# What is Chinchilla Scaling Laws? Definition & Technical Mechanics

> **Source:** [SPIDITS AI AI Glossary](https://spidits.com/ai-glossary)  
> **Category:** Theoretical AI  
> **Canonical Citation:** [https://spidits.com/ai-glossary/chinchilla-scaling-laws](https://spidits.com/ai-glossary/chinchilla-scaling-laws)

## Definition
Chinchilla Scaling Laws are empirical guidelines stating that for optimal model performance, parameter size and training token volume should be scaled in equal proportion. This challenged prior practices of building massive models trained on insufficient datasets.

## Architectural Mechanics & Deep Dive
Chinchilla Scaling Laws are empirical rules developed by Google DeepMind that describe how to scale LLM pre-training parameters and token counts optimally under a fixed compute budget. The laws demonstrated that previous models were over-parameterized and trained on too few tokens, proving that scaling parameter size and dataset size in equal proportions is the most compute-efficient strategy.

## Industry Applications & Real-World Use Cases
Training budget allocation, dataset sizing, and pre-training configuration

## Frequently Asked Questions

### What did the Chinchilla paper prove?
That many models (like GPT-3) were over-parameterized and under-trained, and that a smaller model trained on more data is cheaper and better.

### What is the training tokens to parameters ratio under Chinchilla?
Roughly 20 tokens per 1 parameter for optimal compute efficiency.

## Related Terminology
- [scaling-laws](https://spidits.com/ai-glossary/scaling-laws)
- [pre-training](https://spidits.com/ai-glossary/pre-training)
- [llama](https://spidits.com/ai-glossary/llama)

---
*From the SPIDITS AI Knowledge Base (300+ terms indexed). Explore full technical definitions: [https://spidits.com/ai-glossary](https://spidits.com/ai-glossary)*