在 New Jersey 的 Buena Vista Township 发生的一起致命车祸，再次引发了对 Tesla 驾驶辅助技术的关注。 2025 年 7 月 6 日，一辆 Tesla Model 3 在停车标志处未停车，撞上一辆 Honda Civic，造成 82 岁 Stephen Field 死亡。尽管最初地方报道将此事归为驾驶员闯停的人为失误，但 Tesla 的内部数据表明情况更为复杂。

根据 NHTSA 的常设命令，Tesla 提交的报告确认事故发生时一套 Level 2 驾驶辅助系统处于激活状态。该报告记录了死亡事故，并显示公司掌握该车的事件数据记录仪和远程信息处理数据，但软件版本、具体事故经过以及道路是否在系统批准的运行区域等关键细节均被 Tesla 以商业机密为由涂黑。

由于基础 Autopilot 被设计为仅用于高速公路车道保持且不会识别停车标志，这些迹象强烈表明车辆可能在运行 Full Self-Driving 软件。无论是 FSD 还是被驾驶员误用的 Autopilot 版本，此事都凸显了系统宣传与实际能力之间的明显差距。报告还记录碰撞前车速仅为 4 mph，这引发了车辆是否在缓行通过或存在驾驶员干预（如误踩油门）的疑问。

Tesla 在事故透明度上的缺失长期备受争议：一方面将软件冠以"Full Self-Driving"，另一方面却封存关键事故数据，助长了对系统过度信任的可能性。随着调查推进，这起案件再次提醒公众——尽管宣传越来越强调自动化，这些仍然是需要持续人工监督的 Level 2 系统。 A fatal car accident in Buena Vista Township, New Jersey, has brought renewed scrutiny to Tesla's driver-assist technologies. On July 6, 2025, a Tesla Model 3 failed to stop at a stop sign, colliding with a Honda Civic and causing the death of 82-year-old Stephen Field. While initial local reports framed the incident as a standard instance of human error involving a driver running a stop sign, internal Tesla data tells a more complex story.

In compliance with a NHTSA standing order, Tesla filed a report confirming that a Level 2 driver-assist system was verified as engaged at the time of the crash. The filing logs a fatality and indicates that the company possesses the event-data recorder and telematics from the vehicle. However, critical details such as the software version, the specific crash narrative, and whether the road was within the system's approved operating area were redacted by Tesla under the claim of confidential business information.

Because basic Autopilot is designed as a highway lane-keeping system that does not respond to stop signs, the circumstances strongly suggest that the vehicle was running the Full Self-Driving (FSD) software. Whether it was FSD or a driver-misused version of Autopilot, the incident highlights a significant gap between the marketing of these systems and their actual operational capabilities. The report notably records a pre-crash speed of only 4 mph, which raises questions about whether the car was performing a rolling stop or if there was potential driver interference, such as unintended pedal misapplication.

The lack of transparency regarding these crashes is a point of contention, as Tesla consistently hides the data that would clarify how its systems behave during emergencies. By branding its software as "Full Self-Driving" while keeping the underlying incident data sealed, the company encourages a level of user trust that may not be warranted. As the investigation continues, this case serves as a stark reminder that these are still Level 2 systems requiring constant human supervision, despite the increasingly automated branding.

Los Angeles 以其独特的建筑景观而著称；这段叙事通过对市内所有仍然矗立建筑的数字化可视化被呈现出来。把每栋建筑以表示其建成年份的方块绘制出来，便能窥见这座大都市随时间扩张的轨迹。值得注意的是，这一视图只反映现存建筑，被拆除的建筑未被计入，因此可视化呈现的是当下的物理足迹，而非完整的历史档案。

时间线从 19 世纪 80 年代开始，记录了基础设施缓慢而稳定的累积，追溯了这座城市在整个 20 世纪直至今天的扩张。用户可以按年代浏览，观察城市密度在不同街区如何变化与转型。从 Downtown 密集的建筑群到 The Valley 广阔的住宅区，地图展示了 Los Angeles 从小型定居点演变为大型都市中心的发展节奏。

可视化工具支持更细致的探索，例如调整高度夸张比例以便在广阔地区更清晰地区分建筑，或聚焦 Wilshire corridor 或 Century City 等特定街区。通过与数据交互，各个建筑年代如何共同塑造城市当下面貌便清晰可见。建筑的高度、占地面积和类型揭示了不断变化的 zoning laws 、经济繁荣周期和建筑潮流在 Los Angeles 天际线上留下的物理印记。

