<?xml version="1.0" encoding="UTF-8" ?>
<rss version="2.0">
<channel>
  <title>Matt Fin's Web Log</title>
  <link>https://mattf1n.github.io</link>
<<<<<<< HEAD
  <lastBuildDate>Thu, 26 Mar 2026 00:00:00 +0000</lastBuildDate>
  <item>
    <title>Blocking the Cisco Secure Client Socket Filter on macOS</title>
    <link>https://mattf1n.github.io/posts/2026-03-26-blocking-cisco-socket-filter-on-macos.html</link>
    <guid>https://mattf1n.github.io/posts/2026-03-26-blocking-cisco-socket-filter-on-macos.html</guid>
    <pubDate>Thu, 26 Mar 2026 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>The Cisco Secure Client Socket Filter installed by the Cisco VPN
client on my laptop (VPN is necessary to connect to USC compute cluster)
was causing problems: sometimes when I started up my laptop I couldn’t
access the internet even when…</p>]]></description>
=======
  <lastBuildDate>Fri, 13 Mar 2026 00:00:00 +0000</lastBuildDate>
  <item>
    <title>My research AI coding workflow</title>
    <link>https://mattf1n.github.io/posts/2026-03-13-my-research-ai-coding-workflow.html</link>
    <guid>https://mattf1n.github.io/posts/2026-03-13-my-research-ai-coding-workflow.html</guid>
    <pubDate>Fri, 13 Mar 2026 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>My eyes generally glaze over reading “how I use AI” posts, but enough
of my colleagues have expressed interest that I think this would be
worth sharing. My overall strategy for coding these days is to have
OpenAI’s Codex agent write and…</p>]]></description>
>>>>>>> 1776c23a60a83e77d1042194aba24ff41895eda4
  </item>
  <item>
    <title>Better math fonts for AI venue submissions</title>
    <link>https://mattf1n.github.io/posts/2026-02-24-better-math-fonts-for-ai-venue-submissions.html</link>
    <guid>https://mattf1n.github.io/posts/2026-02-24-better-math-fonts-for-ai-venue-submissions.html</guid>
    <pubDate>Tue, 24 Feb 2026 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>Most AI conference venues have Times (or Times New Roman) as their
font for submissions. Unfortunately the math in the templates is set in
computer modern (or latin modern, not sure). Improve your paper’s
aesthetics by using Times for the…</p>]]></description>
  </item>
  <item>
    <title>Some papers that caught my attention at NeurIPS 2025</title>
    <link>https://mattf1n.github.io/posts/2025-12-08-papers-that-caught-my-attention-neurips-2025.html</link>
    <guid>https://mattf1n.github.io/posts/2025-12-08-papers-that-caught-my-attention-neurips-2025.html</guid>
    <pubDate>Mon, 08 Dec 2025 00:00:00 +0000</pubDate>
    <description><![CDATA[<p><a href="https://arxiv.org/abs/2506.19004">Broken Tokens? Your
Language Model can Secretly Handle Non-Canonical Tokenizations</a> shows
that language models are fine with however you tokenize your doc. Can we
hide information in the tokenization…</p>]]></description>
  </item>
  <item>
    <title>Do language models lexicalize phrases?</title>
    <link>https://mattf1n.github.io/posts/2025-04-24-lexicalization-questions.html</link>
    <guid>https://mattf1n.github.io/posts/2025-04-24-lexicalization-questions.html</guid>
    <pubDate>Thu, 24 Apr 2025 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>Do langauge models “lexicalize” certain multi-token words or phrases
(i.e., treat them as atomic units)? How would we measure lexicalization
in LMs?</p>]]></description>
  </item>
  <item>
    <title>LoRA weight decay thoughts</title>
    <link>https://mattf1n.github.io/posts/2025-04-22-lora-weight-decay-thoughts.html</link>
    <guid>https://mattf1n.github.io/posts/2025-04-22-lora-weight-decay-thoughts.html</guid>
    <pubDate>Tue, 22 Apr 2025 00:00:00 +0000</pubDate>
    <description><![CDATA[<p><a href="https://irhum.github.io/blog/lorawd/">Irhum makes an
observation</a> that I have thought about from time to time:
LoRA-adapted model weights decay to the original model during training
(not 0). How much of LoRA’s success could be attributed…</p>]]></description>
  </item>
  <item>
    <title>Some reading on recursive reasoning transformers</title>
    <link>https://mattf1n.github.io/posts/2025-04-04-recursive-reasoning-transformers.html</link>
    <guid>https://mattf1n.github.io/posts/2025-04-04-recursive-reasoning-transformers.html</guid>
    <pubDate>Fri, 04 Apr 2025 00:00:00 +0000</pubDate>
    <description><![CDATA[<figure>
<img src="../files/recurrent-reasoning.svg"
