<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>BPE on Thamme Gowda</title><link>https://gowda.ai/tags/bpe/</link><description>Recent content in BPE on Thamme Gowda</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Mon, 30 Mar 2026 20:30:00 +0000</lastBuildDate><atom:link href="https://gowda.ai/tags/bpe/index.xml" rel="self" type="application/rss+xml"/><item><title>From O(N) to O(log N): A Faster BPE Training Algorithm, Buried and Rediscovered</title><link>https://gowda.ai/posts/2026/03/faster-bpe-learn/</link><pubDate>Mon, 30 Mar 2026 20:30:00 +0000</pubDate><guid>https://gowda.ai/posts/2026/03/faster-bpe-learn/</guid><description>I wrote a fast BPE training algorithm in 2020, buried it in a Python codebase, and forgot about it. Five years later, I rewrote it in C++ and benchmarked it: up to 11× faster than SentencePiece. The trick? A max-heap with lazy deletion instead of periodic linear scans.</description></item></channel></rss>