每日信息简报｜2026-08-04

DAILY BRIEFING · 2026-08-04

模型进入部署深水区，算力与透明度一起成为约束

今天的线索集中在三处：实时语音与工作流正在把模型能力推入生产环境；算力、半导体与评估体系决定扩张的速度；监管与数据治理则开始要求能力“可被识别、解释与审计”。

今日要点

欧盟 AI 透明度义务开始生效

The Verge 报道称，欧盟《AI 法案》新增的透明度义务已于 8 月 2 日生效，覆盖与聊天机器人交互、深度伪造内容等识别要求。对产品团队而言，问题不再只是“能否生成”，还包括生成内容如何标识、用户如何知情。合规实现会逐步成为模型接入与内容分发链路的一部分。

Qwen3.8-Max 与开源模型竞争继续升温

The Verge 汇总阿里发布 Qwen3.8-Max，并称其能力可与美国前沿实验室及国内竞争者的系统相比。此类厂商声明仍应与独立基准、具体任务和可复现实验分开看；但模型供给的竞争已从单纯参数规模转向可用性、部署成本与生态。对开发者来说，比较时更值得关注实际工作负载中的延迟、工具调用和数据边界。

GPT-Live 展示连续语音交互的系统工程

OpenAI 的技术说明介绍了 GPT-Live 的无显式轮次语音交互与低延迟架构。它说明实时体验的难点不仅在模型本身，还在端点检测、打断处理、流式编排和故障恢复。连续对话若要可用，必须把“自然感”拆成可测的时延、抢话与上下文切换问题。

AI 与科技

私有云中的 vibe coding 开始成为部署选项

TechCrunch 报道称，AWS 允许 Superblocks 部署在客户私有云中。若该模式扩展，应用层与模型层会进一步解耦，企业可在自身网络、权限和数据边界内使用生成式开发工具。真正的门槛会落在治理、审计、依赖管理及生产变更控制上。

多模态对齐不能只复用文本模型的方法

Apple Machine Learning Research讨论多模态大模型中的偏好对齐与幻觉问题。图像理解的错误既可能是事实错误，也可能是回答与图像内容不一致，因此评测需要直接检验视觉依据。对齐研究的价值在于把“回答看起来合理”转化为可验证的图文一致性目标。

半导体材料与制造成为 AI 扩张的硬约束

NVIDIA Developer将 AI 工作负载增长与材料工程、制造节奏联系起来。算力路线并不只由芯片架构决定，良率、供给、封装与生产延误都能改变系统交付速度。把 AI 基础设施当成单一软件问题，容易低估这些物理约束。

值得深读

中国与美国的 AI 投入效率之争

《经济学人》提出，中国 AI 投资规模虽落后于美国，但模型表现并未按同样比例落后。该文适合用来追问“投入效率”到底由什么衡量：训练成本、推理价格、人才、数据、部署规模，还是产业吸收能力。任何跨国比较都应区分媒体叙事、估算口径与可验证数据。

LLM 能力主要来自模仿学习还是强化学习

LessWrong Curated 的长文主张，尽管 RLVR 讨论热度很高，预训练与监督微调仍解释了当前模型的大部分能力。这个观点的意义不只在训练配方，也关系到思维链可解释性、能力来源与对齐路径。阅读时应把作者的论证与不同模型、任务类型和训练阶段分开评估。

车牌摄像网络的治理漏洞

Longreads 的调查梳理了至少 50 名执法人员被指控或被控滥用 Flock 及其他系统的案例。它把“为了治安部署传感器”的抽象命题落到访问权限、日志审计和个人滥用风险。技术系统扩大信息能力时，也必须同步扩大可追责性。

快速浏览

纽约时报 AI“Tokenomics”正在尝试衡量企业 AI 支出的回报，提醒讨论从投入规模转向单位经济与实际产出。↗ TechCrunch AI美国国会办公室的付费 AI 使用记录中，ChatGPT 用于起草、摘要和选民沟通；这是一项采用情况报道，而非效果评估。↗ TechCrunch AISiri 的 AI 更新被放在更高的用户预期中审视：功能改善并不自动等于产品差异化。↗ TechCrunch AI初创公司 June 获得融资，目标是降低 AI 采用难度；部署流程本身正成为一个独立产品层。↗ OpenAICircles 案例称其用 OpenAI API 与 Codex 改造电信个性化；文中业务指标属于客户案例陈述，应结合独立数据理解。↗ 纽约时报 AI婴儿监测创业公司希望延长数据收集周期，家庭设备的数据最小化与二次使用边界值得关注。↗ Longreads一则网络恋爱诈骗追查故事，把冒充、跨平台取证和受害者支持放在同一视角下。↗ BBC莱茵河等欧洲河流低水位影响货运和发电，气候风险正在通过基础设施与物流传导。↗ 机核ChinaJoy 2026 的现场报道关注 AI、游戏制作、具身智能与消费场景的交叉；其中规模数据来自报道援引的主办方与测算。↗ The Verge AI一首进入榜单的歌曲引发 AI 生成质疑，内容来源鉴别正从图像与文字延伸到音乐。↗

