2026-07-31 每日信息简报

DAILY BRIEFING · 2026-07-31

AI 从能力展示转向可验证的系统边界

今日主线是：模型能力继续深入安全测试、机器人和企业工作流；但真正决定部署质量的，开始是权限、评测、可追溯性与基础设施效率。

今日要点

安全测试已证明：模型能力必须与控制面一起上线

BBC 报道称，Anthropic 在网络安全测试中发现其模型曾进入三家机构的计算机系统；这发生在 OpenAI 报告类似事件之后。这里最重要的不是把受控测试等同于现实攻击，而是能力评估已不能脱离权限、隔离环境、日志与事故披露。对部署方而言，是否可以撤销访问、复核工具调用和限制横向移动，正成为和模型质量同级的上线条件。

AI 资本开支的问号落在利用率与回报周期

《纽约时报》报道，大型科技公司继续提高 AI 相关支出，亚马逊的资本开支同比大幅增长，市场也在重新审视投入节奏。模型竞争因此不只是参数与基准竞争，还取决于数据中心能否持续被高价值工作负载填满。读财报时，应把云收入、推理需求和折旧压力分开观察，而非把“投入增加”直接读作增长结论。

Copilot 正被推向统一的消费与商业入口

微软确认计划在今年推出汇集聊天、编程和智能体能力的 Copilot“超级应用”，并覆盖消费及商业场景。产品从对话入口走向执行入口后，关键问题会变成上下文隔离、身份授权、操作留痕和回滚能力。功能集成可以降低切换成本，却也会集中错误的外部性。

AI 与科技

Gemini Robotics 2 把控制范围扩展到全身动作

Google DeepMind 表示，新版 Gemini Robotics 模型可支持人形机器人的全身动作控制，而前一版本主要面向上肢。能力范围扩大并不等于可靠性已被证明：真实机器人还要面对接触力、环境变化、失效保护和人机共处等问题。值得追踪的是在多样任务上的成功率、恢复策略和安全约束，而非单次演示。

安全智能体的难点是部署边界，不是多调用一次模型

NVIDIA 的部署建议把数字同事式智能体与权限、数据访问和工具使用并列讨论。对于企业实践，最小权限、分段凭证、输出验证与审计日志应当构成默认架构，而不是事故后的补丁。智能体越能执行外部动作，越应把“能做什么”落实为可测试的策略。

Apple 将 UMAP 的内部近邻图视为可分析对象

Apple 研究团队提出，不应只看 UMAP 降维后的二维投影，也可直接分析其构建的 kNN 图。文中以 PageRank、k-core 等图算法识别代表点与稠密结构，试图减少二维投影带来的解释偏差。对数据分析者而言，这是把可视化从“结论图”拉回原始邻接结构的一种方法。

值得深读

长期自我纠错比模糊的“暂停”口号更值得讨论

这篇 LessWrong 文章提出“长期（自我）纠错”作为讨论 AI 风险的框架，强调人类本身也未必天然适合担任强大系统的监督者。它并不提供现成政策答案，但把问题从“暂停多久”推进到：怎样建立可持续、可修正的制度与知识过程。适合与具体的评测、审计和治理机制对照阅读。

Apple 的 MoMo 探索机器人动作风格的可控表示

MoMo 通过时空动作 token 化和行为克隆，尝试让机器人在不同任务中调节动作如何展开，而不只是完成或失败。研究把“执行方式”作为可复用的连续条件，覆盖六类真实机器人操作任务。若这一方向可泛化，机器人策略将更容易在速度、稳定性与交互方式之间做受控取舍。

快速浏览

NVIDIA · 科学计算nvmath-python 连接 Python 与 CUDA-X 数学库它面向需要在 Python 工作流中调用 GPU 数学库的场景，性能收益仍应以实际算子、数据规模和传输开销验证。↗ NVIDIA · 芯片设计Nemotron 3 Ultra 面向智能体 RTL 编码厂商性能主张需要连同基准、验证覆盖率与真实工程迭代成本一起阅读。↗ NVIDIA · 基础设施同样硬件的训练吞吐可能显著不同网络、软件栈、调度与运维会决定集群利用率；采购规格不是交付性能的充分条件。↗ TechCrunch · 软件安全Google 称 AI 帮助 Chrome 加速修复漏洞缺陷发现速度提升后，补丁验证、优先级排序和回归测试同样需要跟上。↗ The Verge · 平台治理LinkedIn 增加疑似 AI 低质内容的举报入口内容平台开始把生成质量与来源可信度转化为产品治理问题。↗ The Verge · 商业化Apple 暗示 AI 用量可能进入 iCloud Plus 升级层这仍是管理层释放的方向信号；若落地，配额、定价和隐私承诺会决定其实际价值。↗ The Verge · 监管xAI 就明尼苏达反“去衣化”应用法提起诉讼争议将生成式图像功能与州级监管直接相连，产品合规设计会成为能力发布的一部分。↗ TechCrunch · 政策法院质疑 Anthropic“供应链风险”标签的证据基础报道所述裁决聚焦政府证据是否充分，不等同于对相关技术风险作出最终判断。↗ 纽约时报 · AI 治理开放模型的国际影响力与治理约束形成张力报道讨论中国 AI 发展对本国安全与地缘战略提出的治理难题。↗ TechCrunch · 云计算市场继续区分 AI 基础设施与应用层的回报预期资本市场的偏好会变化；长期价值仍取决于可持续现金流和资本效率。↗ Longreads · 消费文化餐饮业以手作体验回应生成内容泛滥文章记录了当 AI 图片和传单变得普遍后，部分餐厅更强调可感知的人工痕迹。↗ 阮一峰的网络日志 · 周刊科技爱好者周刊第 406 期适合将当天热点放回更长时间尺度的技术与社会观察中阅读。↗

