Same Screen, Different Worlds: Karpathy, Pocock, and the Widening AI Divide
Same technology. Different experiences. Radically different realities. Why AI is advancing faster than our understanding of what it can do—and why access alone isn't enough.
Two people open ChatGPT. Same interface. Same little text box. Same blinking cursor. One asks it to rewrite an email, gets something that sounds suspiciously like a corporate HR department wrote a birthday card, and concludes that artificial intelligence is mostly overhyped autocomplete.
The other is orchestrating multiple AI agents to research, write software, analyze data, debug problems, and execute projects that would have taken a team weeks to complete.
Same screen. Completely different worlds. And apparently, the distance between those worlds is getting bigger.
A recent observation from Andrej Karpathy, paired with a fascinating conversation between developer Matt Pocock and AI educator Nate Herk, helps explain what's happening. But I think there's another piece to this puzzle that deserves attention.
The technology is advancing faster than most people's understanding of what the technology can actually do.
And we're beginning to see the consequences.
Karpathy's Funnel: Eight Billion People, Four Different Realities
On October 11, Andrej Karpathy revisited an observation he'd made in April 2026.
Karpathy is an OpenAI cofounder, former Tesla AI director, and the guy credited with coining the term vibe coding in February 2025. So he's not exactly a casual observer of this stuff.
His April observation was that people were increasingly talking past each other about AI capabilities. Some were experiencing extraordinary advances while others saw little more than occasionally useful chatbots.
Six months later, he returned to that idea.
Except now, the gap wasn't just growing. It had become what he described as a sharp funnel. His admittedly rough, back-of-the-napkin estimates look something like this:
Roughly 6 billion people have barely interacted with large language models at all.
Another 1–2 billion use free chatbots occasionally. They see a somewhat better search engine, a writing assistant, maybe something that can help plan a vacation.
Around 20 million people use frontier-grade AI professionally, particularly in software development and mathematics. They're watching agents complete complex projects that previously required weeks or months.
And at the narrowest end, perhaps 5,000 people inside frontier AI laboratories have access to experimental capabilities the rest of us have barely glimpsed.
Karpathy describes research environments involving enormous groups of agents collaborating on software, cybersecurity, mathematics, and scientific discovery. These numbers aren't census data. They're illustrative estimates, not a scientifically measured breakdown of humanity.
But the fundamental observation is fascinating.
At one end of the funnel, somebody asks a chatbot for a chicken recipe.
At the other, researchers are figuring out how to manage and verify work being performed by increasingly autonomous AI systems. And both groups are forming opinions about the state of artificial intelligence.
No wonder we can't agree on anything.
Just Because You Have a Fast Car Doesn't Mean You Can Drive It
Here's a gross oversimplification that I think actually helps explain this.
Just because you have access to a fast car doesn't mean you know how to drive it. And just because you're an extraordinary driver doesn't mean you have access to a fast car. Hand somebody the keys to a Formula 1 car without the necessary training, and you're probably going to have a very expensive afternoon. Meanwhile, an experienced driver can squeeze remarkable performance out of a vehicle with a fraction of the horsepower.
AI works similarly.
Karpathy is largely describing the capability side of this equation.
Different models. Different tools. Different access levels. Different amounts of computational horsepower. But capability alone doesn't determine what someone can accomplish. There's also the matter of knowing how to use it. And that brings us to Matt Pocock.
Matt Pocock and the Skill Nobody Realizes They Need
In an October 16 interview with Nate Herk, developer and educator Matt Pocock described how dramatically his own workflow has changed. He spent years writing software the traditional way.
Character by character. Function by function. All the familiar programming rituals we've been performing for decades.
Then AI coding agents became capable enough that he effectively stopped writing code manually.
Now he describes managing multiple agents working on separate projects, checking their work, and producing software at a pace that would have been almost unimaginable a couple of years ago. But what's interesting isn't simply that Pocock has access to powerful models. It's what he believes makes someone effective at using them.
Communication.