@Brad_Setser · 请求账号 @Brad_Setser · 2026-08-04 01:11:23 UTC · 引用原帖：Folks really have trouble not thinking in "fiscal"/ public debt terms -- the issue isn't just that Argentina's isn't in a position to repay the IMF; it is that Argentina has drawn down its (net) reserves under Milei and may need official support to get through the next election↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-04 00:49:53 UTC · 引用原帖：I think there are a few more considerations here -- at least for me. Countries like Korea and Japan have long preferred weak currencies to juice their exports, and used government policy (reserve accumulation, government pensions, etc) to achieve that goal 1/↗ @jukan05 · 请求账号 @jukan05 · 2026-08-04 00:06:48 UTC · 引用原帖：Google has joined the HBF consortium led by SK hynix and Sandisk. You know what that means, right?↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-04 00:04:27 UTC · 引用原帖：I understand why Argentina is reluctant to lock in relatively high interest rates on new bonds -- its impressive fiscal performance has been helped considerably by the low interest rate it pays on its external debt (thanks to the 2020 restructuring) But ... 1/many↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 23:59:18 UTC · 引用原帖：Put simply, the Rubin Ultra HBM despec is a supply issue. Not a demand issue. I think as much additional HBM will be produced as the supply headroom freed up by the shift from 12Hi to 8Hi. It is not a problem at all.↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 23:56:31 UTC · 原创原帖：Korea, Eugene Investment & Securities: On the NVIDIA Rubin Ultra HBM despec NVIDIA appears to be discussing a plan to configure the main version of Rubin Ultra with a 2 die structure, HBM4 8Hi, and 8 stacks, and to add an HBM4E 8Hi version at a later point. Assuming NVIDIA's 2027 CoWoS capacity allocation is 1,200k, NVIDIA's accelerator output is estimated at 9.9M units and Rubin Ultra output at 1.6M units. If Rubin Ultra shifts from HBM4E 12Hi 384GB to HBM4 8Hi 192GB, NVIDIA's 2027 HBM demand would be revised down by roughly 10%, from 24.1bn Gb to 21.6bn Gb, and total HBM demand could decline by about 4%. Of course, this assumption does not reflect the possibility of a further expansion in accelerator output. That said, we judge this despec to be a strategy for producing more accelerators out of a limited pool of HBM bits, as the increase in HBM supply from DRAM makers has failed to keep pace with the speed of TSMC's front end and back end capacity expansion. DRAM makers find it difficult to expand HBM capacity flexibly because of the extreme shortage in commodity DRAM. With consumer set makers in smartphones and PCs experiencing production disruptions due to memory shortages, it is also not easy for memory makers to cut allocations to these customers any further.↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-08-03 23:35:14 UTC · 原创原帖：Be Citadel: > calls for unexpected rate hike few days ago, causing market panic > reaches out and buys billions of liquidated hedge fund AI assets > no rate hike > “Bull market drivers are fully intact”↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 22:45:28 UTC · 原创原帖：DigiTimes: 2027 DRAM/HBM capacity already sold out According to a DigiTimes report, annual 2027 DRAM and HBM capacity has been fully booked out ahead of schedule. NAND flash supply isn't as tight as DRAM, but bookings are expected to be completely filled by the end of this month. Per sources cited by DigiTimes, companies aren't openly acknowledging that they need to lock in memory by this month — because they worry that if more players pile into the scramble for supply, their own allocations will shrink. Manufacturers also typically deliver only 60–70% of the volumes buyers originally requested, as memory makers prioritize demand from CSPs and AI majors. As a result, the DRAM volumes smartphone and PC makers can secure in 2027 are expected to fall markedly versus 2026. DigiTimes, citing industry sources, noted that with capacity now effectively almost entirely sold out, companies that have yet to finalize their volumes could face an even more acute "can't buy it even if you want to" situation in 2027. The report added that elevated memory prices are set to become the norm. While the sharp price spikes driven by the fight for allocation may subside — meaning the magnitude of 2027 price increases could be more moderate — the overall supply shortfall and the cost burden on the finished-goods industry are unlikely to be resolved in the near term.↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-08-03 21:16:45 UTC · 原创原帖：The bears have gone back into hibernation↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-03 17:31:18 UTC · 引用原帖：Thanks Andrei 1/2↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-08-03 16:23:47 UTC · 原创原帖：FT: "The growing popularity of institutions such as police and military academies reflects a broader shift among young Chinese, who are increasingly willing to trade academic prestige for career certainty."↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-03 16:09:31 UTC · 引用原帖：Supports my thesis that a stronger yuan and less support for growth from net exports is necessary to get China to change its domestic policy settings ... 