该信息库来自 LARIAC 2020 building outlines 和 LA County Assessor roll，充当一个动态档案。它既是建筑研究的工具，也是居民理解和解读周边环境的方式。把每栋建筑置于其建成年份的语境中，人们可以感受到构成这座城市的历史层次，认识到现代城市是过去规划决策和历代建设实践积累而成的成果。 Los Angeles is defined by its unique architectural landscape, a story told through a digital visualization of every building currently standing in the city. By plotting individual structures as boxes appearing in the year they were built, the data offers a glimpse into the growth of the metropolis over time. It is important to note that this view reflects only the surviving city, as demolished structures are not included, meaning the visualization captures the present-day physical footprint rather than a complete historical record.

The timeline begins in the 1880s, documenting a slow but steady accumulation of infrastructure that traces the city's expansion through the twentieth century and into the present. Users can navigate this history by decade, observing how urban density has shifted and transformed across different neighborhoods. From the dense clusters of Downtown to the sprawling residential sectors of The Valley, the map illustrates the rhythmic pacing of development as Los Angeles evolved from a smaller settlement into a massive urban center.

Visual tools allow for granular exploration, such as adjusting height exaggeration to better distinguish between structures in sprawling areas or focusing on specific districts like the Wilshire corridor or Century City. By interacting with the data, it becomes clear how various eras of construction contributed to the city's current character. The heights, footprints, and building types associated with these structures reveal the physical manifestation of shifting zoning laws, economic booms, and architectural trends that have shaped the Los Angeles skyline.

This repository of information, sourced from the LARIAC 2020 building outlines and the LA County Assessor roll, acts as a dynamic archive. It serves as both a tool for architectural study and a way for residents to contextualize their surroundings. By seeing every building in the context of its birth year, one can appreciate the layers of history that define the city, recognizing that the modern city is a culmination of past planning decisions and construction efforts still standing today.

在 2003 年 1 月发给 Microsoft 高管的一封电子邮件中，Bill Gates 对 Windows 及公司网络基础设施日益恶化的可用性表达了强烈不满。他详述了一次下载 Windows Movie Maker 的失败尝试，把整个过程形容为长达一小时的折磨：网站响应缓慢、导航混乱、安装步骤难以理解。他批评系统要求不必要的重启，并在程序列表中出现一些莫名其妙的测试文件，称整个体验一团糟，反映出公司在关注用户需求方面存在系统性疏忽。

领导层迅速回应并承认这些投诉是合理的。 Will Poole 指出 Gates 的挫败感有其道理，强调必须明确责任人来修复公司网站、 Windows Update 以及操作系统本身。包括 Amir Majidimehr 和 Dave Fester 在内的其他高管开始协调，商定谁将负责解决这些反复出现的可用性问题，并认为这些问题应成为未来产品发布时的正式验收项目。

内部讨论揭示了更深层的结构性问题，比如通常掌控网站发布的营销团队与工程团队之间存在摩擦。 John Martin 指出，下载流程的混乱在很大程度上源于营销方面的疏忽，他主张公司应优先提供无缝统一的客户体验，将下载和安装视为一个直观连贯的过程，而不是一系列支离破碎的技术障碍。

来自 Ian Mercer 等团队成员的反馈还指出了 Windows Update 的具体不足，例如它无法有效向最终用户推广新功能，也无法提供简洁、非侵入性的软件更新路径。讨论凸显出公司内部各团队各自为政，导致用户体验不一致。最终，这次交流反映了 Microsoft 内部的一次自我反思，领导者们开始探讨如何重新组织工作，以确保连下载软件这样简单的任务也不会成为用户体验的败笔。 In a January 2003 email to senior Microsoft executives, Bill Gates expressed intense frustration regarding the deteriorating usability of Windows and the company's web infrastructure. Gates detailed a failed attempt to download Windows Movie Maker, describing the process as an hour-long ordeal characterized by slow website performance, confusing navigation, and an incomprehensible installation sequence. He criticized the system for requiring unnecessary reboots and cluttering his program list with obscure test files, labeling the entire experience an absolute mess that reflected a systemic lack of attention to user needs.

The response from leadership was swift and acknowledged the validity of his complaints. Will Poole noted that Gates's frustration was reasonable and emphasized the need to identify owners for fixing the company's websites, Windows Update, and the operating system itself. Other executives, including Amir Majidimehr and Dave Fester, began coordinating to determine who would be responsible for resolving these recurring usability issues, which they agreed needed to become a formal sign-off item for future product releases.

