2025：企业编码之年

2025: The Year of Enterprise Coding

奥利维耶·戈德芒 Olivier Godement · Unsupervised Learning · 2025-12-10 · 约 58 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

Olivier Godement 探讨 AI 模型的演进、企业应用以及对科学研究的影响。

Olivier Godement discusses the evolution of AI models, enterprise adoption, and the impact on scientific research.

要点 · TL;DR

• 2025 年企业编码将开始实质性采用。
  Enterprise coding adoption will meaningfully start in 2025.
• 科学和药物发现是被低估的 AI 用例。
  Science and drug discovery are underhyped AI use cases.
• 降低成本是释放巨大 AI 需求的关键。
  Cost reduction is key to unlocking massive AI demand.

核心观点 · Key points

• 2025 年是企业级编程之年，有意义的采用已经开始。
  2025 is the year of coding in the enterprise, with meaningful adoption starting.
• 科学和药物发现是被低估的 AI 用例，潜力巨大。
  Science and drug discovery are underhyped AI use cases with huge potential.
• 降低成本是释放大量未开发 AI 需求的关键。
  Cost reduction is critical to unlock massive untapped demand for AI.
• 企业需要严格的评估和数据基础设施才能成功应用 AI。
  Enterprises need rigorous evaluations and data infrastructure for AI success.
• 模型提供商将同时提供模型和框架作为标准蓝图。
  Model providers will offer both models and harnesses as standard blueprints.

反共识 · Contrarian takes

• 模型疲劳真实存在；对于非平凡用例，热切换模型已不再有效。
  Model fatigue is real; hot-swapping models no longer works for non-trivial use cases.
• 模型并非一切；框架和数据同样重要。
  The model is not everything; harnesses and data are equally important.
• 大多数企业知识存在于人脑中，而非文档中，这使得评估变得困难。
  Most enterprise knowledge is in people's brains, not documented, making evals hard.
• 语音 AI 尚未通过图灵测试；自然度是下一个前沿。
  Voice AI hasn't passed the Turing test yet; naturalness is the next frontier.
• 持续学习（根据反馈更新权重）将是智能体的重大突破。
  Continuous learning (weight updates from feedback) will be a major unlock for agents.

本期章节 · Chapters（共 30）

• 引言与2025预测 Introduction and 2025 predictions
• 科学加速的迹象 Glimmers of scientific acceleration
• 听众问答介绍 Introduction to listener Q&A
• OpenAI角色与生态伙伴 OpenAI's role vs ecosystem partners
• 企业反馈与ChatGPT通用界面 Enterprise feedback and ChatGPT as universal interface
• 当前能力与未来飞跃 Current capabilities vs future leaps
• 下一代模型前沿 Next model frontiers
• 智能体实习生类比 Agent as Intern Analogy
• 产品市场契合类别 Product-Market Fit Categories
• 未来领域与企业采用 Future Domains and Enterprise Adoption
• 跨用例的脚手架模式 Scaffolding Patterns Across Use Cases
• 智能体架构与标准化 Agent architecture and standardization
• 成本作为限制因素 Cost as a limiting factor
• 模型调优的阶跃变化 Step change in model tuning efficacy
• 多数企业会用RL吗？ Will most enterprises use RL?
• 对初创公司的影响 Impact on startups
• 模型选择的关键因素 Key factors for model choice
• 模型疲劳与评估挑战 Model Fatigue and Evaluation Challenges
• 语音AI与图灵测试 Voice AI and the Turing Test
• Codex与软件工程领域 Codex and the Software Engineering Space
• 企业采用智能体编码工具 Enterprise Adoption of Agentic Coding Tools
• 开源框架与行业演进 Open-sourcing the harness and industry evolution
• AI新手企业速查表 Cheat sheet for enterprises new to AI
• 企业AI挑战：隐性知识与评估 Challenges in Enterprise AI: Tacit Knowledge and Evals
• Sora API的企业用例 Enterprise Use Cases for Sora API
• 快问快答：AI中过度炒作与低估 Quickfire: Overhyped and Underhyped in AI
• 观念转变：模型非万能，框架与数据重要 Changed Mind: Model is Not Everything, Harnesses and Data Matter
• 反思OpenAI的错误 Reflecting on OpenAI's Mistakes
• 反思过往失败与实验 Reflections on past failures and experiments
• 对Gemini 3的看法 Thoughts on Gemini 3

阅读全文双语转录 →