- Thoughts
The faster technology changes, the more we need to study what endures
Even if tools and syntax change, the principles of the problem remain for a long time. This is why we study the basics: to better learn and judge new technologies.
I participated in the Google Machine Learning Bootcamp. This was before the current AI boom. At the time, studying machine learning meant starting with linear algebra, probability, and calculus, learning the principles of machine learning and deep learning, and then going through the process of training models yourself. That is how I studied AI.
However, just a few years later, some of the technologies I learned as being important back then are not widely used in the field today. It is not because AI disappeared.
Rather, it is because it evolved too quickly.
Over the past few years, new models, frameworks, and methodologies have emerged incessantly. Technologies that were treated as important until recently are suddenly replaced by others, and functions that once had to be implemented manually are now solved by a single model or API. Those who joined companies after studying AI recently are likely experiencing something similar.
Both those who learned a few years ago and those who learned recently must feel that the technologies they studied are changing rapidly. It is an era where new technology emerges after a single night's sleep.
The faster technology changes, the more I look toward what has survived the longest
That is why I believe that the more we study new technologies, the more we should learn the old ones. To be precise, we must learn what has survived the longest.
It is easy to understand if you think about web development. In web development, numerous frameworks and libraries have appeared, including Angular, Vue, and React, and the way of developing has continued to change with each one.
If you only know the syntax and usage of a specific framework, you have to relearn a great deal every time a new technology emerges.
However, the story changes slightly if you understand what HTTP is, how browsers and servers communicate, why state must be managed, and why asynchronous processing is necessary.
Even when a new framework appears, you are not learning a completely new world.
This is because, while tools and syntax change, the problems we are trying to solve remain longer than one might think.
I believe AI is similar. Learning how to use currently popular models or frameworks is certainly necessary. I am not saying we should study only the essence without using new technologies.
But if you study only those, the value of your knowledge shakes significantly every time the technology changes.
Conversely, understanding how to view data, why a model works well on training data but fails on new data, how to measure performance and how much to trust those numbers, and what role probability and optimization play within a model does not easily disappear.
New technology does not suddenly appear out of nowhere.
Most of it emerges to solve the limitations of previous technologies. Therefore, if you know the principles and problems of previous technologies, when you encounter a new technology, it doesn't end with simply saying, "I heard this is good these days." You can see why it emerged, what it solved, and what was sacrificed in exchange.
I believe this is the biggest reason to study the essence.
Studying the fundamentals does not mean everyone must build a model
I believe this distinction is especially important now. Just because you study AI does not mean everyone must build massive models themselves. Creating and training a competitive large-scale model from scratch requires immense data, computing resources, capital, and manpower.
In reality, the work most developers will be tasked with is likely closer to understanding existing models and applying them to real-world problems than building a model from scratch.
If that is the case, it might seem as though knowing the essence is unnecessary. Since good models and APIs already exist, is there a need to know the principles inside them?
I believe the opposite. As tools become more powerful, the things we have to implement ourselves decrease, but the task of judging which tool to choose, how much to trust the results, and why it failed remains. The ability to build a model yourself is different from the ability to properly apply a good model to a real-world problem.
And even in the latter, what is ultimately needed is not simply knowing many of the latest tools, but understanding the principles upon which those tools operate and where they fail.
I am not saying we should not learn new technologies. Rather, I am saying we should study the old things to learn new technologies better.
Trendy technologies tell us what to use, but the principles that have survived the longest tell us why we should use them.
A Trip to Suncheon and Yeosu, Day 3
The last day, ending with some studying and a run.

A Trip to Suncheon and Yeosu, Day 2
The second day of the trip, from a clear morning by the sea to sunset and dusk on Baekyado Island.

A Trip to Suncheon and Yeosu
A trip to Suncheon and Yeosu.


Comments 0
Share your thoughts on this article.
Loading comments…