Learning AI by Hand: Tom Yeh on Slowness, Foundations, and What AI Cannot Change
EO Korea
Tom Yeh is a professor of computer science at the University of Colorado Boulder and the founder of AI by Hand, which he describes as a global education initiative to make what happens inside AI's "black box" accessible and approachable. His method is to write out the math of AI models by hand, and his aim is to show that AI "is not a big mystery" but something anyone can understand. In this short talk he asks why people should work through calculations by hand when AI can do them very well. His answer covers what learning is for, why foundations outlast tools, why cheating persists, and what employers should actually look for.
A Transformer, One Box at a Time
Yeh opens with an example of his approach. A transformer, he explains, is meant to process the individual words in a sentence. He might start at, say, token number four, and then draw a box to show that each token is actually represented by multiple numbers. In his simplified case he uses three numbers per token. The point of the drawing is to make each step of the model's math visible and traceable on paper.
An "Old Professor" Relearning Deep Learning
The project grew out of his own gap in knowledge. Yeh says he "missed deep learning entirely" as a student because he was a bit too old. He studied support vector machines and other traditional machine learning methods. After he became a professor, suddenly everyone around him was doing deep learning, and he had to learn it from scratch.
AI by Hand, he says, is his way of sharing that learning journey, including his own struggles with the math and algorithms of AI models. The only way he could really understand the material was to draw or write it on paper, and that was when he would think, "oh, I actually get it." He started sharing his drawings and his way of mapping out the math. Many people responded, commenting that they liked how he broke things down by hand. That feedback gave him the name: he decided to "just call it AI by hand."
A Semester of C++ on the Blackboard
Yeh traces his belief in handwork to an experience teaching introductory programming. Students told him he was going too fast. He wanted to share a lot of knowledge, and with slides and live coding the pace had gotten away from him. So he taught an entire semester of C++ programming on the blackboard instead of live coding, writing his notes out by hand.
As the semester went on he noticed three benefits. First, he could only go as fast as he could physically write. Second, students could only be asked to learn at a humanly possible pace, following along as he wrote. Third, when students copied his notes into notebooks by hand, their hands were not on their keyboards checking Instagram messages, which helped them focus.
For Yeh, "by hand" is about reconnecting with our humanity. Using your hands keeps you at a human speed. Over time he came to see the value of that "old school" approach.
What Is Learning For?
This leads to his central question: what is the purpose of learning? Is it a physical or digital artifact that proves you learned something, or is it something you have internalized and actually own? AI can give you an answer right away, he says, but "having an answer doesn't mean you know it." People can buy degrees and certificates. In his view, how much you own and value a piece of knowledge is proportional to how much time you spent acquiring it. He argues that each person first has to define what learning actually means to them.
Matrix Multiplication as the Evergreen Core
To illustrate what lasts, Yeh recalls learning linear algebra as an undergraduate because it was required for a CS degree. He learned matrix multiplication with no idea why it mattered. Over the years, it kept coming back:
- Computer graphics became popular after Jurassic Park put CGI up front, and CGI uses a lot of matrix multiplication.
- A few years later came the big data movement, which also needed matrix multiplication for processing.
- Then everyone was supposed to become a machine learning engineer instead of a data scientist, and machine learning is again matrix multiplication.
- Today the push is to be "AI native" and even send children to AI school. It is still matrix multiplication.
- In a few years, he suggests, the conversation might be about quantum computing, which he says would once again involve matrix multiplication.
The tools keep changing, he says, but something core and foundational stays the same. It is "evergreen." You could revisit it in one or two years and it would still be relevant. He contrasts this with passing topics. DeepSeek was popular at one point, but less so now. He then mentions a newer, very popular tool (rendered in the transcript as "Common Crawl") and says nobody knows if it will still be popular in two months. He is, however, "pretty confident" that transformers will remain a relevant topic.
Gyeongbokgung and Rebuilding on Stone
Yeh tells a story from a visit to South Korea a couple of summers earlier, when he toured Gyeongbokgung palace. He describes it as a beautiful palace with a long history. What struck him, he says, was learning that the whole complex had burned down around the 1500s except for its foundation of solid rock, and that in the 1800s it was rebuilt on that same foundation.
