OpenAI 深度研究的先驱之一 Issa Falford 分享了这款使用推理和网络浏览完成多步研究任务的代理产品的起源故事。
Issa Falford, one of the pioneering minds behind OpenAI's Deep Research, shares the origin story of this agentic product that uses reasoning and web browsing to complete multi-step research tasks.
要点 · TL;DR
Deep Research 使用强化学习综合信息用于知识工作，而不仅仅是浏览。 Deep Research uses RL to synthesize information for knowledge work, not just browsing.
理想代理结合研究与行动能力，并具备强大的安全护栏。 Ideal agents combine research and action with strong safety guardrails.
强化微调数据高效且能跨任务泛化。 Reinforcement fine-tuning is data-efficient and generalizes across tasks.
核心观点 · Key points
Deep Research 使用强化学习从多个来源综合信息，适用于知识工作。 Deep Research uses reinforcement learning to synthesize information from many sources for knowledge work.
理想的智能体将结合研究和行动能力，并配备强大的安全护栏。 The ideal agent will combine research and action-taking capabilities with strong safety guardrails.
强化微调数据效率高且能跨任务泛化，但需要高质量策划的数据。 Reinforcement fine-tuning is data-efficient and can generalize across tasks, but requires high-quality curated data.
智能体记忆和上下文管理对于需要数小时或数天的长期任务至关重要。 Agent memory and context management are critical for long-horizon tasks that take hours or days.
当智能体可以访问私人数据并采取现实行动时，安全挑战会升级。 Safety challenges escalate when agents can access private data and take real-world actions.
反共识 · Contrarian takes
研究等只读任务比交易型智能体用例更有野心和实用性。 Read-only tasks like research are more ambitious and useful than transactional agent use cases.
Deep Research 的幻觉比以往任何 OpenAI 模型都少，但引用仍然至关重要。 Deep Research hallucinates less than any previous OpenAI model, but citations remain essential.
模型可以学会未明确教授的规划和智能策略，例如规避限制。 The model can learn to plan and use smart strategies not explicitly taught, like evading restrictions.
Deep Research 在编码和数据分析方面出奇地有效，而不仅仅是浏览。 Deep Research is surprisingly effective for coding and data analysis, not just browsing.
用户不应决定模型思考多久；模型应推断最佳思考时间。 Users should not have to decide how long the model thinks; the model should infer optimal thinking time.
本期章节 · Chapters（共 19）
0. 引言与起源故事Introduction and origin story
1. 选择只读任务而非事务性任务Choosing read-only tasks over transactional ones
2. 从演示到训练：数据与工具创建From demo to training: data and tool creation
3. 最喜欢的测试任务与早期使用Favorite test tasks and early usage
4. 数据与工具创建细节Details of data and tool creation
5. 当前工具与未来扩展Current Tools and Future Expansion
6. 智能体的护栏与信任Guardrails and trust in agents
7. 深度研究改进与统一智能体Deep research improvements and unified agents
8. OpenAI 的产品与研究Product vs research at OpenAI
9. 用例与专家验证Use cases and expert validation
10. 学习用户偏好与智能体记忆Learning user preferences and agent memory
11. 挑战与通往 AGI 之路Challenges and Path to AGI
12. 智能体能力的重大障碍Big Blockers to Agent Capabilities
13. 深度研究与其他模型的用例Use Cases for Deep Research vs. Other Models
14. 训练深度研究的意外成败Surprising Wins and Failures in Training Deep Research
15. 平衡思考时间与用户体验Balancing thinking time and user experience
16. 未来深度研究任务耗时数天Future deep research tasks taking days
17. 一年内智能体的惊人能力Surprising agent capabilities in a year