What Is Artificial Intelligence? A Practitioner's Primer

The Transformer Revolution

Introduced by Vaswani et al. in the landmark 2017 paper “Attention Is All You Need,” the Transformer architecture replaced sequential recurrent neural networks (RNNs) with parallelizable self-attention mechanisms.

Self-attention allows a neural network to calculate mathematical relationship scores between every word (token) in a document simultaneously, regardless of how far apart they appear. This breakthrough enabled training massive models across trillions of tokens on GPU clusters.

The Three Stages of Modern Frontier Model Training

1. Pre-Training: Raw self-supervised learning where the model learns syntax, world knowledge, and reasoning by predicting the next token across massive web and book corpora.
2. Supervised Fine-Tuning (SFT): Training on thousands of curated question-and-answer pairs to transform raw token predictors into helpful assistants.
3. Reinforcement Learning from Human/AI Feedback (RLHF/RLAIF): Training reward models to penalize toxic or incorrect answers and reward truthful, concise, and logically coherent derivations.