He reads this as a metaphor for technology. If you have a foundation in something like matrix multiplication, you can apply it to AI or whatever comes next, and it doesn't matter much if a particular tool becomes obsolete. You rebuild your skills on the solid base. If you focus only on surface features and tools and neglect the foundation, he warns, you will keep rebuilding your house without ever having a foundation to rebuild on.
The Skill of Acquiring Skills
Yeh then turns the idea toward personal experience. Think about a skill you became good at while growing up, perhaps piano, soccer, or chess, and how you acquired it. That ability to master something difficult is, in his words, part of your identity that doesn't change as new AI tools appear every day. Knowing "I can acquire this, I can become really good at it" is what lets you apply yourself to new AI tools as well.
He offers himself as an example. He fell behind on deep learning for quite a while, but he had learned the skill of breaking down difficult topics by patiently writing everything on paper. With that skill he says he eventually caught up, even though he started far behind people who had worked in deep learning for a long time.
So, he tells listeners, piano or soccer skills are not useless. Skipping one particular AI tool is "absolutely fine," because nobody knows what will come next. But skipping piano practice or soccer practice and giving up on it is not fine, since that is how you eventually find out who you are. It is not about "this just one tool."
Willingness, Not Memory
At this point in his career as an educator, Yeh says he cares less about whether students learn a specific piece of math. He expects that a year after a lesson, students may remember almost nothing of it. What he thinks lasts is the experience of having understood something, and the willingness to show up, study foundations, and open the black box. That willingness sets people apart. Others never try or take on the challenge.
What differentiates people, he argues, is not how well they remember the equations of the transformer or the attention mechanism. It is the memory of having once worked hard on something difficult, spending hours in the library, and succeeding, which gives them confidence for the next learning challenge. He sees solid foundational knowledge as a sign that someone invested fully in the process. Its absence, in his view, suggests a lack of effort or unwillingness to invest time in a challenging task.
Chegg, AI, and the Real Cause of Cheating
Yeh returns to his introductory programming course. He says he spent a lot of effort writing new assignments every semester because of Chegg, where solutions were shared. His team tried technical countermeasures such as checking IP addresses and even setting traps on the site so they would know who accessed it. At the time he hoped Chegg would go out of business or be shut down by the government.
"My dream came true," he says. AI became the new cheating tool and Chegg went out of business, but the problem remained. That made him think about the source. Chegg and AI, he argues, are only symptoms. The underlying cause is why people feel they have to cheat in the first place. Chegg is gone and people still cheat, and he bets they would find ways to cheat even without AI. In his view, AI cheating distracts from the larger problem of society's incentive system: why students feel compelled to cheat, and why the system doesn't reward real learning that takes time.
Hiring for Fundamentals, Not "AI Native"
Yeh closes on hiring. When he considers hiring someone, he says, what he cares about is basic: good work ethic, strong problem solving, and whether the person is a team player who communicates and works well with others. Those are the qualities that make you want to keep someone as an employee, and he sees AI skill as a byproduct.
His reasoning is that a real problem solver will learn AI automatically because they need it to solve problems, without being told. If an organization feels it must force employees to become AI native, he suggests, it may not have hired the right people, perhaps because it forgot to prioritize problem solving. Likewise, a genuine team player will learn to use AI to support collaboration on their own. His advice is to trust your instincts and keep hiring such people, because they will adopt AI naturally.
The reverse does not hold. If someone is not a team player, always looks out for their own interest, and does not respect others, AI will not fix that. His final point: "AI cannot change people," but people can change AI.
My name is Tom Yeh. I'm a professor of computer science at the University of Colorado in Boulder. I'm also a founder of AI by Hand. It's a global education initiative to make AI inside the black box accessible and approachable. Writing out all the math by hand and doing so understanding that AI is not a big mystery. It's something we all can understand.
What is the purpose of learning? Having an answer doesn't mean you know it. People can buy degree, buy certificate. Do you have ownership of this big idea? Core and foundational that doesn't change is evergreen. AI cannot change people, but you can change AI.
A transformer is meant to process individual words in your sentence. And I can start with, say, where at the token number four in my sentence. And then this is a box I drew to show that, well, each token actually has multiple numbers. In this case, simple case, I say three numbers.