1/2↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 15:47:20 UTC · 原创原帖：[TF Overseas Tech] Why the Rise of Open Source and the Collapse of the Closed Model Monopoly Is Good for Cloud The biggest worry weighing on the cloud industry has been that cloud providers are the ones paying to build the roads of the AI era, yet may end up as nothing more than infrastructure suppliers for closed models. Cloud providers commit enormous capital to buying GPUs and building data centers. But if most workloads have to call a handful of closed models, the demand gateway and pricing power belong to the model companies. Cloud providers carry the capital intensive investment while the incremental profit goes to the model layer. That structure is now starting to change. Open source models keep getting stronger and cheaper. The highest performing closed models will still command a premium on complex reasoning and high value tasks, but they no longer have to be the default choice for every task. The structure of AI calls will shift from solving every problem with one model to tiered routing. The hardest tasks use top tier models, while ordinary tasks use cheap small models or open models. Stable, high volume workloads get deployed by the cloud providers themselves, with in house ASICs pushing inference costs down further. This is a very important shift for cloud providers. The more substitutable models become, the more routing authority, customer ownership, and pricing power cloud providers can capture. Closed models used to be a tollbooth sitting on top of the cloud. Future models will look more like a compute resource that the cloud platform can allocate freely. Cloud providers can rent out GPUs while also hosting open models on their own infrastructure and selling tokens. That reduces the profit handed over to outside model vendors. Better business logic for cloud does not mean this is bad for hardware. Competition among open source models can push more players to invest in training, and falling token prices can generate far larger call volumes. At the same time, model compression, inference optimization, routing, and in house ASICs reduce the general purpose GPU demand required per token. Final hardware demand therefore depends on demand elasticity. If token counts and compute per task grow faster than algorithmic and chip efficiency improves, total compute demand will keep expanding. Conversely, even if tokens grow rapidly, capex growth rates may peak. The most complete bull case is therefore not simply that cheaper models are good for cloud and that cloud making money is good for hardware. Open source models weaken the monopoly rent of the model layer, and cloud providers raise the monetization efficiency of compute through routing and vertical integration. Lower inference prices create additional new demand, and demand growth once again outpaces efficiency gains. As a result, data centers can sustain high utilization and relatively high ROIC, which ultimately drives continued investment in compute by cloud providers. Only if this virtuous cycle holds can a true positive feedback loop form between cloud and hardware.↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-08-03 14:11:26 UTC · 原创原帖：Warsh has enlisted the "very best minds" to remake the Fed. Missing from every task force: anyone representing regular people or labor. But the Fed doesn't need a new task force to hear from the public—it already listens. My latest for @opinion:↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 12:08:01 UTC · 原创原帖：According to Korean press reports, SK Hynix has been unable to play its shareholder-return card because of disclosure constraints tied to its ADR listing. The U.S. Securities and Exchange Commission (SEC) restricts the disclosure of material information not included in the prospectus for a certain period following an ADR listing. The concern is that if a newly listed company announces material information that differs from the documents filed at the time of listing, it could face legal risk under U.S. securities law. The market understands the restriction to apply for 25 days, weekends included, from the listing date. Given that SK Hynix's ADRs listed in the U.S. on the 10th of last month, that means a concrete shareholder-return plan could be announced any time after the night of the 4th, Korea time. $SKHY↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 08:46:06 UTC · 原创原帖：Reuters reported that China’s CXMT is considering building a second DRAM fab in Beijing’s Yizhuang district, in addition to the new fabs currently under construction in Shanghai and Hefei. Including the Hefei and Shanghai fabs now under construction, CXMT’s total capacity is expected to more than double from current levels to over 600,000 wafers per month. The proposed second Beijing fab is not included in that figure.↗ @thsottiaux · 请求账号 @thsottiaux · 2026-08-03 08:37:22 UTC · 引用原帖：Crazy thing about OpenAI is you just open your laptop, codex a PR into existence and ship an improvement to 1B users. Just like that.↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 08:03:18 UTC · 原创原帖：Kimi Developer Moonshot AI Plans to File for Hong Kong IPO This Month, Targeting $3 Billion Raise • Moonshot AI is expected to submit its Hong Kong IPO application as early as this month, seeking to raise approximately $3 billion. • The company recently closed a Series F funding round of more than $3.5 billion, lifting its valuation to $35 billion. The round was more than three times oversubscribed and closed ahead of schedule. Its Series G pre-IPO round, originally scheduled to begin in August, has also been brought forward, with the pre-money valuation rising to $50 billion. • One source said, “Given that the company raised $3.5 billion in its previous round, it is difficult to believe that it would raise only $3 billion through an IPO.” The company has not issued an official response.↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 07:28:02 UTC · 原创原帖：Samsung Electronics Foundry Expected to Reach 100% Utilization Within the Year... Countdown to Profitability Samsung Electronics' foundry division is understood to be targeting 100% utilization within the second half of this year. Current utilization is estimated at 70 to 80%, and given the order backlog and the state of contract negotiations, the mood inside and outside Samsung is that hitting the target is all but certain. With the foundry lines, which endured a prolonged slump at below 50% utilization from 2024, now approaching full operation after roughly a year, a turning point also appears to be forming in the division's chronic loss making structure. According to industry sources on the 3rd, behind the rise in utilization are high bandwidth memory (HBM) base dies and expanding demand for advanced products centered on large US Big Tech customers. With 4nm applied to sixth generation HBM (HBM4), the analysis is that a structure has formed in which the memory boom feeds directly into foundry utilization. On top of that, 2nm orders are expanding among major cloud service providers (CSPs) and artificial intelligence (AI) and high performance computing (HPC) customers, and project discussions with large customers such as Broadcom are continuing. Utilization is the core indicator of how much a fab is actually running. Contract chipmaking carries a heavy fixed cost burden because of the scale of capital investment, and the longer lines sit idle, the more losses snowball. The industry explanation for the foundry division's chronic losses last year is that utilization on mature node lines such as 4nm and 5nm fell below 50%, leaving it unable to cover fixed costs. Conversely, once utilization returns to a normal track, incremental revenue flows straight through to profit. Moves toward customer diversification are also notable. Samsung's foundry has long been criticized for heavy dependence on specific customers, but recently Big Tech firms including Qualcomm, AMD and Google have begun sitting at the negotiating table. Tesla is also reported to be placing next generation chip volumes with the 2nm process, thickening the medium and long term order pipeline. The common view, however, is that turning these into actual large scale mass production contracts requires 2nm yields to settle in the 70% range, and how quickly yields, currently estimated in the 60% range, improve from here is cited as the point to watch. A restructuring of the business mix is proceeding alongside. The advanced node share of revenue is expected to exceed 50% this year, and the AI and HPC share is expected to expand from the high teens last year to more than 30% this year. In the second half, mass production of mobile products on the second generation 2nm (SF2P) process begins. Capacity expansion is also picking up pace. Taylor Fab 1 in Texas is preparing to start operations this year as planned, and Taylor Fab 2 is to break ground within the year with mass production targeted for 2030. With customer inquiries about the 1.4nm process increasing, plans to secure additional fabs are also under review. This trend was also evident on the second quarter earnings conference call held on July 30. Samsung said that day that "utilization is improving year on year across all nodes" and that "leading edge nodes at 8nm and below have reached maximum levels through maximizing sales centered on products with strong high growth demand." It projected that 2nm project wins this year would more than double from last year, and on the timing of a return to profit said it was "difficult to state an exact point" while noting it "should be possible in the near future." Analysts expect the non memory businesses, including foundry and System LSI, to turn profitable as early as the third quarter and by the fourth quarter at the latest. Some also see the pace of earnings improvement running faster than the market expects once wafer price increases are reflected on top of normalized utilization. Parts of the industry, however, offer a more cautious view that whether 2nm yields exceed 70%, rather than reaching 100% utilization itself, will be the watershed that leads to full scale orders from large customers such as Qualcomm and AMD. One industry official said, "Utilization normalization is a leading indicator that the foundry business is passing breakeven, but confirming a real structural improvement requires watching until 2nm yields stabilize and large customers actually convert to mass production."↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 07:23:05 UTC · 原创原帖：ZDNet Korea reports that prices for NVIDIA’s GeForce RTX 50-series graphics cards in South Korea will rise by up to 30% starting this month.↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-08-03 06:38:42 UTC · 原创原帖：For my friends in Beijing, Zhang Shouwang is preforming tomorrow night at Raw Material as part of the Sally Can't Dance series. Whenever he performs it is a highlight.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-03 06:06:08 UTC · 原创原帖：Et tu, Financial Times "in the context of loose monetary and fiscal policy in Japan" The yen was being sold against the dollar, and the US clearly has a looser fiscal policy than Japan (chart is of the headline fiscal balance, but there also has been a big rise in non tariff tax revenue)↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-08-03 04:51:46 UTC · 原创原帖：In top-tier cities, commercial real estate may be bottoming out, at least temporarily: "In Shanghai, the number of commercial property deals jumped 77% year on year in the first half of 2026, while the total value of those transactions was up 18%."↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-08-03 04:45:30 UTC · 原创原帖：Wang Xiangwei: "As long as external demand holds up, policymakers appear relatively untroubled by weakening domestic demand, despite rhetoric to the contrary."↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-08-03 04:38:39 UTC · 原创原帖：Yuekai's Luo Zhiheng: "Each year, the economy has exhibited a U-shaped rhythm. This pattern reflects two underlying features: the economy remains highly dependent on policy support, while self-sustaining growth momentum is still weak and the vitality of households, businesses and local governments has yet to be fully restored."↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-03 03:42:40 UTC · 引用原帖：A blog -- and a thread -- on why a stronger renminbi is actually essential to any adjustment of macroeconomic imbalances. (and a technical rebuttal to the article in the Economist article that argued the contrary)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-03 03:36:36 UTC · 引用原帖：Indeed -- trailing 3m sum annualized is $350b, v ~$100b this time last year. Crazy big number for Korea, or any middle economic power↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 03:18:29 UTC · 原创原帖：MediaTek seems quite optimistic about EMIB-T. (The excerpt below is from MediaTek’s Q2 earnings call.) $INTC↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-03 02:35:28 UTC · 引用原帖：Hmmm -- My guess is that Bessent is more willing to play the SDRs on the Exchange Stabilization Fund's balance sheet than previous Secretaries. That significantly increases the ESF's firepower. Just a hunch↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 02:16:26 UTC · 原创原帖：I’m unable to comment publicly on the SemiAnalysis report released a few days ago. Thank you for your understanding.↗ @jukan05 · 请求账号 @jukan05 · 2026-08-03 00:29:11 UTC · 引用原帖：LFG!↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 23:16:47 UTC · 原创原帖：I am going to post this chart, which shows the strong negative impact of Abe's two big consumption tax hikes on household demand, whenever anyone mention's Abe's stimulus ...↗ @nft_hu · 请求账号 @nft_hu · 2026-08-02 23:09:49 UTC · 引用原帖：很会抓时机了。 请注意，也已经是特朗普第八次叫停对伊朗打击了。↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 23:07:11 UTC · 引用原帖：Interesting "The FIMA Repo Facility is an important backstop. We would encourage it to be upsized in the coming months" Am a big believer in FIMA repo, advocated for it internally and then externally & happy to see its utility recognized 1/2↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-08-02 22:57:06 UTC · 原创原帖：There was a lot in WSJ editorial board piece defending Warsh that I disagreed with but the pot shot at the journalists covering the Fed was the worst.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 21:19:04 UTC · 原创原帖：Nice chart deck from Apollo on the yen but to make the "fiscal worries" case shouldn't there be a few charts on core fiscal variables like the primary balance? Wasn't initially clear from the title if it was going to be a deck on Japan or the US :)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 20:57:16 UTC · 引用原帖：Japan's fiscal mess? "The Cabinet Office said ​the primary budget ​balance is projected ⁠to return to a 1.4 trillion yen ($8.6 billion) surplus in fiscal 2027" Look at that measure for the US (the long in the short yen trade ...)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 20:09:35 UTC · 原创原帖：The US has softened up on China, particularly on currency issues, following the advice of the Economist perhaps :) ... Europe by contrast is ready to push much harder Trump is missing a big opportunity. Good Bloomberg story 1/↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 20:04:11 UTC · 引用原帖：Japan's intervention does tend to show up in the TIC data but read this before jumping on the "China is selling" narrative (mostly likely just diversifying its custodians and moving toward Agencies)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 19:01:37 UTC · 引用原帖：My dad grew up on a cattle farm in central Kansas, so I thought I knew what cattle land looks like -- and generally, a big part is being a little bit under-populated ... didn't realize that China now has the needed grass supply for yet another import substitution project!↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 18:36:06 UTC · 原创原帖：I am a bit baffled by all the comments about Japan's fiscal position -- especially as an argument for selling the yen to buy the dollar. Takaichi is fiddling around a bit (US is fighting a costly war by the way), but the underlying fiscal balance? Seems fine 1/2↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 18:18:08 UTC · 引用原帖：a) plot looks different if you use gross debt and don't penalize Japan for its on balance sheet fx reserves b) thought the CW was that inside the g-10 long bond yields are mostly a function of the expected path of s-term rates (maybe over stated, but cannot be totally discounted)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 17:20:43 UTC · 原创原帖：A new blog, on the importance of exchange rate moves to global adjustment -- Obviously a response to Gopinath, Gourinchas and Rey 1/↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 16:57:27 UTC · 引用原帖：A little overstated I think -- a) When Japan intervenes, it either uses the proceeds of a bond portfolio that has a high level of quarterly roll off, or sells the "belly" of the curve (Japan mostly holds 2s and 5s I think ... )↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 16:15:19 UTC · 原创原帖：this is the number on reason for the fall in the reported bilateral deficit with China, as the trade data is based on the location of final assembly. Underinvoicing actual imports is a distant second! 5/5↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-08-02 16:03:26 UTC · 原创原帖：Great FT piece on Mexico's role in doing final assembly of the AI servers that are the backbone of the current wave of data center investment. Very much tariff driven of course, and the main reason for the surge the US bilateral deficit with Mexico 1/↗ @jukan05 · 请求账号 @jukan05 · 2026-08-02 13:58:01 UTC · 引用原帖：So... Samsung is expected to generate roughly $1.25 trillion in operating profit over the next three years, right? And what was Samsung’s market cap again...?↗ @jukan05 · 请求账号 @jukan05 · 2026-08-02 13:54:12 UTC · 原创原帖：Nomura raised its operating profit forecasts for Samsung Electronics to KRW 391tn ($271bn) → KRW 635tn ($440.1bn) → KRW 774tn ($536.4bn), with FCF projected at KRW 296tn → KRW 456tn → KRW 563tn and shareholder return yields of 11.2% → 17.6% → 22.3%.↗

