From: Cameron Otsuka <cameron@otsuka.haus>
Date: Mon, 27 Jan 2025 22:15:08 +0000 (-0800)
Subject: add paper link
X-Git-Tag: v1.0.0~62
X-Git-Url: https://git.otsuka.systems/?a=commitdiff_plain;h=ba650866cc3aa42633f570e8da7d434e80234d17;p=cotsuka.github.io

add paper link
---

diff --git a/content/articles/r1-vs-o1-ai-as-commodity-or-moat/index.md b/content/articles/r1-vs-o1-ai-as-commodity-or-moat/index.md
index 237dd80..cbf4de6 100644
--- a/content/articles/r1-vs-o1-ai-as-commodity-or-moat/index.md
+++ b/content/articles/r1-vs-o1-ai-as-commodity-or-moat/index.md
@@ -26,7 +26,7 @@ DeepSeek trained R1 for an estimated $5.6MM on A100 and H800 GPUs (not the lates
     <figcaption>Source: <cite><a href="https://x.com/RnaudBertrand/status/1881709223152878000/photo/1">@RnaudBertrand</a></cite></figcaption>
 </figure>
 
-Despite less capital expenditure, R1 has comparable performance to o1 due to several technical innovations. I don't pretend to deeply understand everything on this front, so instead I'll direct you to this post by XX, who made a good breakdown of what's happening under the hood. Ultimately, R1 is essentially a drop-in replacement for o1.
+Despite less capital expenditure, R1 has comparable performance to o1 due to several technical innovations. I don't pretend to deeply understand everything on this front, so instead I'll direct you [to this paper](https://arxiv.org/abs/2501.12948), which breaks down what's happening under the hood. Ultimately, R1 is essentially a drop-in replacement for o1.
 
 <figure>
     {% image "./r1-o1-benchmarks.jpg", "Accuracy / Percentile (%) by Benchmark" %}