• Nowe wąskie gardło AI: pamięć zamiast GPU [00:00:06]

  • Wraz z przechodzeniem sztucznej inteligencji z fazy trenowania do masowej inferencji (obsługa zapytań, długie okna kontekstowe, agenci), głównym problemem staje się przepustowość i pojemność pamięci [00:03:13].
  • Niewystarczająca pamięć sprawia, że drogie akceleratory marnują moc obliczeniową, czekając na transfer danych [00:01:33].

• Ograniczenia pamięci HBM (High Bandwidth Memory) [00:01:43]

  • HBM jest niezwykle szybka, lecz bardzo droga, fizycznie ograniczona przestrzennie obok GPU oraz trudna w zaawansowanym pakowaniu warstwowym [00:01:51].
  • Deficyt HBM pozwala producentom (Samsung, SK Hynix, Micron) dyktować wyższe ceny, co odbija się również na podwyżkach i ograniczonej dostępności standardowego DRAM-u [00:02:11].

• HBF (High Bandwidth Flash) jako nowa warstwa pamięci [00:00:16]

  • Nowy standard rozwijany wspólnie przez SK Hynix i SanDisk (w ramach Open Compute Project), angażujący także firmy takie jak Google, Meta czy Tenstorrent [00:04:14].
  • HBF łączy ideę pionowo ułożonych warstw komórek NAND (flash) z architekturą umożliwiającą masowy odczyt równoległy [00:04:32].
  • Specyfikacja przewiduje pakiety o pojemności do 512 GB (w stosach 8 lub 16 warstw), przepustowość od 0,4 do 3 TB/s oraz interfejs UCIe [00:05:16].
  • Szacuje się, że HBF zaoferuje 8–16 razy większą pojemność niż HBM przy porównywalnym koszcie pakietu [00:06:04].

• Wielopoziomowa architektura pamięci [00:06:57]

  • HBF nie ma na celu zastąpienia HBM, lecz wprowadzenie pośredniej warstwy pamięci: HBM dla danych krytycznych pod kątem latencji, HBF dla dużych wag modeli i pamięci podręcznej (KV cache), a niżej DRAM i tradycyjne SSD [00:07:06].
  • Inne koncepcje rynkowe (np. Samsung) to m.in. pionowe układanie pamięci bezpośrednio nad GPU (zHBM) czy obliczenia w pamięci (LPDDR5X-PIM) [00:08:01].

• Horyzont czasowy i wdrożenie [00:10:04]

  • Projekt pierwszego układu HBF zaliczył tape-out w sierpniu 2026 r., a pierwsze próbki urządzeń spodziewane są w 2027 r. [00:10:13].
  • Jednocześnie SK Hynix inwestuje ponad 4 mld USD w nowy zakład HBM w Indianie (planowane uruchomienie w 2029 r.), co potwierdza wieloletnią koegzystencję obu technologii [00:10:49].

【数据】Cohere的复合验证系统覆盖CPU和GPU，通过三个关键问题确保安全性：硬件真实性、安全配置和预期代码运行。这种多层次的验证机制提供了比单一组件验证更强的安全保障，是机密计算环境验证的最佳实践。

Why Restarting Your Router Fixes the Internet

• Routers are specialized computers [00:01:23]

  • They run their own operating system, CPU, RAM, and multiple background services (firewall, DHCP, NAT, DNS, management UI).
  • Over time, memory leaks, software bugs, or hung processes can degrade performance while hardware indicators still look normal [00:01:51].
  • A reboot clears volatile RAM and restarts all software daemons in a clean state [00:02:09].

• Rebuilding WAN connection state [00:03:08]

  • Local connectivity (LAN/Wi-Fi) may remain intact even when the upstream link to the ISP fails.
  • Restarting forces the device to renew its DHCP lease or re-establish a PPPoE session from scratch [00:03:17].
  • While IP addresses may occasionally change, the actual benefit comes from renegotiating a valid session state with the ISP [00:04:18].

• Flushing NAT and stateful connection tables [00:05:18]

  • Modern home networks run hundreds of simultaneous connections across streaming devices, laptops, and phones [00:04:42].
  • The router's NAT and stateful firewall track every open socket; if tables saturate or become corrupted, packets stall.
  • A restart purges exhausted connection tables and lets client devices open fresh connections [00:05:59].

• Resetting local DNS caching [00:06:52]

  • Domestic routers often function as intermediate DNS proxies and local caches.
  • Corrupted cache entries or unresponsive DNS forwarding can make web browsing fail even when IP routing works [00:07:08].
  • Rebooting rebuilds the DNS resolver state.

