[{"content":"Coding has become a very accessible commodity. In software engineering, code has become the cheapest output. This is due to LLMs being very good at producing structured text. And code is structured text. It s very well structured in fact. It follows type definitions, grouping, clear hierarchies. But does that mean we don\u0026rsquo;t need software engineers? Well we do. Firstly, coding and programming are different. I\u0026rsquo;ve briefly explained this in my previous post. Secondly, frontier models are very expensive to run in terms of energy. The models we use are largely subsidised. So, it will gradually become infeasible to access such models current prices and at current volumes if they raise the prices.\nSo one of two things may happen.\nScenario 1. The price of compute remains the same, and we will have to use smaller and locally run models. But the locally run models are orders of magnitude smaller than the frontier models. You could say they are short stories compared to a fable(pun intended). When people are forced to use locally run models, the picture will be different compared to what we see today. Software engineers will be needed more than ever. But that is a discussion for a another time. I\u0026rsquo;ll create a writeup about this later.\nScenario 2. Compute will become cheaper. If thats the case, then we will keep using such powerful models and the coding will remain accessible.\nI\u0026rsquo;ll share my opinion on what to do if scenario 1 plays out at another time. For now let us think about scenario 2, which is more or less the same as the current state of the world in terms of AI usage. Software is cheap and easy to make right now. So, what should we do? How does a fresh graduate stand out in the market? People don\u0026rsquo;t hire freshers now(which is a very bad idea). Until the organisations get a reality check and start hiring again what can the aspiring engineers do?\nI think in this age, sadly, being a software engineer is not enough. You have to be a \u0026ldquo;Software Influencer\u0026rdquo;.\nWhat do I mean by that? I mean that you shouldn\u0026rsquo;t just create toy projects anymore. You have to influence, you have to make impact. More simply, you have to create things that are useful to people. You have to make sure they are usable(can\u0026rsquo;t emphasise this enough). To do this, you need to look for problems to solve, no matter how trivial or simple they seem. Building a solution that helps people will help you be in the limelight. This ensures exposure that is critical for an aspiring engineer. I\u0026rsquo;ll be more specific. Lets think of a simple use case, making notes to remember different things. I am giving this example because I tend to forget things. You can create a note taking tool/reminder tool that is easy and intuitive to use. If it is good people will use it. That is impact. You can also solve your own problems and just publish the solutions. Statistically, those solutions will be used by other people who have the same preferences/problems you have. This will result in you getting more exposure. And your solution may even become a product that generates revenue.\nSo, lets start software influencing everybody.\n","permalink":"https://sakibbai.com/posts/2026-08-25-the-age-of-software-influencers/","summary":"\u003cp\u003eCoding has become a very accessible commodity. In software engineering, code has become the cheapest output. This is due to LLMs being very good at producing structured text. And code is structured text. It s very well structured in fact. It follows type definitions, grouping, clear hierarchies. But does that mean we don\u0026rsquo;t need software engineers? Well we do. Firstly, coding and programming are different. I\u0026rsquo;ve briefly explained this in my previous post. Secondly, frontier models are very expensive to run in terms of energy. The models we use are largely subsidised. So, it will gradually become infeasible to access such models current prices and at current volumes if they raise the prices.\u003c/p\u003e","title":"The Age of Software Influencers"},{"content":"Deep Neural Networks are usually robust to hyper parameter choices. The training will be reasonably good with various reasonable combinations of hyper parameters.\nHowever, Generative Adversarial Neural Networks(GANN) do not follow this pattern. GANNs are used to generate synthetic data, such as real looking fake images. I won\u0026rsquo;t go deep into what a GANN is right now. I\u0026rsquo;ll provide some practical facts. Training such networks require specific hyper parameters and architectural choices.\nFor example:\nDCGAN needs BatchNorm in both generator and discriminators(not in the first and last layers), leaky ReLU activation function Adam Optimizer Strided Convolutions There is another problem in training these networks. Its called Mode Collapse. It is where the generator keeps generating a subset of data because its easy to do.\nTo demonstrate how hyperparameters effect training convergence I taught an AI model how to draw digits. These are videos of a computer learning how to draw English digits. Initially it just generates random noise and gradually digits take shape. These numbers are drawn by a computer and not a human.