Internal discussions among the staff highlighted deeper structural challenges, such as the friction between marketing teams, who often controlled web releases, and engineering teams. John Martin pointed out that the chaotic state of downloads was largely due to marketing oversight, arguing that the company should prioritize a seamless, unified experience for customers. He suggested that downloading and setup should be treated as a single, intuitive process rather than a fragmented series of technical hurdles.

Further feedback from team members like Ian Mercer identified specific shortcomings in Windows Update, such as its inability to effectively promote new features to end-users or provide a streamlined, non-intrusive path for software updates. The conversation underscored a broader consensus that the company's internal teams were siloed, leading to inconsistent user experiences. Ultimately, the exchange reflected a moment of self-reflection within Microsoft, as leaders debated how to reorganize their efforts to ensure that even simple tasks like downloading software did not result in a failure of user experience.

BZip3 是一款高性能压缩工具，作为 BZip2 的精神继任者。它使用 order-0 的 context-mixing 熵编码器和基于 suffix arrays 的快速 Burrows-Wheeler 变换，因而在压缩比和处理速度上都明显优于前代。它还结合了 RLE 、 Lempel–Ziv 与预测（Prediction）等处理步骤，采用类似 LZ77 的字符串匹配和类似 PPM 的上下文建模。与 BZip2 一样，BZip3 对文本和源代码的压缩进行了特别优化。

在以 Perl5 源代码为语料的对比基准测试中，BZip3 在文件体积缩减和速度之间表现出良好的平衡。配置适当的块大小和线程数后，它持续生成比 BZip2 和 Zstandard 更小的文件。此外，当与诸如 lrzip 之类的长距离去重工具配合使用时，BZip3 的压缩效果非常具有竞争力，优于单独使用 LZMA 或 BZip2 的情况。

尽管 BZip3 为了可靠性已在包括各类 ARM 、 MIPS 和 x86 配置在内的广泛架构上进行了大量测试，但开发者仍附带重要免责声明：由于底层算法复杂且存在罕见的边缘情况错误，除非能够接受理论上存在但概率极低的数据丢失风险，否则不应将其用于关键任务的数据。

该软件对编译器选择高度敏感：在 Linux 上使用 clang13 构建的版本在每线程的压缩和解压吞吐量上表现尤为出色。安装简便，支持 autotools 、 CMake 等常见构建流程，并可通过 Homebrew 等包管理器在多种系统上获取。项目采用 LGPLv3 许可证，并对第三方库和作者在 Burrows-Wheeler 变换与熵编码等组件上的贡献予以致谢。 BZip3 is a high-performance compression tool designed as a spiritual successor to BZip2. By utilizing an order-0 context mixing entropy coder and a fast Burrows-Wheeler transform that leverages suffix arrays, it achieves significantly better compression ratios and faster processing speeds than its predecessor. It also incorporates RLE with a Lempel Ziv and Prediction pass, relying on LZ77-style string matching and PPM-style context modeling. Like BZip2, BZip3 is particularly optimized for compressing text and source code.

In comparative benchmarks using a corpus of Perl5 source code, BZip3 demonstrates a strong balance between file size reduction and speed. When configured with specific block sizes and thread counts, it consistently produces smaller files than BZip2 and Zstandard. Furthermore, when combined with long-range deduplication tools like lrzip, BZip3 achieves highly competitive compression results that outperform standalone LZMA or BZip2 implementations.

While BZip3 is designed for reliability and has been extensively tested across a wide array of architectures, including various ARM, MIPS, and x86 configurations, the developer includes a notable disclaimer. Due to the complexity of the underlying algorithms and the potential for rare edge-case bugs, users are cautioned that they should not use the program for mission-critical data unless they are prepared for the slight, albeit theoretically possible, risk of data loss.

Performance of the software is notably sensitive to the choice of compiler, with Linux builds using clang13 showing impressive throughput in both compression and decompression per thread. Installation is straightforward, supporting standard build processes like autotools and CMake, and the project is available for various systems via package managers like Homebrew. The project is licensed under the LGPLv3, with contributions from various third-party libraries and authors acknowledged for components such as the Burrows-Wheeler transform and entropy coding logic.

