Ranjgith: Attention Please! – My Dive into the Transformer Architecture

ranjgith@blog:/post/transformers-architecture

ranjgith@portfolio:/post$cat post/transformers-architecture.md

#Attention Please! – My Dive into the Transformer Architecture

author: Ranjgith

date: Jul 16, 2025

read: 3 min read

Ever been amazed at how AI can write essays, translate Shakespeare into Spanish, or generate a picture of a cat surfing on Mars? (Okay, that last one is oddly specific — but possible!) It all started with a single sentence: "Attention is all you need" — the name of a paper that shook the foundations of machine learning in 2017. So here’s me, trying to make sense of it — and writing this blog for you. 😁

GenAI • Technology • Machine Learning

Attention Please! – My Dive into the Transformer Architecture

#The Problem Before Transformers

Once upon a time (a few years ago), we had Recurrent Neural Networks (RNNs) and LSTMs. They were great at processing sequences like text — reading one word at a time like a snail on a sugar rush. The downside? Slow. And they had memory problems worse than my morning brain without coffee.

They couldn't remember earlier words in a sentence very well, especially in long sentences. You know that feeling when someone tells you a long story and you forget the start midway through? Yeah, that's LSTMs.

#The Game-Changer: Attention

Enter the Transformer. The brainchild of Vaswani et al. in 2017. The idea was revolutionary:

Instead of reading word-by-word in order, why not let the model pay "attention" to all the words at once?

Imagine reading a sentence like:

"The cat, which had been hiding under the sofa, suddenly jumped onto the table."

When processing the word "jumped", you intuitively know "the cat" is the one doing it. The Transformer gets this too — thanks to attention.

#What is Self-Attention?

Self-attention lets the model look at every other word in the sentence when processing each word. Think of it as a high school group project (ugh). You’re working on your part but constantly peeking at everyone else's work to understand the whole picture. That’s self-attention in action.

Each word gets transformed into a vector (just a fancy list of numbers), and we compute how much attention each word should pay to the others. The result? A weighted mix of word meanings.

#Encoder, Decoder & More

The original Transformer had two main components:

Encoder: Reads the input (like English).

Decoder: Generates the output (like French).

Each is made of layers. These layers have:

Multi-head self-attention: Multiple peeks at the data from different angles.

Feed-forward neural networks: Just some crunching to process the vectors.

Positional Encoding: Since we're not reading sequentially anymore, we add information about the word positions using sine and cosine functions. Fancy math!

#Why Transformers Rule the World

Parallelism: Unlike RNNs, Transformers read everything at once. Much faster.

Scalability: Stack more layers, train with more data, voila — GPT and BERT.

Contextual Understanding: They "get" the meaning better by attending to relevant words.

#Where Are They Used?

ChatGPT, GPT-4, etc. – You’re reading one now. 😄

Google Translate – making your broken French sound fluent.

BERT, T5, ViT – names from the Transformer family tree.

DALL•E, Midjourney – even image generation now uses similar ideas.

#My Thoughts

I’ve always believed that the future belongs to the curious. When I first tried to decode Transformers, it felt like reading Egyptian hieroglyphics. But with every paper I read, every toy model I built, the pieces started clicking. The architecture doesn’t just process text; it understands it (well, kind of). It was like learning how the mind of a digital oracle works.

It made me rethink what "learning" means — and what "understanding" could become.

#Let’s Wrap This Up

Transformers are elegant, smart, and just plain brilliant. Like Tony Stark with a whiteboard. They changed the game of AI, and there’s no looking back.

So next time ChatGPT crafts a beautiful haiku about your cat, remember to whisper a thank you to the humble attention mechanism.

As always, let me know your thoughts, ideas or arguments — I'm all ears (and attention). 📩