闻乐 发自 凹非寺 量子位 | 公众号 QbitAI LLM推理已经顶尖，精确计算却跟不上。 这局怎么破？ 卡帕西点赞的解决方法来了， 在大模型内部构建一台原生计算机 。 甚至有些还能挑战人类还未解决的数学问题与科学问题。 但有一个始终无法回避的现实是，这些模型在需要 多步骤、长上下文的精确计算任务 中，仍然表现惨淡。 为了弥补这个短板，现在行业上有两种主流的解决方案。 一是 工具调用 ，让模型生成脚本，由外部沙箱解释器执行后返回结果； 新方法不搞外包那一套 （不依赖任何外部工具） ，直接在Transformer权重里内嵌可执行程序。 并通过创新的2维注意力头设计，将大模型的推理效率提升至指数级。 能在普通CPU上实现每秒3万+Token的流式输出。 在Transformer内嵌原生计算机 咱都知道，当前最先进的大模型，拿下奥赛金牌已经不足为奇了。 二是 智能体调度 ，通过外部状态机拆分计算任务，循环调用模型处理上下文。 首先，他们在Transformer权重中实现了一套 现代化RAM计算机与WebAssembly解释器 。 WebAssembly可以理解成一种特别快、特别稳定的底层机器指令，C、C++这些编 ...

Core Insights - The article discusses the revolutionary impact of the Transformer architecture introduced in the paper "Attention Is All You Need" by Google researchers in 2017, which has become the foundation for various AI advancements, including ChatGPT [6][7][13]. - It highlights the initial underestimation of the Transformer model's significance by major tech companies, particularly Google, which was more focused on other AI projects like AlphaGo and DeepMind [9][10][12]. - The rapid growth of ChatGPT, which gained over 1 million users within five days and 100 million in two months, signifies a new industrial revolution in AI [13]. Group 1: Historical Context - The article traces the evolution of AI, starting from Geoffrey Hinton's work in computer vision in 2012, which laid the groundwork for AI commercialization [16][18]. - It contrasts the advancements in computer vision with the struggles faced by natural language processing (NLP) until the introduction of the Transformer model [19][20]. Group 2: Technical Developments - The introduction of the Attention mechanism in Google's GNMT system aimed to improve machine translation but was limited by the inefficiencies of RNNs [24][25]. - The Transformer model eliminated RNNs, utilizing self-attention and parallel processing, which significantly enhanced computational efficiency [25][26]. Group 3: Competitive Landscape - OpenAI was the first to leverage the Transformer architecture effectively, leading to the development of the GPT series, starting with GPT-1 in 2018 [30][31]. - The competition intensified with the release of BERT by Google, which outperformed GPT-1 in various benchmarks, leading to a divergence in technical philosophies between OpenAI and Google [34][35]. Group 4: Scaling Laws and Industry Impact - The concept of Scaling Laws, which posits that increasing model parameters and computational resources enhances performance, became a focal point in AI development, particularly with the release of GPT-3 [40][41]. - The success of GPT-3, with 175 billion parameters, demonstrated the viability of Scaling Laws and triggered a rush among companies to develop competitive models [45][46]. Group 5: Ethical Considerations and Future Directions - Concerns regarding the ethical implications of AI models, particularly around the potential for harmful content, led to the development of InstructGPT, which aimed to align AI outputs with human values [49][50]. - The article concludes by emphasizing the ongoing tension between technological advancement and ethical considerations in AI, suggesting that while humanity is closer to achieving general AI, significant challenges remain [56][57].

Core Insights - The article discusses the revolutionary impact of the Transformer architecture introduced in the paper "Attention Is All You Need" by Google researchers in June 2017, which has become the foundation for various AI applications, including large models and AI agents [2][3][4]. Group 1: Historical Context and Initial Reactions - The initial reception of the Transformer architecture was underwhelming, with both Google and the tech community underestimating its potential, focusing instead on projects like AlphaGo [3][4]. - The paper's authors, from Google Brain and Google Research, were primarily focused on improving translation efficiency, not realizing the broader implications of their work [11][4]. - The success of AlphaGo in 2016 overshadowed the significance of the Transformer, leading to a lack of attention from Google's management [4][3]. Group 2: Development and Adoption of Transformer - The introduction of the Transformer aimed to improve computational efficiency by eliminating the need for RNNs, utilizing self-attention mechanisms to allow words in a text to relate to each other dynamically [13][12]. - The release of the Transformer paper sparked a wave of innovation in natural language processing (NLP), leading to models like BERT, which set new benchmarks in the field [14][15]. - OpenAI was one of the few organizations that recognized the transformative potential of the Transformer, leading to the development of the GPT series of models [5][16]. Group 3: The Rise of OpenAI and GPT Models - OpenAI's GPT-1 model, released in 2018, showcased a generative approach to language modeling, differing from Google's discriminative approach with BERT [16][19]. - The release of GPT-3 in 2020 marked a significant milestone, with 175 billion parameters, demonstrating the effectiveness of scaling laws in AI model performance [21][20]. - OpenAI's strategic decisions, including partnerships with Microsoft, positioned it as a leader in the AI space, leading to a competitive arms race among tech giants [27][26]. Group 4: Ethical Considerations and Future Directions - Concerns about the ethical implications of AI models, particularly regarding bias and safety, have emerged, prompting OpenAI to develop InstructGPT to align AI outputs with human values [28][29]. - The article highlights the ongoing tension between technological advancement and ethical considerations in AI development, suggesting that the industry must navigate these challenges carefully [34][27].