Not memorizing elaborate prompt structures. Not collecting magical incantations from LinkedIn. Not necessarily having the newest, biggest, most expensive model. Communication.
Pocock argues that the best results come from clearly expressing what you care about, what you're trying to accomplish, and what constitutes a successful outcome. Which sounds almost insultingly obvious until you realize how spectacularly bad humans can be at explaining what they actually want. Ask any software developer who's built exactly what a client requested, only to discover that it wasn't remotely what the client intended.
Pocock even developed a tool called Grill Me, which essentially instructs an AI to interrogate the user until the missing requirements, assumptions, and potential problems start revealing themselves. It's a fascinating reversal.
Instead of humans learning to speak computer, we're increasingly teaching computers to help humans figure out what they meant in the first place.
And that's only part of it.
Pocock emphasizes the importance of creating environments where agents can perform reliably. Good software architecture. Tests. Feedback loops. Consistent terminology. The ability to recognize and correct mistakes.
He describes experimenting with specification-driven development, watching his codebase deteriorate, and eventually realizing that improving the environment in which the agents operated was more important than simply generating increasingly elaborate specifications.
In other words, the fundamentals still matter. The typing just matters a whole lot less. And this is where the fast-car analogy starts getting interesting.
Karpathy is describing horsepower.
Pocock is describing driving ability.
But I think we've got a third problem. A whole lot of people don't even know how fast the car has become.
Yesterday's Limitations Become Today's Assumptions
Think about how we've traditionally learned technology. You buy a computer. Learn the operating system. Become proficient. Maybe a new version comes along every few years, and you adapt.
The internet spread remarkably quickly by historical standards, but meaningful adoption still unfolded over many years. AI is operating on a very different timeline. Models improve. New tools appear. Capabilities emerge. Entire workflows that were impractical six months ago become increasingly accessible. And even people who are deeply immersed in this technology are struggling to keep up.
I work with AI. Build with AI. Develop AI systems. Write about AI. Spend an almost certainly unhealthy amount of time thinking about AI. Seven days a week. AI from the firehose. Somebody should probably stage an intervention, but they'd have to get past my agents first. And even I struggle to keep pace.
So what exactly are we expecting from someone who runs a restaurant, practices law, manages a small business, or sells real estate? They have jobs. Families. Responsibilities. Lives that don't revolve around the latest model release or agent framework.
Here's where the problem gets particularly interesting.
Someone tries AI in January. They ask it to do something complicated. It fails. Fair enough. They conclude AI can't do that particular task. Six months later, the technology may be perfectly capable of accomplishing it.
But that person doesn't know. And why would they try again? They already decided it doesn't work. Their understanding of AI has effectively been frozen at the moment they stopped experimenting.
The technology keeps updating. Our mental models don't necessarily update with it.
And the faster AI improves, the more consequential that mismatch becomes.
This isn't merely an adoption problem. It's a perception problem. People are forming long-term opinions about a technology whose capabilities can change dramatically in a very short period. Sometimes the original criticism was perfectly valid. It just isn't valid anymore. But the person holding that belief has already moved on. And that's one of the mechanisms I suspect is contributing to Karpathy's widening funnel. ##
Siri May Have Accidentally Trained Us to Talk to AI Incorrectly
There's another dimension to this that I've been observing firsthand.
Talking to AI is a skill. And I don't think we appreciate how unfamiliar that skill can be.
For years, traditional voice assistants taught us to communicate with computers through carefully structured commands.
"Hey Siri, set a timer."
"Hey Siri, call somebody."
"Hey Siri, remind me to get coffee."
That was about the extent of the meaningful relationship. Try explaining a complicated, half-formed idea to the Siri most of us grew up with, and you'd probably get an apology, an irrelevant search result, or some spectacularly useless response.
So we adapted.
We learned to speak to computers like computers. Short commands. Keywords. Carefully phrased instructions. And now we're putting sophisticated conversational AI in front of people and telling them to just talk naturally.
Except many people have no idea what that means.
During the week of October 12, 2026, I spent time with about a dozen people discussing AI and helping them understand how to use it.
