<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>by Arlo</title><description>关于智能系统、市场、复杂性与自由的笔记。Notes on intelligent systems, markets, complexity, and freedom.</description><link>https://byarlo.org/</link><language>zh-cn</language><item><title>What Would It Mean for Chinese AI Chips to Catch NVIDIA?</title><link>https://byarlo.org/posts/chinese-ai-chips/</link><guid isPermaLink="true">https://byarlo.org/posts/chinese-ai-chips/</guid><description>A structural look at what &apos;catching up&apos; in AI chips actually requires—and why the answer depends heavily on how you define the benchmark.</description><pubDate>Fri, 20 Mar 2026 00:00:00 GMT</pubDate></item><item><title>工作消耗的不是时间，而是注意力</title><link>https://byarlo.org/posts/attention-not-time/</link><guid isPermaLink="true">https://byarlo.org/posts/attention-not-time/</guid><description>时间是可以规划的，注意力却不是。区分这两者，可能是理解知识工作生产力的核心问题。</description><pubDate>Sat, 14 Feb 2026 00:00:00 GMT</pubDate></item><item><title>为什么 FLOPS 越来越难以预测大模型的真实推理性能</title><link>https://byarlo.org/posts/flops-reasoning/</link><guid isPermaLink="true">https://byarlo.org/posts/flops-reasoning/</guid><description>算力不再是推理能力的充分代理指标。本文梳理几个让 FLOPS 预测失效的机制：稀疏激活、KV cache、推测解码与 test-time compute。</description><pubDate>Sat, 10 Jan 2026 00:00:00 GMT</pubDate></item></channel></rss>