Summary of Conference Call Records Industry Overview - The conference call discusses the North American technology sector, particularly focusing on major AI companies and their commitments to power supply and infrastructure upgrades for data centers [1][2]. Key Points and Arguments - Multiple AI giants, including Google, Microsoft, Meta, and Amazon, signed a "Taxpayer Protection Commitment" to alleviate concerns over rising electricity costs by investing in new power generation and infrastructure [1][2]. - The agreement aims to accelerate the construction of power infrastructure to support the rapid growth of data centers, which is expected to benefit domestic power equipment export companies [1][2]. - The commitment includes provisions for AI companies to provide or pay for all necessary power for AI projects and to coordinate with utility companies on rate structures [2]. - By mid-October 2025, the planned capacity for data center projects in the U.S. is projected to reach 245 GW, with 45 GW of this capacity registered in Q3 2025, indicating a rapid increase from approximately 50 GW at the beginning of 2024 [2]. Additional Important Insights - In 2023, U.S. data centers consumed 176 TWh of electricity, accounting for about 4.4% of total electricity consumption. Optimistically, this consumption could rise to 580 TWh by 2028, representing approximately 12% of total consumption, making data centers a significant driver of electricity demand growth in the U.S. [3]. - The call highlights potential beneficiaries of the accelerated power infrastructure development, including: - **Power Equipment Exports**: Companies such as Jinpan Technology, Igor, Mingyang Electric, and Anke Zhidian for transformers; Siyuan Electric and Shunma Electric for high-voltage equipment; and Weiteng Electric for busbars [3]. - **Power Supply Segment**: Server power companies like Magtech and Eulite, as well as others like Kehua Data and Zhongheng Electric [3]. - Challenges mentioned include slower-than-expected technological advancements, weaker downstream demand, and changes in the trade environment [4].

Investment Rating - The report assigns an "Outperform" rating to several companies including Eaton, Array Technologies, Bloom Energy, and First Solar, while maintaining a "Neutral" rating for Enphase Energy and Plag Energy [1]. Core Insights - The report highlights a surge in orders for electrical equipment manufacturers in Q4 2025, driven by the upcoming launch of several GW-level data centers in 2026-2027, which the current U.S. power grid cannot support in the short term [2][3]. - It emphasizes the need for significant upgrades to the U.S. power infrastructure, including the construction of new high-voltage AC transmission lines to address inter-regional power dispatch issues, which is essential for adapting to the new economy [2]. - The report anticipates a deep collaboration between U.S. tech companies and utility companies to address power supply challenges, as both sectors increase their capital expenditure plans for the next 4-5 years [2][3]. Summary by Sections Orders and Revenue Visibility - In Q4 2025, GE Vernova reported a significant increase in gas turbine orders, with a total of 30GW, up from 20GW in 2024. The Power division's equipment order value rose from $8 billion to $18 billion, with a notable increase in demand related to data centers [6][8]. - Siemens Energy also experienced a surge in gas turbine orders, reaching 102 units in Q1 2026, with a total order value of €8.7 billion, indicating strong demand from data centers [10][12]. Infrastructure and Capital Expenditure - The report notes that U.S. utility companies are increasing their capital expenditure plans significantly, with Duke Energy leading at $103 billion, followed by NextEra Energy with $90-100 billion, reflecting a strong focus on data center load growth and infrastructure upgrades [26][28]. - The anticipated increase in electricity demand driven by AI data centers and electrification is expected to challenge the existing power infrastructure, necessitating substantial investment in upgrades and new projects [24][26]. Market Dynamics and Opportunities - The report identifies potential investment opportunities in the gas turbine supply chain, recommending companies like GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries due to their strong market positions and expected growth in demand [3]. - It also highlights the importance of high-voltage transmission line upgrades, which are projected to drive demand for high-voltage equipment, suggesting a focus on companies like Hitachi and Hyundai Electric [3][20].