从初级软件工程师成长为资深工程师，作者遇到了意想不到的挑战。职业早期以高速学习和巨大的精神消耗为特征，而现在的工作则截然不同：日常任务变得例行化，许多工作被 AI 所介入。尽管生产力提升了，但作者感到思维越来越迟缓、懒散、缺乏深度。这种停滞又被随手可得的低投入数字娱乐放大，使得感受无聊、培养深度专注几乎不再可能。

精神状态变差的一个明显信号是阅读习惯的变化。作者从大学时的贪婪读书者，变成如今难以看完一本书的人，究其原因是无尽刷屏和各种数字干扰。尽管在过去一年尝试养成手写和写博客等习惯，但作者仍觉得现在的认知状态比起十年前那种更敏锐的头脑有明显差距。

为了解决这个问题，作者利用最近一次乡间度假刻意断联，实践慢生活。把自己置于自然、家庭活动和桌游之中，营造出一个远离数字过载的环境。在这样的氛围里，作者重新投入阅读，先是迅速读完了许多历史和物理书籍，随后又转向文学经典和个人成长类书籍。

其中最重要的变化之一是对数学和物理兴趣的自发重燃。读完一本物理史之后，作者主动深入复习微积分和三角学。他们摒弃了在电子设备上读教科书那种糟糕体验，转而寻找更适合网页阅读的资源。重新投入这些复杂学科的学习，带来了像当年学编程时那种纯粹为兴趣而学习的愉悦感。

这次"让大脑复苏"的尝试能否带来长期的认知改善还有待观察，但作者已经感觉到日常思维不那么迟缓了。对于这些学术兴趣未来如何融入职业发展、尤其在动荡的科技行业中会有何影响，作者还不确定；目前的首要目标只是更深入地理解那些更高深的物理概念。 The transition from being a junior software engineer to a seasoned professional has brought unexpected challenges for the author. While early career years were defined by rapid learning and intense mental exhaustion, the current landscape of work feels different. Daily professional tasks have become routine and are often mediated by AI, which, while increasing productivity, has left the author feeling as though their thoughts are becoming slower, lazier, and less profound. This stagnation is compounded by the constant availability of low-effort digital entertainment, making it nearly impossible to experience boredom or cultivate deeper focus.

A primary indicator of this mental decline is the author's changing relationship with reading. Having moved from a voracious reader during university years to someone who struggles to finish books today, the author identifies doomscrolling and digital distractions as the culprits. Despite attempting to integrate habits like handwriting and blogging over the past year, the author felt a persistent gap between their current cognitive abilities and the sharper intellect they possessed nearly a decade ago.

Seeking a remedy, the author used a recent vacation to the countryside as an opportunity to disconnect and engage in intentional, slow living. By surrounding themselves with nature, family activities, and board games, they created a setting that discouraged digital overstimulation. This environment allowed for a return to reading, leading the author to fly through historical and physics texts, eventually moving on to literary classics and personal development books.

Perhaps the most significant development during this period was the spontaneous rekindling of a passion for mathematics and physics. After finishing a book on the history of physics, the author felt compelled to dive deeper into calculus and trigonometry. They moved past the subpar experience of reading textbooks on digital devices by finding specialized web-native resources. Engaging with these complex subjects, much like learning to code in the early days, provided a refreshing sense of learning for the sheer joy of it.

Whether this experiment in de-brainrotting will provide a long-term shift in mental clarity remains to be seen. However, the author notes that they are already feeling less intellectually sluggish in their daily life. While they remain uncertain about how these academic pursuits might fit into their broader career trajectory, especially given the unpredictable nature of the current tech industry, the primary goal for now is simply to attain a deeper understanding of advanced physical concepts.

公共厕所小便池的设计一个多世纪以来几乎未曾改变，但尿液飞溅问题一直存在。尿液飞溅到地面和使用者身上不仅带来严重的卫生隐患和异味，还需要频繁且昂贵的清洁维护。全球数以百万计的小便池在用，这一看不见的低效每天造成大量的水和人力浪费，并增加使用者接触细菌的风险。

研究人员将流体力学和微分方程的原理应用于这一长期未改的问题。通过分析液流的飞溅动力学，研究发现冲击角度是决定飞溅产生的关键因素。具体来说，团队发现当入射流与受撞表面夹角不超过 30°时，飞溅几乎可以被完全抑制。基于这一发现，研究团队运用等角曲线问题的数学方法，设计出新的优化小便池几何形状，确保尿液撞击盆面时的入射角不超过该临界值。

