<feed xmlns="http://www.w3.org/2005/Atom"> <id>https://mradassaad.github.io/</id><title>Assaad's Blog</title><subtitle>I write about recent things I learned and found interesting. I obtain a Ph.D. in eco-hydrology and environmental science from Duke University in 2020. I now work as an ML Engineer at Capital One Shopping.</subtitle> <updated>2026-05-04T00:21:14-04:00</updated> <author> <name>Assaad Mrad</name> <uri>https://mradassaad.github.io/</uri> </author><link rel="self" type="application/atom+xml" href="https://mradassaad.github.io/feed.xml"/><link rel="alternate" type="text/html" hreflang="en" href="https://mradassaad.github.io/"/> <generator uri="https://jekyllrb.com/" version="4.4.1">Jekyll</generator> <rights> © 2026 Assaad Mrad </rights> <icon>/assets/img/favicons/favicon.ico</icon> <logo>/assets/img/favicons/favicon-96x96.png</logo> <entry><title>Why SSMs Struggle in Parameter Golf: A Structural Analysis at 25M Parameters</title><link href="https://mradassaad.github.io/posts/why-ssms-struggle-in-parameter-golf/" rel="alternate" type="text/html" title="Why SSMs Struggle in Parameter Golf: A Structural Analysis at 25M Parameters" /><published>2026-05-03T00:00:00-04:00</published> <updated>2026-05-04T00:11:39-04:00</updated> <id>https://mradassaad.github.io/posts/why-ssms-struggle-in-parameter-golf/</id> <content type="text/html" src="https://mradassaad.github.io/posts/why-ssms-struggle-in-parameter-golf/" /> <author> <name>Assaad Mrad</name> </author> <category term="ML Systems" /> <category term="Research" /> <summary>TL;DR Over ~3 weeks of experimentation on an SSM-based submission to OpenAI’s Parameter Golf, I converged on a legal Mamba-3 hybrid at post-quant+TTT 1.1456 bpb, the best SSM submission in the 16MB track. Despite this, a persistent gap to the transformer SOTA remained. The key contribution of this writeup is not the submission itself but two structural handicaps that empirically cap SSMs in th...</summary> </entry> </feed>