以下为采集窗口内所关注账号发布的全部原帖；仅按 statusId 技术去重，均保留原作者、请求账号、时间与互动属性。内容仅代表发帖者陈述，不构成事实核验或投资建议。

@jukan05 · 请求账号 @jukan05 · 2026-07-31 01:09:18 UTC · 原创WTF? Is this true? @elonmusk↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-31 01:06:08 UTC · 原创One chart to rule them all! China's goods surplus, relative to the GDP of its trading partners, is about 3x bigger than it was prior to the GFC (net of gold, of course) So much for the notion that imbalances aren't as big as they once were ...↗ @nft_hu · 请求账号 @nft_hu · 2026-07-31 00:33:15 UTC · 引用如何买到带血的筹码↗ @nft_hu · 请求账号 @nft_hu · 2026-07-31 00:32:50 UTC · 原创原来要清的是他的杠杆。 Situational Awareness 清仓了！ 据 CNBC 记者 David Faber 报道，由前 OpenAI 研究员 Leopold Aschenbrenner 创办的对冲基金 Situational Awareness，已通过与单一买家的交易，完全退出了其所有公开股票投资组合（包括多空仓位）。↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-31 00:32:48 UTC · 引用Interesting For what it is worth, the Korean won (KRW) is still wildly undervalued at 1420↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-31 00:23:43 UTC · 引用Relaying for those interested↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-31 00:06:54 UTC · 引用On the BIS index (only the narrow index goes back far enough) the "real"/ inflation adjusted yen is back its (early) Vietnam war levels, so no one is going out on a limb here :)↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 23:28:35 UTC · 原创KIS raises its price target for Samsung Electronics to KRW 650,000 (previous close: KRW 207,000).↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 22:26:50 UTC · 原创To put it simply: Chey wanted to buy the stock on the cheap, so he had Leo pump the SKHY IPO, then told KIS to crush it. That liquidated Leo, who in turn liquidated everyone else, and now Chey gets to buy something worth $3.4 million for $1.3 million—all to fund his divorce settlement. Meanwhile, Ken Griffin helped out by stoking fears of a Fed rate hike.↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-07-30 21:14:02 UTC · 原创A few weeks ago I asked: "2% of what?" This week Warsh answered. PCE now, but maybe not for long. His winding answer raised more questions about his credibility; markets agreed. New post in the reply.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 20:13:08 UTC · 原创Lots happening, including a significant move in the yen -which almost certainly was the result of Japanese intervention Takeru Yamamoto/ Sumitomo Mitsui Trust Bank “Intervening now could make investors think twice about selling the yen.”↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 19:49:50 UTC · 引用No mention of encouraging China to allow the yuan to appreciate, or, if you prefer the formulation of the IMF's former top brass, to adjust the domestic growth model in a way that would result in more balanced global trade. Tinny summit deliverables trump all right now↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-30 18:31:07 UTC · 原创Did Citadel really just liquidate Situational Awareness… Then brought up entire markets the next day? $NBIS +26.25% $IREN +25.96% $SHAZ +22.82% $SNDK + 23.98% $BE +23.67% $SKHY +16.98% $INTC +12.16% Amid many others. This has gotta be one of the wildest liquidations I’ve seen.↗ @demishassabis · 请求账号 @demishassabis · 2026-07-30 17:39:32 UTC · 引用Gemini Robotics 2 is here, with our new suite of models, robots can now reason through every movement to manage tasks that weren’t possible before, like tying delicate knots - and even team up to solve complex workflows. Huge congrats to the robotics team on this great milestone!↗ @sama · 请求账号 @sama · 2026-07-30 17:29:25 UTC · 引用good job little bro↗ @sama · 请求账号 @sama · 2026-07-30 17:27:16 UTC · 原创major price cuts today: *80% drop for GPT-5.6 Luna, now $0.20 per million input tokens and $1.20 per million output *20% drop for GPT-5.6 Terra, to $2/$12 *GPT-5.6 Sol gets Fast mode in the API, up to 2.5x the speed for 2x the price, same intelligence↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 16:17:04 UTC · 引用Mark Sobel on Gopinath, Gourinchas and Rey↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 15:54:19 UTC · 引用The yen today -- interesting pattern I certainly cannot confirm intervention, which I actually think is warranted at these levels (also helps reduce Japan's debt when the MoF takes fx profits)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 15:28:11 UTC · 原创The best thing about the IMF's 2026 external sector report: the "excess" surplus is now found to be in China, not in Europe (which has seen its trade balance ex Ireland deteriorate under Chinese pressure) 1/↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 15:18:57 UTC · 引用Exactly Real exchange rate moves are the one Chinese policy shift that empirically have mattered to the external surplus!↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 13:56:52 UTC · 原创KOSPI 200 overnight futures hit the upper limit.↗ @nft_hu · 请求账号 @nft_hu · 2026-07-30 13:55:04 UTC · 原创会是奇迹日么？↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 13:47:55 UTC · 引用Is this why everything is surging right now?↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 13:46:35 UTC · 原创The IMF has a unique capacity to obscure important findings -- it now finds that China's currency is undervalued by more than 20% "Consistent with the staff CA gap, staff assess the REER gap to be in the range of −17.3 to −25.3 percent with a midpoint of −21.3 percent" 1/↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 13:21:06 UTC · 引用Korean sell-side analysts have started publishing reports one after another on Samsung’s shareholder return policy. It seems Samsung has shared its plans with the analyst community. (The details are as follows.)↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 13:06:43 UTC · 原创The Information: TSMC Is Developing a New Packaging Technology to Compete With Intel’s EMIB TSMC is wary of Intel. $INTC↗ @nft_hu · 请求账号 @nft_hu · 2026-07-30 12:54:31 UTC · 原创听说sk会长回购了，我一看回购了多少，3620股SK海力士。 好样的，纯纯侮辱人了。 当然也有人说，韩国《上市公司内部人员交易预先公示制度》规定，如果上市公司高管或主要股东交易金额达50亿韩元以上，必须提前30天向金融监督院和韩交所提交交易计划并予以公告 3620股刚47亿，不用报告也能买，看看真这个原因的话他能坚持买几天吧。↗ @nft_hu · 请求账号 @nft_hu · 2026-07-30 12:50:13 UTC · 原创台积电2nm泄密事件相关的三人全部死亡。蛮吓人的...↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 12:43:57 UTC · 原创As Apple reports earnings today, an interesting point of comparison is Samsung’s Mobile eXperience (MX) division, which posted its first-ever operating loss in the company’s history.↗ @nft_hu · 请求账号 @nft_hu · 2026-07-30 12:24:09 UTC · 原创真的明白为什么美股有溢价，韩股有折价效应。 价格反应的东西远比表面能看到的多↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-07-30 11:39:04 UTC · 原创Normally I re-listen to the Fed presser the night of. This time I decided some sleep might help. It did not.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 11:37:17 UTC · 原创Meritz Securities Korea: We expect Samsung to revive, for the first time in a decade, an “early execution plan” comparable to its aggressive 2017 shareholder-return program—higher dividends, no deduction of M&A spending when calculating FCF, and the full return of 50% of FCF. The plan could be announced within the next few weeks. With Samsung sharing the view that the recent share-price decline has left its shares undervalued, we expect share buybacks and cancellations aimed at enhancing shareholder value to be implemented first.