• Rebooting all-in-one hardware components simultaneously [00:08:36]

  • Consumer ISP gateways typically combine a modem/ONT, Wi-Fi access point, network switch, router, and firewall in one enclosure.
  • Power cycling reboots every layer of the edge infrastructure in one step [00:09:22].
  • If separate ONT/modem units are used, power cycling only the router may not fix upstream line issues [00:09:48].

• Diagnosing recurring connection drops [00:10:23]

  • Periodic reboots (e.g., once every few months) are normal, but daily or weekly reboots indicate underlying issues.
  • Outdated firmware can contain unresolved software bugs and security flaws [00:10:32].
  • Legacy hardware may be bottlenecked by modern smart home devices and high-bandwidth traffic [00:10:59].
  • Inadequate ventilation and thermal throttling can destabilize router hardware [00:11:30].

Esta parte ahora tiene más sentido para mí porque durante las clases hemos usado herramientas que antes no conocía, como Fossil, HedgeDoc y VSCodium. También he entendido que escoger una herramienta no es solamente mirar si funciona, sino pensar qué permite hacer, cómo maneja la información y qué tanto control tenemos sobre ella.

Este término me gusta mucho porque siento que se puede estructurar el conocimiento abstracto y darle forma en un soporte digital o físico.

OpenAI 官方（在同一天的驳回动议中）对"产品差异性"的正面表态，与前一句 Bloomberg 的独立判断相互印证。EP.97 专题05 依赖这类一手/准一手信息说明：这场诉讼的攻防焦点正在从"谁挖了谁的人"转向"谁的产品形态才代表 Agent 时代的硬件未来"。

Bloomberg 的这句判断被 MacRumors 直接引用来支撑 OpenAI 的核心抗辩——如果产品形态本身与 Apple 现有或在研产品线明显不同，那么"窃取商业机密来做同款产品"的指控在产品逻辑上就站不住脚。这句话把 EP.97 专题05 的法律争议和硬件形态两条线索连接了起来。

这是判断 OpenAI 硬件路线的关键产品定义：无屏幕、纯语音交互的"曲奇饼干"形态。EP.97 专题05 用这条信息论证 OpenAI 押注的是"calm computing"（无屏优先）路线，与 Apple 一直以来的软硬件集成、屏幕中心化路线正面对撞——这也是这场诉讼背后"下一代个人计算终端定义权"之争的产品层证据。

• Market Disruption and Competition: GoPro is struggling heavily due to being outcompeted by Chinese rivals, specifically DJI and Insta360, which offer superior hardware features and video quality at a significantly lower price point.
• Loss of Monopoly Advantage: GoPro originally dominated the action camera space during a period of near-zero competition, relying heavily on brand recognition while failing to rapidly innovate or adjust its high pricing model.
• National Security Implications: Commenters highlight that GoPros are heavily utilized in the Western aerospace and defense sectors. Shifting to Chinese alternatives like Insta360 introduces potential geopolitical risks and hardware attack vectors.
• Industrial Policy Differences: A debate emerged around China's highly successful industrial strategy. While some attribute the rise of DJI and Insta360 to aggressive local ecosystem competition and long-term state support, others point to intense labor exploitation (e.g., 996 work culture) and strategic overproduction of STEM graduates.
• Corporate Bureaucracy: Internal organizational issues within GoPro, such as slow-moving bureaucracy, high risk aversion, and an inability to match the agile, enterprising product cycles of its modern competitors, accelerated its decline.

• Identical Architecture: An iFixit teardown reveals that the $499 Trump Mobile T1 phone is structurally almost identical to the 2024 HTC U24 Pro, a mid-range Taiwanese-branded smartphone manufactured in China.
• Interchangeable Parts: The component layout, chip placement, and screw patterns match so closely that iFixit successfully swapped the main motherboard from an HTC U24 Pro into the T1 chassis, and the phone booted and functioned perfectly.
• Shared Specifications: Both devices share a Qualcomm Snapdragon 7 Gen 3 processor, 12GB of RAM, and 512GB of internal storage, though the T1 uses a Micron memory chip instead of the HTC's SK Hynix chip.
• Functional Differences: The only significant hardware variance is the battery. The T1 utilizes a larger 5,000mAh battery (manufactured in the Philippines) compared to the HTC's 4,600mAh cell, but the T1's charging speed is capped at 30W, which is half of the HTC's 60W fast-charging capability.
• Cosmetic Tweaks: Aesthetic changes on the T1 include a gold-painted exterior, an American flag graphic on the back, a slightly repositioned camera flash, and a speaker grille featuring seven circular holes instead of six pill-shaped ones.
• Manufacturing and Assembly: Despite early marketing claims of being an American-made device, the phone is designed and manufactured in China. The "Assembled in the USA" label likely refers to final assembly in Florida from around 10 pre-fabricated imported modules.
• Repairability and Value: While the T1 offers decent component value for its $499 price point, it received a low provisional repairability score of 3 out of 10 from iFixit due to a total lack of public service manuals, official spare parts, and guaranteed long-term software support.