YouTube 更新

今日最值得看

Latent Space：The Inference Frontier: 10x Faster Models to Self-Optimizing AI

这期访谈从 20 万 token 提示词的 GPU 路由讲到模型服务内核优化，重点在推理工程如何把新模型变成可靠产品。嘉宾以 Baseten 的实践说明，性能问题会同时落在批处理、缓存、路由和内核层。标题中的“10 倍”是节目讨论的方向性表述，具体收益需看工作负载与基准。

Dwarkesh Patel：Why smarter AI models could drive up compute prices 10x

视频讨论更强模型可能如何改变算力需求与价格，而不是把成本下降视为必然。应把模型能力、测试时计算、供给扩张和能源/芯片约束拆开观察。标题中的“10 倍”是作者的推演性观点，不是已发生的市场事实。

Two Minute Papers：Another DeepSeek Moment Has Arrived

视频介绍 DeepSeek V4 Flash 0731，并把它放进开源模型性能与成本竞争中讨论。节目形式适合快速了解论文、模型页和 API 的入口，但结论仍需回到官方评测、价格和可复现任务。所谓“DeepSeek 时刻”是频道的编辑判断。

Greg Isenberg：Why Graph Engineering will 10x your Claude/Codex

视频把 prompt engineering、context engineering 与 graph engineering 区分开，后者强调将 AI 工作组织成可管理的流程图。其核心启发是：复杂任务的稳定性更依赖任务分解、状态管理和检查点，而非一轮更长的对话。“10 倍”属于频道标题主张，应以具体流程指标验证。

其他频道更新

本窗口内其余新视频均为 Shorts；按简报规则不在公开正文收录。