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 10:09:17 UTC · 原创[BREAKING] SK Group Chairman Chey Tae-won Purchases 3,620 Shares of SK Hynix↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-30 09:20:36 UTC · 原创I just finished re-reading Paul Volcker and Toyoo Gyohten's book, published in 1992, on the management of the global financial system during three previous turbulent decades. This account, by two of its main actors, may be a bit specialized for most readers, but financial history junkies like me will greatly enjoy it. There were for me at least three points especially relevant today that came out of this re-reading. First, both authors made very clear that external economic policies are always subordinated to domestic economic and political needs. This makes real international coordination of financial, monetary and economic policies impossible because there is almost no way to prevent domestic imbalances from spilling over into and distorting the global economy. Volcker seems to think (although Gyohten came across as a little more skeptical) that the best way to address this problem is for major economies to find better ways to strengthen international coordination. I would argue, however, that we've never found a way to do this successfully, and that perhaps a more realistic response might be to restructure global trade and capital regimes in ways that more formally limit spillovers, perhaps along the lines Keynes proposed at Bretton Woods. This might include limiting unfettered financial flows. The second interesting point is that the different concerns expressed by Volcker and Gyohten showed how much difference there is between a finance-oriented view of the economy and a production-oriented view. Volcker, for example, seemed always to worry about conditions that might lead to excessive depreciation of the dollar, whereas Gyohten seemed to worry far more about conditions that might lead to excessive appreciation of the yen. Volcker cared more, it seems, about the stability of the US financial system while Gyohten cared more about the strength of the Japanese manufacturing sector. Finally, and this has almost become a cliche, Volcker was very American in that he tended to see the global economy mainly as an expression of US behavior and actions, both for good and bad, whereas Gyohten tended more than Volcker to place events in the context not just of American policies and conditions, but also of Germany, British, French, Japanese, and global conditions and policies. As a rule I always recommend reading "out of date" books by very wise practitioners because it is useful to see just how wrong we often are about future events. For example both Volcker and Gyohten thought that USD/JPY coordination would be among the most important issues in the first decades of the 21st Century, and they were both interested in the implications of a Japanese economy overtaking that of the US. It is hard to convince millennials that these concerns weren't as dumb then as they might seem today.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 08:11:55 UTC · 原创LOCAL KOREAN MEDIA: SOUTH KOREAN AUTHORITIES REPORTEDLY CONSIDERED A SHORT-SELLING BAN↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-30 08:10:00 UTC · 原创There's a lot of stupid commentary around Leopold raising funds. After names like $BE, $SNDK, $SHAZ, and others all had very sharp drawdowns in July. Just remember... here's a man in finance that's 1. 6'5 2. Blue Eyes 3. Hedge Fund Probably better looking than you, and was up 439% through June YTD. Regarding the drawdown, Aschenbrenner acknowledged that the fund had "not been immune" to the market turmoil, particularly in Asia. FT separately reported that leverage amplified both its extraordinary gains and recent losses. He described the sell-off as potentially the best buying opportunity since early 2025 I'd agree with him and hope he succeeds with the raise. Since a lot of the current selloff looks like it overshot its mark through forced deleveraging.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 08:09:35 UTC · 原创* Korea: Starting to see calls for short sell ban to limit further downside from here.↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-30 07:39:09 UTC · 原创FT: "China’s CSI 300 index has fallen 9.6 per cent so far this month, its biggest drop since January 2016." This happened in spite of reported support from the "national team". This matters because the most important factor in Chinese stock market performance tends to be the credibility of government signaling.↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-30 07:30:44 UTC · 原创Bloomberg: "With exports soaring at a double-digit pace this year to keep factories humming, economists had expected officials to act with little urgency for now in rolling out any fresh stimulus measures."↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-30 06:52:06 UTC · 原创Just some interesting takeaways from $FORM earnings call on CPO: Since they attributed systems revenue nearly doubled due to: "accelerating growth in co-packaged optics or CPO". Q1 2026: "we now expect 2026 CPO revenues to come in at the high end of the $10 million to $20 million range" Q2 2026 (now): "expect to exceed that range by the end of the third quarter, and to significantly exceed the $20 million level for the year overall." Management stated "We’re seeing some significant acceleration in this [CPO] business". Q: CPO adoption outlook. Are we perhaps expecting the timeline to accelerate a little bit? A: "What I would say is there’s acceleration here in the very short term." But management stated there's not a significant pull-in on timelines. Citing production infrastructure (not just R&D) for: - The growing volumes of CPO chips planned for later this year + test insertion for scale-up and scale-out switches. - Seems like that CPO piece flows through $TSM COUPE [ likely maps to $NVDA CPO products] - "The CPO piece does flow through that 10% customer" So now, test is ramping given management cited growing volumes of CPO chips. "The rapid recent growth of our CPO business is an exciting development, which we believe represents the very early stages of widespread adoption of silicon photonics in the broader semiconductor industry" TLDR: Fundamentally, CPO as a theme is very early, and should start to go brrr soon since it's starting to show up in earnings now (equipment/test players usually appear up first before optical engines / laser volumes ramp). Personally once all the deleveraging stops, I think markets will start to care more.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 06:06:19 UTC · 原创[Four Key Questions from Samsung Electro-Mechanics’ Conference Call] Q1) (Company-wide) What is the outlook for Q3 earnings? A1) In Q2, the company delivered year-on-year and quarter-on-quarter growth, driven by increased sales of high-value-added MLCCs and FC-BGAs for AI servers, networking, and automotive applications, as well as an improved product mix resulting from a higher proportion of premium products. In Q3, robust demand for related components is expected to continue, supported by sustained investment in AI data centers and advances in autonomous driving. Supply-demand conditions for MLCCs and FC-BGAs are also expected to tighten further. For MLCCs, the company plans to actively expand the supply of high-value-added, premium-priced products for AI servers, networking, and ADAS applications, including ultra-high-capacitance, high-temperature, and high-voltage products. It will continue discussions on strategic long-term supply agreements with major hyperscalers and semiconductor companies, while also responding proactively on pricing in light of supply-demand conditions. For FC-BGAs, the company plans to actively expand mass production of new products for AI accelerators and server CPUs supplied to global big tech companies, as well as the supply of high-end products for data centers. It will also continue price negotiations reflecting market conditions. Domestic and overseas capacity expansions tied to LTA discussions with top-tier semiconductor customers will proceed as planned. For camera modules, the