大多数人认为高性能AI芯片必然伴随着高能耗和散热挑战，但作者认为微软的Maia 200在提供强大计算能力的同时实现了惊人的能效优势，仅消耗Nvidia Blackwell B300 Ultra一半的功率。这一反直觉的发现挑战了AI领域'性能与能耗成正比'的传统认知，暗示了专用AI芯片架构设计的创新突破。

大多数人认为AI在硬件领域的自主发展和创新还很遥远，但作者认为AI在硬件领域可能遵循与软件工具相同的轨迹，因为软件工具的自主开发已经从看似荒谬变成了现实。这是一个挑战行业共识的观点，暗示了AI可能更快地实现对物理世界的直接控制。

大多数人认为高带宽内存(HBM)是AI加速器的最佳选择，但Qualcomm声称其新的高带宽计算(HBC)技术能在每瓦带宽上提供6倍的提升，这一性能优势挑战了当前数据中心AI加速器的行业共识，暗示传统HBM技术可能面临被颠覆的风险。

大多数人认为苹果会通过软件更新让所有兼容设备都能获得完整的AI功能，但作者指出苹果将Siri AI的完整功能限制在特定高端机型上，这挑战了苹果过去通过软件更新让旧设备获得新功能的传统做法。这种策略暗示了AI功能可能与硬件限制紧密相关，而非纯粹的软件升级。

AI芯片竞赛反映了行业对计算能力的迫切需求。初学者应关注这一趋势对成本和可用性的影响，并考虑采用模型优化、蒸馏等技术来减少对高端硬件的依赖。了解不同硬件架构的特性有助于做出更明智的技术选择。

从零开始设计ASIC展示了专用硬件的优势，但也强调了与AI研究人员紧密合作的重要性。初学者应理解，硬件设计必须与算法需求紧密结合。对于非专业团队，考虑使用GPU或TPU等现成解决方案可能更实际，除非有特定性能需求。

计算资源短缺是AI行业面临的核心挑战之一。初学者应认识到专用芯片不仅是性能优化，更是应对计算资源限制的战略选择。了解不同工作负载下的硬件权衡对于资源规划至关重要，特别是在预算有限的情况下。

从概念到芯片仅用9个月的时间展示了当前AI硬件开发的加速趋势。这对行业初学者意味着芯片设计迭代周期显著缩短，但也可能带来潜在的质量和兼容性问题。建议关注长期稳定性和可扩展性，而非仅追求速度。

大多数人认为苹果会通过新系统更新淘汰较旧设备以刺激硬件销售，但苹果却选择支持5年前的iPhone 11，并承诺显著性能提升。这与苹果通常推动用户升级硬件的策略相悖，表明苹果在软件支持策略上更加用户友好，而非纯粹的商业驱动。

Chuwi Minibook X Review Summary

• Overview: The Chuwi Minibook X is a highly compact, 10.5-inch budget sub-ultrabook priced around $350 that serves as a fun "knock-around" utility or hobbyist machine.
• Core Specifications:
  • Intel N150 Twin Lake processor (4-core/4-thread, 3.6GHz).
  • 16GB LPDDR5-6400 RAM (soldered) and a 512GB upgradable NVMe drive.
  • 10.51-inch IPS 2K 16:10 touchscreen display.
  • Weight is ultra-portable at 912 grams (under 1 kg).
  • Ports include two USB-C ports, one of which supports Power Delivery (PD) charging.
• Hardware Quirks & Power:
  • Ships with an odd 12V/2A USB-C charger, but runs perfectly fine on standard 20V PD chargers.
  • Battery life holds up reasonably well, lasting roughly 6 hours looping a film in VLC.
  • Thermals are solid; under stress testing, the hottest parts of the chassis stay around 32°C (90°F).
• Linux Compatibility:
  • Runs Linux "boringly well" with functioning suspend, sleep, touchscreen, Wi-Fi 6, and Bluetooth (requiring non-free Intel blobs).
  • The Big Flaw: The screen panel is sourced from a cheap tablet and is mounted sideways. Fixing the 270-degree clockwise orientation requires adjustments across the whole stack: bootloader configuration, a video panel orientation kernel parameter in initrd, and desktop environment overrides.
• Downsides: The 2K screen runs at a low 50Hz refresh rate, the diving-board trackpad lacks physical buttons, the audio is tinny, and the keyboard is highly finicky, only registering strokes when pressed directly in the center.
• Verdict: Because it is affordable and disposable enough not to matter if it breaks, it acts as a perfect modern sandbox for testing new software environments like NixOS, RiverWM, or Steam.

