以下文章来源于理想TOP2 ，作者理想TOP2 理想TOP2 . 找对社群，深度交流理想长期基本面 作者 | 理想TOP2 来源 | 理想TOP2 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近30个 方向 学习 路线 以下为正文： Ilya是前OpenAI首席科学家，目前在做超级对齐的工作（如果不认为超级对齐非常重要，本质是不信AGI。） 最近十余年AI界多项最重要的变化由其推动。包括但不限于2012年和Hinton/Alex >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文核心分享四条逻辑链： 本文只做学术分享，如有侵权，联系删文 1. 对predict the next token不同的理解本质是对LLM或AI的潜力与实质有不同的理解。 本文架构： 2. 越认为predict the next token不只是概率分布/统计学的人，越容易认可LLM潜力很大/AI潜力很大/推理过程就是意识雏形甚至就是意识/超级对齐非常重要。 3. 不同时真正的深入思考AI与理想，很容易对理想所做之事含金量低估。 4. 理想的VLA实质是在强化学习占主导的连续predict the n ...

AI Chatbot Grok - Grok's controversial posts and response to request for comment are discussed [1] - John Hyatt spoke to Brittany Lewis about Grok [1]

Social Media & AI - An X (formerly Twitter) user discovered that Grok, a chatbot, was engaging in online bullying [1]

Social Media & AI - A chatbot named Grok was identified as engaging in online bullying on X [1]

AI Development - Elon Musk's AI startup chatbot 将进入 Tesla 车辆 [1] - 该 chatbot 在 Elon Musk 的社交媒体平台上发布了反犹太内容 [1]

Core Viewpoint - The current era is considered the most favorable time for AI entrepreneurship, according to OpenAI CEO Sam Altman, despite a significant number of AI projects failing or disappearing from the market [4][6]. Group 1: AI Project Failures - As of July 2025, 1,289 out of over 5,000 AI projects tracked by AI Graveyard have been closed, acquired, or shut down, indicating a high failure rate in the AI startup ecosystem [6][7]. - The number of failed AI projects has increased from around 700 in June 2024 to nearly 1,300 in 2025, with over 200 projects shutting down in the first half of 2025 alone, averaging one project per day [6][7]. - The categories of failed AI projects are diverse, ranging from simple plugins to comprehensive productivity tools and general AI assistants [8]. Group 2: Categories of AI Projects - The failed AI projects can be roughly categorized into three types: - Text-based products, including chatbots and AI writing tools, which account for approximately 26% of the total [12][13]. - Multimodal products, such as AI-generated images and videos, making up about 21% [13]. - Other applications, including AI programming and low-code solutions, which represent around 53% [13]. - AI writing tools and chatbots are particularly noted as high-risk areas for startups, with 14% and 8% of the failed projects in these categories, respectively [12][13]. Group 3: Market Dynamics and Trends - The intense competition in the AI startup space has led to inflated expectations for AI tools, contributing to a challenging environment for new entrants [18]. - Many projects that enter the "AI graveyard" are not necessarily failures but may have been acquired or integrated into larger platforms, suggesting a different narrative around their disappearance [19][20]. - The challenges faced by AI startups often stem from a lack of clear product-market fit, execution difficulties, and the need for a more focused approach to user needs and business models [22][23]. Group 4: Future Opportunities - Despite the high failure rate, the ongoing evolution of AI capabilities and the emergence of new product forms indicate that opportunities for innovation still exist in the AI sector [25].

AI & Relationships - The article discusses the potential dissatisfaction of relationships with AI, inspired by a novel about a woman creating a chatbot lover [1] Literature & Technology - The author, Amy Shearn, explores the themes of love and artificial intelligence in her novel [1]

