科技藏珍

 

  https://semianalysis.com/2025/01/31/deepseek-debates/ SemiAnalysis 深入分析了 DeepSeek 在人工智慧產業中的迅速崛起，重點關注其與 High-Flyer 的策略性關係、 GPU 投資，以及訓練模型的真實成本。 DeepSeek 與 High-Flyer ·         High-Flyer （中國對沖基金）早期認識 AI 的變革潛力，不僅限於金融領域。 2021 年投資 10,000 張 A100 GPU ，預測 AI 需求增長。 2023 年 5 月成立 DeepSeek ，深入 AI 研究。 High-Flyer 與 DeepSeek 共享人力與計算資源， GPU 投資總額超過 5 億美元，即使面臨出口管制。 GPU 資源與投資 ·         DeepSeek 擁有約 50,000 張 Hopper GPU （ H800 、 H20 ），符合監管標準。 這些 GPU 用於交易、推理、訓練及研究。 總伺服器資本支出約 16 億美元，運營成本約 9.44 億美元。 訓練成本與效率 ·         DeepSeek 公開聲稱模型訓練成本約 550 萬美元，但 SemiAnalysis 指出該數字誤導。 綜合評估顯示，實際訓練成本可能高達 16 億美元，包括基礎設施、硬體及運營費用。 市場影響 ·         DeepSeek 的崛起引發 GPU 需求討論。 儘管部分觀點認為其效率將降低對 GPU 的需求，但 SemiAnalysis 指出，這反而提升了市場需求。 H100 和 H200 GPU 價格因此受到影響。 總結 ·         D...

Introduction:  Nvidia神話不再？！AI行業大變天⚠今次真係東升西降☀ 實測Deepseek R1: 淺白解說其特別之處 | 為什麼會吸引全球高度關注？ DeepSeek震撼美股！ The Engineering Unlocks Behind DeepSeek | YC Decoded Development: Deep-Agent / R1-V DeepSeek R1 GAVE ITSELF a 200% Speed Boost - Self-Evolving LLM Did DeepSeek R1 Invent a Language Humans CAN'T Understand? DeepSeek R1 Cloned for $30?! PhD Student STUNNING Discovery Politics: 城寨新聞 III 27/1/2025: (9:50) 城寨新聞 I 29/1/2025: 中共人工智能  (21:52) 城寨咖啡室： 2/2/2025: DeepSeek 〈蕭若元：蕭氏新聞台〉2025-01-28 〈蕭若元：蕭氏新聞台〉2025-01-30 《蕭若元：蕭氏新聞台》2024-01-31 《蕭若元》2025-02-02 《蕭氏新聞台》2025-02-04 What DeepSeek is Teaching us about U.S. National Security: A breakdown DeepSeek遭多地政府禁用 改變國運的突破？DeepSeek引爆中國社群媒體 Alexandr Wang Is DeepSeek lying? Scale AI CEO Alexandr Wang on U.S.-China AI race Legal and Copyright Issues OpenAI says DeepSeek used its models illegally Microsoft and OpenAI investigate   Deepseek懶人包 AI War DeepSeek異軍突起 能否助中國勝出美中科技戰？－ BBC News 中文

Research Papers  Research efforts have explored methods to enable LLMs to refine their outputs through self-reflection and iterative processes. A study titled "Large Language Models Can Self-Improve" demonstrated that LLMs could enhance their reasoning abilities by generating high-confidence, rationale-augmented answers for unlabeled questions and fine-tuning themselves using these self-generated solutions. arxiv.org Another approach, detailed in the paper "SELF: Self-Evolution with Language Feedback," involves a framework where LLMs engage in self-reflection and refinement, iteratively improving their performance without human intervention. This method draws inspiration from human learning processes, emphasizing the potential for LLMs to evolve through self-assessment and feedback. arxiv.org

  Bernard Stiegler  ( 1 April 1952 – 5 August 2020) was a French philosopher. Gilbert Simondon  ( 2 October 1924 – 7 February 1989) was a French philosopher best known for his theory of  individuation  and his work on the field of  philosophy of technology .  Yuk Hui  is a Hong Kong philosopher and Professor of Philosophy at Erasmus University Rotterdam. He is known for his writings on philosophy and technology. Bruno Latour  ( 22 June 1947 – 9 October 2022) was a French philosopher, anthropologist and sociologist.  He was especially known for his work in the field of  science and technology studies  (STS).   Andrew Feenberg  (born 1943) is an American philosopher. He holds the  Canada Research Chair  in the Philosophy of Technology in the School of Communication at  Simon Fraser University  in Vancouver. 

Genesis: Artificial Intelligence, Hope, and the Human Spirit Thus far, the history of computing has followed a trajectory of ever greater integration and interaction between humans and machines. Fashioning our tools to fit us ever more closely—in line with millennia of human practice— we have not previously considered developing tools that were unsuited to our anatomy or intellect, instead remaining guided only and wholly by the limits of our biology. But now the advent of AI may persuade at least some of us to contemplate a reverse mission: In a case where our tools appear to outpace our capabilities—as AI sometimes does already—might we consider engineering ourselves so as to maximize the tools’ utility and thus ensure our continued participation in shared endeavors like those outlined in previous chapters? Biological engineering efforts designed for tighter human fusion with machines are already underway. Starting with physical interconnects by means of chips in the human brain, ...

Genesis: Artificial Intelligence, Hope, and the Human Spirit As machines increasingly assume human qualities (and if some humans enhance themselves to take on machinelike qualities), lines will become hazy. What is AI and what is human will change and, in some cases, merge. In coming to judge how we must keep pace with AI, humans will therefore need to assert more clearly what it is that distinguishes us from machines. How, then, will we compile and compress the entire range of human experience for easy comprehension by AI? To preclude either our demotion or our replacement by machines, some will want to lay claim to difference via our proximity to divinity. Others will wish to ascertain more tactical conclusions: which kinds of decision-making can be delegated to machines and which cannot. We propose the articulation of an attribute, or set of attributes, that most of humanity could rally behind and orient around: one that will supply a floor underneath what is preferable but not a c...