Unfortunately, while I was a student, I missed deep learning entirely. I was a bit too old. So, I studied support vector machines, traditional machine learning methods. So, then I became a professor, and all of a sudden all these people are doing deep learning. I have to learn deep learning all over again. So, what I'm doing today with AI by Hand is to share my learning journey. How I as an old professor trying to learn deep learning from scratch.
All of this I'm sharing, this being about to show my own struggle with understanding AI model math and algorithms. Only way I can get it is I get to draw or write on the paper. This is where, oh, I actually get it. So, I want to share my drawing, share my way to map all the math. And then a lot of people resonated. And people started to comment on, hey, I really like your approach to breaking down by hand. I started, hey, maybe I just call it AI by Hand. There's a reason that people resonate with this, connect with this thing that I'm doing by hand.
Why do we like to calculate this by hand when AI can do this very well? When I was teaching introduction to programming, I had feedback that I was going too fast, I was going through slides, and so on, just too fast. 'Cause I really like to share a lot of my teaching and knowledge with my students. So, I decided to teach an entire semester of C++ programming on the blackboard instead of doing live coding. So, I decided to do that. So, I had a whole semester of writing, just my notes, I wrote it down on a piece of paper.
As the semester was progressing, I see a few benefits. One is that I can only go at the humanly possible speed of my writing. I cannot go any faster than they can write. Second, students can only learn at a humanly possible speed. They can only follow how much I write. And number three is that if I get my students to use a hand to copy my notes on a notebook, their hands are not on their keyboard checking their Instagram messages. So, it helps with focus as well.
The by hand is really about a way to connect back to our humans. So, using your hand, you can go at a human speed. It comes to you in a human way. And over time I learned this, I started to see this value. So, going back to the old school by hand.
What is the purpose of learning? Is it about that physical or digital artifact that proves they have learned, or something you feel that you actually internalize, you actually own this kind of stuff? Do you have ownership of this big idea? Well, AI gives me the answer right away. Having an answer doesn't mean you know it. People can buy degree, buy certificate. Whether you own something, you value something, it's actually proportional to how much time you spend acquiring that piece of knowledge. You have to first define what learning actually means to you.
I remember when I was an undergrad, learning linear algebra as part of a requirement for getting a CS degree. We learned linear algebra and we learned matrix multiplication. I had no idea why it is even important. Then it turns out over time, computer graphics became really popular because of Jurassic Park that put CGI up front. And people thought, "Oh, hey, everybody needs to learn CGI." And CGI uses a lot of matrix multiplication.
And then after a few years, there was the big data movement. Then it turns out you also need matrix multiplication to do some sort of processing. And then you move into machine learning. Again, forget about data science, we should be machine learning specialists, engineers. Again, matrix multiplication. And today's AP is AI. AI. Everybody needs to be AI native. We should raise our kids and send them to AI school. Matrix multiplication. In a few years, all we talk about is quantum computing. It's possible, right? And guess what? Matrix multiplication again.
So, you see this trend that every time there's something, the tools keep changing, but there's always something that's core and foundational that doesn't change. It's evergreen. You could revisit it a year from now, 2 years from now, it's still relevant. People still care a lot about it. What is DeepSeek? That was popular at the time, but it's been a while now. So, DeepSeek is not as popular as before. There's a new thing, for instance, like this Common Crawl. Super popular. But see, in 2 months, is it still popular? We don't know. But I'm pretty confident the transformer topic is still going to be popular.
A couple summers ago, I had an opportunity to visit South Korea and I got to visit where everybody else would go, the historical Gyeongbokgung. This is the palace that has thousands of years of history. Very beautiful palace. And then what struck me when I learned a bit more history, the entire thing was burned down in like 1500, except for the foundation that was made in solid rock. So in the 1800s, they rebuilt the entire palace based off the same foundation.
But I did have to talk about the story because that reminds me of how this technology has been changing over and over again. But if you have a foundation on matrix multiplication, you could just apply it to AI. It doesn't really matter. Rebuild your skill based on your solid foundation. So that's why I'm focusing on foundation, because I believe there's something you can rebuild. It doesn't really matter whether the new tool is obsolete. If you keep focusing on the surface features, the tools, and forget about the foundation, you just have to keep rebuilding your houses. You still never have your foundation to rebuild it on.
