Generative AI Project Lifecycle

Generative AI project lifecycle

1. Define the scope
2. Choose an existing model (or pre-train a model)
   • scaling choices for pre-training
     • goal: maximize model performance
     • constraints: compute budget
     • scaling choice
       • increase dataset size (number of tokens)
       • increase model size (number of parameters)
     • what is found: increasing training dataset size is more important than increasing model size
3. Prompt engineering: Prompt Engineering
4. Fine-tuning (supervised learning/supervised fine-tuning): LLM Finetuning
5. Align with human feedback (safety tuning)
   • goal: HHH = Helpfulness, Honesty, Harmlessness
   • e.g.Reinforcement learning with human feedback (RLHF)
6. Evaluation: Evaluate AI Model & System
7. Model optimization and deployment:
   • Model-level optimization: LLM Optimization (Pre-deployment, optimizing the model itself)
   • Inference-level optimization: AI model Inference Optimization (Post-deployment, optimizing model serving and runtime)
8. Augment model and build LLM-powered applications: LLM-powered Application