两种由此产生的设计被命名为 Cornucopia 和 Nautilus，并通过高速成像与定量质量测量实验得到验证。与常见商业小便池相比，这些新型模型显著减少了飞溅，其中 Nautilus 表现尤为出色。除了流体动力学上的优势外，Nautilus 还通过更低、更友好的边缘高度，提升了对儿童与轮椅使用者等更广泛人群的无障碍性。

研究表明，精确而非激进的几何调整即可大幅改善公共卫生并提高可持续性。若广泛采用这些无飞溅设计，可减少与清洁相关的用水与清洁剂消耗，带来显著的环境和经济效益。研究人员还指出，用于抑制飞溅的物理模型具有更广泛的应用前景，他们甚至打趣地提出了一种反向设计，称为 Urine-no，通过使入射角达到 90°来刻意最大化飞溅，以此威慑公共场合随地小便。 Public restroom urinals have remained largely unchanged in design for over a century, despite the persistent issue of splashback. This phenomenon, which involves urine droplets splashing onto floors and users, creates significant hygiene problems, generates foul odors, and requires intensive, costly cleaning efforts. With millions of urinals in use globally, this invisible inefficiency results in massive daily waste of water and human labor, alongside increased exposure to bacteria for facility occupants.

Researchers have now applied principles of fluid physics and differential equations to address this stagnation. By analyzing the splash dynamics of liquid streams, the study established that impact angle is a critical factor in splash generation. Specifically, the team determined that when an impinging stream hits a surface at an angle of 30 degrees or less, splashback is almost entirely suppressed. This finding allowed the team to use the isogonal curve problem to mathematically design new, optimized urinal geometries that ensure urine impacts the basin at or below this critical threshold.

The resulting designs, dubbed the Cornucopia and the Nautilus, were validated through both high-speed imaging and quantitative mass-measurement experiments. Compared to common contemporary commercial urinals, these new models demonstrate a dramatic reduction in splashback, with the Nautilus design proving particularly effective. Beyond its fluid dynamics performance, the Nautilus is also designed to improve accessibility for a wider range of users, including children and those in wheelchairs, by featuring a lower, more accommodating rim height.

These findings suggest that precise, non-drastic geometric adjustments can lead to substantial improvements in public sanitation and sustainability. By eliminating the constant need for cleaning-related water and solvent usage, the widespread adoption of these splash-free designs could offer significant environmental and economic benefits. The researchers also noted that the underlying physics model for splash suppression has broader applications, and they even humorously proposed a inverse design called the urine-no, which utilizes a 90-degree impact angle to intentionally maximize splash as a deterrent against public urination.

在 European Union 针对智能手机和平板电脑的可维修性法规实施一年后，合规情况仍然令人担忧。 Right to Repair Europe 联盟报告称，目前市场上超过 80% 的设备未能向用户提供法规要求的维修信息。尽管这些规则自 2025 年 6 月生效，要求制造商公布维修说明和备件定价，但对 European Product Registry for Energy Labelling 的审查显示，大多数公司并未履行这些义务。

该组织对超过 2,300 款智能手机型号进行检查，发现只有 18% 的机型真正提供了能让消费者查到维修说明或备件的可用网站。其余条目中很大一部分字段为空，还有些则把用户引导到与维修无关的支持页面。个别情况下，厂商甚至把用户指向 Temu 或 AliExpress 等第三方市场去寻找零件，暴露出透明度不足和对法规精神的漠视。

尽管未能满足这些基本要求，许多厂商仍为自己打出高分的可维修性评分。允许企业自评的做法遭到倡导者批评：即便厂商在形式上合规，所提供的信息往往难以查找。对于主要品牌，想要查到单个零件的具体价格常常需要反复点击，且有时所列价格范围宽泛到几乎无法为消费者提供准确的维修成本参考。

Right to Repair Europe 现在质疑 EU 当前执法策略的有效性。他们指出，监管机构似乎很少对厂商提交到注册表的数据进行核查。批评者认为，在监管方要求公开支持这些自评分数的完整文件之前，这一体系仍然容易被操纵。

展望未来，预计还会有进一步的监管变动以强化消费者权利。从 2027 年 2 月起，新的移动设备将被要求配备用户可更换电池，目的是延长硬件使用寿命。尽管针对可穿戴设备等专业产品存在例外，但这一变化仍体现了 EU 为减少电子垃圾、提升消费电子长期可用性所做出的持续且充满挑战的努力。 One year after the implementation of the European Union's repairability regulations for smartphones and tablets, compliance remains alarmingly low. The Right to Repair Europe coalition reports that more than 80 percent of devices on the market currently fail to provide the mandatory repair information to owners. Although the rules, which took effect in June 2025, require manufacturers to publish repair instructions and spare parts pricing, a review of the European Product Registry for Energy Labelling indicates that most companies have not followed through.

