The First Typo

I recently asked an AI assistant about AI researcher Andrej Karpathy. The answer was largely accurate, but one detail immediately caught my attention: it referred to him as “Karpovy” instead of Karpathy. The mistake was small enough that the rest of the response still read smoothly, yet significant enough to make me question what else I should verify.

The interesting part was not the typo itself. People misspell names all the time. What changed was my level of trust. Once I noticed the error, I found myself slowing down and checking details that I might otherwise have accepted without question.

My first reaction was curiosity. The assistant clearly knew who I was talking about. It referenced the correct work, the correct companies, and the correct accomplishments. Yet somewhere along the way, the exact name became distorted.

The experience reminded me of comparing notes with friends after a night out.

• One person remembers the conversation at dinner.
• Another remembers the music.
• Another remembers the joke that had everyone laughing.
• Someone else remembers an awkward moment that nobody else noticed.
• Occasionally it feels like everyone attended a slightly different event.

Yet we rarely conclude that our friends are lying. We understand that each person experienced the same evening through a different perspective and later reconstructed it from memory.

The event was shared. The representation was personal.

That observation may be useful as more people begin experimenting with generative AI.

Many of us were introduced to AI through headlines that implied something close to perfection. AI would revolutionize education. Transform business. Accelerate science. Change the world.

Those claims may eventually prove partly true. But somewhere along the way, many people unconsciously adopted another expectation: If AI is this powerful, surely it should not make typos.

Then the first typo appears. Or a name is misspelled. Or two different AI tools provide two slightly different answers to the same question.

For some of us, that missed expectation initially made GenAI easier to dismiss. Of course, some never give it a second chance.

Perhaps the first typo should not make us dismiss the tool.

During a vibe coding session today, I noticed that my coding agent had to fix some bugs due to typos. But this time, before I got upset, I realized that history is filled with remarkable people who made mistakes.

• Scientists published papers containing errors.
• Engineers shipped software with bugs.
• Authors produced books containing typos.
• Teachers forgot names.
• Parents accidentally called one child by another child’s name.

Nobody looked at these moments and concluded that intelligence itself had failed. Instead, we learned something more practical: capability and perfection are not the same thing.

One observation I keep returning to is: Humans and AI can both produce imperfect reconstructions from limited information, but through fundamentally different processes.

Human memory is not a neural network, and a neural network is not a human brain. Still, both can produce summaries, associations, abstractions, and reconstructions that are sometimes useful and sometimes wrong.

Most of the time, those representations are good enough. Sometimes they differ. Sometimes they drift. Sometimes they contain mistakes.

The important question is not whether mistakes exist. The important question is whether we understand the kinds of mistakes a system makes and how to work with them.

A typo is not evidence that AI is useless. It is not evidence that AI is conscious. It is not evidence that AI is deceiving us.

More often, it is a reminder that a fluent answer still deserves verification.

As AI becomes more common in our homes, schools, and workplaces, we will need a new form of literacy. Not literacy about transformers, embeddings, or tokenization. Most people do not need to know those details any more than they need to understand fuel injection to drive a car.

Instead, we need literacy about strengths, weaknesses, verification, and trust.

We need to learn when a typo matters. When it doesn’t. When to verify. And when to continue the conversation.

The first typo is not the end of the story. It may be the beginning of understanding how these systems work.