」🧐
2026-07-20 08:30:19
風水
@Mimiwftt
2026-07-20 15:37:41
好朋友好粉丝们，你们的留言我都看到了。你们耐心复盘我之前的语音space音频，不要光看文字。 精炼技术，多观察，少操作，中长线方向耐心持股，一定会有收获。 中长线不需要我每天安抚！反而助长了短线情绪。 不要给我留言了！爱你们💕
2026-07-20 14:50:27
RT by @Mimiwftt: 果然太阳底下没有新鲜事，记得很早的时候，在蚂蚁财富社区里，一个博主也是喜欢分享观点，挺客观的，后面关注的人多了，就建了群，免费的，后面也是“课代表”之类的搞事情，建了自己的小群之类攻击人家，🤣🤣🤣,后面这个博主就让这个课代表在群里自由发挥，结果亏的一塌糊涂。
2026-07-20 14:43:47
我真不在意这些外在的东西，名利都没有跟真诚的人在一起最开心！ 这个号我可能注销不了，收不到验证码，推文留给你们慢慢学习吧！
2026-07-20 14:35:45
黑粉儿其实倒还好！他们不了解我，无非就是为了黑而黑。 你是老人了小西瓜！最恶心的就是那些虚伪的人，坏到骨子里的身边人！
2026-07-20 14:06:20
以后你们自己玩儿吧！别他妈的再来我这里找存在感！
2026-07-20 13:46:22
回落了吧！
2026-07-20 13:41:36
去年刚来X我就说了，一般人是接不住我的真诚的。
2026-07-20 13:39:25
我他妈收了谁的钱了！傻逼
2026-07-20 13:28:07
Image
2026-07-20 13:19:42
薛伯伯说过：“想拯救散户的都是傻逼”！ 我一没收学费，二没给你们挖坑，抄底逃顶一个不落，还耐心教你们技术逻辑，带着你们走正路！ 那些“叛徒”你们良心上过得去吗？
2026-07-20 12:47:14
你们别猜了！不是黑子，是我曾经很信任的一帮人。去年的事儿一直延续至今。 有一些人是真的又蠢又坏！分不清谁才是值得信赖的。得不到就毁掉呗！ 我看人很准，只要这个人有问题，我不想跟他多讲一句话！ 我离开是有原因的，因为我看到了那些人身上人性的恶！ 我可以接受一个人笨，能力有限，这个是与生俱来的无法改变，但我接受不了一个人坏！尤其是骨子里的坏！
2026-07-20 12:35:20
嫉妒，真的会让人面目全非。 我是真不理解有些人，天天盯着我，恨不得拿放大镜找我哪里说错、做错。既然这么不喜欢我，那还天天追着看我的观点干嘛？甚至录屏、逐句分析，图什么？ 吃着我的、喝着我的，最后还反过来砸我的锅。 你们见过我天天跑去评论哪个博主吗？我从来不跟别人较劲，我只跟昨天的自己比。 我唯一在意的，就是今天的自己有没有比昨天更好一点。
2026-07-20 12:18:13
是不是我太少出错了，就恨不得抓住我的小辫子黑？
2026-07-20 12:16:45
5/7大盘没破位。看多有错吗？ 大盘破位4000了让大家谨慎，有错吗？ 天天追着我的过去的观点有意思吗？ 看看时间，兄弟！真他妈的搞笑！
2026-07-20 11:59:02
他妈的听不懂人话！不知道什么叫走一步看一步！
2026-07-20 11:58:00
傻逼！给我滚！
2026-07-20 11:57:51
你们想玩儿就自己好好玩儿！别他妈的总跟我比！有病吧！
2026-07-20 11:56:56
我并没有说过我看大盘看到什么点位！别他妈的给我造谣！傻逼！
2026-07-20 11:50:34
以后我尽量语音聊，纯文字容易引起误会。 你们也尽量多听语音，文字版的话个人理解力不同也会有偏差。
2026-07-20 11:42:37
RT by @Mimiwftt: 今日录音~ 我用夸克网盘给你分享了「video_to...1140」，点击链接或复制整段内容，打开「夸克APP」即可获取。 /~907d3Zfp7A~:/ 链接：https://pan.quark.cn/s/8b041ed0dd87
2026-07-20 11:00:40
x.com/i/spaces/1AKEmmLdaNbKL
2026-07-20 10:54:26
双创方向短期不会有大行情，没有主升浪。 高位震荡是短期最好的走势。 不听劝的尽管去撞“南墙”
2026-07-20 10:36:01
我就没有这种“偏财”运。
2026-07-20 10:17:55
不知道低吸哪里的去低吸大金融。
2026-07-20 10:08:50
底部板块没必要高抛低吸，只有低吸的价值。 高位板块别在恐慌中割肉，后面还会有诱多拉升，到时择机高抛离场。
2026-07-20 10:06:08
这不是最重要的！ 重要的是我看多航空装备，高位套牢比较多的商业航天今天其实并没有很大涨幅。 要记住，博弈就是，资金会向阻力最小的方向流入。
2026-07-20 10:00:26
煤炭7/2开始看多，总有低点低吸的，如果涨起来追，那你就是韭菜！
2026-07-20 09:56:57
继续延续周五行情
2026-07-20 09:53:00
今天大涨的都是我复盘语音给你们的看多的方向
2026-07-20 09:50:12
教学贴，别追了！ 昨天语音中短线的选股方法，交易买点，你们看看吧。
