Original video in Chinese.
Key Takeaway
- In the AI era, with large language models as containers of knowledge, humans cannot win on breadth of knowledge; the only solution is to “Go Fundamental” and go deep into basic principles.
- Basic principles are relatively stable and universal, so they can handle rapid technological updates and be applied across multiple fields.
- Humans’ unique advantages in the AI era lie in abstract thinking, cross-domain associations, and creative thinking.
- Combining these uniquely human strengths with mastery of basic principles can lead to tremendous innovation and achievement.
- Mastering basic principles can help individuals become more competitive in the AI era and use AI as leverage to unlock greater value.
Full Content
If you know how to learn in the AI era, then no matter how powerful large language models get, you won’t panic.
Why do large language models make everyone so anxious?
Because the essence of large language models is that they are containers of knowledge.
The training process of a large language model is the process of compressing knowledge.
The usage process of a large language model is the process of decompressing knowledge.
This container is large, easy to use, and anyone can use it. So in terms of breadth of knowledge, there is no way we humans can win.
I think the only solution is: Go Fundamental.
That is to say, in the AI era, if you want to learn, no matter which field it is, you absolutely must go deep into the level of basic principles.
Why do I say that? I’ve summarized two fundamental reasons.
First, technology changes very quickly, but basic principles are relatively stable. And basic principles are usually universal, so they can be applied across multiple fields.
For example, in computer science, algorithmic complexity analysis can be traced all the way back to Turing’s era. Even today, this very old theory is still used to evaluate model efficiency.
What is algorithmic complexity analysis?
For example, a customer service center needs three people and one day to make calls to 50 customers. When the number of customers increases to 5,000, how much manpower and time will it require?
By the same logic, in the field of large language models, when the amount of data increases, how will training time change? How will the required computing resources change?
So algorithmic complexity analysis focuses on how fast the time or resources needed to complete a task grow as the task scale increases. This is still valid in today’s AI era, and especially important.
Let me give another example: optimization theory in mathematics is at the core of algorithms. For instance, training a neural network is essentially an optimization process, with the goal of finding network parameters that minimize prediction error.
So do you know where this optimization theory originally came from?
Euclid discussed the problem of finding the shortest distance between two points in The Elements, which counts as one of the earliest thoughts on optimization problems. But back then it was only the beginning of the idea, and a systematic theory had not yet formed.
From the late 18th century to the early 19th century, driven by physics problems, the foundations of this theory began to take shape. By the mid-to-late 20th century, with the development of computer science, optimization theory became an independent and important branch of mathematics.
That is why I say basic principles basically do not change, and they are universal.
Leave all the ever-changing stuff to AI to learn. We just grasp the most basic, lowest-level theories and build the most solid understanding.
I believe everyone must have heard this: when you climb to the top of a mountain in a certain field, you realize that different fields are actually connected—you can actually see that other mountain from this mountaintop. It’s basically the same idea, except that instead of saying we’re climbing upward, it’s better to say we’re going downward, all the way into the basic principles, and then everything becomes interconnected.
OK, that’s the first reason. The second reason is that compared with AI, humans’ unique strengths are abstract thinking, cross-domain associations, and creative thinking. When these unique strengths are combined with mastery of basic principles, they often produce unexpected and extremely huge results.
Let me give a few awesome examples:
Da Vinci combined knowledge of art and anatomy to create accurate and beautiful drawings of the human body.
Picasso fused elements of African art with European painting traditions and created Cubism.
Einstein applied mathematical knowledge to physics and proposed the theory of relativity.
Watson and Crick combined knowledge of biology and chemistry and discovered the double-helix structure of DNA.
In my view, all of these are successful cases of combining uniquely human strengths like abstract thinking, cross-domain associations, and creative thinking with basic principles.
Just imagine: if you really can do Go Fundamental, and then try to combine those principles with knowledge from other fields, to collide, expand, and diverge from them, I believe an individual like that will be extremely competitive in the AI era. He definitely won’t be anxious about AI; instead, he can use AI as leverage—such broad knowledge reserves are like a super long lever, and he himself is the fulcrum. The super system formed by such a super individual and AI will definitely be able to move something very heavy and achieve very high accomplishments.
OK, that’s all for this episode. What we talked about today is a bit abstract, but I think it’s even more valuable than introducing a specific technology or tool as I did before. If you want to discuss further, come find me on newtype—I’m there. See you next time!