Transcription
In the last six months, I've been approached by three different groups wanting to replace me with an AI. One person said it would be like a magic gumball machine. >> This is a wowy zoey gumball. >> I'd seed it with my previous videos and scripts, turn some knobs, and then watch endless [music] finished products come out. My first thought was, if you have this magic machine, what do you need me for?
[music] What's the role of my thought? Socrates faced this exact question 2400 years ago. He wasn't worried about AI. He was worried about writing, fearing it would replace our thoughts and said it would create the show of wisdom without the reality. Because reading a thought isn't having it. It must be spoken.
[music] Looking at this technology today, he might have been way ahead of his time. >> Hello, I'm here. Oh, [snorts] >> hi. >> Hi. [music] >> What Socrates felt was that there is a fundamental difference between hearing something and saying something. He wouldn't hand you answers. He forced you into a dialogue.
[music] He made you generate the idea yourself. The struggle to find the words is the mechanism of thinking. And so you could design dialogues using the right questions [music] which would unfold creating the potential for this knowledge to grow inside someone's mind.
[music] Ancient religious traditions understood this. For example, the Talmud isn't a list of answers. It's [music] an encoded debate. The oral Torah had to be spoken, discussed, regenerated each generation.
[music] Buddhist monks did the same thing with coons, [music] paradoxes designed to force students to wrestle with questions, not memorize answers.
[music] Millennia later, modern research proved exactly why. In the 1920s, psychologist Liv Vigotsski had this beautifully simple idea that thinking is talking. First with people, then with ourselves.
>> Crying is the first step. When your child cries, he cries all over. His arm and leg muscles contract, his face screws up into a knot, his mouth opens, and out comes the cry. The sound of human speech. But a certain amount of crying is necessary if he will talk someday.
>> So he observed babies closely, specifically their babbling faces. >> A portion of his time is spent making pleasing sounds, or more accurately, comfort sounds. During these early months, your baby must be given the time to practice [music] new sounds privately. Sounds that he makes will begin to feel and sound familiar.
>> It's how the machinery of thought is built by talking out loud. >> Selftaught gives our children words to think with.
>> First, babies [music] practice syllables, then words, then they glue words together, eventually forming sentences that others can understand. This [music] outward babbling eventually turns inward and becomes what we call inner dialogue, the voice in your head. The struggle to form words out loud creates the circuitry for thinking.
>> Winnie the Pooh lives in a house in the forest. Who's there? >> Winnie the Poo. >> Winnie the Poo.
>> In 1978, it got that name, the generation effect. You need to do it to know it. One key experiment was showing people flash cards with word pairs on them. Later on, they asked people to recall the words they read. Almost nothing stuck. But then they changed one thing. They showed them a word with a partially generated word. All of a sudden, memory skyrocketed. No new information was given to them. The only difference was the struggle to generate the information that made it stick. And further experiments proved how this [music] effect scales. It's not just about memorizing words. It's about grasping an idea.
>> The smell. >> I know. It smells. It smells good. >> Smells good. >> Where did you get it? >> Smells good.
>> And brain scans confirmed why. When you generate a thought from scratch, you activate multiple regions at the same time, which encodes information in your mind. But when you just read or hear, those circuits barely light up. And this applies to both actions and words. Take drawing. Most people cannot draw, but they can trace. Tracing creates the illusion of skill because the result looks perfect. But take the original away and [music] you cannot reproduce it. If you only ever trace, you never learn to draw.
[music] To learn to draw, you have to practice drawing from scratch with sketches. Once you learn how to do that, you have a special power. You can bring anything to life on a blank page. A skill that may become even more valuable in the future. The first sketch of an [music] idea. Ex. The intelligence you build, whether with action or words, requires you to generate from scratch. Whenever we skip this step, we pay a price.
[music] This brings us to the paradox of technology. It accelerates our progress by doing jobs for us. But in doing so, it prevents us from knowing how we did them. Speaking of the struggle to generate thoughts, here's something I've been using that really gets this right. Boot.dev. Most coding tutorials just give you the answer. Boot.dev makes you build it. Every lesson is hands-on, so you're writing actual code, making mistakes. That friction I talked about, they built it in. And here's the clever part. Their AI tutor boots doesn't just solve problems for you. It asks you questions, forces you to think through the solution yourself. It's the Socratic method for learning to code. If you're serious about learning back-end development, Python SQL Go, this is how you build that capability, not just simulate it. Go to boot.dev and use code art for 25% off your first year or use this QR code.
and reliance on any technology kills that specific part of your mind. In the 2000s, neuroscientists studied London taxi drivers and found that the region of their mind for storing spatial knowledge was physically larger than average. But drivers who switched to GPS, that gray matter actually shrank. A recent paper in the Lancet found that doctors using AI assistance for just 4 months had a weakened ability to spot cancer on their own. By outsourcing seeing to the tool, they atrophied the fundamental visual skill needed to diagnose. The most damning evidence comes from Kristoff Van Neman.
[music] He had two groups solve logic puzzles. Group A got helpful software highlighting valid moves as they worked through it. Group B got [music] nothing. They struggled alone. Of course, group A solved the puzzles faster. But then when the help was removed for future puzzles, group A collapsed. They quote [music] aimlessly clicked around, he reported. Group B had no issue solving the puzzles. And until recently, this atrophy was limited to specific skills.
