Matei Zaharia discusses the convergence of coding agents and custom agents into a unified platform called Omnigents, highlighting the need for portability, security, and collaboration.
要点 · TL;DR
AI 智能体和模型将通过利用数据重写传统软件。 AI agents and models will rewrite traditional software by leveraging data.
开源推动生态系统增长和网络效应。 Open source drives ecosystem growth and network effects.
统一存储层无需单一查询引擎即可解决 HTAP 问题。 Unified storage layer solves HTAP without a single query engine.
核心观点 · Key points
将数据放在正确的位置是关键；AI模型和智能体已经足够好，可以重写传统软件。 Getting data in the right place is key; AI models and agents are good enough to rewrite traditional software.
开源促进生态系统和网络效应，正如Spark和OmniGen所示。 Open source fosters ecosystem and network effects, as seen with Spark and OmniGen.
统一存储层（L-TAP）解决了HTAP挑战，无需单一查询引擎。 Unified storage layer (L-TAP) solves HTAP challenges without a single query engine.
安全性和治理对企业采用至关重要，需要上下文策略。 Security and governance are critical for enterprise adoption, requiring contextual policies.
随着基础模型和强化学习的改进，定制模型将变得越来越容易。 Custom model customization will become easier over time, driven by better base models and RL.
反共识 · Contrarian takes
传统软件将通过将AGI置于数据之上来重写，而不是构建新堆栈。 Traditional software will be rewritten by slapping AGI on top of data, not by building new stacks.
正确的HTAP不是单一引擎，而是统一存储；查询引擎可以保持分离。 HTAP done right is not a single engine but unified storage; query engines can remain separate.
向量数据库不应成为独立类别；它们只是通用存储。 Vector databases should not be a separate category; they are just general storage.
过度适配少数客户比试图包罗万象更好；这能带来真正的产品。 Overfitting to a few customers is better than boiling the ocean; it leads to real products.
企业偏好开放格式并避免锁定，这与专有方法相反。 Enterprises prefer open formats and avoid lock-in, contrary to proprietary approaches.
本期章节 · Chapters（共 29）
开场与赞助商消息Introduction and Sponsor Message
Databricks产品发布与OmnigentsDatabricks Product Launches and Omnigents
架构并行与开放共享Architecture Parallels and Open Sharing
互操作性与Vibe编码Interoperability and Vibe Coding
开源与托管服务Open Source vs. Managed Services
Agent框架的通用APICommon API for Agent Harnesses
计算沙箱与架构Compute Sandboxing and Architecture
Databricks运营规模Scale of Databricks Operations
Omnigen：安全与控制Omnigen: Security and Control
Token最大化与安全Token maxing and security
参与OmnigenGetting involved with Omnigen
创业机会Startup opportunities
OLTP与数据库历史OLTP and database history
CDC痛点CDC Pain Points
HTAP梦想与妥协HTAP Dream and Compromises
L-TAP：统一存储L-TAP: Unifying Storage
L-TAP如何成为可能How L-TAP Became Possible
从辩论到原型文化Debate to Prototype Culture
企业客户与技术公司客户Enterprise vs. Tech Company Customers
梦想引擎愿景The Dream Engine Vision
性能关键维度Key dimensions for performance
渐进演进与硬切换Incremental evolution vs hard cut
Databricks与Snowflake关键差异Databricks vs Snowflake: key differences
竞争理念：Snowflake vs DatabricksCompeting philosophies: Snowflake vs Databricks
Ali的角色与Mosaic收购Ali's role and Mosaic acquisition
专用模型与通用模型Specialized models vs general models
数据价值与模型定制Data value and model customization
数据中心的AI与软件重写Data-Centric AI and Rewriting Software