Intro to AI for Physicists

A short, modern introduction to ML, deep learning, and large-scale AI — for people who already know calculus, linear algebra, and a bit of statistics.

This is a short, modern introduction to AI/ML/DL written for physicists. It is not a survey of classical machine learning — there are already excellent books for SVMs, decision trees, and kernel methods, and they are no longer the bottleneck for understanding what is happening in AI today.

Instead, this book focuses on the ideas that drive modern AI:

• Why scale matters. Why a single recipe — transformers + large compute — has eaten most of the field, and what that means for the next decade.
• The transformer, attention, and the residual stream — the architecture behind nearly every frontier model.
• Scaling laws as empirical physics: power-law behavior of loss in compute, data, and parameters; and what Chinchilla actually said.
• Large language models — pretraining, fine-tuning (SFT, LoRA), and post-training (RLHF, DPO, GRPO, RLVR).
• Reasoning and test-time compute — chain-of-thought, RL on verifiable rewards, and why this changed the trajectory.
• Diffusion and score-based generative models — the part of modern AI most directly continuous with statistical physics.
• Inference — KV cache, speculative decoding, batching: the engineering that decides whether a model is usable.

The treatment is mathematical where it helps and intuitive where it doesn’t. Code is included sparingly and only when it is the clearest explanation. Equations use the conventions a physicist already knows — sums over indices, derivatives, expectations — and avoid the notational tics of ML papers when they don’t add anything.

0.1 Who this is for

Anyone comfortable with:

• Linear algebra (matrices, eigendecompositions)
• Calculus and basic optimization (gradients, chain rule)
• Probability (expectations, KL divergence, sampling)

You do not need prior ML background. If you have written optimizer.zero_grad() in your life, you may skim the early chapters.

0.2 What this is not

• Not a research survey. Citations are minimal and chosen for pedagogy, not credit assignment.
• Not a software manual. We will not teach PyTorch, JAX, or distributed training systems in depth.
• Not opinion-free. Where the field has converged on a view, this book takes that view rather than presenting all historical alternatives equally.

0.3 How to read it

Chapters build on each other but most can be read out of order if you already have the basics. The dependency graph is roughly:

landscape → dl_primer → transformer → scaling_laws → pretraining → finetuning → rl → reasoning
                                                  ↘
                                                   diffusion (mostly independent)
                                                  ↘
                                                   inference

0.4 Source and contributions

Source for this book lives at github.com/cfpark00/intro-to-ai-for-physicists. Issues and PRs are welcome.