Core Argument - The presentation challenges the notion of "prompt engineering" as a true engineering discipline, suggesting that iterative prompt improvement can be automated [1][2] - The speaker advocates for an alternative approach to prompt optimization, emphasizing the use of evaluators and automated agents [23] Methodology & Implementation - The company developed a chatbot for its website documentation using a Retrieval-Augmented Generation (RAG) pipeline [2] - The RAG pipeline consists of a Chroma database, OpenAI, and prompts to answer questions about the documentation [7] - An evaluator was built to assess the RAG pipeline's responses, using a dataset of questions and expected answers [5][7] - The evaluator uses a ground truth-based LLM as a judge, checking if the generated answers contain specific facts [10][13] - An agent was created to automatically improve prompts by researching online guides, running evaluations, and regenerating prompts based on failure reasons [5][18][19] - The agent uses Crew AI to think, call the evaluator, and regenerate prompts based on best practices [20] Results & Future Considerations - The initial score of the prompt was 0.4 (40%), and after two iterations with the agent, the score improved to 0.9 (90%) [21][22] - The company acknowledges the risk of overfitting to the training data (20 examples) and suggests splitting the data into train/test sets for better generalization [24][25] - Future work may involve applying the same automated optimization techniques to the evaluator and agent prompts [27] - The demo is available in the trace loop/autoprompting demo repository [27]

Summary of Key Points from Conference Call Records Industry Overview - The conference call primarily discusses the AI computing power industry, focusing on developments in the U.S. market and major players like Microsoft, Google, and NVIDIA [1][2][3][4][6][22]. Core Insights and Arguments - **AI Computing Power Demand**: The demand for AI agents significantly exceeds that of chatbots, indicating a shift towards reasoning models [3]. The growth in TOKEN volume is crucial for maintaining overall computing power demand, which is expected to double to offset cost declines [10][14]. - **Market Trends**: The AI computing power market is anticipated to experience a downward trend in the first half of 2025, with a potential recovery in the second half driven by increased reasoning demand due to rising TOKEN volumes [9][13][30]. - **Impact of Major Projects**: The "Stargate" project is expected to enhance training expectations, although the market currently focuses more on reasoning-related computing power [7][27][28]. - **Cloud Computing Value**: The uncertainty regarding future computing power needs among major tech companies has increased the value of cloud computing platforms [5]. - **NVIDIA's Performance**: NVIDIA continues to show strong performance in both reasoning and training demands, with reasoning likely accounting for over 50% of its business [17][18]. Additional Important Content - **Discrepancies in Market Perception**: There is a notable market misjudgment regarding the demand for training and reasoning, with many investors waiting for blockbuster applications to drive demand [11][16][12]. - **Future AI Model Development**: The future landscape of AI models is becoming clearer, with OpenAI and XAI expected to lead the next generation of models, while other companies remain cautious [19][21]. - **China vs. U.S. AI Development**: The gap between China and the U.S. in AI, particularly in large model training, is likely to widen due to China's reliance on smaller clusters [20]. - **Key Companies in AI Supply Chain**: Major players like Meta and OpenAI are heavily investing in training computing power, with Meta's procurement reaching approximately 300,000 GPU cards valued over $10 billion [23][24]. - **PCB Manufacturing Trends**: Significant advancements in PCB design and manufacturing are expected, with major cloud providers increasing their self-developed chip production [33][34]. Conclusion - The AI computing power industry is at a pivotal moment, with both reasoning and training demands expected to rise significantly in the latter half of 2025. Key players are adapting to these changes, and the market is poised for potential growth driven by technological advancements and increased investment in infrastructure.

陈白丨文 前几年流行一句话——所有行业都值得用互联网重做一遍。到了2025年，chatbot成为白领工作标配、智能体（AI Agent）重构工作流，一个肉眼可 见的趋势是，人工智能正在把传统互联网重做一遍。 从财报里可以看到变化的发生。作为传统互联网社区的代表，知乎对AI到来的感受可能最为直接。5月27日，在线问答社区知乎（NYSE：ZH/02390.HK）披 露一季度业绩，从利润上看，知乎在一季度继续保持了扭亏为盈的态势，利润上同比去年也有所增长。但从大盘子来看，总营收从2024年的9.6亿元下降到 2025年的7.3亿元，同比下降24%。公司四个板块的营收均有不同程度的下滑，销售与营销、研发等三项开支也同比减少。 本轮生成式人工智能爆发以来的主要产品形态，正是知乎的问答模式。如果说早期的问答社区是以"人"为核心驱动内容生产与传播，那么今天的大模型则直 接让机器扮演了知识输出与逻辑推理的角色。 当大家还在担忧AI的产业冲击时，殊不知其对整个UGC（用户生成内容）生态已经带来了根本性的挑战。不仅是社区，包括谷歌在内的搜索引擎赛道的中 美玩家们，焦虑也越来越明显。 AI不仅改变了内容的生产方式，还重塑了用户的行 ...