The ML Engineer’s
Cookbook
2026 Edition
52 recipes — GRPO, Flow Matching, World Models, and everything in between
A cookbook of modern training methodologies for AI engineers who build. Centered on the recipes that actually matter right now — spanning 7 domains — with pipeline diagrams, compute estimates, and paper references.
What’s Inside
Language Models
12 recipes progressing from reward-based RL methods (GRPO, DAPO, RLVR) through alignment techniques (Preference Optimization, Constitutional AI, Process Supervision) to self-improving systems (Recursive Self-Improvement, Synthetic Curriculum, Interaction Models) and alternative architectures (State-Space Models, Linear Attention).
12 recipes
Vision
7 recipes covering the shift from diffusion-based generation to flow matching and rectified flow, with preference optimization for visual quality, reward-guided fine-tuning, and self-training loops for continuous improvement.
7 recipes
3D Generation
8 recipes spanning multi-view diffusion, Gaussian splatting supervision, mesh generation, neural field training, scene-level generation, world-state prediction, synthetic 3D pretraining, and animation distillation — one of the fastest-moving areas with few consolidated references.
8 recipes
Speech
5 recipes tracing the arc from self-supervised speech tokenization (HuBERT, EnCodec) through codec language models (VALL-E, AudioLM) to speech RL fine-tuning, multi-speaker distillation, and few-shot voice cloning.
5 recipes
Robotics
7 recipes from latent world models (Dreamer) and action diffusion through sim-to-real transfer, behavior cloning, offline RL, and interactive correction loops — covering the full training stack for embodied AI.
7 recipes
Agents
7 recipes building from tool-use RL through web agents and computer-use models, memory optimization, skill distillation, hierarchical planning, and multi-agent coordination — the practical training stack for deploying LLM agents.
7 recipes
Synthetic Data
7 recipes covering the full synthetic data pipeline: self-instruction, evolutionary expansion, constrained generation, judge/critic training, quality filtering, curriculum synthesis, and production data flywheels for continuous improvement.
6 recipes
Every Recipe, One Format
Instead of summarizing papers, each recipe extracts the training methodology: the problem it solves, the pipeline, architecture decisions, compute requirements, and where it works best.
On-Policy Distillation
Instead of learning from teacher answers, let the student generate responses and have the teacher correct those trajectories.
Training Pipeline
Get Your Free Copy
52 training recipes with pipeline diagrams, compute estimates, and paper references. Delivered as a PDF to your inbox.