Look Inside a Large Language Model to Become a Better Java Developer
summary
Shows how a working LLM can be expressed in Java using DeepLearning4J and JTok, with tokenization by byte-pair encoding, token embeddings, positional embeddings, and attention implemented via query/key/value matrices; points out GPT2-like sizes (≈52,000 tokens, claims ~175 billion dimensions) and uses gradient descent plus GELU-based feed‑forward nonlinearity (GELU). Demonstrates a translated Karpathy mini-model that trains on a laptop in ~1 minute to generate plausible baby names, and notes production-scale training needs CUDA/GPU and Project Babylon’s heterogeneous accelerator toolkit.
AI-written summary · about
chapters (25)
- 0:00Large language models intro
- 1:45Introduction
- 3:15Why Java for large language models
- 5:47DeepLearning4J implementation
- 6:08How LLMs predict tokens
- 8:13Understanding tokenization in language models
- 10:25Byte pair encoding
- 14:44Neurons and neural networks
- 17:57Vectors matrices dot products
- 21:11Neural network architecture
- 22:58Token embeddings
- 23:41Training neural networks
- 24:23Gradient descent
- 26:33Vectors and meaning
- 27:38Vector arithmetic
- 30:08Positional embedding
- 31:13Attention mechanism
- 34:33Attention word relationships
- 37:55Training attention matrices
- 40:04GPUs for attention
- 41:06Feed forward networks
- 44:00Universal approximation theorem
- 46:36GELU functions
- 47:43Java small LLM
- 48:31Java LLM demo
related talks
46mGenerative AI in Practice: Concrete Use Cases for an LLM in Java
52mAI for Java Developers - From Buzzword to Code
30mKeynote: LLM Research and Development at National Institute of Informatics
46mUnder the Hood of LLMs — all those questions you never dared ask about them
01h 00mInside GPT – Large Language Models Demystified
43mBeyond the Hype: A Realistic Look at Large Language Models
44mHow to build your own Tiny Language Model from scratch
50mGenerative AI in Practice: Concrete LLM Use Cases in Java with the PaLM API