The Three Scaling Laws of AI: Pre-Training, Post-Training, and Test-Time

Jensen’s opening framework — “AI Scales Beyond LLMs” — identified three distinct scaling laws, each requiring enormous compute. There is no plateau in sight.

Scaling Law #1: Pre-Training (2015-2022)

The original scaling law that drove the initial AI revolution. BERT, Transformers, GPT series — teach models to understand through massive data exposure. More data, more parameters, more compute yields more capability. This is the foundation: raw knowledge acquisition at scale.

Scaling Law #2: Post-Training (2022-2024)

RLHF and skill acquisition that enabled ChatGPT. Transform raw language capability into useful, aligned behavior through human feedback. This enabled commercial AI products — models that could be helpful, harmless, and honest. Post-training turned capability into usability.

Scaling Law #3: Test-Time (2024+)

The breakthrough Jensen highlighted as “revolutionary.” OpenAI’s o1 model introduced reasoning at inference — models that think before responding. They use more compute at runtime to solve harder problems. DeepSeek R1 proved this capability can be open-sourced. This is System 2 thinking: deliberate reasoning, not just pattern matching.

The Curves Keep Climbing

Each scaling law requires enormous compute, and the curves continue climbing simultaneously:

• Pre-training scales knowledge
• Post-training scales alignment
• Test-time scales reasoning

Strategic Implication

Test-time compute is the current frontier. Models that “think longer” perform better — but consume more inference compute. This creates new infrastructure demand beyond training: inference clusters that support extended reasoning chains. The industry moved from training-dominant to inference-dominant compute requirements.

---

This is part of a comprehensive analysis. Read the full analysis on The Business Engineer.