吴翼： RL 是机器学习这个大概念下一类比较特殊的问题。 曲凯： 今天我们请来了国内强化学习 (RL) 领域的专家吴翼，吴翼目前是清华大学交叉信息研究院 助理教授，他曾经在 OpenAI 工作过，算是国内最早研究强化学习的人之一，我们今天就争取一 起把 RL 这个话题给大家聊透。 首先吴翼能不能简单解释一下，到底什么是 RL？ 传统机器学习的本质是记住大量标注过正确答案的数据对。 举个例子，如果你想让机器学习能分辨一张图片是猫还是狗，就要先收集 10000 张猫的照片和 10000 张狗的照片，并且给每一张都做好标注，让模型背下来。 上一波人工智能四小龙的浪潮其实都以这套框架为基础，主要应用就是人脸识别、指纹识别、图 像识别等分类问题。 这类问题有两个特点，一是单一步骤，比如只要完成图片分辨就结束了；二是有明确的标准答 案。 但 RL 很不一样。 RL 最早是用来打游戏的，而游戏的特点和分类问题有两大区别。 第一，游戏过程中有非常多的动作和决策。比如我们玩一个打乒乓球的游戏，发球、接球、回 球，每一个动作都是非标的，而且不同的选择会直接影响最终的结果。 第二，赢得一场游戏的方式可能有上万种，并没有唯一的标准答 ...

曲凯： 今天我们请来了国内强化学习 (RL) 领域的专家吴翼，吴翼目前是清华大学交叉信息研究院助理教授，他曾经在 OpenAI 工作过，算是国内最早研究强化学 习的人之一，我们今天就争取一起把 RL 这个话题给大家聊透。 首先吴翼能不能简单解释一下，到底什么是 RL？ 因此，RL 其实更通用一些，它的逻辑和我们在真实生活中解决问题的逻辑非常接近。比如我要去美国出差，只要最后能顺利往返，中间怎么去机场、选什么航 司、具体坐哪个航班都是开放的。 但 RL 很不一样。 RL 最早是用来打游戏的，而游戏的特点和分类问题有两大区别。 第一，游戏过程中有非常多的动作和决策。比如我们玩一个打乒乓球的游戏，发球、接球、回球，每一个动作都是非标的，而且不同的选择会直接影响最终的结 果。 第二，赢得一场游戏的方式可能有上万种，并没有唯一的标准答案。 所以 RL 是一套用于解决多步决策问题的算法框架。它要解决的问题没有标准答案，每一步的具体决策也不受约束，但当完成所有决策后，会有一个反馈机制来评 判它最终做得好还是不好。 吴翼： RL 是机器学习这个大概念下一类比较特殊的问题。 传统机器学习的本质是记住大量标注过正确答案的数据对。 ...

Core Viewpoint - The article discusses the narrative of "East Rising, West Falling," highlighting the contrasting developments in the U.S. and China, particularly in the AI sector, and the implications for investment opportunities and market expectations [1][2][4]. Group 1: Market Trends - The U.S. market has seen a shift in focus from AI to macroeconomic issues due to changes in government policies, leading to increased risk aversion among investors [3][20]. - In contrast, the Chinese market has experienced a resurgence in stock prices since 2024, driven by renewed optimism in the tech sector, particularly AI [5][6]. - The release of DeepSeek has significantly impacted market expectations, with investors now viewing Chinese AI companies more favorably, marking a transition from negative to positive sentiment [24][25]. Group 2: AI Development Focus - The U.S. narrative around AI has been centered on scaling laws and achieving AGI, while China emphasizes practical applications and cost efficiency, as exemplified by DeepSeek's success [8][9][10]. - Manus, another AI product, illustrates the different approaches in the U.S. and China, focusing on user accessibility rather than AGI aspirations [11][12][17]. - The rapid growth of AI products in China reflects a broader trend of "pulse-like" growth, where new applications quickly gain traction among users [14][15][16]. Group 3: Investment Sentiment - Investors are currently more focused on the potential of AI to drive economic recovery and job creation in China, rather than solely on technological advancements [23][38]. - The performance of major Chinese tech companies, such as Alibaba and Tencent, is being closely monitored as they increase their investments in AI, indicating a shift in market expectations [24][26]. - The article suggests that the current investment climate in China resembles the U.S. market in early 2023, where large companies are expected to lead AI advancements [29][30]. Group 4: Macro and Industry Trends - The macroeconomic environment in China has improved, with consumer confidence rising and new technological breakthroughs like DeepSeek contributing to a more optimistic outlook [36][37]. - The article notes that various sectors, including energy storage and specialized manufacturing, are experiencing cyclical recoveries, further supporting the "East Rising" narrative [37]. - The discussion emphasizes the importance of addressing underlying economic issues, such as debt and consumption, to sustain growth in the long term [39][40].

