02. The Mystery of Emergence: Why Does Quantity Lead to Quality?

The Alchemist’s Confusion

Imagine you are an ancient alchemist. You throw stones, wood, and water into a cauldron. No matter how much you stir, they remain just a mixture.

But when you heat the cauldron to a precise degree, or add the 1001st ingredient, a golden light flashes, and the mixture transforms into a completely new, sentient substance.

This is the sense of awe and confusion AI scientists have felt over the past few years.

Recently, the development of Large Language Models (LLMs) has sparked a popular saying: “Humans don’t truly understand why, after pouring enough data and parameters into a model, it suddenly develops reasoning abilities.”

This sounds terrifying, as if we are creating a monster we cannot control. But this isn’t pure “unknown”; it is a scientific phenomenon known as Emergence.

1. No One Taught It to “Reason”

First, we must break a misconception: No engineer ever wrote a single line of code to teach ChatGPT how to perform logical reasoning.

The training goal of a model has always been deceptively simple: Next Token Prediction.

It’s like playing a word-association game:

• You say: “The cat sat on the…”
• Model predicts: “mat.”

It’s as if you asked a child to memorize every book in the world. At first, it’s just rote memorization.

• You ask: “1 + 1 = ?”
• He answers: “2.” (Because he’s seen that line in a book)

But when the volume of reading reaches a certain threshold (say, all the math books ever written), something magical happens.

• You ask: “12345 + 67890 = ?”
• This specific problem has never appeared in any book.
• Yet, he provides the correct answer.

He is no longer “reciting”; he has learned the “Rule of Addition.” This jump from mechanical memory to mastering underlying patterns happens naturally during the process of predicting the next word. To predict more accurately, the model is forced to understand the logical structure of context—this is the prototype of reasoning.

2. The “Aha!” Moment: Quantitative to Qualitative

As model scales grow, scientists observe a non-linear phenomenon:

• Small Models (e.g., 1B parameters): Almost incapable of multi-step reasoning; terrible at math.
• Medium Models (e.g., 10B parameters): Limited improvement; still hallucinates frequently.
• Large Models (e.g., 100B+ parameters): Like a switch was flipped, performance in math, logic, and coding leaps significantly.

This capability doesn’t grow linearly (it’s not 10% -> 20% -> 30%). It’s discontinuous (0% -> 5% -> 90%). These “unlocked” abilities are called Emergent Abilities.

3. The Physics Explanation: Phase Transition

Is this truly inexplicable black magic? Not really. If you’ve studied physics, this looks familiar.

Think about water freezing.

• When water temperature drops from 20°C to 1°C, it remains liquid; its properties barely change.
• But when it crosses the critical threshold of 0°C, it suddenly turns into ice. Its density, hardness, and form undergo a qualitative mutation.

Emergence in LLMs is an “Intelligence Phase Transition.”

• Adding parameters and data is like pumping energy into the system.
• When the complexity reaches a critical threshold, the system undergoes a qualitative change. Isolated bits of knowledge suddenly connect, forming a higher-order, composable internal representation.

Reasoning isn’t handled by a specific “reasoning neuron”; it’s a collective behavior of billions of neurons. Just as a single water molecule isn’t “hard,” but billions of them together as ice are.

4. What do we “not understand”?

So, we aren’t completely in the dark. We know it’s a phase transition; we know it comes from scale.

But we do have many engineering unknowns:

• Why this specific threshold? Why 100B parameters and not 50B? Currently, we can only find out through trial and error.
• What’s the next ability? We cannot predict what will emerge when we scale 10x again. (Self-awareness? Deception?)
• Controllability: Can we unlock “Reasoning” without unlocking “Manipulation”? Currently, we can’t.

5. Summary: It’s Mapping, Not Thinking

While we use the word “Reasoning,” we must stay sober: AI reasoning is not the same as human thinking.

Human reasoning often involves consciousness, emotion, and intuition. In LLMs, reasoning is more accurately described as: Structured, composable pattern transformation in high-dimensional vector space.

It isn’t “thinking”; it is executing an extremely complex but highly stable mathematical mapping. It finds the shortest path from “Question” to “Answer” in a space of trillions of dimensions.

Conclusion: Emergence isn’t a supernatural event or an out-of-control accident. It is the natural byproduct of Scale, Data, and Diversity.

We are in the early stages of understanding this new science. We know the direction (scale up), but we cannot yet precisely predict every turning point. This is not a cause for fear, but a sign that this science is just beginning.