Hacker News Discussion

• The "Awful But Great" Dynamic: Multiple owners echoed that while the device has clear hardware limitations, its minute size makes it incredibly charming and practical for specialized situations like tiny train or airline tray tables.
• The Beater Device Rationale: Users appreciated having an inexpensive machine for travel where theft or invasive customs inspections are potential concerns, ensuring their primary, expensive data-heavy laptops remain safe at home.
• The Keyboard/Trackpad Debate: Some commenters noted that the keyboard issues described in the article would completely ruin the experience for serious work, while another owner claimed they couldn't reproduce the center-only key registering issue with their own fingers due to the small size of the keys.
• Used Alternatives vs. Form Factor: Several participants suggested that a used enterprise laptop (like a ThinkPad X270 or Dell XPS 13) offers better build quality, performance, and value for under $100. However, counterarguments highlighted that those alternatives are fundamentally heavier and bulkier than a sub-1kg, 10-inch form factor.
• Nostalgia for Small Tech: The thread sparked widespread yearning for extinct ultra-portable form factors, with many fondly recalling the 11-inch MacBook Air and the 12-inch MacBook, wishing modern manufacturers would build high-quality, ultra-compact hardware again.

大多数人认为Nvidia在ARM架构上的Windows设备尝试已经失败，历史不会重演，但作者暗示这次Nvidia的RTX Spark芯片是'一个完全不同的野兽'，更强大而非更弱小，挑战了人们对ARM架构Windows设备失败的固有认知。

Jeff Dean的评论表明AlphaEvolve已经从软件层面深入到硬件设计，能够提出违反直觉但高效的电路设计，直接集成到TPU芯片中。这展示了AI系统在硬件设计领域的突破性应用，可能改变芯片设计范式。

大多数人认为AI系统的硬件设计需要人类专家精心设计，但作者认为AI本身可以设计出比人类更高效的硬件电路。这挑战了传统硬件工程领域的共识，暗示AI可能在最底层的硬件设计上超越人类专家的直觉和经验。

Summary of "Googlebook" Discussion (Hacker News)

• The Concept: Googlebook is presented as a high-end hardware play—a "premium upgrade" to the Chromebook—intended to shift the user experience from traditional apps to an LLM-native environment powered by Gemini.
• The Vision: Some users see this as a "moonshot" to make apps irrelevant, replacing them with raw data feeds and AI-driven visualizations, potentially creating a "sci-fi" style user interface.
• Execution Skepticism: A dominant theme is doubt regarding Google's ability to execute. Commenters cited current frustrations with Gemini (cutting off mid-sentence, poor Maps integration, and hallucinations in Sheets) as evidence that the tech isn't ready for a primary OS role.
• Target Audience: Discussion suggests Google is targeting "Chromebook natives"—students who grew up in the Google ecosystem and are now entering university or the workforce and want a "pro" version of their familiar tools.
• Public Sentiment on AI: There is a sharp divide in perceived AI adoption. Some argue "normies" (non-tech users) love AI for daily tasks and creative fun, while others claim people increasingly resent AI being "forced" into every product.
• Competitive Landscape: Comparison was drawn to historical shifts (like the move to Mosaic/Web browsers). While some see this as an inevitable evolution, others believe Apple’s approach—gradually improving Siri and app intents—is more practical than trying to eliminate apps entirely.

https://www.facebook.com/groups/705152958470148/posts/1264728369179268/

Custom gothic typeface on a Royal 10 made for supply/hardware settings. Uppercase is primarily fractions and abbreviations including: Dz, Gr, Bx, Bl, lb, In, Ft, Pr, Cs, Pc, Yd, and GL.

Fractions include odd numbers for the halfs, thirds, fourths, sixths, eighths, twelfths, and sixteenths,

非GPU设备无法使用带音频输入的Prompt API，这是初学者在使用前需要注意的技术限制。

明确提到Trainium2芯片将在第二季度上线，而Trainium3芯片将在今年晚些时候上线，提供了具体的时间节点。这一数据点显示了芯片技术迭代的快速节奏，以及Anthropic与AWS在硬件路线图上的紧密合作。这种快速迭代能力对于保持AI模型的竞争力至关重要，但也带来了基础设施规划和成本控制的挑战。

使用超过100万颗Trainium2芯片的数据，展示了Anthropic在AI硬件部署上的巨大规模。这一数字不仅反映了计算能力的投入，也显示了与AWS在芯片定制上的深度合作。对于AI模型训练而言，百万级芯片的部署规模是行业顶尖水平，表明Claude可能需要大量计算资源进行训练和推理。