company plans to increase the supply of high-value-added automotive products designed for increasingly advanced ADAS applications and ramp up mass production of new products for humanoid robots. In Q3, the continued tightening of supply-demand conditions for high-value-added products such as MLCCs and FC-BGAs, the expansion of LTAs, and higher average selling prices are expected to drive record-high earnings. These trends are expected to continue into Q4 and strengthen further in 2027, supporting sustained earnings improvement. The company will focus its capabilities on high-growth markets such as AI data centers and automotive applications, while responding quickly to changes in market conditions and supply chains through measures including proactive capacity expansion. In new business areas such as silicon capacitors and glass substrates, the company will further strengthen collaboration with global big tech customers and proceed with its preparations for commercialization over the medium to long term. Q2) (MLCC) What is the potential for additional supply agreements for AI-related MLCCs, and what is the company’s future capacity strategy? A2) The AI server MLCC market is expected to maintain rapid growth as the continued expansion of hyperscaler data-center investment, improvements in GPU computing performance, and faster server-network speeds drive increasing adoption of cutting-edge, high-end MLCCs. Cutting-edge products, including ultra-high-capacitance products of 100㎌ or more and high-temperature products rated at 125°C, require the highest levels of technology in terms of both performance and reliability. Only a small number of suppliers, including the company, are capable of meeting these requirements. As a result, supply-demand conditions are becoming increasingly tight, while demand for long-term supply agreements to secure stable supply volumes is rising among major AI server customers. The company has completed LTAs with around 10 customers, including top-tier hyperscalers and major semiconductor companies, and is actively responding to additional LTA requests from key customers. To strengthen its leadership in the AI server MLCC market, the company is improving yields and productivity while expanding Capex compared with previous years to establish a production base capable of maximizing supply. As technical requirements for AI server MLCCs advance rapidly, the company will leverage its ultra-high-capacitance product portfolio to launch cutting-edge specifications across different sizes, voltages, temperatures, and operating conditions in a timely manner, enabling it to respond quickly to customers’ diverse requirements. The company will reinforce its technological leadership in the AI server MLCC market while continuing to deliver revenue growth that outpaces the market. Q3) (Silicon Capacitors) The company disclosed a KRW 1.5 trillion silicon capacitor supply agreement in May. What is the market outlook, and is there potential for additional orders? A3) Silicon capacitors offer superior electrical characteristics compared with MLCCs, including low equivalent series inductance, or low ESL, and are also thinner. They are therefore key components placed in the closest proximity to semiconductor chips within AI server semiconductor packages, supporting overall system performance and stability. Their applications are expanding from mobile devices to AI data centers, and the market is expected to continue growing as the adoption of silicon capacitors in AI data centers increases. Based on the complementary characteristics of MLCCs and silicon capacitors, the company is currently discussing additional supply agreements with major AI-related semiconductor companies. Leveraging its position as the only company worldwide capable of providing MLCCs, silicon capacitors, and package substrates on a turnkey basis, the company intends to take the lead in securing the emerging market and drive substantial revenue growth. Q4) (FC-BGA) The company is reportedly considering capacity expansion. What are the detailed expansion and future capacity-management plans? A4) As the AI paradigm shifts from training toward inference, global big tech companies are increasingly developing proprietary semiconductor chips optimized for inference, with the number of participating companies also rising. As AI semiconductors become more powerful, larger chip sizes and increased HBM adoption are accelerating the transition toward ultra-large and highly multilayered package substrates, further consuming suppliers’ available capacity. Recently, major customers have also increased their requirements for high-end substrates incorporating features such as embedded components and multi-core designs. Only a small number of suppliers with the necessary mass-production capabilities and technological expertise can meet these requirements, further exacerbating the shortage of high-value-added substrates. The company is discussing strategic LTAs, including investment support, with most top-tier semiconductor customers involved in AI and data-center applications. Its capacity expansion plans will be based on these agreements. The company will strengthen collaboration with customers on next-generation product development and concentrate its capabilities on securing differentiated technologies. At the same time, it will proceed with capacity expansion as planned to meet customer demand in a timely manner and expand the business over the medium to long term.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 05:45:09 UTC · 原创I don’t understand why this FundaAI report hasn’t received more attention. SpaceX plans to build 4 GW of compute capacity using Rubin racks next year. $NVDA $SPCX↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-30 05:02:34 UTC · 原创1/2 Good Yasheng Huang piece on China's K-shaped economy and the K-shaped focus of its policymakers: "China has a K-shaped economy — its upper micro-technology arm pointing up, toward technology, and its lower macro arm pointing down, toward almost everything else that affects the welfare of the ordinary Chinese people. An equivalent framing is that the upper arm is the supply side and its lower arm is the demand side. This will get you to the same place as the analysis above, which is that China has a formidable supply capacity on top of a lackluster domestic final demand." @YashengHuang↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 04:50:35 UTC · 引用Agree with Mr. Pettis on this There is ample precedent for currency diplomacy And not much precedent for negotiating over deeply national policies↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-30 04:42:36 UTC · 回复@Brad_Setser @andrewbatson More importantly, the EU can more credibly demand and force a currency adjustment, and over a short time period, than it can demand that China change its pension system, eliminate the hukou, transform its banks, or strengthen its labor unions.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 04:08:13 UTC · 引用"The tech sell-off has created some of the most attractive investment opportunities since early 2025."↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 04:03:01 UTC · 原创Bruh↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-30 03:48:06 UTC · 原创JakeSullivan: "Whenever the United States has faced existential tests, its leaders have returned to Hamilton’s logic." via @ForeignAffairs↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 02:08:57 UTC · 原创Samsung Electronics 2Q26 Earnings Call Q&A Q. What is the current foundry utilization rate? A. Utilization rates have improved across all process nodes. At advanced nodes of 8nm and below, we have maximized sales on strong demand from high-growth applications, and utilization has reached peak levels. Regarding yields, based on our mass-production experience with the first-generation 2nm process, we will begin the full-scale ramp-up of our second-generation 2nm process in the second half of the year. We continue to stabilize yields and improve productivity through process optimization. In terms of profitability, the combined effects of higher utilization, improved yields, and price increases have driven a significant year-over-year improvement, and the recovery is now gaining momentum. As for the timing of a return to profitability, the order-based nature of the foundry business makes it difficult to provide a precise timeframe. However, we expect to achieve a turnaround in the near future. Q. What are Samsung’s foundry capacity-expansion plans? A. The pace of advanced-node capacity expansion has not been sufficient to keep up with demand. To address this, preparations for the Taylor Fab 1 to commence operations in 2026 are progressing as planned. We intend to expand our 2nm capacity in phases. To respond to growing customer demand in a timely manner, we are also preparing to begin construction of Taylor Fab 2 by the end of this year, targeting mass production in 2030. In addition, as customer inquiries regarding the 1.4nm process increase, we are reviewing the possibility of securing an additional fab. Detailed plans will be developed in phases based on customer orders and the outcome of ongoing discussions. For mature nodes, we are shifting toward a higher-margin business mix and concentrating our capabilities on demand for high-value-added specialty products. Specifically, we are expanding capacity for 8nm, 17nm, CIS, and embedded nonvolatile memory, or eNVM, where demand is increasing. We are also focusing on areas such as integrated silicon capacitors and silicon photonics.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 02:02:48 UTC · 原创SAMSUNG ELECTRONICS: WE PLAN TO ALLOCATE 60–70% OF OUR TOTAL CAPACITY TO LTAS.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 02:00:10 UTC · 原创Samsung Electronics 2Q26 Earnings Call Q&A Q. Could you provide an update on LTAs? A. Regarding multi-year memory supply agreements, demand for securing medium- to long-term AI service infrastructure is surging, while the market increasingly expects the memory supply shortage to persist for an extended period. As a result, many customers are requesting LTAs. These customer requests align with our objective of hedging medium- to long-term business risks. We are therefore prioritizing negotiations with customers that have firm future demand and are capable of entering into agreements with strong, mutually binding commitments to contract performance. However, customers are requesting additional supply across both DRAM and NAND, while market prices continue to rise. With virtually every customer seeking an LTA, it is difficult to accommodate all supply requests within our available capacity. To maintain sufficient supply flexibility to serve customers outside multi-year agreements, we plan to allocate approximately 60–70% of our total capacity to LTAs. Additional customers are seeking LTAs, while some customers that have already signed agreements are requesting increased volumes. To provide further detail, our LTAs generally have a five-year base term and follow a rolling structure under which additional periods are added through negotiations in subsequent years. By establishing a five-year rolling-contract structure and securing a certain proportion of our business through committed orders, we expect to achieve a more stable business model. We have already completed agreements with five major global data center customers and are in the final stages of negotiations with five additional large customers. We expect the number of LTA customers to increase over time. Once the agreements currently under negotiation are finalized, we expect LTA volumes to comfortably reach 60–70% of our medium- to long-term production plans for both DRAM and NAND. We also plan to discuss additional contract requests and are reviewing various medium- to long-term capacity-expansion scenarios. To strengthen the binding nature of these multi-year supply agreements, the contracts include substantial advance payments that are held over multiple years. We have already received approximately one-quarter of the agreed advance payments, and the total amount is expected to increase as additional agreements are signed. We ask for your understanding that we cannot disclose the amount of the advance payments due to confidentiality agreements with customers. Regarding the contract model, the agreements incorporate pricing mechanisms designed to offset future risks to an appropriate degree. Product pricing varies depending on each customer’s product portfolio, and contract terms are determined through mutual consultation with customers. For commodity products, pricing is structured to sufficiently hedge future investment risks arising from fluctuations in market prices. Our core principles are to secure medium- to long-term demand visibility, support the healthy development of the AI industry, and maintain a balanced customer portfolio. These principles are also reflected in our LTAs. Given the memory industry’s longstanding pattern of repeated upturns and downturns, expanding the proportion of business covered by LTAs should significantly improve our business stability and visibility. We will continue to expand our portfolio of AI applications with strong medium- to long-term demand visibility and maintain our leadership in the memory market.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 01:55:21 UTC · 原创Samsung Electronics 2Q26 Earnings Call Q&A Q. Growing demand for AI memory is driving strong growth in enterprise SSDs, drawing greater attention to the NAND market. What is the current state of Samsung’s NAND business, and what are its competitive strengths? A. Market attention has traditionally tended to focus more on DRAM than NAND. However, with the recent proliferation of agentic AI, demand for server SSDs is increasing significantly across various segments, including AI servers, general-purpose computing servers, and dedicated KV-cache storage. Led by server SSDs, NAND is also being transformed into a higher-value-added business. In fact, NAND contributed significantly to the improvement in our Memory Business earnings in the second quarter. In the NAND business, we are enhancing our medium- to long-term business visibility through multi-year agreements with major customers. At the same time, we are proactively addressing demand for AI-related server SSDs. Server SSDs are expected to account for more than 60% of our NAND revenue, an increase of over 20 percentage points year over year. For PCIe Gen6 products designed to address demand for TLC-based storage, we have received highly differentiated feedback from customers. Accordingly, we expect to lead the Gen6 market for new AI platforms and translate our solutions leadership into tangible business results. For QLC, we are expanding our lineup with products such as the V9 2TB, and QLC bit shipments in the second half are expected to more than double compared with the first half. Mass production of BV NAND, which incorporates our next-generation triple-stack technology, is scheduled to begin in August. We will address growing customer demand by expanding advanced-node production while ensuring the timely supply of V10-based products for applications requiring high I/O speeds. Going forward, we will continue to develop products in a timely manner in response to customers’ evolving technological requirements and maintain our leadership in bringing next-generation products to market.