Three of them really got it. They could see the potential, understood the collaborative nature of the interaction, and started imagining how it could improve their work. The others struggled. Not necessarily because they lacked intelligence, curiosity, or technical ability.
They struggled with the interaction itself.
Some had a microphone sitting right there, capable of turning a spoken conversation into a genuinely productive exchange with AI. And they were still pecking away at the keyboard. Because talking to a machine feels weird.
I get it. I never developed much enthusiasm for talking to Siri either. Years of disappointing interactions didn't exactly make me want to develop a conversational relationship with my phone. But modern conversational AI is a different experience. I can hit the microphone and start talking. I can wander through an idea. Change direction halfway through. Introduce something seemingly unrelated. Contradict myself, reconsider, and come back around. And the system can often reconstruct what I'm trying to communicate.
It can turn my brain slop into something coherent. That's extraordinary. Not because it understands perfectly every time. It doesn't. But because I no longer need to arrive with a fully formed instruction before beginning the interaction. The conversation itself can become part of the thinking process. We explored this earlier in Prompting Is Dead. Long Live the Conversation.
The point wasn't that structure or precision had become irrelevant.
It was that communicating intent had become more important than memorizing the perfect incantation. And that takes practice.
You Have to Do the Reps
This is where Pocock's observations become especially valuable. He talks about climbing a trust ladder with AI agents. Start with manageable tasks. Observe what the system does. Learn where it succeeds and fails. Gradually increase the responsibility you're willing to delegate.
Essentially, learn to drive before entering the race.
Karpathy, meanwhile, encourages people to challenge AI with their biggest, hardest problems.
Those ideas might sound contradictory. I don't think they are. Karpathy wants people to discover the ceiling of the technology. Pocock wants people to develop the skills and judgment required to operate safely and effectively beneath that ceiling.
One reveals what's possible. The other helps you become capable of doing it. And both require experience.
You can't become proficient at working with AI by attending a presentation, watching a demonstration, or collecting a PDF containing fifty prompts guaranteed to transform your life.
You have to use it.
Talk to it.
Challenge it.
Let it challenge you.
Learn what it does well, what it does badly, and how to recognize the difference.
For some people, this comes naturally. For others, particularly those whose previous computing experiences involved rigid commands and unforgiving interfaces, it's a substantial adjustment.
That's not an intelligence problem. It's an acclimation problem. And I think we're underestimating how much education and hands-on experience will be required to help people make that transition. Not simply training on a particular model. Training people to work with systems that can converse, reason through problems, use tools, and increasingly execute sophisticated tasks.
That's a new kind of computer literacy.
The Divide Is Feeding Itself
Here's where Karpathy and Pocock's observations really collide.
Karpathy describes the widening gap in exposure to AI capabilities.
Pocock explains why exposure alone doesn't translate into effective utilization. But there's a third variable connecting them.
What we believe AI can do determines how we use it. And how we use it determines what we discover it can do.
Think about the feedback loops. Someone has a disappointing experience with AI. They conclude it isn't particularly useful. They stop experimenting. Their skills don't develop, and their understanding of its capabilities becomes increasingly outdated.
Meanwhile, someone else keeps experimenting. They discover something new.
That enables them to tackle a bigger problem. Which teaches them something else. Which unlocks another capability. And the cycle continues.
Two people can begin with the same technology and end up with radically different capabilities simply because one kept exploring while the other stopped.
Now add rapidly improving models to the equation. The active user benefits from each new advance. The inactive user might not even realize those advances have occurred. One feedback loop compounds experience. The other compounds outdated assumptions. And the divide widens.
We're not just dealing with unequal access to technology. We're dealing with unequal rates of adaptation to it. That's a much harder problem to solve.
The Dunning-Kruger Adjacent Problem
There's also an uncomfortable cognitive component here. People don't know what they don't know. And AI introduces an interesting twist.
Someone can be genuinely knowledgeable about a technology as it existed six months ago and confidently misunderstand its current capabilities.