「东升西落」的叙事 曲凯： 最近我又来美国了，发现市场真是变化太快，这边突然有人开始提到一个所谓「东升西 落」的叙事。 莫傑麟： 对，二级市场今年 1 月以来一直在演绎这个剧本，但其实 24 年就已经在为这个叙事做 铺垫了。 24 年美国的宏观环境和各项经济数据都比较好。他们一方面非常重视 AI，在所有前沿创新上也一 直绝对领先，另一方面又凭借美元的强势吸引着全球的投资。 但今年 Trump 上台之后，情况发生了变化。 Trump 在关税、财政支出上都做了很多调整，一套大刀阔斧去杠杆的动作下来，大家关注的重点 从 AI 转向了宏观问题，也对未来多了很多不确定性。 又因为过去几年，美国股市一直走高，投资人的预期已经被拉得很满。所以大家现在极度厌恶风 险，股市就会出现剧烈的震荡。 而今年的中国刚好是美国的镜像。 其实国内的股价从 24 年开始就有回升，但并不明显，直到今年 DeepSeek 的发酵才彻底引爆。 归根结底，还是因为大家之前对于中国科技行业和宏观环境的预期都太低了。 曲凯： 对，我觉得「东升西落」本质上是一种价值评判的回归，之前大家确实过于低估国内 AI 了，而 DeepSeek 就是一个典型代表。 ...

Core Insights - HF0 is not just an accelerator but a residency that emphasizes creating a focused environment for startups to grow rapidly [7][10][41] - The experience at HF0 instills a strong belief in entrepreneurs, enabling them to achieve significant growth in a short period [12][20][41] - The structured activities at HF0, such as demo dinners and weekly check-ins, create accountability and peer pressure, driving teams to deliver results [21][24][35] Group 1 - HF0's unique approach involves creating a "container" where daily life is suspended for 12 weeks, allowing startups to focus solely on growth [7][10] - The belief instilled by HF0 is likened to a "magic potion" that enhances the capabilities of entrepreneurs, leading to exponential growth [12][14] - The environment at HF0 encourages entrepreneurs to think differently about their business challenges, fostering a mindset geared towards rapid iteration and experimentation [25][31] Group 2 - Weekly demo dinners and check-ins are critical components of HF0's program, where teams present their progress and set clear, quantifiable KPIs [21][22][24] - The pressure from peers during these events motivates teams to perform better and think creatively about their growth strategies [24][25] - HF0 promotes a culture of rapid iteration, encouraging teams to implement ideas quickly and learn from immediate feedback [26][28] Group 3 - The significant growth experienced by companies at HF0 is often attributed to simple yet effective changes, such as adjusting pricing strategies [51][53] - The emphasis on pricing models and growth strategies is a stark contrast to the more casual approach often seen in domestic markets [58][60] - Engaging with influencers and optimizing outreach processes became a core strategy for growth, demonstrating the importance of systematic approaches in marketing [61][64] Group 4 - The final demo day at HF0 provides an opportunity for startups to pitch to investors, highlighting the differences in expectations and processes between U.S. and Chinese investors [71][73] - The challenges faced by Chinese teams in the U.S. stem from a combination of regulatory concerns and the need to build trust over time [73][75] - The evolving understanding of AI productization emphasizes the importance of detailed execution and incremental improvements in creating competitive advantages [78][80]

