Understanding the lifecycle of an AI model — from training to deployment — helps you understand why different models behave differently and what's possible.

Pre-Training: Learning the World

The first stage is pre-training, where the model learns general knowledge from massive datasets:

• Data: Trillions of tokens from the internet — books, websites, code, Wikipedia, forums
• Objective: Predict the next word (for language models) or reconstruct masked words
• Cost: $10-100+ million in compute for frontier models
• Time: Weeks to months on thousands of GPUs
• Result: A "foundation model" with broad knowledge but no specific task training

This is like a student reading the entire library — they know a lot, but haven't been taught how to have a conversation or follow instructions.

Fine-Tuning: Specialization

After pre-training, models are fine-tuned for specific tasks:

• Instruction tuning — Training on examples of following instructions (Q&A pairs, task completions). This turns GPT into "ChatGPT."
• RLHF (Reinforcement Learning from Human Feedback) — Human raters rank different model outputs, and the model learns to prefer higher-ranked responses. This improves helpfulness, safety, and quality.
• Domain fine-tuning — Training on specialized data for specific industries (legal, medical, financial).
• LoRA/QLoRA — Efficient fine-tuning techniques that modify only a small subset of parameters, making it affordable to customize large models.

Inference: Using the Model

Inference is when you actually use the model — sending a prompt and getting a response. Key concepts:

• Tokens — AI models process text as tokens (roughly ¾ of a word). "Hello, world!" = 4 tokens.
• Context window — The maximum number of tokens the model can consider at once (128K for GPT-4o, 1M+ for Gemini).
• Temperature — Controls randomness. Low temperature (0.1) = predictable, focused answers. High temperature (0.9) = creative, varied answers.
• Top-p (nucleus sampling) — Another way to control output diversity by limiting which tokens the model considers.

Latency & Cost

• Larger models = higher quality but slower and more expensive
• Smaller models = faster and cheaper but less capable
• This is why companies offer model tiers (GPT-4o vs GPT-4o mini)

Open Source vs. Closed Source

• Closed source (GPT-4, Claude, Gemini) — You access via API. Can't see the architecture, weights, or training data. The company controls everything.
• Open source (Llama, Mistral, Gemma) — Weights are publicly available. You can download, modify, fine-tune, and run locally. More freedom but requires technical expertise.

The open-source community is closing the gap rapidly, with models like Llama 3 matching or exceeding some closed-source models on benchmarks.