Their expertise hasn't necessarily disappeared. Their information has aged.
This isn't strictly the Dunning-Kruger effect, which concerns people's ability to assess their own competence. But it creates a similar problem: confidence can outlast the evidence that originally justified it. And here's the part we need to be careful about. The same thing happens at the other end of the funnel.
Just because someone has access to extraordinarily capable AI doesn't automatically mean they're making better decisions. They might simply be producing more impressive mistakes. We've explored this in Caterpillars of Cognition and the Burden of Being a Butterfly.
AI can expand human cognition. It can also become an extraordinarily efficient substitute for exercising it. A thousand agents confidently executing a terrible strategy aren't necessarily an improvement over one human making a mistake. They're just a thousand times more enthusiastic about it.
Which is why Pocock's emphasis on verification, judgment, and engineering discipline is just as important as Karpathy's observations about capability.
Horsepower is wonderful. Brakes are also pretty damn important.
The Next Billion Users Need More Than Access
In March 2026, we explored the enormous disparity in AI adoption in 84 Percent of Humanity Has Never Used AI. Let That Land.
That earlier article used a different estimate and classification of AI users, but the underlying concern was similar. An enormous portion of humanity had barely experienced this technology. And many of the people who had experienced it were barely scratching the surface.
Karpathy's funnel puts that disparity into sharper focus. Pocock's experience helps explain why simply expanding access won't eliminate it. And my recent experiences working with people suggest something else.
We need to rethink what AI education actually looks like. Not another webinar about writing better prompts. Not another list of ten incredible things ChatGPT can do. Those can be useful introductions, but they don't develop lasting fluency. What people need is an opportunity to experience what AI can accomplish in the context of their own lives and work.
A business owner exploring a workflow they've always assumed required several employees. A teacher discovering how to create individualized learning experiences. A real estate professional learning to research, organize, and evaluate opportunities in ways that previously required substantial manual effort. Someone who's never written software discovering they can build a working prototype around an idea they've carried for years.
And throughout that process, learning when the system can be trusted, when it needs supervision, and when its answers require independent verification. Because the goal isn't to convince everybody that AI is magical. It isn't.
The goal is to help people develop an accurate, continuously updated understanding of what the technology can actually do. And perhaps more importantly, what they can accomplish with it.
That requires more than access.
It requires experience. It requires curiosity. And it requires the willingness to reconsider something we thought we already understood.
Same Screen. Different Worlds.
I think this is what makes Karpathy's observation so important.
We're having a global conversation about AI as though everybody is evaluating the same technology.
We're not.
One person remembers an unreliable chatbot from an earlier model generation. Another is using sophisticated agents to accomplish work that previously required an entire team. Someone inside a frontier laboratory may be watching systems operate at a level of capability neither of those people has encountered. All three can be describing their actual experiences. And all three can walk away with radically different conclusions.
Pocock reveals the other half of the problem. Even when people gain access to advanced capabilities, they don't automatically know how to use them. Communication, judgment, context, verification, and experience still matter. And there's no guarantee that giving someone a more powerful model will magically supply those skills.
Which brings us back to the two people sitting at their computers. Same screen. Same blinking cursor. One sees a chatbot. The other sees a collaborator, a research team, a software development department, or an entirely new way to turn ideas into action.
Neither is necessarily seeing the whole picture. But one of them may have stopped looking. And that's the divide we should be paying attention to. Not because everyone needs to become an AI researcher, developer, or seven-day-a-week AI obsessive. Trust me, the firehose isn't a lifestyle recommendation. But because we're approaching a point where yesterday's experience with AI may be a remarkably poor guide to what it can accomplish today.
The technology keeps changing. Our assumptions don't necessarily change with it. And the longer those assumptions go unchallenged, the farther apart our experiences become.
The problem isn't that billions of people haven't learned what AI can do.
It's that millions of people think they already have.
While we're busy arguing about what AI couldn't do the last time we tried it, the technology has already moved on.
Same screen. Different worlds. And the distance between them keeps growing.
Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.
Comments