与市面上多数"通用Agent"不同，OpenAI 的 Deep Research 从诞生那一刻起就被锁定在一件事上—— 通过强化 学习，将搜索、浏览、筛选与整合信息的能力内化为模型的原生技能，直接训练进参数里，而不是仅靠 Prompt工程和外部工程组合 。 那么，OpenAI 是如何把这套复杂技能训练进参数里的？他们在数据筹备、强化微调、安全与记忆管理上又摸 索出了哪些最佳实践？ OpenAI Deep Research团队核心成员Isa Fulford最近在一个访谈中做了分享： 我们认为这个访谈提供了一个透视 OpenAI 构建旗舰智能体 Deep Research 的独特视角，并提供了一些开发实 践经验，因此锦秋基金（ 微信公号锦秋集ID：jqcapital）对本文进行了编译。 01 Deep Research 的起源与目标 OpenAI 团队在强化学习算法刚刚显露锋芒时，放弃了订汉堡、订花那条看似容易衡量的交易型赛道， 转而攻克浏览与知识整合——他们认为整合知识是AGI 必不可少的前置技能， 也因为"纯读取"比"直接 下单"更安全。 数据的质量比数量更重要。 Deep Research 倾向"小而准"： ...

Core Insights - The article emphasizes the rapid acceptance and integration of AI models in the automotive industry, particularly focusing on the development of intelligent agents and their applications in vehicles [2][4][7]. Group 1: Current Trends and Developments - All major manufacturers have reached a consensus on the application of agents in vehicles, marking a significant shift in the industry's approach to AI technology [4][7]. - The acceptance speed of large model technology by manufacturers has exceeded expectations, with a clear consensus forming among mainstream automakers by early 2024 [8]. - The focus of applications has shifted towards intelligent voice enhancement, multimodal interaction breakthroughs, and the integration of visual foundational models in intelligent driving [8][9]. Group 2: Challenges and Technical Bottlenecks - Key challenges include high inference latency, online inference costs, and the need for significant development to adapt existing hardware for large models [10][12][16]. - Data collection across the vehicle remains difficult due to the current centralized architecture, which leads to inefficiencies in data transmission and limits model training [11][12]. - The existing chips are not designed for large models, leading to computational bottlenecks and challenges in deploying models effectively in vehicles [12][16]. Group 3: Core Capabilities of AI Agents - AI agents are expected to autonomously complete tasks, significantly enhancing user experience compared to traditional assistants [18][20]. - The agents exhibit multimodal perception and understanding, enabling them to recognize various environmental factors and user states [19][22]. - The interaction style has shifted towards voice-driven commands, reducing reliance on complex app interfaces [20][22]. Group 4: Future Directions and Integration - The future of automotive AI will focus on creating a unified AI model that supports both cabin interaction and intelligent driving functions, leading to a more integrated vehicle experience [9][68]. - The development of a central computing architecture will facilitate deeper information sharing and functional collaboration between cabin systems and intelligent driving systems [67][68]. - The industry is moving towards an AI-defined vehicle paradigm, where AI will reshape the entire automotive ecosystem from design to service delivery [69][70].