<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Aniket Ghosh — Writing</title><description>Aniket Ghosh&apos;s writing. Research on what goes on inside neural networks, and essays on most other things. Whatever I publish, wherever it lands, gets tracked here.</description><link>https://itsme-aniketghosh.github.io/blog/</link><item><title>Linear probes tell you where quantization will hurt</title><link>https://www.lesswrong.com/posts/oJJyYDgPD95jEfvQx/linear-probes-tell-you-where-quantization-will-hurt</link><guid isPermaLink="true">https://www.lesswrong.com/posts/oJJyYDgPD95jEfvQx/linear-probes-tell-you-where-quantization-will-hurt</guid><description>Probing where a capability lives in the residual stream predicts which layers survive low-bit quantization and which ones you can&apos;t touch.</description><pubDate>Sat, 25 Jul 2026 00:00:00 GMT</pubDate><category>interpretability</category><category>quantization</category><category>mech-interp</category></item></channel></rss>