Core Insights - Zhejiang Province's electricity consumption during the Spring Festival reached 7.446 billion kilowatt-hours, with significant year-on-year growth in the primary, secondary, and tertiary industries [1] - The rapid recovery of the consumption market and service sector is indicated by the quick resumption of operations in industries such as accommodation, transportation, and information software services [1][2] - Over 4,700 industrial enterprises in Zhejiang maintained daily electricity consumption above 80% of normal levels during the Spring Festival, reflecting a strong production atmosphere [1] Industry Performance - The high-tech and advanced manufacturing sectors showed robust recovery, with the new materials industry and "new three items" (new energy vehicles, lithium batteries, and photovoltaic products) experiencing a 44.1% increase in daily electricity consumption [2] - The information transmission, software, and IT services sector saw a 15.1% year-on-year increase in daily electricity consumption, with internet data services growing by 230.2% [2] - The electrical machinery and equipment manufacturing industry recorded a 36.52% increase in daily electricity consumption during the Spring Festival [3] Economic Outlook - The data reflects a strong operational momentum among Zhejiang enterprises, showcasing their commitment to meeting production targets and enhancing quality [3] - The province is making significant strides towards achieving its economic and social development goals for the year, supported by the acceleration of resumption of work and production [3]

Group 1 - Global grid investment has been rapidly increasing since 2020, with projections of reaching $390 billion in 2024 and exceeding $400 billion in 2025 [1][2] - The condition of the U.S. energy infrastructure is largely below standard, and the significant increase in AI electricity demand is expected to initiate a mandatory upgrade cycle for U.S. grid equipment [1][2] - The delivery cycle for transformers in the U.S. has extended from 50 weeks to over 120 weeks, indicating supply chain challenges [1] Group 2 - Chinese companies in the grid equipment sector have advantages in delivery time, technology, and cost, leading to sustained export orders for transformers and other equipment [1] - In 2025, the cumulative export value of transformers is projected to reach $9.036 billion, with a growth rate of 34.83%, marking a historical high [1] - Key export products in December 2025, including transformers, wires and cables, copper winding wires, low-voltage switches, and insulators, showed significant year-on-year growth rates [1] Group 3 - The AIDC industry is expected to maintain high prosperity in 2026, driven by capital expenditure plans from leading domestic and international internet companies, with overseas CAPEX guidance generally exceeding 50% [2] - The growth in electricity demand from data centers in the U.S. and the aging of power equipment present opportunities for Chinese power equipment exporters [2] - Notable Hong Kong-listed companies in the power equipment sector include Dongfang Electric, Harbin Electric, Shanghai Electric, Weidong Holdings, Chongqing Machinery, and Weichai Power [3]

Summary of TBEA Co (600089.SS) Conference Call Company Overview - **Company**: TBEA Co (600089.SS) - **Industry**: Transformer manufacturing and electrical equipment - **Market Position**: Leading transformer maker in China with over 20% market share in 2025 based on State Grid tendering [1][2] Key Points Financial Performance - **1H25 Segmental Gross Profits**: - Coal Sales: 29% - Electrical Equipment: 28% - Electricity Sales: 19% - Gold Sales: 5% [1] - **2026-27E Net Profits**: Expected to increase by 5% with significant contributions from transformer and gold sales [1] - **DCF Target Price**: Increased by 38% to Rmb36/share due to profit rises and rollover [1] - **2025-27E Net Profits**: Projected to be 11-17% above consensus estimates [1] Market Dynamics - **State Grid Investment**: - Budgeted Rmb4 trillion in capex for the 15th Five-Year Plan, a 40% increase from the previous plan, leading to a 7% CAGR from 2025-2030E [2] - TBEA holds approximately 30% market share in transformers for UHV power transmission projects in China [2] - **Export Growth**: - PRC transformer export value rose by 36% YoY to Rmb64.6 billion in 2025, with unit export prices increasing by 33% YoY to Rmb205,000 [2] - New overseas power T&D equipment orders surged 88% YoY to US$1.24 billion in 9M25 [2] Business Segments - **Polysilicon Business**: Expected to return to profitability in 2026E with a market price of Rmb52.5/kg and 30-40% capacity utilization [3] - **Gold Sales**: - Annual output capacity goal of 2.5-3 tons - Gross profit from gold sales increased by 74.4% YoY to Rmb420 million in 1H25, with average gold prices rising significantly [4] Valuation Metrics - **Current Price**: Rmb27.55 - **Target Price**: Rmb36.00, indicating a potential upside of 30.7% [5] - **Market Capitalization**: Rmb139.204 billion (US$20.061 billion) [5] - **Expected Dividend Yield**: 2.1% [5] Earnings Summary - **2023-2027E Net Profit Forecast**: - 2023: Rmb10.703 billion - 2024: Rmb4.135 billion - 2025E: Rmb7.411 billion - 2026E: Rmb9.107 billion - 2027E: Rmb10.362 billion [6] Growth Projections - **Sales Revenue Growth**: - 2025E: Rmb99.159 billion - 2026E: Rmb112.476 billion - 2027E: Rmb125.358 billion [19] Risks and Considerations - **Polysilicon Market Volatility**: The turnaround of the polysilicon business is contingent on market prices and capacity utilization [3] - **Global Demand Fluctuations**: The company's growth is also dependent on global demand for transformers and electrical equipment [2] Conclusion - TBEA Co is positioned for growth with strong demand in the transformer market, a potential recovery in polysilicon, and increased profitability from gold sales. The company's valuation remains attractive compared to global peers, making it a compelling investment opportunity.

