Thirteen decks covering the twelve topics — 1,033 slides in total. The syllabus says what each topic covers and which lab goes with it, and most topics were also taught on camera: where a recording exists, it is linked beside the deck.
Topics 1–4 build the foundation (what AI is good for, what quantitative macro needs, Python/PyTorch, core numerical methods). Topics 5–6 add machine learning and apply it to macro models. Topics 7–10 are the classical core — dynamic programming, accuracy, perturbation and projection, parallel computing. Topics 11–12 put it together on heterogeneous-agent and OLG models. Topic 3A is a standalone session on AI-assisted coding and can be read at any point.
The paradigm shift from coder to Research Architect; the evolution of AI; where it helps economic research and where human judgment stays essential.
What quantitative macro is; the workhorse models; the theory behind computing the optimal growth model; which numerical techniques you will need.
Python and PyTorch from the ground up: data types, functions and classes, control flow, NumPy, and working with an AI assistant on numerical code.
A practical Research Architect workflow: separate algorithm design from implementation,
iterate Design → Implement → Validate → Extend, and organize a project with
CLAUDE.md, skills, agents and rules. Pairs with the
starter project.
Function approximation, optimization, root-finding, interpolation, differentiation and quadrature; discretizing AR(1) shocks with Tauchen and Rouwenhorst.
PyTorch for economists — tensors, autograd, torch.nn, training loops;
deep learning fundamentals; custom losses that encode economic constraints.
Value-function approaches (deep VFI, actor-critic) and Euler-equation approaches (residual minimization as supervised learning), with the accuracy and scalability trade-offs against classical methods.
Discrete-time dynamic programming, the contraction mapping theorem, value function iteration with its grid and interpolation choices, and time iteration on Euler equations.
Howard's policy improvement, MacQueen-Porteus bounds, the endogenous grid method, and how to tell whether a solution is actually right — Euler equation errors and Den Haan-Marcet statistics.
Perturbation around the steady state (first and second order, Blanchard-Kahn, pruning) and global projection methods (Chebyshev, Smolyak, finite elements, collocation versus Galerkin) — and when each is the right tool.
New to the site. Serial versus parallel computing, Amdahl's law and
scaling, MPI with mpi4py, PyTorch DataParallel and DDP, and running
jobs on an HPC cluster under SLURM.
Aiyagari-Bewley-Huggett foundations and the canonical stationary pipeline, then Krusell-Smith with modern methods — DeepHAM's value/policy networks against DEQN's Euler-residual training, and what each buys you.
New to the site. Life-cycle models two ways: classical nested fixed point with backward induction and a transition-matrix distribution, then a neural network with an embedded market-clearing layer, Fisher-Burmeister constraints and homotopy training for stability.
Those topics used to sit at the end of this syllabus and now have a course of their own: AI for Economic Research: Dynamic Models, Language, and Agents. See the retired topics page for where each one went.