HackerNews中文版

我在使用苹果的新 iOS 26 Foundation Models 框架时遇到了一个常见问题。由于训练截止日期的原因，所有主要的语言模型（如 GPT-4、Claude、Gemini）对此一无所知。为了不必等几个月进行模型更新，我花了两个小时构建自定义训练数据：研究：利用 Gemini 的深度研究，爬取所有可用的文档、论坛、GitHub 仓库、Reddit 讨论串和 YouTube 字幕，寻找“iOS 26 Foundation Models Framework”。优化：让 Claude 将所有内容重组为干净的、层次分明的 Markdown 格式，以便于语言模型的处理。实施：将其加载到 Claude Projects 中，作为自定义知识层。结果：从“我没有相关信息”变成了对前沿 API 的专家级指导。我的开发工作流程从反复试验转变为顺畅的 AI 辅助实施。这项研究非常全面，甚至引用了我前一天在苹果开发者论坛的帖子——形成了一个奇怪的递归循环，我在用我刚刚贡献的知识训练 AI。这种方法适用于任何新的框架或 API。这个模式是可预测的：每次重大发布都会产生一个临时的知识空白，而自定义训练数据可以填补这个空白。技术写作及方法论详见：https://rileygersh.medium.com/how-i-gave-claude-gemini-knowledge-of-ios-26s-foundation-models-03395d7e905c

查看原文

I ran into a common problem building with Apple's new iOS 26 Foundation Models framework. Every major LLM (GPT-4, Claude, Gemini) knows nothing about it due to training cutoffs.Rather than wait months for model updates, I spent 2 hours building custom training data:Research: Used Gemini's deep research to crawl all available docs, forums, GitHub repos, Reddit threads, YouTube transcripts. Looking for "iOS 26 Foundation Models Framework".Optimization: Had Claude restructure everything into clean, hierarchical markdown optimized for LLM ingestion.Implementation: Loaded into Claude Projects as a custom knowledge layer.Result: Went from "I don't have information about that" to expert-level guidance on bleeding-edge APIs. My development workflow shifted from trial-and-error to smooth AI-assisted implementation.The research was so thorough it even referenced my own Apple dev forum post from the day before—creating a weird recursive loop where I was training AI on knowledge I'd just contributed.This approach works for any new framework or API. The pattern is predictable: every major release creates a temporary knowledge gap that custom training data can fill.Technical writeup with methodology: https://rileygersh.medium.com/how-i-gave-claude-gemini-knowledge-of-ios-26s-foundation-models-03395d7e905c

我为过时的最先进大语言模型（SOTA LLMs）创建了一个新的训练数据集。