<?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</title><link>https://ashishmalik.in/post/</link><description>Recent content in Posts on</description><generator>Hugo -- gohugo.io</generator><language>en</language><lastBuildDate>Sat, 20 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://ashishmalik.in/post/index.xml" rel="self" type="application/rss+xml"/><item><title>I Wanted My LLM to Reason. I Only Had a 4GB GPU</title><link>https://ashishmalik.in/post/tiny_reason/</link><pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/tiny_reason/</guid><description>This weekend I was thinking about what reasoning models and thinking models actually are, and whether the difference is just marketing or something real.
The way I understand it now: reasoning models (o1, DeepSeek-R1, and similar) are built or shaped to produce long internal chains before they commit to an answer. They pause, backtrack, and sometimes correct themselves mid-trace. Thinking models, or plain instruct models like Qwen2.5-Instruct, will answer in one pass.</description></item><item><title>I Wanted to Win Kaggle. I Built a Memory System Instead.</title><link>https://ashishmalik.in/post/autokaggle_journey/</link><pubDate>Wed, 25 Mar 2026 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/autokaggle_journey/</guid><description>I did not set out to build a memory system. I set out to win Kaggle.
That sounds naive written down, but it is the honest origin. I kept reading the same advice in winner writeups (stratified K-Fold, target encoding, pseudo-labeling, weighted blends) and thinking: this is not magic, it is a checklist. Top competitors are not discovering new mathematics every week. They are applying known tricks faster, with better feature creativity, and with CV discipline that most beginners skip.</description></item><item><title>Optimized CNNs using Kaiming normal distribution</title><link>https://ashishmalik.in/post/optimizing-cnns/</link><pubDate>Sun, 10 Aug 2025 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/optimizing-cnns/</guid><description>Optimized CNNs using Kaiming normal distribution Convolutional Neural Networks (CNNs) have revolutionized the field of computer vision, enabling breakthroughs in image recognition, object detection, and many other visual tasks. However, training deep CNNs effectively presents several challenges. One critical aspect that significantly impacts CNN performance is weight initialization. In this blog post, we&amp;rsquo;ll explore how the Kaiming normal distribution, introduced by He et al. in their groundbreaking paper &amp;ldquo;Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,&amp;rdquo; optimizes CNN performance.</description></item><item><title>How Autoencoders Move from Pixel Space to Latent Space</title><link>https://ashishmalik.in/post/pixel-to-latent/</link><pubDate>Wed, 30 Jul 2025 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/pixel-to-latent/</guid><description>How Autoencoders Move from Higher Dimensions (Pixel Space) to Lower Dimensions (Latent Space) Autoencoders are fascinating neural network architectures that learn to compress high-dimensional data (like images) into lower-dimensional representations and then reconstruct the original data. This process of moving from pixel space to latent space is fundamental to many modern AI applications, from image generation to data compression. In this post, we&amp;rsquo;ll explore the mathematical foundations and examine different autoencoder variants like Variational Autoencoders (VAE) and Vector Quantized GANs (VQ-GAN).</description></item><item><title>Optimizing Matrix Multiplication</title><link>https://ashishmalik.in/post/faster_multiplication_of_matrix/</link><pubDate>Sun, 20 Jul 2025 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/faster_multiplication_of_matrix/</guid><description>&amp;ldquo;Matrix multiplication is at the heart of modern AI and computer vision.&amp;rdquo; - Yann LeCun
Faster Matrix Multiplication: Methods, Complexity, and Code Matrix multiplication is a fundamental operation in scientific computing, machine learning, and engineering. Optimizing it can lead to significant performance gains. In this post, we explore several methods for matrix multiplication, their time complexities, and provide code samples for each.
1. Naive (Triple Loop) Method Time Complexity: $O(n^3)$</description></item><item><title>Equivariance vs. Invariance in Neural Networks</title><link>https://ashishmalik.in/post/equivariance_vs_invariance/</link><pubDate>Sun, 29 Jun 2025 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/equivariance_vs_invariance/</guid><description>&amp;ldquo;Machine Intelligence is the last invention that humanity will ever need to make.&amp;rdquo; - Nick Bostrom
This blog post examines the fundamental mathematical concepts of equivariance and invariance in the context of neural network architectures. We explore how these principles have evolved from theoretical constructs to practical implementations across various domains, with particular focus on computer vision applications. Through analysis of recent advancements in equivariant architectures such as Group-Equivariant CNNs, TFNs, and SE(3) Transformers, we demonstrate how explicitly encoding symmetry constraints in neural networks leads to improved sample efficiency, generalization capabilities, and interpretability.</description></item><item><title>Benchmarking Analysis of PyTorch Models using torch.compile()</title><link>https://ashishmalik.in/post/torch_compile/</link><pubDate>Sun, 22 Jun 2025 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/torch_compile/</guid><description>Accelerating PyTorch with torch.compile(): When and Why It Works PyTorch 2.0 introduced torch.compile(), a powerful feature aimed at accelerating PyTorch models with minimal code changes. In this post, I&amp;rsquo;ll analyze benchmark results across different model architectures to understand when torch.compile() delivers significant speedups and why its performance improvements vary dramatically across different model types.
