在一次广泛的讨论中，风险投资家埃里克（Eric）分享了他对科技发展格局，尤其是在人工智能及其对传统商业模式影响方面的见解。他质疑现有SaaS公司能轻易适应AI时代的说法，指出它们的竞争前沿已经彻底转移。过去通过精心执行计划所创造的价值，如果公司未能适应AI，现在反而可能毁于一旦。 埃里克从他公司的投资中汲取经验，重点提到了人工智能基础设施公司 **Fireworks**。他解释说，运行大型AI模型异常困难。像 Fireworks 这样的专业公司，即使使用相同的开源模型和硬件，也能比主要的云服务提供商实现5倍的速度和显著更高的吞吐量。这种效率源于他们在优化AI模型执行方面深厚的专业知识，这证明它并非商品化业务，而是需要独特技能的领域。 埃里克将过去**云计算**的普及与**当今AI**的采用进行比较，指出了显著的相似之处和差异。早期的AWS曾面临质疑，但市场规模最终被证明远超预期，导致了主要参与者的寡头垄断，以及众多市值达1000亿美元的“较小”赢家。他预测AI也会出现类似结果，即基础模型提供商形成寡头垄断，同时也会有许多成功的专业公司。一个关键区别在于当前企业对AI的拥抱：与早期对云的怀疑不同，蓝筹企业将AI视为机遇和威胁，积极进行实验并寻求“AI向导”（AI Sherpas）来指导其采用。 在AI应用方面，埃里克提到了**Sierra**公司，它从客户服务自动化起步，现在正在构建“长效代理”（long-running agents）。Sierra 堪称产品开发中的一个新范式，其中理解“AI能力的不平坦边缘”至关重要。产品经理现在必须深入理解模型的细微差别、优势和不足，而不仅仅是客户问题。这需要持续的演进和“建造沙堡”（build sandcastles）的意愿，即拥抱持续的淘汰。 新时代需要**敏捷性**，并要求重新评估传统的商业策略。例如，旧的销售模式举步维艰，因为AI驱动的产品常常销售的是“魔法”，超出了传统的配额能力。赢家是那些敏捷、技术深厚，并愿意从第一性原理质疑所有假设的人。 在讨论市场瓶颈时，埃里克认为对智能的需求是“无限的”，但他对**能源供应**表示担忧。他提出，模型将算力转化为智能，而对智能的巨大需求意味着对能源的巨大需求。全球能源短缺可能会限制AI的增长，导致智能减少或AI服务成本更高，他认为这“非常糟糕”。 他与**Cerebras**（一家旨在为深度学习构建专用芯片的硬件公司）的经验，凸显了硬件投资的巨大难度。尽管关于GPU不适合某些AI工作负载的最初洞察是正确的，但将其转化为功能性产品却涉及应对复杂的物理学问题、错综复杂的供应链以及多年不懈的优化。他认为，AI代表着一种新的、巨大的工作负载，类似于过去的几代计算技术（CPU、GPU、移动），每一代都催生了新的数十亿美元公司。他甚至预见到专用CPU来运行AI生成的代码的新机会。 在**机器人技术**方面，埃里克认为关键挑战是从受控环境转向现实世界任务，而这需要AI。当前的瓶颈是机器人缺乏互联网规模的数据，这类似于大型语言模型（LLMs）从人类文本数据中自我发展起来的方式。解决方案包括垂直整合的机器人-模型-数据收集系统，以及利用高价值数据。他引用了Sunday Robotics的例子，该公司使用专用手套收集洗衣折叠等任务的精确数据，专注于训练和模型改进的“飞轮效应”（flywheel），而不仅仅是特定任务本身。 回顾他作为投资者的角色，埃里克强调了与创始人的**伙伴关系**和深厚的默契。他寻找那些能真正向企业家学习并与他们合作的机会，经常是引导他们巩固自己的信念，而不是直接指挥他们。他的投资理念包括以下问题：“我能否说服我在乎的人加入这家公司？”“我会在周六晚上9点接电话吗？”以及“如果我们做对了，有人会在意吗？”——这确保了个人投入和市场影响力。 他还解释了公司决定设立**增长基金**的原因，他认识到由于现代成果的巨大规模，高现金回报倍数的机会领域已经超越了早期投资。 最后，埃里克谈到了当前的内部讨论，尤其是围绕**AI生态系统中的价值积累**。他驳斥了“零和思维”，认为由于AI市场的巨大规模，基础模型、基础设施、应用和硬件等多个层级都将蓬勃发展。他还引用了杰夫·辛顿（Jeff Hinton）关于AI取代放射科医生的预测，并以此为例提出了一个警示，强调数据可用性、体制惯性、责任归属等现实世界因素会显著减缓AI的普及，并将AI的角色从替代转变为辅助，这也预示着更广泛的AI驱动的失业潮可能遵循类似轨迹。

In a wide-ranging discussion, Eric, a venture capitalist, shares his insights on the evolving landscape of technology, particularly in AI and its impact on traditional business models. He challenges the notion that existing SaaS companies will easily adapt to the AI era, stating that their competitive frontier has completely shifted. What once built equity – meticulously executing a plan – now risks destroying it if companies fail to adapt to AI. Drawing lessons from his firm's investments, Eric highlights **Fireworks**, an AI infrastructure company. He explains that running large AI models is exceptionally difficult. Specialized companies like Fireworks can achieve 5x speed and significantly higher throughput than major cloud providers, even when using the same open-source models and hardware. This efficiency stems from deep, specialized expertise in optimizing AI model execution, proving that this isn't a commodity business but one requiring unique skill. Comparing the adoption of **cloud computing** in the past to **AI today**, Eric notes striking parallels and differences. Early AWS faced skepticism, but the market proved much larger than anticipated, leading to an oligopoly of major players and numerous $100 billion "smaller" winners. He predicts a similar outcome for AI, with an oligopoly of foundational model providers and many successful specialized companies. A key difference is the current enterprise embrace of AI: unlike the initial skepticism towards cloud, blue-chip enterprises view AI as both an opportunity and a threat, actively experimenting and seeking "AI Sherpas" to guide their adoption. For AI applications, Eric points to **Sierra**, which started with customer service automation and is now building "long-running agents." Sierra exemplifies a new paradigm in product