<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Justin T Chiu</title><description>Justin Chiu</description><link>https://justinchiu.netlify.app/</link><item><title>Character prefix completion</title><link>https://justinchiu.netlify.app/blog/completion/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/completion/</guid><description>Sampling given character prefixes</description><pubDate>Thu, 09 Jan 2025 00:00:00 GMT</pubDate></item><item><title>Differentiating through optimization with the IFT</title><link>https://justinchiu.netlify.app/blog/ift/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/ift/</guid><description>Differentiating through optimization</description><pubDate>Thu, 26 Mar 2026 00:00:00 GMT</pubDate></item><item><title>Definitions of Memorization</title><link>https://justinchiu.netlify.app/blog/memorization/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/memorization/</guid><description>Information-theoretic definitions of memorization</description><pubDate>Mon, 16 Jun 2025 00:00:00 GMT</pubDate></item><item><title>RL should be the same shape as pretraining</title><link>https://justinchiu.netlify.app/blog/opinionatedrl/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/opinionatedrl/</guid><description>Token-level policy gradient is correct, average over batch tokens, and a packed batch is a replay buffer</description><pubDate>Wed, 10 Jun 2026 00:00:00 GMT</pubDate></item><item><title>Differentiating through an associative parallel scan</title><link>https://justinchiu.netlify.app/blog/pscan_diff/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/pscan_diff/</guid><description>Some calculus busywork</description><pubDate>Mon, 08 Jan 2024 00:00:00 GMT</pubDate></item><item><title>First post</title><link>https://justinchiu.netlify.app/blog/setup/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/setup/</guid><description>First post</description><pubDate>Wed, 20 Dec 2023 00:00:00 GMT</pubDate></item><item><title>SFT is bad RL</title><link>https://justinchiu.netlify.app/blog/sftrl/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/sftrl/</guid><description>RL is better than SFT</description><pubDate>Mon, 14 Jul 2025 00:00:00 GMT</pubDate></item><item><title>What objective is STaR optimizing?</title><link>https://justinchiu.netlify.app/blog/star/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/star/</guid><description>Everything is a latent variable model</description><pubDate>Sat, 27 Apr 2024 00:00:00 GMT</pubDate></item><item><title>Sum-and-Sample RL: Replay Buffers</title><link>https://justinchiu.netlify.app/blog/sumsamplerl/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/sumsamplerl/</guid><description>The math behind replay buffers</description><pubDate>Sat, 09 Aug 2025 00:00:00 GMT</pubDate></item><item><title>Tree RL</title><link>https://justinchiu.netlify.app/blog/treerl/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/treerl/</guid><description>An unbiased policy gradient estimator for branching rollouts</description><pubDate>Tue, 13 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Variational Autoencoders from First Principles</title><link>https://justinchiu.netlify.app/blog/vae/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/vae/</guid><description>Brute-forcing VAE math</description><pubDate>Mon, 04 Aug 2025 00:00:00 GMT</pubDate></item><item><title>Variational inference: Ising models</title><link>https://justinchiu.netlify.app/blog/vi/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/vi/</guid><description>Variational inference for a simple Ising model</description><pubDate>Sat, 30 Dec 2023 00:00:00 GMT</pubDate></item><item><title>Why Routing Replay in MoE RL is Necessary</title><link>https://justinchiu.netlify.app/blog/weakening/</link><guid isPermaLink="true">https://justinchiu.netlify.app/blog/weakening/</guid><description>Unf***ing Your RL Gradient Estimator</description><pubDate>Wed, 03 Dec 2025 00:00:00 GMT</pubDate></item></channel></rss>