So how can you apply this to your own situation? Think about the way you grew up. Maybe your parents sent you to soccer, or maybe the piano, and you have some skill you have become good at. Is it piano? Is it chess? And think about the way you acquired the skill, and that is actually something that doesn't change. So as AI tools come every day, the fact that you can acquire a very difficult skill, that is something that is part of your identity that doesn't change. If you continue to focus on that and you have the ability to realize, hey, I can acquire this, I can learn this skill, I can become really good at it, that's when you could continue to apply that skill to a new AI tool.
So I'm a good example. I was falling behind on deep learning for quite a while, but I have learned skills really trying to break down difficult topics by patiently writing everything down on paper. So with that skill I was able to eventually catch up. I caught up on deep learning from a position way behind people who actually have been working on deep learning for a long time.
So for you, your piano skill, your soccer skill is not useless. That will be a skill to help you eventually, once we figure out all this crazy stuff. And this one tool, if you just skip this tool, I think it's absolutely fine. Who knows what you're going to do next. But if you skip your next piano practice, you skip your next soccer practice and you give up on that, that's not fine, because that is going to be, in the long run, how you find who you are. But not this tool. Not just this one tool.
At this moment in my career as an educator, I have started to care more about not that they learn this math. A lot of times when I teach, I say, oh, you showed up, you listened to me, you're trying to go through it. But I bet maybe a year from now you don't remember anything. But what you can remember is more about that you were able to understand this. At the moment you are willing to come here to understand the foundation, you are willing to open up the black box. That willingness is what sets you apart from others. Others would never try, never attempted, never took on a challenge.
That's what differentiates. It's not really about how much you remember the equations about the transformer, about the attention mechanism. It's really about: there was once upon a time I tried hard to memorize this. I stayed in the library for hours and studied. I was successful. So next time when there's a learning challenge, I can learn this. So that is more important, what differentiates. People with a foundation implies the process, 100% invested in learning it. So that is what I value, versus a person who never really learned the foundation, which implies the lack of effort, the lack of willingness to invest time and effort to take on a challenging learning task.
When I was teaching intro to programming, I would really spend a lot of effort making new assignments every semester because of Chegg, places where people share solutions. And we had all these technical solutions, we were trying to check the IP addresses and whether people were accessing this, and we even put some kind of a trap on the site. If somebody accessed that, we know they accessed this. But at the time I was like, "Hey, well, I hope Chegg could go out of business. Maybe get shut down by the government. And then that'll help us educators."
And my dream came true, because AI became the new cheating tool and Chegg went out of business. And then I realized, "Hey, Chegg is out of business. My dream came true, but the problem is still there. What's going on?" So, it reminds me again, we should keep going back to the source. There was a reason why people had to cheat in the first place. That is the main cause. Chegg and AI are just the symptoms. So, Chegg is gone. People still cheat. I bet when AI is gone, people can still find ways to cheat. It's this AI cheating that distracts us from the bigger fundamental problem of society's incentive system. Why are students compelled to cheat? Why is it that the system doesn't encourage real learning that has actually had time spent on it?
So, when you hire somebody, what I care about is: does the student have a good work ethic? What I care about is: is the student a good problem solver? Is this person a team player that can actually communicate well and work with others? So, when you hire somebody, you can go back to those basics. That's what you actually care about. You want to keep those people as employees. And this AI thing is just going to be a byproduct of this.
Think about it. When we hire someone because they're a problem solver, in order to solve problems, that person is automatically just going to learn AI. You don't have to tell them. The reason why you had to force them to be AI native, you have to go back. Maybe you didn't hire the right person. You forgot to emphasize the person being a problem solver. Similarly, if you are hiring this person because this person is a team player, because that person is a team player, the person will learn how to use AI to facilitate collaboration. So, you don't even have to tell the person, the person will automatically do that as well.
Trust your instinct. Continue to hire people like that, because those people will automatically adopt AI. If you're not a team player, AI is not going to make you a team player. If you always look out for your own interest, you do not respect others, AI is not going to fix that. How can AI fix that? AI cannot change people. Only you. But you can change AI.
Article published