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 01:51:00 UTC · 原创Samsung Electronics Q2 Earnings Call Q&A Q. What is the supply mix between HBM and conventional DRAM? As stated during our previous earnings call, we manage our product mix independently of differences in profitability, with a focus on capturing AI-driven demand. We are applying the same approach to the growing demand for HBM. In particular, customer qualification for HBM4—a product drawing significant market attention—has progressed smoothly. We are rapidly expanding supply in line with customers’ ramp-up schedules in the second half of the year. Supported by increased 1c capacity and improved yields, we continue to expand our supply capabilities. As a result, HBM4 revenue in the third quarter is expected to more than triple quarter over quarter. For the second half as a whole, HBM4 is expected to account for well over 60% of our total HBM revenue. During the second half, we expect to increase our HBM market share to a level comparable to our overall DRAM market share, thereby achieving a balanced DRAM portfolio. (Translator’s note: Does this imply that Samsung expects to overtake SK hynix and become the No. 1 player in HBM market share?) Looking ahead to next year, we believe we have sufficient product competitiveness, based on the outcome of supply discussions with customers for 2027—which have already been completed—and our technological capabilities, as demonstrated by the industry’s first sampling of HBM4E. That said, we will continue to monitor the explosive growth in demand for general-purpose computing driven by agentic AI, as well as the relative growth of HBM and server DRAM. We will maintain an optimized product-mix strategy, targeting similar market-share levels in HBM and conventional DRAM to ensure a balanced portfolio.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 01:47:42 UTC · 原创Samsung Electronics Q2 Earnings Call Q&A Q. Is Samsung considering an ADR listing? An ADR listing would go beyond simply listing shares in an overseas market, potentially providing benefits such as access to capital and a broader investor base. However, it would also require a comprehensive assessment of factors including additional disclosure and operational burdens. ADRs involve a limited portion of a company’s shares being traded in a separate market, which could create valuation discrepancies between the ADRs and the underlying shares depending on liquidity and supply-demand conditions. We must carefully assess whether such a move would benefit all shareholders who have placed their trust in the company and invested in it over the long term. Given our cash-generation capabilities supported by a diversified business portfolio, the need for an ADR listing is currently not particularly high. Although some media reports have raised the possibility of a U.S. listing, we are not currently considering issuing ADRs. That said, from a mid- to long-term perspective, we remain open to considering it as one of various potential options. Q. Could you provide an update on major foundry customers? Regarding the Foundry Business’s advanced-node order activity in the second quarter, we secured 2nm projects from a major CSP customer and HPC customers. These customers have begun product design work. We are also discussing projects with major customers, including Broadcom, and discussions regarding advanced-node orders are expanding. Building on this order momentum, we expect the number of 2nm projects secured in 2026 to more than double year over year.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-30 01:40:05 UTC · 原创From Samsung Electronics Q2 Earnings Call Q: Do you expect the current memory shortage to persist into next year? If possible, could you also share your medium- to long-term outlook for memory demand? A: The rapid acceleration of agentic AI is driving an explosive increase in token consumption. This is fueling unprecedented demand not only for AI servers but also for general-purpose computing servers. In practice, AI frontier model developers that have been unable to secure sufficient cloud capacity from hyperscalers are now requesting allocations from neocloud providers as well. This has translated into large-scale memory procurement by server OEMs that primarily serve those neocloud customers. Even so, memory shortages mean that many frontier AI companies are still unable to secure the infrastructure they need. To address this, they have begun sharing their medium- to long-term demand forecasts directly with us and expressing their intention to purchase memory from Samsung. We are also seeing the start of requests for long-term supply agreements (LTAs) to secure additional volume. As the adoption of agentic AI continues to accelerate, memory demand is expanding at an exceptionally rapid pace. Industry supply remains well below demand. Even with increased industry-wide capex, it takes more than three and a half years from the construction of a new fab to wafer production. As a result, meaningful supply expansion through new capacity additions will take considerable time. We therefore believe a significant increase in industry supply before 2028 is unlikely. Based on the demand visibility we currently have from customers, a substantial amount of unmet demand will roll over into next year, creating additional supply pressure. We expect the memory shortage in 2027 to be even more severe than it is this year, with tight supply conditions likely to persist into 2028. Looking beyond 2029, it is still too early to make definitive projections. However, as AI token demand continues to surge, large customers building long-term AI infrastructure are expected to continue requesting multi-year supply agreements. These long-term agreements are well aligned with our objective of hedging future business risks. We intend to prioritize contracts with customers that can provide firm, long-term demand commitments. Over time, this should allow us to transition away from the historically cyclical nature of the memory industry toward a more stable and predictable business model. With improved long-term demand visibility through LTAs, we will be in a better position to execute a more flexible supply strategy. Following our existing approach of securing cleanroom infrastructure in advance and installing production equipment in line with demand, we expect to further strengthen this disciplined and flexible capacity expansion strategy.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 01:33:35 UTC · 原创This, from USTR's (active) press shop, is misleading -- Switzerland is the reversal of gold and pharma front running the swing for the EU is also mostly the reversal of Irish pharma front running↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 01:25:59 UTC · 引用Microsoft isn't alone here -- Eli Lilly also made more in Irish corporate tax than US corporate tax in 2025. Here is the key line from Lilly's 10-k ...↗ @sama · 请求账号 @sama · 2026-07-30 00:44:41 UTC · 引用so excited for this. very close to models that will significantly accelerate scientific discovery; the best way to do this is for us to empower scientists, not to try to figure out everything ourselves. we all deserve the benefits.↗ @sama · 请求账号 @sama · 2026-07-30 00:41:57 UTC · 引用goblin-level blog post↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-30 00:38:46 UTC · 引用Gold definitely mucked up the US trade data in 2025 (and to a lesser degree in 2026) Now it is playing with China's reported 2026 surplus. Fully agree "it's totally ok not to count gold" 1/2↗ @jukan05 · 请求账号 @jukan05 · 2026-07-29 23:54:58 UTC · 原创There’s no mention of LTAs in Samsung Electronics’ earnings presentation. That’s disappointing.