ML / AI Deep Dive

The core concepts behind modern LLMs, RAG pipelines, and production AI systems. Read in order if you’re new; jump to whichever concept your interview questions target.

Concepts at a Glance

#	Concept	One-liner
1	Transformer Architecture	Encoder + Decoder blocks; replaces RNNs through parallelisation.
2	Self-Attention	Q.K.V matrices decide which tokens attend to which.
3	Multi-Head Attention	Multiple attention heads in parallel; each learns a different relationship.
4	Positional Encoding	Inject order into a parallel-only architecture (sinusoidal, learned, RoPE).
5	Tokenization	BPE / WordPiece / SentencePiece substrate that converts text to integer IDs.
6	Embeddings	Dense vectors where semantic similarity = geometric proximity.
7	Cosine vs L2	Pick the angle metric for text similarity, the distance metric for magnitudes.
8	RAG	Retrieve -> Augment -> Generate; beats stale or proprietary knowledge alone.
9	Fine-Tuning vs RAG	Update weights vs inject data vs prompt. Pick by data stability and cost.
10	RLHF	Post-pretraining alignment with human preference signals.
11	Gradient Descent & Backprop	The chain-rule loop that trains every neural net.
12	LSTM vs Transformer	RNNs scale linearly; self-attention scales in parallel and gets long context.
13	Hallucinations & Defenses	Probabilistic next-token prediction requires layered defences.
14	Embedding Drift	New embedding model = new vector space = re-index required.
15	HNSW	Approximate nearest-neighbors graph that makes 10M-vector search ms-latency.

How These Layers Compose

       tokens (ints)
           |
           v
+--------------------+    +--------------------+    +--------------------+
|  Token + Positional|    |  Multi-Head        |    |  LayerNorm +       |
|  Embedding         | -> |  Self-Attention    | -> |  FeedForward       |
+--------------------+    +--------------------+    +--------------------+
                                                          |
                                                          v
                                                  (stack N times)
                                                          |
                                                          v
                                                  Next-token logits
                                                          |
                                                          v
                                                Human feedback (RLHF)

The same backbone powers GPT-4, Llama, Claude, Mistral, Gemini — only the data and scale differ.

Tip: If asked “what changed in the last 5 years of AI?”, 80% of the answer is scale + attention + RLHF.