DSAI4207 · Introduction to Large Language Model
Explore the ideas and techniques behind large language models through interactive visualizations.
Lecture 3 · Neural Network Foundations
▶ Interactive Scalar Accumulation
Build a = w₁x₁ + w₂x₂ + b term by term and see the same result geometrically.
▶ Interactive Scalar Activation
Switch activations to see how one weighted sum z becomes different outputs φ(z).
▶ Interactive What a Linear Map Changes
Watch W reshape a grid and point cloud; singular axes show where it stretches most.
▶ Interactive A Feature Map Makes XOR Linear
Add the feature x₁x₂ to lift XOR into 3D, where a plane separates all four points.
▶ Interactive Scalar Computation Graph
Build a scalar graph and trace values forward, then gradients backward through the chain rule.
Open Demo →
▶ Interactive Forward, Backward, Update
Train an XOR network one phase at a time: forward values, backward gradients, parameter update.
Open Demo →Lecture 4 · Self-Attention
▶ Interactive Sequence Convolution
Slide one kernel across a sequence to see local context and shared weights build each output.
Open Demo →
▶ Interactive Backpropagation Through Time
Unroll an RNN to trace states forward and gradients back to shared parameters.
Open Demo →
▶ Interactive Self-Attention Matrix
See each token's scores become a normalized row of weights that mixes source values.
Open Demo →
▶ Interactive Multi-Head Attention
Compare parallel attention heads and see how their different outputs combine.
Open Demo →Lecture 5 · LLM Architecture
▶ Interactive Absolute Position Encoding
Explore how sinusoidal signals add position to token representations.
Open Demo →
▶ Interactive 
▶ Interactive 
▶ Interactive LayerNorm, BatchNorm & GroupNorm
Compare which values each normalization method groups together.
Open Demo →
▶ Interactive Transformer Encoder–Decoder
Explore how the encoder, decoder, cross-attention, and masks work together.
Open Demo →Lecture 6 · Generation, Decoding & Inference
▶ Interactive Decoding Lab
Compare greedy decoding, beam search, temperature, top-k, and top-p to see how scores become generated text.
Open Demo →
▶ Interactive KV Cache Lab
Step through generation with and without a KV cache, then explore prefill, decode, and the compute–memory trade-off.
Open Demo →
▶ Interactive Speculative Decoding Lab
Draft tokens with a small model and verify them with a larger model; explore acceptance, correction, and speed-up.
Open Demo →