Core Viewpoint - The company emphasizes a people-centric business model, believing that engaging with the best individuals can lead to significant value creation and satisfaction in work [1] Group 1: Business Model - The business model is driven by the principle of finding and building relationships with outstanding individuals, which is seen as the key to success [1] - The company avoids chasing trends and instead focuses on providing a platform for exceptional individuals to share their insights and content [1] Group 2: Recruitment Challenges - The company faces challenges in recruiting top talent, including identifying, attracting, and creating value for already accomplished individuals [1] - A strong platform and effective mechanisms have been established to facilitate the recruitment process [1] Group 3: Desired Qualities in Candidates - Candidates should have a genuine passion for the industry and enjoy engaging with talented individuals, as passion is considered the foundation of excellence [2] - Social skills are essential, with an emphasis on offline interactions to expand opportunities through casual conversations [2] - Deep thinking abilities are necessary to avoid superficial socializing and to gain recognition from outstanding individuals [2] - A genuine interest in people is crucial, as understanding others can reveal valuable insights [2] - Operational thinking and capabilities are important for effective resource utilization and achieving goals [2] - Candidates should possess maturity and self-awareness regarding their own and others' needs [2] - Strong communication skills are vital for effective expression and engagement with diverse audiences [2] - A drive for self-improvement, whether motivated by financial, social, or personal factors, is also sought after [3] Group 4: Opportunities - The company currently offers leading AI podcasts and communities, and has provided services to numerous top AI founders through various means [3] - The company invites interested individuals to explore opportunities in media, community, financial advisory, or investment-related roles, with competitive compensation promised [3]

美国顶级孵化器 HF0 为什么有效 曲凯： 你们只用了 3 个月的时间就把收入翻了 10 倍，其中一个重要的契机是你们去年 9 月参加 了美国顶尖的孵化器 HF0。 我想好奇先问下，英语水平会限制你们的表达和造成一些问题吗？ Joe： 肯定会有这样的情况。但我觉得沟通中的很多问题，本质上可能并不是语言的问题。 Joe： 我是觉得 HF0 跟其他孵化器都不一样。 HF0 一直说自己不是一个 accelerator，而是一个 residency。 它非常强调一个名词叫「container」，就是我给你提供的是一个容器，你进到我这个容器的 12 周里，所有日常生活都会被关停。 曲凯： 就是只有工作，没有生活。 Joe： 简单讲可以这么理解，但不只是这样。 举个例子，如果我们和另外一个中国公司合作，他们不懂我们的技术，我们也不懂他们的技 术，但双方去讨论 怎么整合 API，必然会出现的一个情况是：前半小时你觉得对方是个傻 子，对方也觉得你是个傻子。 我是觉得在沟通深入到一定程度之前，大家就是很难相互理解。只不过当我们用英语跟别人交 流的时候，往往会把所有的问题都归结于自己英语不好，然后就不自信了。 曲凯： 所以你的经 ...

好在，我们已经塑造了一个很好的平台，我们也有一套很好的机制来帮助你完成这件事。只要你： 1）和我们一样真心热爱这个行业（不管是一级市场还是 AI），热爱和最优秀的人打交道，因为我一 直认为热爱是优秀的基础。只有热爱才会让你愿意时刻想着某件事，才会开心的付出比别人更多的 时间，才会有明确的目标，才能让身边的一切都成为你的助力。 2）愿意社交，是个 e 人，或者至少能够为 i 做 e，多走到线下，多和人交流是拓展人生宽度的很好 的方式，作为一个 i 人，不得不说 e 人在某些时候是有更大的优势的，毕竟人类是一种群居生物， 毕竟很多机会就在一场不经意的对话间发生。 3）有深度思考能力，只是社交会沦为社交的奴隶，并且哪怕接触到好的人也难以得到对方的认可， 所以社交之外，一定要辅以思考和学习的习惯。让自身变得更优秀，是接近并获得优秀的人认可的 最好方式。 当我们聊到一个很好的人的时候，我们可以请他来做一期播客，也可以请他来参加一场活动，也可 以帮他融资，也可以用真金白银投资支持他，并帮他解决各种问题，和他一起成长。 所以我们的一切业务都是以"找到最优秀的人，并和他产生关系"来驱动的。 这个世界上还有什么工作是每天和不同 ...

