The New Literacy

“Once you learn to read, you will be forever free.”
— Attributed to Frederick Douglass

What does it mean to race towards upskilling humanity?

I talked to a family friend who moved to a rural town in Texas. She has never used ChatGPT. Her husband is an engineer at a startup, and they are both afraid he is going to lose his job. I told her the thing I tell everyone now: he has got to become a power user of AI.

She looked at me and said, "I don't even know what that means. Describe it to me like a child." I think that most people don’t know what it means. And yet we can’t aspire towards mass literacy if most people aren’t even able to understand what that literacy looks like.

We are good at measuring literacy in a language we already speak. Every one of us speaks at least one actual language, so we know exactly how lost we are in a language we don't speak. You know what fluency feels like, and so you know what it feels like to not have it. But AI literacy is harder to feel, because most people have no fluent version to hold up against. Seeing what literacy looks like clearly is itself the beginning of literacy. So let’s try to do that.

I think about it in five levels. The levels are actually a spectrum, but they give you a sense of how deep the rabbit goes, what people do at each level, and where the real leverage lives. What level are you at? Imagine what you could do if you were one level deeper.

Level 1: Prompt Engineering

This is the shallow end of the pool. Prompt engineering is about communicating effectively with AI to generate what you want. Tell the AI what role to play. "Imagine you are a world-class marketer with a knack for understanding target audiences. Your writing is informative, entertaining, and inspiring." Be specific. Don't ask "What are some good business ideas?" Ask: "Generate five innovative business ideas in sustainable technology, each with a brief description of the product, the target market, and the first wedge."

LLMs get more powerful by the month. You can now hand it a two-hundred-page contract and ask where you are exposed. You can upload a year of messy bank statements and ask where your money actually goes. You can paste a dense research paper, the kind that used to require a graduate degree to parse, and say, explain this to me like I am twelve. You can ask it to research a question the way a roomful of analysts would. It goes out, reads dozens of sources, and comes back with a cited report.

All of this from a chat box. No code. No training. Just words, aimed well. That is the floor. But, fundamentally, you’re still just chatting with a bot. Prompt engineering is the keyboard literacy of this era. Everyone needs it. It’s incredible. It’s also the thinnest layer of the stack.

Level 2: Tools, Pipelines, Agents and Vibes

The second layer is where the power users of AI tools get into real workflows.

A marketer building a campaign uses AI to define audience segments, draft variant copy across each segment, produce dozens of ad variants, schedule placements across platforms, watch the results in real time, and adjust the campaign mid-flight. The workflow between each of these AI tools is automated. They call that a “pipeline.” It used to take a team of eight. Now it takes two who know the pipeline: the creative and the analyst.

A designer uses Canva's Magic Studio, or Figma plugged into a frontier model, to go from a product requirements document to a wireframe to a polished mockup in an afternoon.

A 3D artist takes a 2D concept image, runs it through an image-to-3D tool, refines the mesh, rigs it, animates it, drops it into a game engine. Each step used to require years of practice. Now the artist's job is to have taste and to guide the tools and the output.

The creator of a video starts by crafting a script with ChatGPT. They then create reference characters, locking in a face, a wardrobe, a look that has to stay consistent across every shot. Then they generate key frames, the anchor images that define each beat of the scene. Then they hand those frames to a video model like Seedance and let it generate the motion between them. Then they pull the clips into an AI-enabled editor like Runway, where they cut, color, score, and stitch the whole thing together. The result is a blockbuster-quality film.

Then there are agents. This is where the Level 2 starts to feel like a different building. A prompt gets you an answer. An agent gets you an outcome. You hand it a goal instead of a question, and it goes off and does the work. It breaks the job into steps. It decides on what tools to use, it checks its own results and tries again when it gets one wrong. It comes back when the thing is done. The shift people describe is from a copilot that suggests to a teammate that executes. The early models waited for you to type the next instruction. The newer ones take a single instruction and carry it across dozens of steps on their own.

Here is a real example: A developer asked his OpenClaw agent to book a dinner reservation. The agent tried the normal way, through the restaurant's booking site. The booking failed. So it reached for an AI voice tool, gave itself a synthetic voice, coded itself the ability to make a phone call, called the restaurant on the phone, talked to a human, and got the table. It did all of this automatically.