大多数人认为AI公司主要依赖通用GPU芯片训练模型，但Anthropic与Amazon的合作表明他们正大规模采用专用AI芯片(Trainium)，这挑战了行业对通用芯片依赖的主流认知。5GW的容量远超大多数AI公司的规模，反映了专用芯片在AI训练中的经济性和效率优势正在被重新评估。

使用超过100万个Trainium2芯片，这是一个惊人的硬件部署规模。这一数字不仅显示了Anthropic与Amazon的深度合作，也反映了训练和运行大型语言模型所需的庞大计算资源。相比其他AI公司，这种规模的芯片部署表明Anthropic正在全力投入AI基础设施。

100万片Trainium2芯片的使用量展示了AI模型训练的硬件规模。这一数量级表明Anthropic正在进行大规模并行计算，这是训练大型语言模型的基础设施要求。与英伟达GPU的采用相比，Trainium芯片代表了云服务提供商在AI硬件领域的差异化竞争策略。

大多数人认为混合不同代际的硬件进行训练会降低性能或效率，但作者认为即使不同代际、不同速度的芯片混合使用，仍能达到与单一芯片类型训练相同的机器学习性能，这挑战了硬件必须同质化的行业共识。

大多数人认为硬件故障会显著降低分布式训练的效率和性能，但作者认为即使在硬件故障率极高的环境下，Decoupled DiLoCo仍能保持88%的有效训练率，而传统方法则暴跌至27%，这挑战了人们对故障容忍能力的传统认知。

通常认为硬件定制化是提高性能的途径，但作者强调通过软硬件协同设计可以大幅提升效率和性能，这与单纯硬件升级的观点相悖。

大多数人认为中国开源AI是出于理想主义或技术自信，但作者认为这实际上是一种战略性的 workaround（变通方法）。由于无法获得美国限制出口的高端芯片，中国通过开放源代码来加速外部反馈循环，弥补计算能力的不足，这是一种务实而非理想主义的策略。

研究采用的H100等效计算方法虽然提供了标准化比较基准，但可能无法完全捕捉不同工作负载下的实际性能差异。这种简化方法在揭示集中趋势的同时，也可能掩盖了AI硬件生态系统的多样性和创新潜力，值得进一步探讨。

研究方法中使用的H100等效转换存在重要局限性，它简化了不同芯片间的性能差异，这可能低估了某些专用架构的实际价值。这种标准化方法虽然在比较中提供了便利，但可能掩盖了AI硬件生态系统的多样性和创新潜力。

令人惊讶的是：即使使用H100-equivalents作为标准测量单位，也无法完全反映不同芯片类型在真实世界中的性能差异，这表明我们对AI计算能力的测量可能存在系统性偏差，影响我们对AI发展速度的准确理解。

这一数据揭示了现代CPU推理的关键瓶颈：内存带宽限制。代理最初尝试的SIMD微优化无法突破这一根本限制，这表明理解硬件特性和系统瓶颈对于有效优化至关重要。这一发现挑战了传统上认为计算是主要瓶颈的观念，强调了内存效率在AI推理中的核心地位。

这个声明具有重要的实际意义，因为它证明了在国产硬件上进行大规模模型训练的可行性。这为在特定硬件环境下进行模型优化和部署提供了实际案例，可能会促进国产AI硬件生态系统的发展。同时，这也展示了该方法在实际应用中的可扩展性和实用性。

令人惊讶的是：即使是大型语言模型，现在只需要8-10GB的显存就能进行微调，这大大降低了AI模型训练的硬件门槛，使更多研究者和开发者能够参与模型定制。

令人惊讶的是：一张消费级显卡竟然能直接运行320亿参数的大模型，这打破了人们对大模型硬件需求的固有认知。过去需要多张高端显卡或专业服务器才能运行的模型，现在单张RTX 4090就能实现，这标志着大模型普及的门槛大幅降低。

令人惊讶的是：AI训练效率的提升速度令人震惊。在短短6年内，语言模型的训练时间从167分钟缩短到不到4分钟，效率提升了40多倍。这种进步远超摩尔定律预测的5倍改进，展示了AI硬件和算法的飞速发展。

令人惊讶的是：AI模型训练速度在6年内提升了约50倍，远超摩尔定律预测的5倍。这种性能提升不仅来自硬件改进，还来自软件优化和算法创新。这一事实打破了人们对技术进步速度的传统认知，展示了AI领域独特的加速发展模式。