alt="My proposed latent reasoning architecture" />
<figcaption aria-hidden="true">My proposed latent reasoning
architecture</figcaption>
</figure>]]></description>
  </item>
  <item>
    <title>LLM have built-in MACs</title>
    <link>https://mattf1n.github.io/posts/2025-04-03-llm-have-built-in-macs.html</link>
    <guid>https://mattf1n.github.io/posts/2025-04-03-llm-have-built-in-macs.html</guid>
    <pubDate>Thu, 03 Apr 2025 00:00:00 +0000</pubDate>
    <description><![CDATA[<p><a href="pico.sh">pico.sh</a> seems like a cool, minimal effort tool
for blogging, but it doesn’t support math, so I’m going to roll my
own.</p>]]></description>
  </item>
  <item>
    <title>Bug in my torch code, random links.</title>
    <link>https://mattf1n.github.io/posts/2024-12-17-bug-in-torch-code.html</link>
    <guid>https://mattf1n.github.io/posts/2024-12-17-bug-in-torch-code.html</guid>
    <pubDate>Tue, 17 Dec 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>Torch code bug that took me an hour to fix: I wrapped a function in
<code>@torch.inference_mode</code> which called another function which
called another function that was trying to call
<code>torch.backward</code>.</p>]]></description>
  </item>
  <item>
    <title>Hello world, MathML in NetNewsWire</title>
    <link>https://mattf1n.github.io/posts/2024-12-05-hello-world-mathml-netnewswire.html</link>
    <guid>https://mattf1n.github.io/posts/2024-12-05-hello-world-mathml-netnewswire.html</guid>
    <pubDate>Thu, 05 Dec 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>I created an RSS feed for my website. I did it by hand, <a
href="https://www.w3schools.com/xml/xml_rss.asp">this template</a>. I
don’t use any frameworks for maintaining my website, I just write
everything by hand. I don’t change my website often…</p>]]></description>
  </item>
  <item>
    <title>The &quot;Right Way&quot; to Ensemble Language Models</title>
    <link>https://mattf1n.github.io/posts/2024-10-21-right-way-to-ensemble-language-models.html</link>
    <guid>https://mattf1n.github.io/posts/2024-10-21-right-way-to-ensemble-language-models.html</guid>
    <pubDate>Mon, 21 Oct 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>Suppose you have <span class="math inline"><em>n</em></span> langauge
models with embedding size <span class="math inline"><em>d</em></span>,
vocabulary size <span class="math inline"><em>v</em></span>, and softmax
matrices <span
class="math inline"><strong>W</strong><sub>1</sub>, <strong>W</strong><sub>2</sub>, …, <strong>W</strong><sub><em>n</em></sub> ∈ ℝ<sup><em>v</em> × <em>d</em></sup></span>
and you want to sample from them as an ensemble. One…</p>]]></description>
  </item>
  <item>
    <title>Research Interest Demo</title>
    <link>https://mattf1n.github.io/posts/2024-10-21-research-interest-demo.html</link>
    <guid>https://mattf1n.github.io/posts/2024-10-21-research-interest-demo.html</guid>
    <pubDate>Mon, 21 Oct 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>I have created a discord server for people interested in
collaborating with the lab. Email me for an invite!</p>]]></description>
  </item>
  <item>
    <title>Obtaining logprobs from an LLM API</title>
    <link>https://mattf1n.github.io/posts/2024-10-21-obtaining-logprobs-from-llm-api.html</link>
    <guid>https://mattf1n.github.io/posts/2024-10-21-obtaining-logprobs-from-llm-api.html</guid>
    <pubDate>Mon, 21 Oct 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>Many LLM APIs give top-<span class="math inline"><em>k</em></span>
logprobs in their outputs. What if we want to obtain <em>all</em> the
logprobs? Here I present two algorithms for obtaining logprobs from an
LLM API. Both of these depend on the API allowing us to add a logit bias
to…</p>]]></description>
  </item>
  <item>
    <title>A differentiable function from binary integer to one-hot representations</title>
    <link>https://mattf1n.github.io/posts/2024-10-21-differentiable-binary-to-onehot.html</link>
    <guid>https://mattf1n.github.io/posts/2024-10-21-differentiable-binary-to-onehot.html</guid>
    <pubDate>Mon, 21 Oct 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>I would like to define a differentiable function <span
class="math inline"><em>f</em> : {0, 1}<sup>log <em>v</em></sup> → {0, 1}<sup><em>v</em></sup></span>
that converts binary number representations of <span
class="math inline">log <em>v</em></span> bits into one-hot vectors.