The campaign group examined over 2,300 smartphone models and discovered that only 18 percent actually provide a functional website where consumers can find repair instructions or spare parts. A significant portion of the remaining entries contain blank fields, while others point users to irrelevant support pages. In some cases, companies have even directed customers to third-party marketplaces like Temu or AliExpress for parts, highlighting a lack of transparency and commitment to the spirit of the legislation.

Despite failing these basic requirements, many manufacturers continue to assign themselves high repairability scores. This practice of allowing companies to mark their own homework has drawn criticism from advocates, who point out that even when manufacturers are technically compliant, the information is often difficult to navigate. For major brands, finding concrete pricing for individual components often requires excessive clicking, and sometimes the listed prices are presented in such broad ranges that they become virtually meaningless for a consumer looking for an accurate repair cost.

The Right to Repair Europe group is now questioning the overall effectiveness of the EU's current enforcement strategy. They note that there appears to be little oversight from public authorities to verify the accuracy of the data manufacturers submit to the registry. Critics argue that until regulators demand the publication of the full documentation supporting these self-declared scores, the system will remain prone to manipulation.

Looking ahead, further regulatory changes are expected to strengthen consumer rights. Starting in February 2027, new mobile devices will be required to feature user-replaceable batteries, a move intended to extend the lifecycle of hardware. While some exceptions exist for specialized products like wearables, this upcoming shift represents a continued, albeit challenging, effort by the EU to combat electronic waste and improve the long-term usability of consumer technology.

在 Telluride Film Festival 上突袭放映的 174 分钟纪录片 You Can See Everything 引发了强烈好奇。影片由 Cult 喜剧演员 Nathan Fielder 和纪录片导演 Lance Oppenheim 联合执导，聚焦 Theranos 前 CEO Elizabeth Holmes，在她入狱服 11 年刑期前不久的生活。该片并非传统调查片，而是一部带有真人秀元素的心理剧，记录了 Fielder 试图探查这位因大规模金融欺诈被定罪的女性内在逻辑的过程。

2023 年，Fielder 和 Oppenheim 在 Elizabeth Holmes 与伴侣 Billy Evans 同住的 Del Mar 住所拍摄。整个采访过程中，Holmes 始终流露出一种令人毛骨悚然的不眨眼的真诚，坚称自己无辜，并继续为其已倒闭的医疗技术辩护。她把 Theranos 的失败解释为一个关于化学的小技术难题，只要给出更多时间本可以解决，而不是刑事欺诈。她对自身叙述的坚定坚持暗示出一种根深蒂固、近乎反社会人格的否认，使 Fielder 的严谨怀疑在很大程度上无效。

影片偏向元叙事风格，时而在现实与表演之间游走。 Fielder 刻意制造了一些尴尬且暴露本质的时刻，例如给 Holmes 看 Abbott and Costello 的经典段子，但她因完全缺乏幽默感且僵硬地按自己的标准定义现实，未能领会其中的意味。随着纪录片推进，片中还加入了曾在迷你剧 The Dropout 中饰演 Holmes 的演员 Amanda Seyfried 。 Seyfried 被安排朗读 Holmes 的访谈与电话记录，这一实验在真实事件与戏剧化再现之间架起了一座桥梁。

尽管影片试图探问 Holmes 行为背后的"为什么"，但明确不为她辩护。伴侣 Billy Evans 的出现为这幅家庭图景增添了毒性，凸显出一种控制欲强且刻薄的性格，使二人关系更加复杂。最终，这部纪录片对生活在自我构建幻想中的个体做出了一次深刻且令人不安的审视，留给观众去消化她那种令人不安的欺瞒本质及其所反映的特权问题的更广泛影响。 The surprise screening of the 174-minute documentary You Can See Everything at the Telluride Film Festival has generated intense curiosity. Directed by cult comedian Nathan Fielder and documentarian Lance Oppenheim, the film centers on Elizabeth Holmes, the disgraced former CEO of Theranos, shortly before she began her 11-year prison sentence. Rather than functioning as a standard investigative piece, the documentary operates as a reality-TV-infused psychodrama, capturing Fielder's attempts to probe the internal logic of a woman convicted of massive financial fraud.