Gemma 4 E2B/E4B 原生预训练 140+ 语言，且强调「超越翻译、理解文化语境」。对 AI 硬件产品而言这个参数意义重大：一个能在设备端离线处理中文、理解文化背景的 2-4B 模型，意味着本地化 AI 硬件（录音笔、学习机、会议设备）无需依赖国内厂商 API，直接用 Gemma 4 就能构建多语言理解能力。

主流观点认为需要高端GPU才能支持长上下文推理的大语言模型，但作者证明TriAttention仅使用消费级单GPU就能部署原本需要高端GPU才能运行的长上下文模型。这一发现挑战了当前对硬件需求的共识，可能使更广泛的开发者能够访问长上下文推理能力。

大多数人认为处理长上下文需要高端GPU或分布式系统，但作者声称他们的方法只需单个消费级GPU就能实现原本需要高端硬件才能处理的长上下文任务。这一观点挑战了人们对长上下文处理硬件需求的普遍认知。

大多数人认为手机硬件需要不断升级才能运行最新的AI功能，但作者认为技术压缩速度如此之快，以至于现有手机在升级前就能运行曾经的顶级模型，这颠覆了人们对硬件更新周期的认知。

大多数人认为支持140多种语言的多模态模型需要大量计算资源，无法在单个GPU上运行。但作者声称这些模型可以全部适配在单个H100 GPU上，这挑战了我们对大型多语言模型资源需求的认知，暗示模型效率可能大幅提升。

大多数人认为AI公司会依赖单一硬件供应商以获得最佳性能，但Anthropic采用多平台策略，挑战了行业共识。这种多元化方法虽然增加了复杂性，但提供了更好的性能和弹性，暗示了AI计算的未来可能更加分散而非集中。

https://web.archive.org/web/20260216112337/https://wccftech.com/western-digital-has-no-more-hdd-capacity-left-out/ Western Digital says consumer sales of HDDs is now just 5% of their sales, everything else going to AI companies. 2026 WD is already sold out. For 2027 and 2028 there is already a few deals in place too.

• comma.ai operates its own $5M data center in-office to handle model training, metrics, and data storage, avoiding the "cloud tax."
• The facility consumes approximately 450kW at peak; power costs in San Diego (over 40c/kWh) totaled over $540,000 in 2025.
• Cooling is achieved using pure outside air with dual 48” intake and exhaust fans, utilizing a PID loop to manage temperature and humidity.
• The compute cluster consists primarily of 600 GPUs across 75 "TinyBox Pro" machines built in-house for cost efficiency and easier repairability.
• Storage is handled by several racks of Dell R630/R730 servers with ~4PB of total SSD storage, favoring speed and random access over redundancy.
• The software stack is kept simple to ensure 99% uptime, utilizing Ubuntu (pxeboot), Salt for management, and "minikeyvalue" for distributed storage.
• By owning their hardware, comma.ai estimates they saved $20M+ compared to equivalent compute costs in a public cloud environment.

Hacker News Discussion

• Users discussed the spectrum of infrastructure, ranging from pure Cloud (low cap-ex, high op-ex) to colocation and on-prem (high cap-ex, high skill requirement).
• A primary concern raised was "brain drain"—on-prem setups can become "legacy debt" if the senior engineers who built the custom systems leave without documenting unwritten knowledge.
• Commenters noted that AWS and other cloud providers are incentivized to keep architectures complex (microservices, serverless) to increase billing, whereas on-prem encourages efficiency.
• There was a debate regarding "software freedom" and the "WhatsApp effect," where small, highly motivated teams can outperform massive corporations by using lean, self-hosted stacks.
• Some users highlighted that while AWS pricing is expected to rise due to hardware costs, the "Quality of Life" and managed services still justify the cost for many startups without comma's scale.

comma-ai #self-hosting #datacenter #hardware-engineering

Sensecap lorawan products at tinytronics Eindhoven, Netherlands

A solar power meschcore node. Ships from Germany at 90 USD but likely there's a proper eu provider to be found too.

so, in short the article is bunk? cloud computing is not cheap with subscriptions for every piece of it. Owning a computer amortised over its years of us willbe cheaper (25 euro/month is 1k laptop every 3 yrs).....although most people hardly use the capabilities of their device.

the AI fever is creating the shortages. The flp side is that when the hype deflates you can pick up compute for low prices

Article hinges on the notion that prices of dRAM, ssd will rise due to shortages.

no, esp not if it's either or (like for some FB is internet access). But fully locked down devices and settings will get a liability quickly too that people can't ignore.

https://web.archive.org/web/20260115090647/https://www.windowscentral.com/artificial-intelligence/jeff-bezos-says-the-quiet-part-out-loud-bezos-envisions-that-youll-give-up-your-pc-for-an-ai-cloud-version (via E)

article on emerging tendency to encourage people to give up PCs and other general purpose computing devices, in favour of cloud and 'dumb' edge hardware. It used to be we aimed to keep all the smart stuff at the edge.

