Understanding AI Through Typhoons

With Shanghai hit by typhoons recently, I've been following the weather news, which shows typhoon paths, wind speeds, impact zones. Meteorologists draw curves on maps, telling us when the typhoon will arrive where, how much damage it will bring.

Then I thought of another question: "Why did this particular raindrop fall on this roof?"

Not "why is it raining here," but "why this raindrop, not that one, fell on this roof, not that one?"

At a time when meteorology is so advanced, we have no answer to this question. But think about it carefully — it reveals a fact we often ignore: we think we understand typhoons, but our understanding has a boundary. What meteorology reveals to us is exactly the right way to understand LLMs.

---

I. The Meteorological Insight: Understanding Principles ≠ Predicting Details

We understand typhoons.

We know that the Navier-Stokes equations describe fluid motion — these equations essentially say that every tiny "water parcel" in a fluid, when moving, is pushed by pressure, dragged by its neighbors, and affected by gravity. The balance of these forces determines whether the water parcel accelerates or decelerates. It's like a crowd of people walking on a crowded street — each person is pushed by those in front, behind, left, and right, while also trying to maintain their own speed, eventually forming the overall flow of the crowd.

We also know that thermodynamics explains energy conversion, that phase transition mechanisms explain cloud and rain formation. We can even predict typhoon paths, rainfall intensity, and impact areas to some extent.

These are not secrets. They are principles written in textbooks.

But if we ask: "Why did this raindrop fall on this roof, not that one?"

A meteorologist will tell you: we don't have a direct answer to this question.

Of course, we know that this raindrop falling here rather than there is certainly not divine protection, nor some mysterious force beyond our understanding.

It's because: understanding the macro principles of a system and predicting each micro detail are two different levels of problems.

A typhoon is a chaotic system. Sensitive to initial conditions, butterfly effect. Even if we perfectly master all physical laws, we cannot precisely predict where each raindrop will fall.

This is not incompetence. This is the nature of the system.

Accepting this is a sign of meteorology's maturity.

Similarly, accepting this is a sign of maturity in understanding LLMs.

---

II. LLMs Are the Same: We Understand Principles, But We Don't Demand Details

People often say LLMs are "black boxes."

But this statement is not accurate.

We understand the Transformer architecture, how attention mechanisms compute relevance, how positional encoding handles sequence order, how gradient descent optimizes parameters, how cross-entropy loss measures prediction quality.

These are not esoteric knowledge. They are public, verifiable, reproducible.

But if we ask: "Why was this word generated, not that word?"

This question is like "Why did this raindrop fall here" — we don't have a direct answer.

LLMs have billions of parameters. The generation of each token is the result of all parameters working together. Even if we fully understand the mathematical meaning of every edge, we cannot precisely predict the choice of a specific token.

It's not because we don't understand the principles.

It's because: the nature of high-dimensional statistical systems is that emergent behavior is difficult to predict point by point.

Accepting this moves us from "mystification" to "understanding."

---

III. Raindrops Are Not Protected by Gods

But there's a dangerous misunderstanding here.

Some say: "Since we cannot predict details, LLMs are unexplainable, have 'spirituality,' and are some kind of wisdom beyond human understanding."

This is wrong.

The exact landing point of a raindrop cannot be precisely predicted, but we know raindrops are not protected by gods. They are the result of gravity, air pressure, humidity, and temperature working together. They are products of physical laws.

Similarly, the output of an LLM cannot be precisely predicted, but we know it's not magic. It's the result of probability distributions, optimization objectives, and training data working together. It's a product of mathematical laws.

Meteorologists don't say: "Typhoons are divine will."

They say: "This is the result of atmospheric circulation, ocean temperature, and the Coriolis force working together."

LLM researchers shouldn't say: "This is machine soul."

They should say: "This is the result of parameter matrices, attention weights, and loss functions working together."

The key to demystification is moving from "divine blessing" to "statistical laws."

---

IV. What We Need Is Mitigation

Meteorology tells us: understanding principles doesn't mean controlling every raindrop.

But meteorology also tells us: while we cannot control details, we can mitigate risks.

We build dams not to stop rain from falling.

We build drainage systems not to eliminate floods, but to give them somewhere to go.

We issue warnings not to change the weather, but to give people time to prepare.

Mitigation is not elimination, it's management.

LLMs are the same.

We cannot control every output, but we can design safety boundaries.

We cannot eliminate all hallucinations, but we can detect and mitigate them.

We cannot stop all misuse, but we can establish usage norms and regulatory frameworks.

What we do is not "perfection," but "reliability."

Meteorology is not about conquering nature, but about coexisting with it.

LLM research is not about creating perfect intelligence, but about enabling AI to coexist with humans.

---

V. From Superstition to Engineering

Attributing raindrop placement to divine protection is superstition.

Understanding meteorological laws and building dams, setting up warnings — this is engineering.

Attributing LLM output to machine soul is superstition.

Understanding Transformer architecture and doing alignment, setting boundaries — this is engineering.

Meteorology took centuries to move from "divine will" to "prediction systems."

LLM research is only a few decades old, but we can complete this transition faster.

Understanding does not equal control, but understanding is the prerequisite for management.

We understand typhoons, so we don't ask where each raindrop falls.

We understand LLMs, so we don't ask why each token was chosen.

We understand meteorological laws, so we do mitigation.

We understand LLM principles, so we do alignment.

This is the mature way to understand complex systems.

It's also the best way to remove the mystery.