development where understanding the "jagged edge of AI capabilities" is paramount. Product managers must now deeply grasp model nuances, strengths, and failures, rather than just customer problems. This requires constant evolution and willingness to "build sandcastles," embracing continuous obsolescence. The new era demands **nimbleness** and a re-evaluation of traditional business playbooks. Old sales models, for example, struggle because AI-powered products often sell "magic," exceeding conventional quota capacities. Winners are those who are agile, deeply technical, and willing to question every assumption from first principles. Discussing market bottlenecks, Eric sees the demand for intelligence as "unlimited," but expresses concern about **energy supply**. He posits that models translate compute into intelligence, and massive demand for intelligence implies massive demand for energy. A global energy shortage could constrain AI growth, leading to less intelligence or more expensive AI services, which he views as "very bad." His experience with **Cerebras**, a hardware company aiming to build specialized chips for deep learning, underscores the immense difficulty of hardware investing. While the initial insight about GPUs not being optimal for certain AI workloads was correct, translating that into a functional product involved navigating complex physics, intricate supply chains, and years of relentless optimization. He argues that AI represents a new, massive workload, akin to past generations of computing (CPUs, GPUs, Mobile), each of which birthed new multi-billion dollar companies. He even anticipates a new opportunity for specialized CPUs to run AI-generated code. In **robotics**, Eric believes the key challenge is moving from controlled environments to real-world tasks, which requires AI. The current bottleneck is the lack of internet-scale data for robots, similar to how LLMs bootstrapped from human text data. Solutions involve vertically integrated robot-model-data collection systems and leveraging high-value data. He cites Sunday Robotics, which uses specialized gloves to collect precise data for tasks like laundry folding, focusing on the "flywheel" of training and model improvement rather than just the specific task. Reflecting on his role as an investor, Eric emphasizes **partnership** and deep chemistry with founders. He seeks opportunities where he can genuinely learn from and collaborate with entrepreneurs, often guiding them to solidify their own convictions rather than directing them. His investment philosophy includes questions like: "Could I talk someone I care about into going to this company?", "Will I pick up the phone at 9 p.m. on a Saturday night?", and "If we're right, will anybody care?" – ensuring personal commitment and market impact. He also explains the firm's decision to raise a **growth fund**, acknowledging that the landscape of high cash-on-cash multiple opportunities has expanded beyond just early-stage investments due to the massive scale of modern outcomes. Finally, Eric touches upon current internal debates, particularly around **value accrual in the AI ecosystem**. He dismisses "zero-sum thinking," believing that multiple layers – foundational models, infrastructure, applications, and hardware – will all thrive due to the sheer size of the AI market. He also offers a cautionary tale from Jeff Hinton's prediction about AI replacing radiologists, highlighting that real-world factors like data availability, institutional inertia, and liability can significantly slow adoption and shift AI's role from replacement to augmentation, suggesting a similar trajectory for broader AI-driven unemployment.