[[The Hardware Hacker by Andrew Huang]], to acquire. Available in Internet Archive: https://archive.org/details/hardwarehackerad0000huan

3:58 "GrapheneOS ist das sicherste, bisher nie von einem Geheimdienst gehackte Betriebssystem."

nein. selbst mit 100% open-source software, die man selber kompiliert mit einem open-source compiler, selbst damit hast du:<br /> closed-source hardware, closed-source firmware, closed-source netzwerktreiber, hardware backdoors in CPUs und GPUs, kompromittierende abstrahlung.<br /> also mit GrapheneOS kannst du deine attack surface nur verkleinern, aber sicher nicht auf null reduzieren.<br /> es ist auch immer die frage, was ist dein threat model, also wie mächtig sind deine gegner, und wie wichtig bist du.<br /> für state-level actors wie NSA oder Mossad bist du praktisch immer angreifbar.<br /> das betrifft praktisch alle consumer-grade hardware, und echte sicherheit hast du nur mit military-grade hardware, aber die kannst du nicht kaufen, genauso wie militärwaffen.<br /> opsec für anfänger... dabei haben wir 1000 andere probleme die jetzt wichtiger sind, vor allem selbstversorgung.

TPD is "Thermal Design Power" — the maximum amount of heat that a computer component (like a CPU, GPU or system on a chip) can generate and that its cooling system is designed to dissipate during normal operation.

https://www.ted.com/talks/matt_ridley_when_ideas_have_sex

Here's the reference for TB, from our discussion

If I may just have a moment of weakness...

This is a hero of the internet that raised me. <3

for - definition - production system ('hardware') - and organizing system ('software') - Arbib and Seba

definition Arbib and Seba - human civilization can be broken down into the interaction between two complimentary systems - the production system - by which we meet fundamamental material needs for food, energy, transportation, water, materials - also called 'hardware' - the organizing system - by which how the production system is governed and managed and includes the economy, polity, security, culture, ideology and values - also called 'software'

comment - A transformation is required in both the hardware and the software to mitigate the worst impacts of our current polycrisis

for - definition- hardware - software - Paul Davies

definition - hardware - software - Paul Davies - In the context of life, - hardware - configuration of matter which harnesses energy from its environment - software - complex information sturctures by which configurations of matter and energy are organized and instructed to self-reproduce

Just fucking kill me.

Despite knowing this had to be the truth - unless God had finally answered at least one of my prayers about intervening directly with The Bluetooth Specification - I'd never actually plugged my QuietComfort 45s into any one of the devices upon which I've had Lossless audio turned on for years... until just now.

Will this follow the same route as the US with NEPA being sidelined?

suboptimal substitution for true 10bit. as for portable monitor this could be an option, but i dont see any advantage over the more cheaper options.

Only 400 nits. No HDR here.

https://www.knapeandvogt.com/

Manufacturer of index followers

https://www.amazon.com/Knape-Vogt-Card-Size-Follower-Anochrome/dp/B0006FKQJ4

https://www.wwhardware.com/catalog/product/view/_ignore_category/1/id/3588/s/kv-index-follower-and-track-kv047/

https://www.leevalley.com/en-us/shop/hardware/storage-and-organization/files/69118-index-followers

Listed as an "Index Follower". Looks to be the same model as https://www.wwhardware.com/catalog/product/view/_ignore_category/1/id/3588/s/kv-index-follower-and-track-kv047/

Oooo...

According to this not-entirely-vetted World Wide Web service, this unit ("Candela per square meter") is 1:1 equivalent to "Nits."

...no, it would seem the latter is just an incorrect name for the same unit. Cool.

Having spent the past two days exploring the potential of the used first-generation HoverBar Duo I somehow managed to acquire for $35, I would emphasize "enough" here as a qualifier at least once over.

Though the following is pure speculation, there were a few highly observable clues to suggest that the previous owner also had a heavy iPad Pro (perhaps a 13,9 as I have,) though - after over-tightening all the adjustment screws (all but stripping threads in the case of the one at the base,) they were ultimately not satisfied with the Duo's management of the weight, even tightened to that state.

(Again, entirely speculative but.)

See my reference Craft document for an index of said shortcuts. (I intend to publish it in more formats soon.)

ICs as hardware versions of AI. Interesting this is happening. Who are the players, what is on those chips? In a sense this is also full circle for neuronal networks, back in the late 80s / early 90s at uni neuronal networks were made in hardware, before software simulations took over as they scaled much better both in number of nodes and in number of layers between inputs and output. #openvraag Any open source hardware on the horizon for AI? #openvraag a step towards an 'AI in the wall' Vgl [[AI voor MakerHouseholds 20190715141142]] [[Everymans Allemans AI 20190807141523]]

.