Fielder and Oppenheim filmed at the Del Mar home Holmes shared with her partner, Billy Evans, in 2023. Throughout the interviews, Holmes maintains an eerie, blinkless sincerity, insisting she did nothing wrong and continuing to defend the viability of her defunct medical technology. She frames the failure of Theranos not as a criminal deception, but as a minor technical hurdle regarding chemistry that could have been solved with more time. This unwavering commitment to her own narrative suggests a deeply entrenched, almost sociopathic level of denial that leaves Fielder's rigorous skepticism largely ineffective.

The film leans into a meta-narrative style, occasionally drifting between reality and performance. Fielder forces awkward, revealing moments, such as showing Holmes a classic Abbott and Costello routine, which she fails to grasp due to a total lack of humor and a rigid need to define reality on her own terms. As the documentary progresses, it incorporates actress Amanda Seyfried, who previously portrayed Holmes in the miniseries The Dropout. Seyfried is tasked with reading transcripts of Holmes's interviews and phone conversations, an experiment that bridges the gap between the actual events and their dramatized portrayals.

While the project explores the "why" behind Holmes's actions, it notably refuses to offer any apologia for her behavior. The presence of her partner, Billy Evans, adds a layer of toxicity to the domestic portrait, highlighting a controlling and abrasive personality that complicates the dynamic. Ultimately, the documentary serves as a profound, unsettling examination of an individual living within a self-constructed fantasy, leaving the viewer to grapple with the disturbing nature of her duplicity and the broader implications of her entitlement.

California Institute of Technology 将于 2026 年 10 月 30 日至 11 月 1 日举办 Caltech Mathathon，这将成为首个完全以研究级数学为主题的 hackathon，具有重要的里程碑意义。此次活动紧随该领域一系列重大突破之后：AI 已成功攻克诸如距今约 80 年的 Erdos planar unit-distance conjecture 、构造 non-sofic groups，以及在 six-sphere 复杂结构方面取得潜在进展等长期难题。

这些快速进展凸显了关于学科未来的关键问题，尤其是人工智能将如何缩短从初步构想到同行评审发表之间的时间。活动还将探讨：在 AI 日益展现出解决复杂猜想能力的时代，人类数学家的角色将如何演变。组织者希望通过汇集全球数学人才，实时应对这一系列系统性变化。

在 40 小时的挑战赛中，100 支队伍将获得超过 200 万美元的 AI 额度，并可使用多款前沿模型。参赛者将专注于攻克未解的猜想并发展新的数学理论。赛后，各队需向由顶尖数学家组成的评审团展示成果，评审团将评估研究内容的实质价值及参赛者对成果的理解深度。

比赛将设一轮现场奖项，以表彰活动期间展示出的最有前景的研究成果；在更广泛的数学界有足够时间验证这些工作的有效性后，将再颁发第二轮奖项。该活动得到众多科技公司和风险投资公司的支持，作为一次实地试验，Mathathon 旨在检验人类直觉与机器智能在推动纯数学发现方面的协作潜力。 The California Institute of Technology is set to host the Caltech Mathathon from October 30 to November 1, 2026, marking a significant milestone as the first hackathon dedicated entirely to research-level mathematics. This event arrives on the heels of major breakthroughs in the field, where AI has successfully tackled long-standing challenges like the 80-year-old Erdos planar unit-distance conjecture, the construction of non-sofic groups, and potential advancements regarding the complex structure of the six-sphere.

These rapid developments highlight critical questions about the future of the discipline, specifically how artificial intelligence can accelerate the timeline from initial ideation to peer-reviewed publication. The event also seeks to explore the evolving role of the human mathematician in an era where AI is demonstrating an increasing capacity to solve complex conjectures. By bringing together global mathematical talent, the organizers aim to address these systemic shifts in real-time.

During the 40-hour challenge, one hundred teams will be equipped with over 2 million dollars in AI credits and access to frontier models. Participants will focus on solving open conjectures and developing new mathematical theories. Following their work, the teams will present their results to a panel of leading mathematicians, who will evaluate both the substance of their findings and the depth of their comprehension.

The competition will feature an initial round of prizes for the most promising results presented during the event. A subsequent round of awards will be issued once the broader mathematics community has had sufficient time to verify the validity of the work. Supported by a wide array of technology companies and venture firms, the Mathathon serves as a practical experiment to test the collaborative potential of human intuition and machine intelligence in the pursuit of pure mathematical discovery.