\nHere are training snapshots from two runs:\nThe authors of DCGAN recommend batch size of 128. I tried to train a DCGAN with both 128 and 1024. The one that has better end-results are trained with batch size 128.\nLink to the GANN training notebook here.\n","permalink":"https://sakibbai.com/posts/2026-08-25-training-gann-models-part-2/","summary":"\u003cp\u003eDeep Neural Networks are usually robust to hyper parameter choices. The training will be reasonably good with various reasonable combinations of hyper parameters.\u003c/p\u003e\n\u003cp\u003eHowever, Generative Adversarial Neural Networks(GANN) do not follow this pattern. GANNs are used to generate synthetic data, such as real looking fake images. I won\u0026rsquo;t go deep into what a GANN is right now. I\u0026rsquo;ll provide some practical facts.\nTraining such networks require specific hyper parameters and architectural choices.\u003c/p\u003e","title":"Training GANN models"},{"content":"I was reading up on Graph Neural Networks(GNN) recently. The way researchers formulated the GNN architectures is really interesting.\nThe easiest way to understand GNNs is to compare it with CNN. Consequently, there is an architecture called GCNN or Graph Convolutional Neural Networks. GNNs and CNNs share some common motivation like, using shared parameters. Without shared parameters, CNN and GNNs both would need infeasible amount of parameters. CNNs solve this problem by exploiting the local spatial correlation of images. GNNs solve this problem using a node\u0026rsquo;s adjacency matrix.\nIn practice, a very interesting trick from graph theory is used for graph convolutions. They represent the graph nodes and edges as an adjacency matrix and find the neighbourhood of a node using simple matrix multiplication. Another cool trick in GNNs that I have found is in training Graphs as batches. In this case, a batch is made of a collection of disjoint graphs in a single matrix.\nI\u0026rsquo;ll probably share an illustrated and in depth post about this later.\n","permalink":"https://sakibbai.com/posts/2026-08-25-graph-neural-networks/","summary":"\u003cp\u003eI was reading up on Graph Neural Networks(GNN) recently. The way researchers formulated the GNN architectures is really interesting.\u003c/p\u003e\n\u003cp\u003eThe easiest way to understand GNNs is to compare it with CNN. Consequently, there is an architecture called GCNN or Graph Convolutional Neural Networks. GNNs and CNNs share some common motivation like, using shared parameters. Without shared parameters, CNN and GNNs both would need infeasible amount of parameters. CNNs solve this problem by exploiting the local spatial correlation of images. GNNs solve this problem using a node\u0026rsquo;s adjacency matrix.\u003c/p\u003e","title":"Graph Neural Networks Are Cool"},{"content":"I have wanted to write publicly for a long time. I have kept notes, half-finished essays, and ideas scattered across too many apps for years. This is the push I needed.\nWhat This Blog Is Three things, roughly:\n1. Technical writing — Notes from my work in machine learning. When I figure something out that took me longer than it should have, I will write it down here. Paper summaries, implementation notes, things that tripped me up. Mostly for myself, but publicly, so maybe it helps someone else.\n2. Research thinking — I spend a lot of time thinking about how neural networks learn and represent knowledge. Sometimes I want to think out loud about that, outside of the constraints of a formal paper.\n3. Everything else — I also write fiction. Personal essays. Thoughts about films, books, games. I am not going to keep that separate from the technical writing. This is one blog, not a brand.\nWho I Am I am a Lecturer at BRAC University and a researcher in deep learning. My MSc thesis at the University of Dhaka looked at how artificial neural networks encapsulate knowledge — how they compress, store, and generalise information from training data. Before that, I spent time as an ML Engineer building practical AI systems: RAG pipelines, LLM reasoning tools, computer vision and ASR fine-tuning.\nI also watch too many films and read too little fiction, which I am trying to fix.\nWhat to Expect No fixed schedule. I will write when I have something to say. You can subscribe via RSS if you want to be notified.\nIf you want to reach me, links are on the About page.\nLet\u0026rsquo;s see how this goes.\n","permalink":"https://sakibbai.com/posts/2026-03-06-hello-world/","summary":"\u003cp\u003eI have wanted to write publicly for a long time. I have kept notes, half-finished essays, and ideas scattered across too many apps for years. This is the push I needed.\u003c/p\u003e\n\u003ch2 id=\"what-this-blog-is\"\u003eWhat This Blog Is\u003c/h2\u003e\n\u003cp\u003eThree things, roughly:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1. Technical writing\u003c/strong\u003e — Notes from my work in machine learning. When I figure something out that took me longer than it should have, I will write it down here. Paper summaries, implementation notes, things that tripped me up. Mostly for myself, but publicly, so maybe it helps someone else.\u003c/p\u003e","title":"Hello, World — Starting This Blog"}]