And then of course, there is vibe coding. A non-engineer pairs with Claude Code or Codex. An hour later they have a working prototype that they can pass to engineers for production. Those who persist can turn that prototype into a real product in months. Bill Gates told me about a financial analyst who couldn’t code but vibe coded a tool that can do his analyst job. He quit the job, sold the tool to his company and turned it into a business.

A teacher who builds a grader that marks essays in her own rubric and her own voice. A nurse who builds the patient-tracker that her hospital's million-dollar system never included. A shop owner who builds the inventory tool that finally fits how her store actually works, instead of bending her store to fit the software. Someone who points an agent at every food vendor they negotiate with on WhatsApp and lets it haggle on their behalf. All without knowing how to code. This is the world that we now live in. People are superhuman.

And yet almost no school teaches any of this.

This is the highest leverage skill that one can learn. It’s not hard and it's a superpower.

Level 3: Creating In Code

It’s popular to say that you don’t need to learn to code anymore. I completely disagree. Someone built every one of the tools I just described above. They did it by architecting, writing and guiding AI to write code. They were profoundly technically proficient.

When people think about “learning to code” they imagine that the job is still typing code.

To be clear, that skill is quickly becoming irrelevant. The AI types the code now, and it types it faster than any human ever could. Think of it as the fastest construction crew on earth. It can pour a foundation, frame a wall, and run the wiring in seconds. What it cannot do is decide what building to make. Tell it to "build me a house" and it will hand you something shaped like a house that floods the first time it rains.

There's a ceiling that no-code and prompting hit, and Level 3 is everything above that limit. The tools in Level 2 run inside something someone else built. But you hit a wall the moment you want something the platform doesn't already do. At Level 3, you build what you want:

You make software talk to other software. You wire a model to your hospital's actual patient records, your store's real payment system, a live sensor feed, 3 APIs that were never designed to work together. You can handle real scale and real stakes. Ten users is a toy. Ten thousand paying users, with real money and real private data, is an engineering problem. You need servers that don't fall over, data that stays secure, a system that survives strangers attacking it. Level 3 is about knowing how to architect software, about building in bite-sized blocks and reading the code rather than one-shot prompts. It’s about knowing which are the right libraries or languages to use for the job, or how to build reusable components. It’s about being able to evaluate whether an open source contribution is robust or malicious.

The people who operate at this level can produce products like OpenClaw. Let’s look at how OpenClaw was built, since it’s one of the most successful open source projects in history. OpenClaw was built by a veteran developer who sold his company for over $100 million. Peter Steinberger built this toy on vacation to fall back in love with coding, and it detonated. The project grew so fast that a friend told him it wasn't hockey-stick, it was a straight line.

This single developer made 6,600 commits in one month. Alone. With no team. In its first five months the project drew ~2,000 contributors and approached ~30,000 pull requests. He reviewed pull requests at a pace that bordered on algorithmic, making the design decisions that kept the project coherent as it grew from a simple agent into a full platform. He describes how he had five to six coding agents running in parallel at all times. He could never have built it without AI and he could never have built it if he could not guide the AI. Hearing him describe how he built it, it’s clear that the man was an engineering machine. It’s just that the definition of engineering has evolved. It’s now machine and man together.

This was one tool. Every tool that I mentioned in Level 2 was built by someone like Peter.

This is honestly the layer where the real power sits.

At risk of spending too much time on the importance of learning to code in a world of AI, I will say a few more things, because, as Jensen Huang, the founder of NVIDIA, describes it saying, “It is hurtful to convince all the young college graduates not to study software engineering because we are going to need more software engineers than ever.”

The bang-for-buck at this layer is staggering. A single person who can wield an API in its raw form can do what would have taken a team of a hundred people a few years ago. In the decades ahead, this will be the ultimate leverage.

As Naval Ravikant, the guru of Silicon Valley, put it: “AI won’t replace programmers, but rather make it easier for programmers to replace everyone else.” In a brilliant podcast, he describes why “every job is going to be disrupted by programmers”

“Does this mean that traditional software engineering is dead? Absolutely not. Software engineers… are now among the most leveraged people on earth…. Software engineers still have two massive advantages on you. First, they think in code, so they actually know what’s going on underneath…. When you have a computer programming for you…it’s going to make mistakes. It’s going to have bugs. It’s going to have suboptimal architecture. So it’s not going to be quite right. And someone who understands what’s going on underneath will be able to plug the leaks as they occur. So if you want to build a well-architected application, if you want to be able to even specify a well-architected application, if you want to be able to make it run at high performance, if you want it to do its best, if you want to catch the bugs early, then you’re going to want to have a software engineering background. The traditional software engineer is going to be able to use these tools much better.”

