以下是对 Harry Stebbings 在 20 VC 播客中对 Bret Taylor 的采访总结。Taylor 是硅谷资深人士，拥有令人印象深刻的履历，包括 Google 地图、Facebook CTO、Quip 创始人，以及现在的 Sierra（一个对话式人工智能平台）的 CEO。他分享了他对当前人工智能格局、代理的未来和领导力的见解。 Taylor 认为目前的人工智能热潮确实是一个泡沫，但它会与互联网泡沫 “押韵”。他认为，虽然存在过度投资的情况，但长期影响将是变革性的，催生万亿美元市值的公司和具有行业定义意义的企业软件。他强调，投资的核心应该是影响力和具有世代价值的公司，而不是仅仅关注倍数。 他驳斥了大型模型将取代所有软件的想法，并将其类比为云市场，基础设施、工具制造商和 SaaS 解决方案并存。他认为，公司更倾向于购买解决方案，而不是自己构建，并认为这种观点在人工智能领域也将成立。他分享了他对 Sierra 的愿景，即该公司将使企业能够构建自己的面向客户的代理，以解决各种问题。 虽然他承认由于缺乏现成的解决方案，人工智能服务公司在短期内会出现激增，但他预计随着 SaaS 解决方案的成熟，这种趋势将会下降。然而，他预测人工智能服务公司将在变革管理方面发挥重要作用，帮助企业适应这场重大变革。 Taylor 意识到人工智能模型的快速商品化，区分了“基础模型”（商品）和“前沿模型”（最佳）。他对大多数公司预训练模型持怀疑态度，认为除非是为了 AGI 研究，否则这是资本密集型的。相反，他认为大多数公司应该利用现有的模型，并针对特定的解决方案进行微调。 关于 AGI，Taylor 提倡负责任的迭代部署，以了解其对社会的影响、获取方式和安全性。他强调了人工智能的可访问性的重要性，以及像 ChatGPT 这样的产品如何为所有人提供广泛的人工智能访问。他认为 OpenAI 内部的企业部门是开发人员围绕该技术构建价值的绝佳途径。 在讨论可持续人工智能商业模式的挑战时，Taylor 强调在产生大量训练成本之前，要专注于产品与市场的匹配。他对由于技术进步而降低的推理成本感到乐观。 在谈到大型超大规模企业的投资时，Taylor 认为，对于公司来说，投资于自己的基础设施是合理的，因为有太多的价值 stake 不能置身事外。这项投资对消费类产品的开发和利用具有重大影响。他认为，开源模型和专有解决方案并存对生态系统来说是净收益。他还强调，对模型构建的投资最终将带来产生收入的机会。 谈到代理，Taylor 将 Sierra 定位为使公司能够构建品牌化的面向客户的人工智能代理，他认为对话是未来人机交互的方式。他用智能手机和触摸屏类比来支持对话式人工智能代理是不可避免的说法。 他讨论了在生成式人工智能中创建确定性的技术挑战，即如何在创造力与业务规则之间取得平衡。他强调了为人工智能代理定义目标和护栏的重要性。这带来了一个独特的设计挑战，即在代理性和控制力之间取得平衡，像人工智能架构师这样的角色将出现来管理对话设计。最终，Sierra 正在帮助公司为所有客户用例构建他们自己的品牌人工智能代理。 最后，Taylor 分享了他令人印象深刻的职业生涯中的一些轶事，并提供了对领导力的见解。他强调，像扎克伯格、贝尼奥夫和佩奇这样的伟大领导者都具有不懈的驱动力、长期思维和沟通愿景的能力。

This is a summary of Harry Stebbings' interview with Bret Taylor on the 20 VC podcast. Taylor, a Silicon Valley OG with an impressive CV including Google Maps, CTO of Facebook, founder of Quip, and now CEO of Sierra (a conversational AI platform), shares his insights on the current AI landscape, the future of agents, and leadership. Taylor believes the current AI boom is indeed a bubble, but one that will "rhyme" with the dot-com bubble. He argues that while there's excess, the long-term impact will be transformative, leading to trillion-dollar companies and industry-defining enterprise software. He emphasizes that the core of investment should be around impact and generational companies as opposed to just multiples. He dismisses the idea that large models will subsume all software, drawing an analogy to the cloud market where infrastructure, toolmakers, and SaaS solutions coexist. He argues that companies prefer buying solutions over building their own and feels that this sentiment will hold for the area of AI. He shares his vision of his company, Sierra, will enable companies to build their own customer facing agents to resolve a wide-range of issues. While acknowledging a short-term spike in AI services companies due to a lack of out-of-the-box solutions, he anticipates a decline as SaaS solutions mature. However, he predicts that AI services companies will have an important stake for change management, to assist enterprises adjust to this large movement. Taylor recognizes the rapid commoditization of AI models, distinguishing between "foundation models" (commodity) and "frontier models" (best-in-class). He is skeptical of most companies pre-training models, deeming it capital intensive unless it's for AGI research. Instead, he believes most companies should leverage existing models and fine-tune them for specific solutions. Regarding AGI, Taylor advocates for responsible iterative deployment to learn about societal impact, access, and safety. He highlights the importance of access to AI and how products like ChatGPT provide broad access to AI for all. He finds the enterprise division within OpenAI as a great way for developers to build value around the technology. Discussing the challenges of sustainable AI business models, Taylor emphasizes focusing on product-market fit before incurring substantial training costs. He's optimistic about decreasing inference costs due to technological advancements. Addressing the investment of larger hyperscalers, Taylor thinks it is rational for companies to invest in their own infrastructure because there's too much value at stake to be outside of the game. The investment has large implications for the development and utilization of consumer products. He sees open source models and proprietary solution existing side-by-side as a net positive for the ecosystem. He also emphasizes that investment in building models will eventually lead to revenue generating opportunities. Moving on to agents, Taylor positions Sierra as enabling companies to build branded customer-facing AI agents, believing that conversation is the future of human-computer interaction. He uses an analogy to smartphones and touchscreens to support the claim that conversations with AI agents are inevitable. He discusses the technical challenge of creating determinism in generative AI, balancing creativity with business rules. He emphasizes the importance of defining both goals and guardrails for AI agents. This poses a unique design challenge, balancing agency and control, with roles like AI architects emerging to manage conversation design. Ultimately, Sierra is enabling companies to build their own branded AI agent for all their clients use cases. Finally, Taylor shares a few anecdotes from his impressive career, offering insights on leadership. He emphasizes relentless drive, long-term thinking, and the ability to communicate vision as key traits of great leaders like Zuckerberg, Benioff, and Page.