2026-07-20 09:37:56
想抄底的今天不能进。等回落。 量能不够，当心回落！
2026-07-20 09:32:26
这就是周末复盘说的普反，后期会分化 绝对不要去高位，低位才有价值。
2026-07-20 09:30:13
科技方向反弹预期降低一些。先高抛出来，等后期筑底再说。
2026-07-20 09:27:20
周五低吸的，把低吸的仓位1840先出来。
2026-07-20 08:20:23
我们先要感谢他能把自己的经历讲出来。 我之前一直压着大家不要去跟风追涨加速段的高位科技，同时让大家拿住已经处在下跌尾声的板块，目的就是防止让你们两头挨打，最后血本无归。
2026-07-20 08:01:56
这个Ai总结的还可以。
2026-07-20 07:54:22
昨天语音复盘重点提到的就是这些。 以后有时间多来听语音，Ai总结的语音复盘有很多不准确的地方。
2026-07-20 07:40:34
那些高喊“高位科技YYDS”的人，和那些把“阿根廷永远的神”挂在嘴边的人，本质上是一类人。 他们买的是信仰，是情绪。 可资本市场从来不是靠信仰运行，而是靠资金、估值、预期和博弈。市场不会因为你相信什么，就朝着你希望的方向发展。 在零和博弈的市场里，最可怕的从来不是波动，而是情绪。恐惧和贪婪会放大每一次涨跌，信仰会遮蔽判断，而真正稀缺的，恰恰是理性和客观。 薛伯伯说过一句话：“忙者不会，会者不忙。” 很多不懂博弈的人，总担心错过下一次暴富的机会，看到别人赚钱就心痒，看到热点上涨就忍不住追，生怕自己慢一步就与财富失之交臂。 可真正懂市场的人，往往没有那么着急。他们愿意为了更高的确定性，主动放弃很多看似诱人的小机会。因为他们知道，股市里永远不缺机会，真正稀缺的是识别机会、把握机会和守住机会的能力。 不会因为一根阳线就改变信仰，也不会因为几天调整就否定逻辑。他们更在意的是赔率、胜率，以及风险收益比，而不是一时的涨跌。 《孙子兵法》有言：“善猎者，必善等待；善战者，无赫赫之功。” 真正优秀的猎人，不会看到猎物就开枪，而是耐心等待最佳时机；真正高明的将领，也不是靠每一场惊心动魄的大胜闻名，而是因为他的决策足够稳健，很少犯错。 投资亦是如此。 很多时候，等待本身就是一种能力，空仓也是一种仓位，不交易也是一种交易。 能够克制冲动，耐得住寂寞，等得到属于自己的机会，比频繁操作更重要。 所以，与其每天追逐热点、沉迷情绪、为信仰买单，不如多学习、多思考、多复盘，建立属于自己的交易体系。 市场每天都有机会，但并不是每一个机会都属于你。 真正成熟的投资者，不是做得最多的人，而是做对得最多的人。 多学习，勤思考；少冲动，少折腾。
2026-07-20 07:23:25
你要知道，国家救的是“市场”，而不是“你”。 这就是要么买宽基指数，要么买“好股票”的理由了。
2026-07-20 07:18:42
这些动作，其实都在预期之内。 前面我就跟大家讲过，托底资金一定会进场。但大家要明白，这里强调的是“托底”，目的在于防范系统性风险，而不是马上开启一轮单边牛市。 很多人一看到国家队入场，就开始幻想高位科技要起飞了。国家队会去高位科技股给你接盘吗？别做梦了。 高位科技板块，如果借着政策刺激出现反弹，更大的意义是给前期被套资金提供减仓、换仓的机会，而不是意味着新一轮主升行情已经开启。 国家队救市从来不是做赔本买卖，他们更倾向于配置大市值、低估值、高股息的核心资产。一方面能够稳定指数、维护市场信心；另一方面，这些资产本身估值安全、分红稳定，也更符合长期资金的投资逻辑，可谓一举两得。 所以，真正值得关注的方向，依然是权重股和红利资产，而不是盲目追逐高位题材。 周末听过我语音直播的朋友应该都知道什么叫“权重搭台，题材唱戏”。 没有权重股把指数稳住，题材行情很难持续；只有权重先搭好台子，市场风险偏好提升之后，题材股才有表现的空间。千万不要把顺序搞反了。
2026-07-20 06:27:40
之前自己做的“手绘版简易飞行棋”，一早儿妹妹让我陪她玩儿。
2026-07-20 06:04:40
不用担心，应该是X的系统问题。
2026-07-20 05:47:37
x是不是出小问题了？
2026-07-20 05:23:02
这就是大部分新股民都会踩过的坑，也是最痛苦的一段经历。 所以我一直压着他们，告诉他们拿住手里的筹码，不要去追高位科技股。不想让他们经历那些亏损、焦虑和后悔。
2026-07-20 05:15:41
你们怕的是自己手里的股票下跌。 上面怕的是系统性风险。
@aleabitoreddit
2026-07-20 14:22:41
