Intro to AI for Physicists
The science of deep learning, for physicists. Neural networks as model organisms — phenomena, methods, and the synthetic experiments that make them tractable.
Contents
Lectures
Deep Dives
What this book is
This book has two aims:
- A quick overview of how AI got here. A walk through the algorithms, philosophies, and economics that drove modern ML to where it is.
- An experimental physicist’s approach to neural networks. A primer on the neuroethological method: networks treated as model organisms, their rich phenomenology surfaced through controlled synthetic experiments and mechanistic probing.
The guiding methodology, introduced in Chapter 1 and revisited in the two “Science of DL” chapters (Chapter 7 and Chapter 11), borrows from the ethological tradition of Tinbergen [1], Lorenz [2], and von Frisch [3]:
- Observe behavior in naturalistic environments.
- Isolate the core phenomenon to study.
- Reproduce the phenomenon in a model system.
- Understand the model system.
- Confirm the understanding on the original system.
Not every chapter walks all five steps cleanly. This is the guiding principle, not a procedure.
The chapters move through the dominant model classes (image classifiers, sequence models, the transformer, large language models pre- and post-training, diffusion, reinforcement learning, world models) and through two pillar chapters on the phenomenology of training dynamics and concept acquisition. The final chapter widens the frame to intelligence beyond task-learning: multi-agent dynamics, evolution, open-endedness.
After the main twelve chapters comes a section of Deep Dives. Each Deep Dive is a full chapter on a single topic that did not fit the main flow but stands on its own. If you want to go deeper on theoretical foundations, LLM systems engineering, interpretability, neuroscience, cognitive science, or any of the other adjacent fields, that is where to look. Treat them as optional but first-class material, not as appendices.
Who this is for
Anyone comfortable with linear algebra, calculus, probability, and a bit of statistical mechanics. No prior ML background is required, but the pace assumes graduate-level fluency with quantitative thinking. The book does not re-derive things you already know.
What this is not
- Not a classical-ML survey. SVMs, kNN, decision trees, kernel methods. Covered well elsewhere and skipped here.
- Not a theory exposition. Generalization bounds, NTK, capacity calculations. Important, but a different book.
- Not a software tutorial. PyTorch, JAX, kubernetes, MLOps. Recipes, not science.
On homework and solutions
A note on the homework, since this often becomes a fraught topic in 2026.
Each chapter has accompanying homework. Each piece of homework also has a reference solution, and the solutions are not hidden, gated, or held back. They live in plain sight in the code/<chapter>/solutions/ directories of the repository, with the same accessibility as the prompts.
There are two reasons. First, in the age of capable generative AI, every solution to every interesting homework problem is effectively addressable anyway. Pretending otherwise is theater. Second, and more important: the homework is for you. Whether you learn from it depends on what you decide to do with it. Some readers learn best by struggling with a problem until it cracks; some learn best by reading a clean solution carefully and then doing variations of it; some learn best by alternating. All of these are legitimate. The book is not policing your study habits.
So: feel free to look at the solutions. Feel free to ignore them. Feel free to work the problems with whatever combination of self-effort, AI assistance, and reference-checking gets you to actually understanding the material. That is the only thing that matters.
Source and contributions
Source for this book lives at github.com/cfpark00/science-of-dl. Errata, suggestions, and plot-quality reports are welcome via the issue templates.