[music] But then came the irony. We invented a new kind of machine by forcing it to learn exactly the way we do, [music] and that freed us from ever having to learn again. Large language models are designed by having it guess the next word billions of times on all written knowledge. Each wrong guess refined its understanding. And unlike previous tools designed for narrow tasks, this was general. A machine that learned a meta language, not just words, but patterns. The structure of arguments, the rhythm of music, the logic of code, the flow of images. It compressed all forms of human expression into a single model. This dramatically increases the potential to atrophy our minds since it can be applied to anything and we already have strong evidence of this.
In 2025, researchers at MIT ran an experiment. They had students write essays under three conditions. A brain only group, a Google search assisted group, and a final group who had chat GBT. They monitored the brain activity with an EEG while they worked and the findings were dramatic. Minutes after finishing their essay, when asked to quote a single sentence from their own work, 83% of the chat GPT group couldn't recall any specific text. And the few who thought they could, 100% of them got it wrong. And the brain activity backed it up. The chat GPT group showed significantly lower brain activity, specifically neural connectivity. Their brains appeared to dim while working. And then when evaluators looked at the essays, the AI assisted essays were technically proficient, grammatically correct, structurally sound, but they were consistently described as hollow or soulless. They seemed remarkably similar to each other. This was confirmed in a 2024 study where they had 300 people write stories, some with AI assistants to seed their ideas, some without. But when they analyzed all the stories together, they found the AI assisted stories were significantly more similar to each other.
This gets us to a cognitive cost beyond the individual. When everyone uses AI in a society, [music] our collective thought begins to narrow. To better understand this effect, imagine all possible thoughts as a massive tree infinite in size.
[music] Each thought, which you can imagine as a long sentence, is a pathway through this tree. [music] Each step in the path is a word. Your mind will only ever explore a tiny microscopic sliver of this tree. But that's your sliver. The pathways you choose are like a fingerprint. But when we look at AI thoughts, such as these generated stories, they tend to cluster much more, both in style and substance. Other researchers wondered if this could be fixed. So they did an experiment where they gave writers 10 different AI personas, each with wildly different cultural perspectives. a Latin American magic realist, a dystopian hard sci-fi writer, etc. And it worked. People creating with these different AI personas seem to create a pool of stories where the diversity was restored. But here's the catch. The diversity of the stories only came from the human design personas in the first place. And other studies showed this. Once you seed an AI with a unique persona, the same problem persists. If you ask it to tell a hundred stories, researchers found the output contain echoes, combinations of the same ideas repeating. So if you see just one output, it may look novel, but after you see many, you realize they're all benol echoes of the same thing. The AI doesn't generate new diversity as well as humans. It remixes what humans feed it. And we can model this by thinking about thoughts as similar to rolling dice at each branch as you select the next word. True random paths are maximally diverse, but they're all meaningless nonsense. AI's dice are weighted heavily biased towards what's been said before in its training data. That's why these paths [music] will make sense, but they will always cluster and echo each other. You can't [music] escape the weighted dice.
But human thought is different. [music] It's like rolling weighted dice, but it's weighted by your unique life experience [music] and what matters to you. And this waiting is always changing. This creates a diverse exploration within thought space that is also meaningful. you will wander into territory nobody else has been [music] and the AI couldn't reach. The key point is a large language model is optimized [music] to continue thought not generate new thought. The creativity is in the seed thought and its [music] continuation the human input AI has no inbuilt curiosity or preferences. know why it works better the more human thought you seed it with [music] and direct it with.
Recent studies also confirm this. AI is weaker than humans at divergent thinking, generating truly novel starting points, but it's highly effective at convergent thinking, refining and executing once you give it direction. And the MIT writing study we saw called this out as their most interesting result. They tested one final group, human first and then access to AI. The students who outlined their thinking first then used chat GBT, their brain connectivity [music] was actually higher than the human only group. The difference was the starting point human [music] first or AI first. So realize that asking common questions to AI constantly will lead to atrophy. Like using GPS every day in your own neighborhood, you'll lose the map in your head. But asking rare questions you've never asked before to an AI, the ones specific to your experiences, the connections only you could make, that's different. Finding the question itself is the work.
So, what about this magic gumball machine? Well, there is no free lunch. The machine can execute from any starting point I give it, but it can't wander the way I do through conversations, wrestling with confusion until I find something I wasn't even looking for. Out of the infinite possibilities, I'll only find a handful of these paths. Those paths have [music] my fingerprint. an AI generating a random script or video based on my past [music] work will be a distant echo, a very blurry fingerprint, because the chance of it finding all my next thoughts is zero. If instead I seed it with my idea, it might seem better for a moment, but it'll fail for the same reason. And ironically, the only AI input that properly reflects me is the one that has been given my entire thought, which only comes from the struggle to have that thought. Because generative AI is all about small inputs and large outputs. When you have small inputs and large outputs, there is simply no way for you to put all of your intention to specify all the things that need to be specified when making art. And so the input is the work. The input is the value. And so think of this. In a world where the cost of answers [music] is dropping to zero, the value of the question becomes everything.
[music] Thank you so much for watching this video. If you've made it this far, then please join my email list. As you know, the algorithm's in charge and I'd love to be able to reach out to viewers directly. I'm working on a new project I'm excited about, and I would like to hear from viewers about potential video ideas you'd like to see in 2026. So, when you fill that out, drop your idea in the box, and I'll reach out.