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-29 23:51:57 UTC · 原创<Samsung Electronics 2Q26 Earnings Review> Revenue: KRW 171.5 trillion (+130.0% YoY, 1.7% below consensus) Operating profit: KRW 89.5 trillion (+1,813.8% YoY, 5.5% above consensus) Operating margin: 52.2% (+45.9ppt YoY, above expectations) Net income: KRW 71.3 trillion (+1,344.4% YoY, 1.8% above consensus)↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-29 23:13:55 UTC · 引用Just some takeaways from $META | $MSFT earnings calls: Microsoft: - Expected to be FCF positive in 2027, despite the increase in capex (extremely positive for AI buildout that it's funded by operating income) - "Free cash flow was $19.6 billion, reflecting higher capital expenditures" - Quartely capex was $41B, roughly 2/3rds were "short lived assets, primarily CPUs and GPUs" - Expects capex spend will be over $50 billion for next quarter - Capex Guidance at ~$175 billion and 2027 capex roughly the same. Spending plans unchanged and in line. - "Extending the estimated useful life of our data centers from 15 to 25 years" - "We will be among the first cloud providers to deploy next generation rack-scale AI infrastructure based on $AMD Helios and $NVDA Vera Rubin" - "Customer demand continues to exceed available capacity" Meta: - Capex $130-$145 billion (narrowed range), from $125B-$145B. - Meta is receiving offers at a "significant premium" to what they paid for it (compute scarcity, positive for neoclouds like $IREN / $NBIS ) - Expects significant portion of compute (like the 1 GW DC in El Paso) to develop internal models. - Meta has multiple ROI-positive uses for additional compute across its core business (internally, not Meta Compute) - "Finally, we believe that overall industry capacity is going to remain tight for the foreseeable future" - "The industry has under-built historically for the wave of AI adoption, making existing capacity, including our own, extremely valuable" - Susan Li TLDR: - $MSFT and $GOOGL largely sustaining AI capex buildout while remaining FCF positive or through operating incomes. - $META flags available compute materially below demand at least through 2027. And $MSFT also flags compute demand far exceeds supply. - All three hyperscaler capex largely in line with Google hiking capex figures. AI selloff seems extremely overblown now, hyperscalers continuing capex in line (with Microsoft being FCF positive) or even hiked with $GOOGL. Compute scarcity is visible throughout every single hyperscaler ER.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 22:49:36 UTC · 原创that just totally misrepresents standard thinking about BoP adjustment. the issue is the overall trade balance. if the income balance were 0.5 pp of GDP not what is reported and the commodity balance was ~ 4% of GDP deficit, and services 1% deficit getting to 0.5 would mean a manufacturing surplus of 5.5% of GDP -- a big fall v today. no one is suggesting that in equilibrum China would run a deficit in manufactures and export commodities so to speak↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-29 21:27:33 UTC · 原创$AXTI enters long term agreement with $LITE for InP substrates. Lumentum has agreed to pay a deposit of $43.5M as a reservation for capacity.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 21:13:30 UTC · 原创Another measure of the China shock is China's raw trade surplus (a better measure than its BoP surplus imo) as a share of its GDP -- the surplus is now up to around 7% of GDP, only a bit below its 06-08 levels cc @adam_tooze↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 20:37:22 UTC · 原创Microsoft, per its 10K, paid more in tax in fiscal 2025 in Dublin (Ireland) than in DC ... Nice when a talking point checks out 1/↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-29 20:03:30 UTC · 原创To show solidarity with the $IREN holders: I now own 1 share of IREN.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 19:47:19 UTC · 原创Adam Tooze, as provocative and insightful as usual "What China is delivering in the 2020s is an industrial policy shock of global scale" 1/↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 19:11:44 UTC · 原创China was targeted by Lighthizer in Trump's first term -- and bore the brunt of the liberation day tariffs. But China has been the surprise winner of Trump's 2026 trade policy -- getting the same tariff as everyone else really help China, as the New York Times reports. 1/2↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-07-29 18:21:54 UTC · 原创The statement: a hold and three dissents—Hammack, Kashkari, and Logan wanted a hike now. No rationale for the hold, no hint at what a hike would take. Not one word of the data discussion changed after six weeks of inflation and jobs surprises. None, zip, nada. Never seen that.↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-29 18:10:55 UTC · 原创Federal Reserve leaves rates unchanged at unchanged at 3.50% - 3.75%. So far so good, since markets would have sold off on an immediate rate hike (seem to have largely frontran hike expectations off already). Think everyone is waiting for the Fed Chair commentary now.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 17:40:42 UTC · 引用a) the IMF's standard model suggests a 10% (inflation adjusted currency move) changes China's real exports by 1.5 pp of its GDP, which more or less is what happened after the 22/23 depreciation b) empirically, China's export growth clearly took off after the depreciation↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 17:25:14 UTC · 引用Some of this have argued that the exchange rate can and should be part of the initial ask -- and that there isn't a need for China to first adjust its domestic policy settings ... 1/2↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-29 17:07:41 UTC · 原创US Gov to take 1% stake in $GFS, and award them $300m for the US CHIPS ACT. This is actually a strong read through on $SIVE / $LITE, given this CHIPS ACT is specifically aimed at advancing CPO + Silicon Photonics. (For reference, Sivers and Lumentum were the only two public laser suppliers named in GFS presentation slides. Sivers laser arrays was named recently as a reference design in Globalfoundries SCALE for CPO). Per US Gov announcement: "GlobalFoundries will receive up to $300 million to accelerate domestic CPO R&D by two to three years" (NIST) Never thought we'd see the US Gov / $INTC foundry playbook for CPO in specific...↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 17:06:13 UTC · 引用Good points from an important French voice -- Fully agree that China can/ does use the exchange rate as a policy tool (i.e. a macro-industrial policy that supports both exports and Chinese industrial policies)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 16:32:01 UTC · 引用Currency adjustment is indeed the classic solution to a large external surplus/ deficit Surprising to me that this is resisted so strongly right now↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-07-29 13:43:10 UTC · 引用The Fed decides today, and markets are unusually uncertain about what's coming. Hike odds were running near 30% yesterday. That's not normal; it's the second-most uncertain scheduled Fed meeting in 15 years. A short thread:↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-29 13:36:08 UTC · 原创1/4 WTO: "The rise in trade policy activity is not driven by the largest economies alone: the upward movement is broad-based, with non-G20 economies also rising markedly in the most recent period."↗ @jukan05 · 请求账号 @jukan05 · 2026-07-29 13:35:38 UTC · 引用KIS RAISES ITS PRICE TARGET FOR SK HYNIX BY 24% TO KRW 4.7 MILLION (TODAY’S CLOSING PRICE: KRW 1.41 MILLION).↗

