Intelligence

A database is a promise about durability and consistency; ML is a bet that patterns in old data predict new data.

On this page

The working table of contents.

What ML actually is — instead of writing rules, you show examples and let the machine find the pattern. Training = finding the pattern. Inference = using it on new data.
The three learning modes — supervised (labeled examples: "this image is a cat"), unsupervised (no labels: "find clusters in this data"), reinforcement (trial and error: "maximize this score").
Neural networks — layers of simple math functions that, stacked deep enough, can approximate surprisingly complex patterns. The key mechanism: backpropagation (adjust weights based on how wrong you were).
The architectures that matter now — CNN (spatial patterns: images), RNN/LSTM (sequential patterns: time-series), Transformer (attention: the architecture behind LLMs). Why attention changed everything (process all positions at once, not one by one).
LLMs — what they are (giant transformers trained on internet text to predict the next token), what they can and can't do, the training stack (pre-training → fine-tuning → RLHF → deployment), prompt engineering as the new interface.
RAG — when the model doesn't know something, give it relevant documents at query time. Embeddings, vector search, retrieval, generation.
Agents — LLMs that can use tools, make plans, and take actions. The loop: think → act → observe → think again.
The honest limits — hallucination, evaluation difficulty, cost, data quality as the real bottleneck, the gap between demo and production.

Going deeper

Classical ML algorithms (linear/logistic regression, decision trees, random forests, SVMs, k-means, PCA).
Deep learning math (loss functions, optimizers — SGD, Adam, learning rate schedules).
CNN architectures (ResNet, EfficientNet, Vision Transformer).
RNN/LSTM/GRU deep dives.
Transformer internals (multi-head attention, positional encoding, layer norm).
Diffusion models (Stable Diffusion, DALL-E).
Training infrastructure (distributed training, mixed precision, DeepSpeed, FSDL).
Fine-tuning methods (LoRA, QLoRA, prefix tuning, RLHF, DPO).
Inference optimization (quantization, distillation, speculative decoding, KV cache).
Evaluation frameworks and benchmarks.
MLOps (model registries, feature stores, experiment tracking, monitoring).
Responsible AI (bias, fairness, interpretability, red-teaming).
Multimodal models (vision-language, audio-language).