cover of episode AGI 2024特辑06 | 对话杨植麟:聚焦 Kimi 一个产品,聚焦生产力场景

AGI 2024特辑06 | 对话杨植麟:聚焦 Kimi 一个产品,聚焦生产力场景

2024/8/26
logo of podcast AI局内人 | AGI Insider

AI局内人 | AGI Insider

AI Deep Dive AI Insights AI Chapters Transcript
People
杨植麟
Topics
杨植麟:月之暗面短期内不会推出Kimi之外的第二款产品,长期战略是将AGI产品最终形态定义为用户的长期伙伴,而非简单的工具。这需要模型具备强大的实用性、可靠性以及与用户建立长期信任的能力。 Kimi选择聚焦生产力赛道,而非娱乐赛道,主要基于以下几个原因:首先,生产力场景更能快速提升模型智商,与公司目标更紧密结合;其次,美国市场发展趋势也支持这一选择;最后,生产力领域仍有巨大提升空间,AI能够带来巨大的新面量。 Kimi成功的关键在于专注于生产力场景,并从底层技术出发,持续优化模型,以解锁更多场景。这需要在技术和产品上聚焦,做出取舍,将有限的资源投入到最能产生价值的地方。 长文本处理(长推理)是实现更复杂任务的关键,但并非唯一条件,还需要更强的推理能力。长推理能力提升了智能增益和用户指令与服务价值交换的比例。长推理能力的应用场景会逐步扩展,从简单的阅读理解到更复杂的规划和多步任务执行。 多模态能力对通用人工智能至关重要,Kimi未来会整合多模态能力,提升用户体验和应用场景。视频生成能力对Kimi的重要性在于其对模型理解能力的提升,而非仅仅是其本身的价值。 长推理的成本未来会显著下降,这得益于工程优化和架构改进。长推理能力将使AI能够执行更复杂的多步任务,并应用于更多场景。长推理能力的应用场景仍在不断探索中,并与用户共同创造新的应用方式。 大模型时代,创业公司需要在组织结构和人才招聘方面进行创新,以适应快速变化的技术环境。创业过程中,快速反应比远期预测更重要。创业状态更像爬山,因为可以清晰地看到进步,并根据实际情况灵活调整。

Deep Dive

Key Insights

Why did Yang Zhilin choose to focus on the productivity scenario for Kimi?

Yang Zhilin chose the productivity scenario because it accelerates the improvement of intelligence. Optimizing intelligence in productivity directly enhances product performance, aligning closely with the company's mission and product roadmap. Additionally, the U.S. market's trends showed that companies focusing on productivity achieved better business scale, funding, and talent attraction.

What is the ultimate goal of Kimi according to Yang Zhilin?

The ultimate goal of Kimi is to become a long-term partner that understands the user deeply, rather than just replicating the user. It aims to be useful by handling increasingly complex tasks and establishing long-term trust and connection with users, ensuring accuracy and reliability in its interactions.

How does Yang Zhilin view the importance of long context in AI development?

Yang Zhilin views long context as essential for advancing AI capabilities, particularly for handling complex tasks that require extended reasoning and planning. He believes that long context is a necessary condition for AI to progress from simple tasks to more complex, multi-step operations, ultimately enhancing the value delivered to users.

What are Yang Zhilin's thoughts on the future cost of long context processing?

Yang Zhilin predicts a significant reduction in the cost of long context processing, potentially by an order of magnitude or more. Advances in engineering optimizations and architectural improvements, such as context caching, are expected to make long context processing more accessible and affordable, enabling broader application in various scenarios.

How does Yang Zhilin describe his entrepreneurial journey with Kimi?

Yang Zhilin describes his entrepreneurial journey as akin to climbing a mountain, where each step brings visible improvements in model capabilities and efficiency. He emphasizes the importance of focusing on a single product, Kimi, and refining it to its utmost potential, rather than diversifying into multiple products.

Chapters
Yang Zhilin, founder and CEO of Yue Zhi Anmian Kimi, explains his strategy of focusing on a single product, Kimi, to achieve excellence. He believes that the needs for entertainment and productivity are not mutually exclusive and will eventually converge within one product. The company prioritizes improving the underlying model to unlock more scenarios, focusing on productivity for faster improvements.
  • Focus on one product (Kimi) to achieve excellence.
  • Entertainment and productivity needs will converge in one product.
  • Prioritize improving underlying model to unlock more scenarios.
  • Focus on productivity for faster improvements.

Shownotes Transcript

他不会推出Kimi之外的第二款产品,至少在很长一段时间内是这样。

回望创业的这一年,杨植麟非常感慨,他告诉我们「创业的状态就像爬山」,每爬一步,模型的能力肉眼可见的变好,每一分算力产出的智能更高。

世界只需要一个Kimi,「娱乐的需求、生产力的需求,可能都没有明确的边界,它们最终都会在一个产品里,只是大家的路径会有不同。」杨植麟说,「我们希望聚焦做一个产品,然后把这个产品做到极致。」

他在 AGI Playground 2024 大会上跟我们把这些话题展开了聊、揉碎了聊:他对产品的定义,对技术路线的选择,对组织形态的思考,以及Kimi的最终目标。

听了本期播客,你就会知道。

嘉宾介绍:

杨植麟,月之暗面 Kimi 创始人 & CEO

时间线:

00:03:45 Kimi 增长这么快,到底做对了什么?

00:07:37 等等,为什么先做生产力赛道?

00:09:38 一个milestone:AI 比你干的更多

00:11:56 Kimi 的终极目标:理解你

00:17:30 长文本的本质是「长推理」

00:21:48 Kimi 会实现多模态统一吗?

00:26:49 未来,长推理的成本必定会下降!

00:35:57 对于创业公司,动态反应比预测更重要

00:42:48 杨植麟:我的创业状态更像在爬山

相关阅读:

对话杨植麟:聚焦生产力,做好Kimi这一个产品)

关于我们

这里是极客公园旗下的科技创业者社区 Founder Park 出品的 AGI 系列播客节目。

「AI局内人」,国绕 AGI 相关领域的技术发展、产品方向以及新的商业模式进行探讨和交流。

Founder Park 正在搭建 Al Native 的产品交流群,群内聚集着致力于在大模型相关领域创业的创业者、产品经理、研发工程师、对大模型技术和场景应用开发进展感兴趣的学者及投资人。

对更多内容感兴趣,欢迎关注公众号「Founder Park」。