<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Posts on Ajay Patel</title><link>https://ajayp.app/posts/</link><description>Recent posts by Ajay Patel</description><language>en-us</language><lastBuildDate>Sat, 11 Nov 2023 15:43:23 -0800</lastBuildDate><atom:link href="https://ajayp.app/posts/index.xml" rel="self" type="application/rss+xml"/><item><title>Learning Interpretable Embeddings via LLMs</title><link>https://ajayp.app/posts/2023/11/learning-interpretable-embeddings-via-llms/</link><pubDate>Sat, 11 Nov 2023 15:43:23 -0800</pubDate><guid>https://ajayp.app/posts/2023/11/learning-interpretable-embeddings-via-llms/</guid><description>Our paper at EMNLP 2023, &amp;ldquo;&lt;em>Learning Interpretable Style Embeddings via Prompting LLMs&lt;/em>&amp;rdquo; used GPT-3 to generate a synthetic dataset used for training interpretable text style embeddings, where each dimension of the embedding vector has a meaningful interpretation to a human.</description></item><item><title>ICLR 2023: T5 Models are Few-Shot Learners</title><link>https://ajayp.app/posts/2022/12/iclr-2023-t5-models-are-few-shot-learners/</link><pubDate>Wed, 21 Dec 2022 21:38:22 -0800</pubDate><guid>https://ajayp.app/posts/2022/12/iclr-2023-t5-models-are-few-shot-learners/</guid><description>Our paper at ICLR 2023, &amp;ldquo;&lt;em>Bidirectional Language Models Are Also Few-shot Learners&lt;/em>&amp;rdquo; surprisingly discovers that older models like T5, that predate GPT-3, were promptable and could perform in-context learning.</description></item><item><title>A Surprising Security Vulnerability on the Google Search Results Page</title><link>https://ajayp.app/posts/2021/01/a-surprising-security-vulnerability-on-the-google-search-results-page/</link><pubDate>Sat, 09 Jan 2021 03:14:20 -0500</pubDate><guid>https://ajayp.app/posts/2021/01/a-surprising-security-vulnerability-on-the-google-search-results-page/</guid><description>Back in 2015, I discovered a surprising security vulnerability that allowed you to run malicious JavaScript code on the Google search engine results page, which might be one of the most secure pages on the internet given how much daily traffic it recieves.</description></item><item><title>Language Comprehension is NP‑Complete</title><link>https://ajayp.app/posts/2020/10/language-comprehension-is-npcomplete/</link><pubDate>Sat, 10 Oct 2020 21:39:13 -0500</pubDate><guid>https://ajayp.app/posts/2020/10/language-comprehension-is-npcomplete/</guid><description>&lt;p>Making sense of language is an NP-complete problem, which seems like something you might intuitively guess, but you can prove this by reducing the well known NP-complete problem of graph $k$-coloring to the anaphora agreement problem in linguistics.&lt;/p></description></item><item><title>Useful Linux Features and Utilities</title><link>https://ajayp.app/posts/2020/09/useful-linux-features-and-utilities/</link><pubDate>Mon, 14 Sep 2020 16:02:11 -0500</pubDate><guid>https://ajayp.app/posts/2020/09/useful-linux-features-and-utilities/</guid><description>&lt;p>There are a number of core Linux features or shell utilities that can help build extremely efficient and effective software tools with minimal effort. I find these particularly useful when building server systems, build utilities, or tools for processing large amounts of data. Also, they are great for quick &amp;ldquo;hacky&amp;rdquo; solutions.&lt;/p></description></item><item><title>Raspberry Pi 4 Kubernetes Cluster with x86 Support</title><link>https://ajayp.app/posts/2020/06/raspberry-pi-4-kubernetes-cluster-with-x86-support/</link><pubDate>Thu, 11 Jun 2020 23:09:51 -0500</pubDate><guid>https://ajayp.app/posts/2020/06/raspberry-pi-4-kubernetes-cluster-with-x86-support/</guid><description>The new Raspberry Pi 4 Model B can perform extremely well in an on-premise server cluster as a replacement for cloud services like AWS. The usual drawback is the lack of x86 support given the Linux ARM package ecosystem is still nascent. I’ve hacked in x86 capabilities resembling Apple’s Rosetta 2.</description></item><item><title>Solving macOS VPN Connectivity Issues</title><link>https://ajayp.app/posts/2020/06/solving-macos-vpn-connectivity-issues/</link><pubDate>Sun, 07 Jun 2020 11:23:06 -0500</pubDate><guid>https://ajayp.app/posts/2020/06/solving-macos-vpn-connectivity-issues/</guid><description>&lt;p>For some reason, when connected to a VPN, macOS can choose to not route your traffic through the VPN and instead route it through the regular connection.&lt;/p></description></item><item><title>Sculpting Thanos to Learn Blender</title><link>https://ajayp.app/posts/2020/05/sculpting-thanos-to-learn-blender/</link><pubDate>Fri, 29 May 2020 03:32:30 -0500</pubDate><guid>https://ajayp.app/posts/2020/05/sculpting-thanos-to-learn-blender/</guid><description>The recent Blender 2.80 release has finally come with a much needed UI overhaul that makes it easier to work through the different stages of 3D modeling and animation and access the various tools Blender has. And with Blender&amp;rsquo;s growing popularity in industry, I finally got around to investing some time into learning it.
While I&amp;rsquo;ve worked with 3D modeling and animation tools years ago, I never really did much beyond merging basic shapes together.</description></item><item><title>Relationship between Cosine Similarity and Euclidean Distance</title><link>https://ajayp.app/posts/2020/05/relationship-between-cosine-similarity-and-euclidean-distance/</link><pubDate>Wed, 27 May 2020 23:59:24 -0500</pubDate><guid>https://ajayp.app/posts/2020/05/relationship-between-cosine-similarity-and-euclidean-distance/</guid><description>&lt;p>For unit-length vectors, both the cosine similarity and Euclidean distance measures can be used for ranking with the same order. In fact, you can directly convert between the two.&lt;/p></description></item></channel></rss>