Benchmark Results I ran comprehensive benchmarks on various model architectures to measure the impact of torch.compile(). Here&amp;rsquo;s what I found:</description></item><item><title>From First Commit to Latest -Tracing the Journey of Open Source Projects</title><link>https://ashishmalik.in/post/navigate_open_source/</link><pubDate>Wed, 12 Jun 2024 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/navigate_open_source/</guid><description>The journey of a thousand miles begins with one step. &amp;ndash; Lao Tzu
Open-source projects are a treasure trove of knowledge, showcasing the collaborative efforts and coding practices of developers worldwide. Analyzing the commit history of such projects can provide valuable insights into their evolution, development decisions, and coding standards. This guide will take you through a detailed process of reviewing an open-source project&amp;rsquo;s commit history using Git, from the initial commit to the latest updates.</description></item><item><title>Effortless Database Connection Pooling in Go</title><link>https://ashishmalik.in/post/connection_pooling/</link><pubDate>Thu, 06 Jun 2024 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/connection_pooling/</guid><description>The processes necessary in opening and shutting a DB connection are not costly enough to be a concern for small-scale operation. However, if your application grows larger, the ongoing opening and shutting of connections costs more and may start to affect how well your application runs. Instead of creating and closing a new connection for every activity, it often makes more sense to figure out a method to maintain existing connections and transfer them from one operation to another as needed.</description></item><item><title>Method Chaining In Go</title><link>https://ashishmalik.in/post/method_chaining/</link><pubDate>Sun, 05 May 2024 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/method_chaining/</guid><description>Method chaining is a programming pattern where multiple methods are called on the same object sequentially, each call returning the object itself, allowing for a fluent and readable syntax. This pattern is commonly used to configure objects, build queries, and perform a series of operations in a clean and concise manner.
We can see a below example in go which is used gorm. an ORM library for Golang.
db := gorm.</description></item><item><title>Writing a Helm chart for an Angular Application at my Org</title><link>https://ashishmalik.in/post/helm_chart_angular/</link><pubDate>Wed, 17 Apr 2024 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/helm_chart_angular/</guid><description>Last week we had an internal hackathon at our org. We were told to find some interesting projects or solve some existing problems. After some discussion, we found around few points like :-
Getting rid of Luigui and having our micro-frontend framework Genrating lighthouse report dynamically Integrating storybook Reducing dependencies from the backend side for the client Creating a mock server that helps at org level to pick the data from redocly All of these ideas were cool enough, but during the time I didn&amp;rsquo;t had the feeling of doing any core frontend related work.</description></item><item><title>Cherry-Pick/Merge changes from one repo to another repo</title><link>https://ashishmalik.in/post/git_cherry_pick/</link><pubDate>Wed, 13 Mar 2024 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/git_cherry_pick/</guid><description>I was working on a random feature of next release on an usual day at my org, since I had already completed all my tasks of the current release. Around 4PM that day, my UI Architect called me for a help regarding cherry-picking few commits from a client repo to our core repo, which sounded little strange initially, but there is a whole another story to that.
I have done cherry-pick a numerous time in a same repo, but this was a first time where I had to this.</description></item><item><title>Implementing a Linux Container in Go and Escaping it!!</title><link>https://ashishmalik.in/post/containers_in_go/</link><pubDate>Fri, 15 Dec 2023 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/containers_in_go/</guid><description>Being a security enthusiast, I always look for ways to tweak around or break the software. Recently I&amp;rsquo;ve been interested in the container escapes and I&amp;rsquo;ve been reading about them for a while now. To understand the concept of container escapes, I&amp;rsquo;ve been reading about the containers underlying technology and tried to implement them in Go for a better understanding. After implementing it, I also tried to escape from it to the host.</description></item><item><title>Google Summer of Code 2021 - Final Report - Casbin</title><link>https://ashishmalik.in/post/casbin/casbin2021/</link><pubDate>Thu, 19 Aug 2021 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/casbin/casbin2021/</guid><description>Organiztion:- Casbin
Projects
Middleware for Kubernetes :- K8s-authz Middleware for Envoy :- Envoy-authz Mentors
Yang Luo ZiXuan Liu Student
Ashish Malik Project Information:- Casbin is an open-source Authorization Library which supports access control models like ACL, RBAC, ABAC for Golang, Python, C/C++ and for many other languages. It provides support for enforcing authorization based on various access control models. There are many casbin plugins which support different tools. Kubernetes being one of the biggest project in Go didn&amp;rsquo;t have this support.</description></item><item><title>Google Summer of Code 2020 - Final Report - OWASP IIDS</title><link>https://ashishmalik.in/post/gsoc/final_report/</link><pubDate>Fri, 14 Aug 2020 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/gsoc/final_report/</guid><description>Organiztion:- OWASP
Project:- Intelligent Intrusion Detection System
Mentor:- Sri Harsha
Student:- Ashish Malik
Project Information:- OWASP&amp;rsquo;s IIDS is an open source software that leverages the benefits of Artificial Intelligence to detect the intrusion and alert the respective network administrator. It is a fully Django based application, which supports multiple ML and NN models. With the increase in the network traffic and the introduction of new applications and attacks, continuous improvement is required to make the IDS to detect those threats.</description></item><item><title>Why AI acts as a double edged sword for cyber security</title><link>https://ashishmalik.in/post/bias_bounties/</link><pubDate>Wed, 15 Jul 2020 00:00:00 +0000</pubDate><guid>https://ashishmalik.in/post/bias_bounties/</guid><description>A double edged sword is something that seems to be good, but that can have a bad effect.
In the ever changing cyber threat landscape, Security has become something of an arms race. With the rising number of threats everyday, just relying on a human resource would not be enough. Also the lack of cyber security proffessionals just adds the existing problem. Well, one good thing is that AI, in turn promises to be a great solution to this.</description></item></channel></rss>