This can be accomplished by using fuzzy logic operators to convert…</p>]]></description>
  </item>
  <item>
    <title>Deep BA Sampling</title>
    <link>https://mattf1n.github.io/posts/2024-10-21-deep-ba-sampling-extending-bat.html</link>
    <guid>https://mattf1n.github.io/posts/2024-10-21-deep-ba-sampling-extending-bat.html</guid>
    <pubDate>Mon, 21 Oct 2024 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>TL;DR: we can use <em>any</em> intermediate LM representation to
prove that a subset of next-token candidates have non-zero
probability.</p>]]></description>
  </item>
  <item>
    <title>Heavy tails and diversity in model distributions</title>
    <link>https://mattf1n.github.io/posts/2025-12-12-heavy-tails-and-diversity-in-model-distributions.html</link>
    <guid>https://mattf1n.github.io/posts/2025-12-12-heavy-tails-and-diversity-in-model-distributions.html</guid>
    <pubDate>Tue, 12 Dec 2023 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>Direct sampling from model output distributions often gives
incoherent outputs. Some have attributed this to a heavy tail, i.e., the
model assigns too much probability to low-probability tokens. My goal is
to test this hypothesis.</p>]]></description>
  </item>
  <item>
    <title>Visualizations</title>
    <link>https://mattf1n.github.io/posts/2023-10-21-visualizations-gallery.html</link>
    <guid>https://mattf1n.github.io/posts/2023-10-21-visualizations-gallery.html</guid>
    <pubDate>Sat, 21 Oct 2023 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>These are some visualizations I have made over the years, both for
academia and for fun!</p>]]></description>
  </item>
  <item>
    <title>The Softmax Function is Linear</title>
    <link>https://mattf1n.github.io/posts/2023-10-21-softmax-function-is-linear.html</link>
    <guid>https://mattf1n.github.io/posts/2023-10-21-softmax-function-is-linear.html</guid>
    <pubDate>Sat, 21 Oct 2023 00:00:00 +0000</pubDate>
    <description><![CDATA[<p><a href="https://x.com/mattf1n/status/1709997521580195963?s=20">I
asked my Twitter followers</a> if they knew that the softmax function is
linear. The result was disbelief.</p>]]></description>
  </item>
  <item>
    <title>ss.py — a CLI wrapper for the Semantic Scholar API</title>
    <link>https://mattf1n.github.io/posts/2023-03-21-ss-py-semantic-scholar-cli.html</link>
    <guid>https://mattf1n.github.io/posts/2023-03-21-ss-py-semantic-scholar-cli.html</guid>
    <pubDate>Tue, 21 Mar 2023 00:00:00 +0000</pubDate>
    <description><![CDATA[<p><code>ss.py</code> is my personal command line tool for searching and
citing academic papers via the Semantic Scholar API. <a
href="/ss.html">About page</a>. <a
href="https://github.com/mattf1n/ss">GitHub</a>.</p>]]></description>
  </item>
  <item>
    <title>Putting word count in Vim statusline for LaTeX files</title>
    <link>https://mattf1n.github.io/posts/2022-11-28-word-count-in-vim-statusline.html</link>
    <guid>https://mattf1n.github.io/posts/2022-11-28-word-count-in-vim-statusline.html</guid>
    <pubDate>Mon, 28 Nov 2022 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>This ftplugin updates the word count in the statusline on every save.
More frequent updates slow Vim down and cause random rendering
problems.</p>]]></description>
  </item>
  <item>
    <title>Configuring Zathura</title>
    <link>https://mattf1n.github.io/posts/2022-10-11-configuring-zathura.html</link>
    <guid>https://mattf1n.github.io/posts/2022-10-11-configuring-zathura.html</guid>
    <pubDate>Tue, 11 Oct 2022 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>I finally got Zathura (the pdf viewer) configured the way I want it
on MacOS. I installed using homebrew following these <a
href="https://github.com/zegervdv/homebrew-zathura">instructions</a>. I
set up an automator script to launch Zathura for me…</p>]]></description>
  </item>
  <item>
    <title>Washington State</title>
    <link>https://mattf1n.github.io/posts/2021-01-20-washington-state-relief-map.html</link>
    <guid>https://mattf1n.github.io/posts/2021-01-20-washington-state-relief-map.html</guid>
    <pubDate>Wed, 20 Jan 2021 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>I learned some Blender and used some open source elevation data to
make a nice looking relief map of Washington State. Check out how I made
it <a href="https://github.com/mattf1n/Relief-Map">here</a>.
:round_pushpin:</p>]]></description>
  </item>
  <item>
    <title>Camping in Cottonwood Wash</title>
    <link>https://mattf1n.github.io/posts/2020-05-14-camping-in-cottonwood-wash.html</link>
    <guid>https://mattf1n.github.io/posts/2020-05-14-camping-in-cottonwood-wash.html</guid>
    <pubDate>Thu, 14 May 2020 00:00:00 +0000</pubDate>
    <description><![CDATA[<p>This last weekend, <a href="http://caitlyndang.com">Caitlyn</a> and I
backpacked up Cottonwood Wash in Utah’s San Rafael Swell, a beautiful
canyon all to ourselves. We even spotted some petroglyphs.
:mountain:</p>]]></description>
  </item>
  <item>
    <title>Course notes from CS183</title>
    <link>https://mattf1n.github.io/posts/2020-04-20-course-notes-from-cs183.html</link>
    <guid>https://mattf1n.github.io/posts/2020-04-20-course-notes-from-cs183.html</guid>
    <pubDate>Mon, 20 Apr 2020 00:00:00 +0000</pubDate>
    <description><![CDATA[<p><a href="/files/lecture.pdf">These are my notes</a> from the course
CS183: Foundations of Machine Learning. They are imperfect and
incomplete but I really enjoyed making them. If you would like to make
edits [email…</p>]]></description>
  </item>
</channel>
</rss>
