前 DeepMind 研究员 Misha 讨论 Reflection AI 通过编程和企业 AI 构建组织超级智能的使命，并推出代码研究代理作为首个里程碑。
Former DeepMind researcher Misha discusses Reflection AI's mission to build organizational superintelligence through coding and enterprise AI, launching a code research agent as the first milestone.
要点 · TL;DR
超级智能将是多个专门系统的集合，而非单一模型。 Superintelligence will be a collection of specialized systems, not a single monolithic model.
编程是超级智能完备问题，解决它就能解锁所有能力。 Coding is a superintelligence-complete problem; solving it unlocks all capabilities.
强化学习产生不均衡的能力，而非全面进步。 Reinforcement learning creates jagged capabilities, not general improvement.
核心观点 · Key points
超级智能将是多种专用智能的集合，而非单一的巨型模型。 Super intelligence will be a collection of specialized intelligences, not a single monolithic model.
编程是超级智能完备问题；解决它就能获得超级智能所需的所有能力。 Coding is a super intelligence-complete problem; solving it yields all capabilities needed for super intelligence.
超级智能的关键构建模块是深度神经网络、Transformer、互联网规模数据和强化学习。 The key building blocks for super intelligence are deep neural networks, transformers, internet-scale data, and reinforcement learning.
组织级超级智能需要深刻理解整个组织，而不仅仅是代码生成。 Organizational super intelligence requires deep comprehension of the entire organization, not just code generation.
强化学习使模型变得有棱角——擅长特定任务，而非全面变强。 Reinforcement learning makes models jagged—good at specific tasks, not generally better at everything.
反共识 · Contrarian takes
通用超级智能将来自多个专用系统的集合，而非单个实验室的模型。 General super intelligence will emerge from a plurality of specialized systems, not a single lab's model.
编程对 LLM 来说比图形界面更直观，因为它们是在互联网代码上训练的，而非鼠标移动。 Coding is more intuitive for LLMs than GUIs because they were trained on internet code, not mouse movements.
RAG 是原始的，在复杂理解中会失败；需要智能体搜索和神经检索。 RAG is primitive and fails for complex comprehension; agentic search and neural retrieval are needed.
SWE-bench 等基准与现实工程脱节；高分并不意味着有用的智能体。 Benchmarks like SWE-bench misalign with real-world engineering; high scores don't mean useful agents.
苦涩的教训是关于用算力和搜索进行扩展，而非关于泛化。 The bitter lesson is about scaling with compute and search, not about generalization.
本期章节 · Chapters（共 18）
超级智能的定义与介绍Introduction and Definition of Superintelligence
超级智能作为组织预言机Superintelligence as an organizational oracle