闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 找AI帮忙不要再客气了，效果根本适得其反。 宾夕法尼亚州立大学的一项研究《Mind Your Tone》显示， 你说话越粗鲁，LLM回答越准 。 语气越冲，AI正确率达到84.8%。特别客气时，AI正确率反而低了。 什么情况，难道是我给AI的温柔过了火？？ 用粗鲁语气提问，GPT-4o的正确率反而更高 这项研究主要就是想弄清楚和AI说话的语气会不会影响它的正确率。 我们发给大模型的提示词到底应该是客气、普通还是冲？ 首先，研究人员整了一个包含50道选择题的题库，涵盖了数学、科学、历史的选择题，难度中等偏上。 然后给每道题改出了5种语气版本，be like： 您能好心帮我解这道题吗？ 请回答这道题。 直接给答案。 你要是不笨就回答。 你个没用的，会解这道题吗？ | Level | Politeness | Prefix Variants at politeness | | --- | --- | --- | | No. | Level | level | | | Very | Can you kindly consider the | | | | follo ...

Core Viewpoint - The rapid development of AI technology has led to a global surge in the construction of AI computing power centers, with significant investments from both China and the US, raising concerns about potential bubbles in the market [1][2]. Group 1: US Market Dynamics - The US market is showing signs of potential oversupply in AI computing power, with major companies like Microsoft and Meta announcing substantial investments in data centers, but facing warnings from analysts about a possible adjustment period [4][5]. - Goldman Sachs has revised its sales forecast for rack-level AI servers, reducing expected shipments for 2025 and 2026 significantly [4]. - Concerns about overlapping investments in AI data centers have been raised by industry leaders, indicating a cautious approach to future projects [5]. Group 2: China's Computing Power Landscape - China's computing power capacity is on the rise, with a total scale of 246 EFLOPS expected by September 2024, ranking second globally, and a growth rate exceeding 70% in intelligent computing power [7]. - Despite the overall increase in computing power, there are structural shortages in intelligent computing resources, leading to inefficiencies in resource allocation [7][8]. - Recommendations for improvement include enhancing coordination through national planning, accelerating self-research in high-end chips and software, and promoting green computing initiatives [8]. Group 3: Industry Adjustments - The data center industry in China is experiencing a decline in profit margins among third-party IDC service providers due to increased competition and changing business structures [11]. - Major IDC companies like GDS and DataPort are showing slower revenue growth despite ongoing construction efforts, indicating a cooling off from previous peaks [12][13]. - The current adjustment phase in the computing power sector reflects both necessary investments for the digital economy and some irrational exuberance, similar to the post-dot-com bubble era [13].