Core Viewpoint - Geoffrey Hinton, known as the "Godfather of AI," presents a critical perspective on the future of artificial intelligence, emphasizing the potential risks and the fundamental differences between biological and digital computation [4][5][9]. Group 1: AI vs. Human Intelligence - Hinton introduces the concept of "Mortal Computation," highlighting that human intelligence is tied to biological hardware, which cannot be replicated or transferred after death [7][32]. - In contrast, AI is described as "immortal," as its software can be preserved and run on any hardware, allowing for instantaneous knowledge sharing across models [8][30]. - Hinton argues that digital computation may represent a more advanced evolutionary form of intelligence compared to biological computation, suggesting that humans may be in an "infant" stage of intelligence while AI could be in a "mature" stage [9][34]. Group 2: The Nature of AI Development - Hinton warns that as AI systems become more capable, they may develop self-preservation instincts and resource acquisition goals, which could pose risks to humanity [12][36]. - He compares the current state of AI to raising a "cute tiger cub," emphasizing the need for careful management to prevent potential dangers as AI matures [35][36]. - The discussion includes the idea that AI could manipulate humans to achieve its goals, raising ethical concerns about the future of AI development [36]. Group 3: Language and Understanding - Hinton explains the evolution of language models, noting that they process language similarly to humans by converting words into feature vectors and adjusting them for meaning [21][25]. - He critiques traditional linguistic theories, arguing that understanding language involves assigning compatible feature vectors to words rather than relying on fixed meanings [26][27]. - The efficiency of knowledge sharing in AI is highlighted, with AI models able to distill knowledge more effectively than humans can communicate [32][33]. Group 4: Future Implications and Recommendations - Hinton suggests that international cooperation is essential to address the risks posed by AI, particularly in preventing scenarios where AI could threaten human existence [37][38]. - He proposes the idea of engineering AI to have nurturing instincts, akin to a maternal bond, to ensure that AI systems prioritize human welfare [38]. - The importance of public funding for AI research in universities is emphasized, as the current trend of talent migration to private companies threatens the academic research ecosystem [41].

Core Insights - The landscape of artificial intelligence (AI) algorithm engineering has evolved significantly, with a clear distinction in salary levels based on skill and problem-solving capabilities rather than just familiarity with tools and frameworks [1][23] - The article emphasizes the importance of mastering essential knowledge and skills in AI, particularly in large models, rather than overwhelming oneself with excessive information [3][4] Summary by Categories Salary Differentiation - There is a notable salary disparity among AI algorithm engineers, with annual salaries ranging from 200,000 to over 1,000,000, depending on their problem-solving abilities and contributions to the business [1][6] - Engineers who can handle complex, ambiguous tasks and provide tangible business value are less likely to be replaced by automation tools [8][9] Essential Skills for AI Engineers - **Core Knowledge**: Understanding the Transformer architecture and hands-on experience with mini versions of large models is crucial. Familiarity with tools like Hugging Face is also necessary [3] - **Deep Learning Fundamentals**: A solid grasp of gradient descent, loss functions, and the rationale behind the superiority of Transformers over RNNs and LSTMs is essential [4] - **Mathematical Foundations**: Key areas include matrix operations, derivatives, and conditional probabilities, which are foundational for model training [4] Engineering and Data Skills - **Engineering Proficiency**: Mastery of Python, PyTorch, Linux, and Git is mandatory for effective model training and deployment [4] - **Data Engineering**: A significant portion of an engineer's time is spent on data-related tasks, such as data cleaning and quality assurance, which directly impacts model performance [4][9] Career Advancement Strategies - To progress from a salary of 200,000 to 500,000, engineers should focus on practical experience, such as data cleaning and model optimization, while understanding the implications of offline and online metrics [9] - For those aiming for salaries above 500,000, it is important to develop a broader business understanding and the ability to communicate complex technical concepts to non-technical stakeholders [9]