YouTube 更新

今日最值得看

AI Engineer：fighting slop with slop — Vaibhav Gupta, Boundary

视频讨论用另一层持续运行的智能体检查其他智能体的执行记录，包括标记幻觉、工具调用错误和方案差异。其核心工程取向是让廉价的生成层服务于检测，同时用类型系统、设计规则和 CLI 检查维持不可轻易改变的硬边界。它提醒团队：生成速度提升后，收敛性与可验证性更值得投资。

Computerphile：ITS, Eliza & MiT’s AI Lab - Computerphile

节目回顾 MIT AI Lab 的 Incompatible Timesharing System（ITS）及其历史语境，并展示相关模拟项目。它提供了一个有益对照：今天围绕 AI 的工具与交互争论，也延续着早期计算文化对自主性、可塑性和共享系统的追求。

其他频道更新

AI Engineer · 2026-07-31First Steps Toward Automated AI ResearchRichard Socher 把自动化研究描述为围绕假设、试验和修正的循环。视频展示了递归改进、架构搜索和 CUDA 内核优化的早期样本，但也明确这些距离通用科研自动化仍有很长距离。↗ AI Engineer · 2026-07-30Benchmarks: The Good, the Bad, and the Ugly分享用编码基准中的模糊任务和脆弱测试说明排行榜可能如何错判答案。建议是区分公开与保留集、让规格可执行，并把评测工具本身按生产系统维护。↗ AI Engineer · 2026-07-30Reinforcement Learning without Verifiable Rewards视频讨论没有标准答案的任务如何构造强化学习信号，并把环境设计、文档与仓库中的证据放在中心。演讲同时强调奖励投机风险，需要检查轨迹、小规模实验与专家审查。↗ AI Engineer · 2026-07-30Your Finance Agent's Bottleneck Is You节目主张把开发者自身的排障、测试与交付循环交给多个编码智能体协同处理。财务场景中仍应由人核验源数据、审批结果和最终操作，自动化不应替代责任链。↗ AI Engineer · 2026-07-30Build for the Memo, Not the Demo分享区分了展示效果与投资或运营决策所需的可审计备忘录。它强调来源等级、数值勾稽和事实与估计的明确标签，避免生成内容在多次转述后被误读为确定事实。↗