推测式解码是对大型语言模型的一种强力优化，用以克服标准自回归解码的效率瓶颈。基线方法在循环中每次只生成并确认一个 token，而推测式解码将流程拆成两个阶段：先由一个轻量级草稿模块提出一串候选 token，随后目标模型在一次前向中验证这些候选项。如果目标模型接受草稿的 token，系统即可一次性提交多个输出，从而显著提升吞吐量。

文章介绍了 vLLM 当前支持的五种具体草稿方法：native Multi-Token Prediction (MTP) 、 Gemma 4 MTP 、 EAGLE-3 、 DFlash 和 DSpark 。这些方法主要在于如何利用目标模型的信息以及如何生成候选 token 存在差异。例如，native MTP 使用模型本身的辅助路径，而 DFlash 和 DSpark 则采用专门的、以目标为条件的网络并行预测整块未来 token 。相比之下，EAGLE-3 依赖一种自回归机制，结合了目标 Transformer 不同阶段的隐层状态。

在 AMD Instinct™ MI300X 和 MI355X GPU 上的实验表明，推测式解码的效果高度依赖具体的模型家族、草稿检查点和目标工作负载。某些配置的吞吐量提升超过 2 倍，但并非所有数据集和模型都能获得一致的性能提高。推测 token 的数量（即 proposal length）是关键调参项：随着草稿 token 数量增加，吞吐量通常会上升，但到达某一点后会趋于平缓甚至下降，因为额外的草稿开销开始抵消验证成功带来的收益。

在实际落地时，推测式解码需要细致的观测和调优。文章建议监控吞吐量、平均被接受长度和各位置的接受率等指标，以确定最优配置。由于不同负载的 token 可预测性差异显著，通常没有一刀切的设置。推荐的流程是从已知配置出发，针对不同的 proposal length 进行扫描，找到最适合当前任务的平衡点。

最后，为新的目标模型训练定制的 speculator 是进一步提升性能的可行途径。该过程包括收集目标模型的代表性隐层状态来训练草稿组件，确保投机器与目标的内部表征对齐。通过让草稿模型的架构和训练数据与目标领域（例如数学或代码生成）相匹配，用户可以获得更高的接受率，从而在生产环境中实现更高的服务吞吐量。 Speculative decoding serves as a powerful optimization for large language models, addressing the efficiency limitations of standard autoregressive decoding. While the baseline approach generates and commits one token at a time in a repetitive loop, speculative decoding separates the process into two distinct phases. A lightweight draft component first proposes a sequence of candidate tokens, and the original target model then verifies these candidates in a single pass. If the target model accepts the draft tokens, the system commits multiple outputs simultaneously, significantly increasing throughput.

The article details five specific drafting methods currently supported in vLLM: native Multi-Token Prediction (MTP), Gemma 4 MTP, EAGLE-3, DFlash, and DSpark. These approaches differ primarily in how they leverage information from the target model and how they generate candidate tokens. For instance, native MTP uses a model-native auxiliary path, while DFlash and DSpark utilize dedicated, target-conditioned networks to predict entire blocks of future tokens in parallel. EAGLE-3, by contrast, relies on an autoregressive mechanism that incorporates hidden states from various stages of the target Transformer.

Experimental results on AMD Instinct™ MI300X and MI355X GPUs demonstrate that the impact of speculative decoding is highly dependent on the specific model family, draft checkpoint, and target workload. While some configurations showed throughput gains exceeding 2x, performance improvements were not uniform across all datasets and models. The number of speculative tokens, or the proposal length, proved to be a critical tuning variable. Throughput generally improved as more tokens were drafted, up to a point, after which it hit a plateau or declined because the overhead of additional drafting work began to outweigh the benefits of successful verification.

Practical implementation of speculative decoding requires careful observability and tuning. The article suggests that users monitor signals such as throughput, mean accepted length, and per-position acceptance rates to identify the optimal configuration. Because different workloads exhibit different token predictability, a one-size-fits-all setting is rarely sufficient. A recommended workflow involves starting from a known configuration and performing a sweep of different proposal lengths to find the balance that best suits the specific task at hand.

Finally, training a custom speculator for a new target model is a viable path for further performance gains. This process involves collecting representative hidden states from the target model to train the draft component, ensuring that the speculator is well-aligned with the target's internal representations. By matching the draft model's architecture and training data to the target's expected domain, such as mathematics or code generation, users can achieve better acceptance rates and, consequently, higher serving throughput in their production environments.