AI Course: From Neural Networks to Agents | Original

Home 2026.04

I offer AI mentoring. 50 CNY per lesson (30 minutes each). Subscribe from 1 to 100+ lessons — customize your own path. I am happy to share what I’ve learned.

My Background

Course Philosophy

Inspired by Yin Wang’s CS video course and his CS course principles, I believe in:

What You’ll Be Able to Do

You’ll understand neural networks deeply, understand transformers, train GPT from scratch, modify nanoGPT, fine-tune models, build AI agents, and build an OpenClaw-like platform.

From math to GPT to AI system. That’s the path.

Who Is This For

Course Format

To enroll, send an email to lzwjava@gmail.com with the subject “AI Course Enrollment.” Please include a brief introduction about yourself and your motivation for learning AI. You can also click here to find out more.

Topics We Can Cover

These are example topics. Pick any subset, or go deeper into any area — we customize the course to your needs.

Neural Networks from First Principles

Understand what a neural network really computes. Scalar, vector, matrix computation. Forward propagation step by step. Backpropagation intuition with manual derivatives. Activation functions, loss functions.

Practice: Print every variable (Zen neural style). Implement a 2-layer NN in pure Python. Train on MNIST.

After this, you understand every number inside a neural network.

From Neural Networks to Deep Learning

Gradient descent, learning rate, convergence. Overfitting vs generalization. Regularization, dropout, batch/mini-batch/SGD.

Practice: Train a 3-layer classifier. Visualize the loss curve. Implement dropout manually.

After this, you understand how deep learning actually trains.

PyTorch Minimal Framework

Tensor fundamentals, autograd, nn.Module design, optimizer mechanics, Dataset and DataLoader.

Practice: Rebuild MLP in PyTorch. Train a CIFAR classifier. Inspect gradients.

After this, you can read any PyTorch model.

Language Modeling Fundamentals

Tokenization (BPE), n-gram models, RNN/LSTM intuition, next token prediction, cross-entropy for language.

Practice: Build a character-level language model. Train on tiny Shakespeare. Generate text.

After this, you understand what GPT predicts.

Transformer Architecture

Self-attention, Q K V explained, multi-head attention, positional encoding, feed-forward blocks, residual connections.

Practice: Implement single-head attention. Build a mini transformer. Visualize attention.

After this, you understand the GPT architecture.

nanoGPT Deep Dive

This is the key lesson. nanoGPT training loop, GPT model architecture, weight initialization, data pipeline, sampling.

Practice: Train nanoGPT on a small dataset. Modify model size. Train on a Chinese corpus. Change the tokenizer.

After this, you can train your own GPT.

GPT Training Engineering

GPU memory optimization, gradient accumulation, mixed precision, checkpointing, multi-GPU training, dataset scaling.

Practice: Train a 100M parameter model. Resume training. Fine-tune a model.

After this, you can run real GPT training.

Instruction Tuning and Alignment

SFT fine-tuning, instruction datasets, RLHF overview, reward model concept, LoRA fine-tuning.

Practice: Fine-tune nanoGPT with a chat format dataset. Build an instruction-following model.

After this, you can build a chat model.

LLM Agents (Claude Code / OpenClaw)

Tool calling, agent loop, planning and execution, memory systems, CLI agent design.

Practice: Build a coding agent. Build a CLI automation agent. Build a multi-step reasoning agent.

After this, you can build an OpenClaw-style system.

Personal AI System (Capstone)

Local model deployment, retrieval (RAG), long context memory, multi-agent systems, CLI toolkit integration.

Final project options: Personal coding AI, research assistant, CLI AI toolkit, or autonomous dev agent.

After this, you have your own AI system.

Original. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjavaWeChat QR · X: @lzwjava · Say hi 👋

Back Donate