AI

从拼凑到整体渲染的PPT制作范式转移，以及如何用代码骨架、资产约束和延迟渲染构建解决风格混乱、幻觉和成本问题的Generative Kernel。

Exploring the paradigm shift from assembling slides to holistic rendering with NBP, and building a Generative Kernel workflow to handle visual consistency, hallucination, and cost.

观察到当代码成本趋近于零时，为一次性决策构建专用工具反而是最优策略。提出AI Native的本质是在信息获取成本坍塌后采用全新策略——从直觉驱动转向高分辨率的数据驱动决策。

Observes that when code cost approaches zero, building disposable tools for one-off decisions becomes optimal. Argues AI Native means adopting new strategies enabled by collapsed information costs—shifting from intuition-driven to high-resolution, data-driven decisions.

AI"偷懒"的本质是LLM输出长度限制导致的注意力分散。Wide Research通过多轻量模型并行处理子任务、主LLM汇总的方式解决，分享为Codex构建该能力的设计思路。

Why AI slacks off on large tasks: LLM output length limitations cause attention drift. Wide Research solves this by parallelizing with lightweight models, then aggregating results with a primary LLM.

记录在API因合规问题被拒后，使用视觉大模型从截图提取财务数据，实现十年手动记账流程的自动化。展示了本地模型、交叉验证和人机协作工作流如何安全处理敏感金融数据。

Documents automating a decade-long manual financial reconciliation process using vision LLMs when API access was blocked by compliance. Demonstrates using local models with screenshots, cross-validation, and human-in-the-loop workflows to process sensitive financial data.

GPT-5是产品升级而非单纯模型升级：新增reasoning_effort和verbosity参数，可控性大幅提升，让开发者能根据场景灵活权衡推理深度和回复长度。

GPT-5 is a major product upgrade, not just a model upgrade: new API parameters for reasoning_effort and verbosity enable unprecedented controllability for building AI-powered products.