= Highlights - Patents granted for unoriginal inventions if prior art outside of the patent literature missed. - Misses most of free and open source software and hardware - number in millions. - = Open Source Hardware Association - created a certification database - centralized prior art. - Novel tool has a semi-automated way of certification from = MediaWiki - websites. - = OSHWA - certification completed on average in 62.5% less than direct form filling.

= TITLE - Towards = open source patents - Semi-automated = open hardware - certification - from = MediaWiki websites

= Policy Position Paper = Executive Summary - Changes in science funders’ mandates - have resulted in open access to data, software, and publications. - Research capacity, however, is still unequally distributed worldwide, hindering the impact of these efforts. - To achieve the SDGs, open science policies must shift focus from products to processes and infrastructure, - including access to open source scientific equipment. - - Conventional, black box, proprietary approaches to science hardware - reinforce inequalities in science and slow down innovation everywhere, - while also threatening research capacity strengthening efforts. - Three policy recommendations to promote open science hardware for research capacity strengthening: - incorporating open hardware into existing open science mandates, - incentivizing demand through technology transfer and procurement mechanisms, - promoting the adoption of open hardware in national and regional service centers.

= TITLE - Equitable Research Capacity - Towards the SDGs: - The Case for = Open Science Hardware

AUTHORS: - Julieta Arancio - https://scholar.google.com/citations?hl=en&user=1bFSyMQAAAAJ - Mayra Morales Tirado - https://scholar.google.com/citations?hl=en&user=d0u_n6UAAAAJ - Joshua Pearce -https://scholar.google.com/citations?hl=en&user=QZ8lPxwAAAAJ&view_op=list_works&sortby=pubdate

I came here after reading a couple articles (one on Super User & one on MS's help forums) recommending X-Mouse Button Control as a general way to disable back buttons on mice.

Note: this doesn't seem to work on Windows 11 for Microsoft USB IntelliMouse Optical.

Drivers run in kernel mode to access the hardware

Ideas ahead of their time in the mainstream (not in the niches). Compare to Lernout and Hauspie wrt early natural language processing 1987-1998 and GPT-3 now.

https://boingboing.net/2018/03/12/the-vim-clutch-a-footpeda.html

By getting to economies of scale faster than other people can?

Smart strategy by Nvidia.

https://blog.jim-nielsen.com/2022/a-web-for-all/

This is a great idea imo.

It looks like they make their business model is based predominantly on hardware over software.

GM of the Hardware part is -33%

Hardware revenue is 7.8%

Ce type de problème de conception est inhérent a la propre finitude de l’être humain : on ne peut pas penser a tout. Cependant, si un système est bien conçu, ce type d’erreur peut souvent être résolu par une mise a jour logicielle (plus ou moins bas niveau). Il est donc possible que la correction de ce bug ne coûte aucune matière additionnelle : pas besoin de remplacer le matériel, mais besoin de beaucoup se creuser le cerveau pour mettre a jour le logiciel.

Flash Translation Layer (FTL)

Overwriting data on SSDs

SSDs writes aren't as fast as they seem to be

2nd difference between SSDs and HDDs

1st difference between SSDs and HDDs

Many thought Volkswagen would crush Tesla as soon as they put their weight behind an electric car initiative. What they didn't consider was that an electric car is more about software than it is about hardware.

Hardware assisted compiler is a fantastic idea. TPUs from Google are essentially this. They're hardware assistance for matrix multiplication operations for machine learning workloads created by tools like TensorFlow.

Herper, M. (2020, July 1). Covid-19 vaccine from Pfizer and BioNTech shows positive results. CNBC. https://www.cnbc.com/2020/07/01/coronavirus-vaccine-from-pfizer-and-biontech-shows-positive-results-report-says.html

Quiet fan

Devices without a TPM module will be always asked for a password (e.g. by BitLocker) on every boot

dualité du régime de vie du projet architectural: la partie hard, la structure, celle dessinée par les architectes, et la partie soft – dynamique, imprévisible, dont s’empare les citoyens.

Internauts build a supercomputer network stronger than 7 most efficient local supercomputers interconnected. The reason is to fight against COVID-19.

You can also join them by using Folding@home software

This journal clearly explains the use of technologies with special aid people how a certain group can leverage it, while also touch basing on what are the challenges which special aid people face financially.

La política debería incluir la garantía institucional de conservar o asegurar la actualización de dispositivos de almacenamiento.

metal L brackets