Ibrahim Haddad 讨论了在关键任务中确定性代码比 LLM 更重要、中国 AI 实验室令人惊讶的开放性，以及 Linux 基金会在 AI 民主化中的作用。
Ibrahim Haddad discusses the importance of deterministic code over LLMs for critical tasks, the surprising openness of Chinese AI labs, and the Linux Foundation's role in democratizing AI.
要点 · TL;DR
中国 AI 实验室开放创新而非抄袭；出口管制反而催生了效率提升。 Chinese AI labs innovate openly, not copy; export controls spurred efficiency.
企业应微调现有开源模型而非预训练，以实现成本效益。 Enterprises should fine-tune open models, not pre-train, for cost-effective AI.
自主 AI 安全需要确定性代码层，而非仅靠模型护栏。 Agentic AI safety needs deterministic code layers, not just model guardrails.
核心观点 · Key points
中国 AI 实验室在创新而非抄袭；像 DeepSeek 的 GRPO 这样的贡献已被全球采用。 Chinese AI labs are innovating, not copying; their contributions like DeepSeek's GRPO are used globally.
出口管制迫使中国创新，催生了更高效的模型和本土芯片发展。 Export controls forced China to innovate, leading to more efficient models and domestic silicon development.
企业应避免预训练自己的模型，而应微调现有开源模型以节省成本。 Enterprises should avoid pre-training their own models; instead, fine-tune existing open models for cost efficiency.
智能体 AI 的安全性必须内建于框架和协议中，而不仅仅是模型，以防止故障。 Safety in agentic AI must be built into frameworks and protocols, not just models, to prevent failures.
LLM 推理通过模式匹配模仿人类推理，而非真正理解；它有用但不等于人类推理。 LLM reasoning mimics human reasoning through pattern matching, not true understanding; it's useful but not equivalent.
采用宽松许可证（Apache 2.0、MIT）的开源模型是 AI 访问民主化的关键。 Open source models with permissive licenses (Apache 2.0, MIT) are key for democratizing AI access.
反共识 · Contrarian takes
中国实验室在研究中比美国实验室更开放，尽管有商业利益仍广泛分享创新。 Chinese labs are more open about their research than US labs, sharing innovations broadly despite commercial interests.
DeepSeek 的成功源于谦逊和集体解决问题，而非个人野心或高薪。 DeepSeek's success is due to humility and communal problem-solving, not individual ambition or high salaries.
AI 并非总是最佳方案；经典 ML 模型在欺诈检测等任务上常优于 LLM。 AI is not always the best solution; classical ML models often outperform LLMs for tasks like fraud detection.
中国转向闭源模型（如 MiniMax）是 IPO 压力驱动，而非技术必要。 The shift to closed models in China (e.g., MiniMax) is driven by IPO pressure, not technical necessity.
世界模型对机器人和具身 AI 更相关，而非仅通过语言实现 AGI。 World models are more relevant for robotics and embodied AI than for achieving AGI through language alone.
智能体 AI 的安全性需要确定性代码层，而不仅仅是模型级护栏，以防止滥用。 Agentic AI's safety requires deterministic code layers, not just model-level guardrails, to prevent misuse.
本期章节 · Chapters（共 30）
引言与Linux基金会角色Introduction and Role at Linux Foundation
背景与经验Background and Experience
中国之行观察Observations from China Trip
AI局限与确定性代码AI Limitations and Deterministic Code
中国开源模型下载量Chinese Open Source Model Downloads
实验室文化差异Lab Culture Differences
中美客服对比Customer service contrast: US vs China
对美蒸馏备忘录的反应Reaction to US distillation memo
揭穿叙事：中国抄袭与出口管制阻碍发展Debunking narratives: China copies and export controls slow development
出口管制与中国创新Export controls and Chinese innovation
开源许可与开放模型框架Open Source Licensing and Open Model Framework
Linux基金会热门项目Hottest Projects under Linux Foundation
Linux内核中的AIAI in the Linux Kernel
开源中的AI与PR过载AI in Open Source and PR Overload
AI作为界面：语音vs UIAI as Interface: Voice vs. UI
Agentic AI的安全性Safety in Agentic AI
LLM推理：真假难辨LLM Reasoning: Fake or Real?
LLM推理与拟人化LLM reasoning and personification
世界模型与Yann LeCunWorld models and Yann LeCun
机器人产业泡沫Robotics industry bubble
企业AI采纳误区Enterprise AI adoption mistakes
企业AI部署常见错误Common mistakes in enterprise AI deployment
构建模块与实验性质Building blocks and experimental nature
企业AI采纳策略AI adoption strategy for businesses
用例与差异化vs生产力Use Cases and Differentiation vs. Productivity
Agentic AI与信任Agentic AI and Trust
标准化Agentic接口Standardizing Agentic Interfaces
Linux基金会的AI使用AI Usage at Linux Foundation
业务与工程团队使用AIBusiness and Engineering Teams Using AI