Some updates on bottleneck timelines: - ABF substrates: 1 year / eg. Ajinomoto (2802) for ABF film, Ibiden, Unimicron, Nan Ya PCB, etc make the substrates - HDI Boards: Over 6 months / Victory Giant (300476), Zhen Ding (4958), Unimicron, Compeq (2313), Meiko (6787) - Multilayer boards: ~6 months / Victory Giant, WUS (002463), $TTMI, Gold Circuit Electronics (2368), - CCL: Over 6 months (severe shortages of materials) / Elite Material (2383), Shengyi (600183), TUC (6274), ITEQ (6213), Nan Ya Plastics (1303), Resonac (4004) Mitsui Kinzoku (5706) / HVLP copper foil. Nittobo / glass fiber/cloth - MLCCs, chip resistors, tantalum capacitors, aluminum electrolytic capacitors: 15-20 weeks to 1 year+ MLCC: Murata (6981), Samsung Electro-Mechanics, TDK (6762), Taiyo Yuden (6976), Yageo (2327), Walsin (2492) Chip Resistors: Yageo, Walsin, KOA (6999), Rohm (6963) and $VSH Tantalum capacitors: Kyocera/AVX, Yageo/KEMET, $VSH again Aluminum electrolytic capacitors: Nippon Chemi-Con (6997) Nichicon (6996) | Timelines sourced from Digitimes, did the honors of adding in related companies. But it's slightly more nuanced since MLCCs and others mix. in commodity types vs. AI DCs. Regardless, just some helpful mapping.
2026-07-20 02:46:43
There’s nothing quite like the scent of more AI capex on a Sunday. Probably some ways to go if leading frontier models are out of capacity from too much demand.
@JonathanDi3614
2026-07-20 11:34:53
3天前预判的电力：粤电涨停，带个止损，博溢价，中午复盘其他机会，下午操作。 很多刚做短线的朋友都有同一个困惑：为什么别人能抓到涨停，自己一买就跌、一卖就涨？同样的行情，别人赚钱，自己亏钱？ 答案很简单——不是天赋问题，是方法论问题。短线交易不是“低买高卖”四个字就能概括的，它是一门需要每天训练和打磨的技术。就像打篮球，你看再多NBA录像，不上场投篮、不练运球、不复盘自己的比赛录像，永远成不了好球员。 复盘，就是短线交易者的“训练场”和“比赛录像回放”。 一、复盘的底层逻辑——你到底在看什么？ 很多人打开软件看一眼账户，涨了开心，跌了郁闷，然后关掉。这不叫复盘，叫“看账户”。 真正的复盘，目的只有一个：找出交易系统里的问题，然后改掉它。 就像球队看回放，不是为了看输赢，是要看清楚：那波团战为什么站位错了？那个技能为什么没躲开？下次遇到同样的情况应该怎么做？ 短线复盘的三个核心问题： 1. 今天的操作，哪些是正确的？标准是什么？ 2. 今天的操作，哪些是错误的？错在哪里？ 3. 如果回到当时的盘面，我应该怎么做？ 二、复盘的时间框架 短线复盘分三个层级： · 盘后复盘（最重要）：每天30-60分钟，系统回顾全天行情，分析每一笔交易。 · 盘前计划（次重要）：每天开盘前15-30分钟，制定当天的交易计划。 · 盘中记录（辅助）：盘中简单记录关键节点的想法，收盘后复盘时再用。 复盘不是形式，是肌肉记忆的训练场。不练，永远上不了场。
@xiaomustock
2026-07-20 10:46:59
均价52.3海力士2x我又来了 梭完了，生死有命富贵在天，买定离手了。
2026-07-20 06:28:30
完犊子，归零
2026-07-20 00:20:04
买了5万刀阿根廷赢，最后一场了想把前面买球输的赚回来。 之前阿根廷那一场对手埃及还是谁进球了，被因为很早之前的踩脚吹掉了， 轮到阿根廷进球踩脚啥事没有。 后面阿根廷打的几局都莫名其妙，对手跟封印了一样只防守不进攻， 感觉是一种无形大手黑幕必须要让阿根廷保送。 巨大的阴谋。 再不赌球了他娘的买啥亏啥，我老是想反买一小博大被爆干，干服了正买又被爆干。🌚 赌球真几把脸黑呐。
2026-07-20 00:00:21
全市场真正唯一害怕的事情就是这些云厂缩减2027年的财务开支，尤其是谷歌亚马逊，它俩家就占了全球总支出一半快4000亿美金。
生成时间：2026-08-04T08:16:02.314Z