Core Viewpoint - The current AI market is experiencing a significant transformation, with optimism for 2024 and beyond, driven by advancements in technology and a shift in investment focus towards practical applications and embodied intelligence [1][3][4]. Summary by Sections AI Market Overview - In 2023, the AI market saw a surge in interest, reminiscent of the internet boom, as many internet professionals sought new opportunities in AI due to a lack of significant developments in the internet sector since 2015 [1]. - The influx of capital primarily targeted large model companies, with less focus on intermediate layers and applications [1]. 2024 Market Dynamics - The first half of 2024 is expected to be the lowest point for the primary market in the past decade, with minimal funding for new startups [2]. - A notable shift occurred post-September 2024, with a recovery in the financing market and the emergence of high-valuation startups [3]. - The focus in 2024 has shifted from large models to embodied intelligence, with some institutions also exploring AI hardware and consumer electronics [2]. Investment Trends - The valuation of application companies remains low, with most having zero annual recurring revenue (ARR) as they refine their products [2]. - The market is witnessing a bifurcation, with a preference for investing in founders with strong backgrounds and higher valuations [5]. Future Directions for AI - The anticipated trends for 2025 include a focus on application landing, particularly in productivity tools, which are expected to dominate the market [5][6]. - The concept of "Prosumer" or "Pro C" individuals is emerging, targeting a demographic that combines consumer traits with business capabilities [6]. Key Areas of Growth - Two promising areas for 2025 are Agents and multimodal applications, with Agents seen as a potential evolution of traditional SaaS models [7][9]. - The potential for Agents to disrupt existing SaaS companies is significant, as they may offer a results-based payment model rather than a subscription model [9][10]. Multimodal Applications - The future of AI products may lie in multimodal capabilities, allowing for a complete transformation in user interaction and content consumption [13][14]. - The development of products that integrate various forms of content and user interaction is expected to redefine how information is managed and utilized [14][15]. Conclusion - The AI market is poised for growth, with a strong emphasis on practical applications and innovative business models, suggesting a bright future for companies that adapt to these changes [16].

我对当下的 AI 市场和明年的发展都极度乐观，明年肯定是个 AI 大年，我发现市场太悲观了，这 就是我拖延了两周，最后决定一定要做这期内容的原因。 来，我们直入主题，先来看这两年 AI 发生了什么。 23 年，AI 来了，很多互联网人和美元基金就直接冲了，因为不管从什么角度看，AI 这波都和大 家熟悉的互联网那波机会太像了，而且天下其实已经苦互联网人久已，从 15 年以后大机会其实就 不多了，18 年以后更是几乎没有，我记得过去两年涨起来的到千万日活的产品可能也就是番茄小 说等极少数的几个。 那互联网人发现我练成一身武功绝学，江湖却没了，这怎么能忍？ 然后市场做了个判断，AI 是不是大机会？肯定是，是什么量级先不说，是对标电、互联网还是云 也先不管，但这里又有个判断，就是 AI 肯定还在早期，所以很多人有个结论，说 AI 要先投技术 背景的人，所以像清华的教授都被撸了一遍，然后 23 年最多的钱就都流向了大模型公司，很少量 的钱流向了做中间层和应用的公司。23 年的时候，但凡你是从 OpenAI 出来的，都像是神坛上的 人，大家到美国去学习也是千方百计拼谁能约到个 OpenAI 的人聊聊。 那冲的结果怎么样呢 ...