AI-ZeroToOne

Goal

• Understand the fundamentals of AI
• Master the AI-related tools and frameworks
• Know the current state of the art, and How it works

Plan

Season 1: Learning Basics Videos

• Machine Learning Basics
• Deep Learning Basics
• GPU & CUDA
• PyTorch

Season 2: Network Structure Videos

• CNN(Convolutional Neural Network)
• RNN(Recurrent Neural Network)
• LSTM & GRU (Long Short Memories)
• Computer Vision & Machine Translation
• Seq2Seq & Attention Mechanism
• ResNet (Shortcut Connections)
• Transformer

Season 3: Content Generation Videos

• VAE(Variational Auto-Encoder)
• GAN(Generative Adversarial Nets)
• Diffusion Models(VI & Score-based)
• Flow Matching & Diffusion(ODE & SDE)
• Conditional Generation (Text-To-Image)
• DiT(Diffusion Transformer)
• Video Generation

Season 4: Language Models Videos

• Review: RNN -> Seq2Seq -> Transformer
• From Word2vec to BERT (Representation Learning)
• GPT Series (Next-token Prediction is Intelligence)
• BERT -> T5 (How to scale-up BERT?)
• From CLIP to Flamingo (Way to Multimodality)

Season 5: Reinforcement Learning Videos

• RL Basics (Agent, Value Function, Policy)
• Markov Decision Process (MP, MRP, MDP, & DP)
• Traditional RL (Model-free Prediction, Model-free Control)
• Deep RL (Value-based RL, Policy Gradient Methods, Actor-critic Methods)
• Learning -> Optimization (TRPO -> PPO)
• RL in Action (LunarLander & Atari)
• RL in Action (Chess AI)
• RL in Action (AlphaGo)

Season 6: Build LLM Videos

Part A: LLM System Concepts

• Engineering Foundation (Compute & Memory & Communication)
• Modern LLM Architecture (RoPE, GQA, SWA, MLA, MoE, AttnRes)
• Data Pipeline (Corpus Cleaning, Tokenization, Chat Templates)
• Distributed Training (DP → PP → TP → ZeRO)
• Post-Training & Alignment (RLHF, DPO, GRPO, LoRA)
• Evaluation (Objective Benchmarks, LLM-as-a-Judge)
• Inference Optimization (KV Cache, FlashAttention, PagedAttention, Quantization)

Part B: Rebuilding NanoChat from Scratch

• Codebase Walkthrough
• Core Module Rewrites (I): Muon Optimizer & Compute-Optimal Scaling
• Core Module Rewrites (II): Sliding Window Attention, Dataloader, KV Cache Engine
• Post-Training Rewrite: SFT Packing-to-Padding Transition
• Local Training: d12 Full Pipeline on Consumer GPU
• Cloud Speedrun: Beating GPT-2 under $100
• Evaluation & Retrospective