Pretrained LLMs (like GPT-4, LLaMA-3, Claude-3) are like monkeys used in animal behavior science, and we now live in an era where most people can interact with these monkeys and play games. This is fantastic! However, rigorous scientists need to think about the underlying "why" to uncover the "universal laws" behind these phenomena, rather than merely studying individual monkeys.

We propose dividing the concept of "intelligence" into multiple dimensions (such as structures, knowledge, reasoning, etc.). For each dimension, we create synthetic data and build an idealized environment for LLM training to understand the theory and push the capability of LLMs in this dimension to the extreme. By performing a large number of controlled experiments, we aim to discover the universal laws of all LLMs, not just a particular version of GPT-4. 

This humorous phrase critiques theoretical physicists for using oversimplified models. However, without idealized environments, one might wrongly assume that iron balls fall faster than feathers. Idealized environments also help discover simple formulas like the ideal gas laws, which have broad applications. 

The same is true in studying LLMs. Commercial LLMs are trained on messy, secretly preprocessed internet data. Training LLMs in a controlled, idealized environment allows us to manage the data and tweak hyperparameters — such as data amount, type, difficulty, and format — to scientifically determine factors affecting LLM performance and suggest improvements.

The law of gravitation (1666-1687) could not have emerged without Kepler's laws of planetary motion (1609-1619), which in turn relied on Tycho Brahe's 20 years of observational data (1577-1597). Studying the universal laws of LLMs also requires extensive observational data, which cannot be achieved by examining merely a few commercial LLMs. Our aim is to derive universal conclusions, regardless of model size or training parameters. How can this be achieved? By breaking down "intelligence" into manageable components, each using synthetic data with controlled size and difficulty. This allows us to repeatedly train many small models under sufficiently varied conditions to initially identify laws, and then test them more broadly.