Diffusion models: restoring Dada's old photographImage generators like Stable Diffusion learn by cleaning up noise. Here is the idea, told through a photo restorer in Anarkali.ML Stories
10 posts
Diffusion models: restoring Dada's old photographImage generators like Stable Diffusion learn by cleaning up noise. Here is the idea, told through a photo restorer in Anarkali.
RAG: giving AI an open-book examRetrieval-augmented generation lets an AI answer from your own documents instead of its memory. Here is how it works, told through a new shop assistant.
Large language models: the autocomplete that grew upChatGPT-style models are trained to do one thing: guess the next word. Here is how that simple job turns into an assistant, and why it sometimes makes things up.
Transformers: the meeting where everyone listens to everyoneThe 2017 idea behind ChatGPT, Claude and Gemini, told as a family meeting where every word decides who to pay attention to.
Tokenization: why AI reads in piecesAI models do not read letters or words. They read tokens, and that small detail explains pricing, limits and why Urdu costs more.
Embeddings: a map where words liveHow AI turns words into coordinates, so that 'biryani' lives near 'pulao' and far away from 'tractor'.
CNNs: how a computer reads your CNICA photo is just a grid of numbers. Convolutional neural networks turn that grid into edges, shapes, letters and finally a name and an ID number.
Overfitting: the student who memorised past papersAhmed scored full marks on every past paper and still failed the board exam. His mistake is the most common way AI models fail too.
Gradient descent: lost in the Margalla fogHow do you find the lowest point of a hill when you cannot see anything? One careful step at a time, which is exactly how AI models learn.
Neural networks, explained by a chai apprenticeA boy learns to make the perfect cup by tasting, adjusting and tasting again. That is almost exactly how a neural network learns.