Boris Cherny, the inventor of Claude Code, says he basically hasn’t written a line of code, even though he is still “coding”. And yet he is building what the vast majority of the population of the planet is fundamentally incapable of building.

Software engineering job postings are rising again, up roughly 11% year over year by one jobs-data estimate, and up much more sharply from the post-ChatGPT trough. The labor market is not simply eliminating coders; it is restructuring around AI-fluent builders.

Sherwin Wu, OpenAI’s head of platform engineering, says that programming is becoming “literally incantations.” You can now “tell Codex” or “tell Cursor exactly what you want to do,” and “it’ll all go do it for you.” The current state of software, he says, is starting to look like The Sorcerer’s Apprentice from Fantasia: Mickey Mouse “finds the sorcerer’s hat” and “tries to do all these things.” The analogy works because the magic is real. “These incantations you can do [are] extremely high leverage,” Wu says, “but you kind of have to know what you’re doing.” In Fantasia, “Mickey goes wild,” “the brooms go crazy,” and “everything’s flooding.” He “sets the brooms off on a task and then goes to sleep.”

So yes, let’s give people the sorcerer’s hat of Level 2. But then we have to immediately start teaching people how to cast these spells properly, so that they can be true sorcerers.

Level 4: Applied AI in Advanced Fields

The fourth layer is where AI literacy meets domain mastery.

Demis Hassabis and his team at Google DeepMind have made some of the most important advances in AI but when he won a Nobel Prize it wasn’t for his work on AI. It was for his work that applied AI to protein folding, AlphaFold. A single protein structure used to take months or years of painstaking lab work to determine, and after decades of effort scientists had solved only about a hundred thousand of them. AlphaFold predicted the shapes of nearly every known protein, around 200 million, in a single year. It’s having a profound impact on pharmaceuticals and our understanding of biology. This is the realm of Level 4.

In drug discovery, AI is scanning vast biological datasets and doing massively parallelized experiments to identify drug targets, predict protein interactions, and shorten development cycles from years to months. In materials science, AI combined with high-throughput experimentation is exploring design spaces no human chemist could have mapped, surfacing new materials for energy storage, electronics, and aerospace. In finance, it is reshaping risk assessment, fraud detection, and algorithmic trading. In aerospace and defense, it is optimizing aircraft components and powering autonomous systems. In energy, it is balancing grids and predicting demand. In agriculture, it is reading soil, weather, and yield data to lift productivity per acre. In telecom, it is optimizing networks, predicting failures, and defending against cyberattacks in real time. In shipping, it is routing fleets, predicting delays, and trimming fuel costs. In construction, a single 40-minute drone flight captures data that would have required three to five days of manual laser scanning, and the high-precision mode reaches sub-10mm accuracy.

At this layer, the most valuable people in the world are not AI specialists. They are domain experts who have made themselves fluent in AI. The geneticist who can wire a model into her wet lab. The structural engineer who can hand a model a sensor feed. The agronomist who can build a fine-tuned crop-yield model. Jensen Huang tells students that if he were young today he would study the physical sciences. This is why. As intelligence becomes abundant, the advances move to wherever intelligence touches the physical world.

I was recently at a Gates Foundation event where I got a glimpse of the frontier of medicine. It was a whirlwind above my head, and I was scribbling notes, but in it, they described that a standard blood test has a 9 million to one compression ratio from the raw data to the data that the doctor sees. For EKGs, it is a 100,000 to one ratio. Computers can process all of the data that humans cannot. AI can look at your EKG today and tell that you had a heart attack years ago, even one you never knew you had, because the scar left a fingerprint in your heartbeat that no human reader could see. That’s the data that we throw away today. In finance, we process massive amounts of data to make billion dollar decisions, but in health we limit it. We trust our airplanes to fly themselves but we don’t allow AI to make decisions in healthcare. As I have heard from multiple people, it should be malpractice for a doctor to not be using AI. And when we see the benefits, the floodgates will open. The people coming through those floodgates with the innovations that save millions will be the people operating at Level 4.

Level 5: Foundational AI Technology

The fifth layer is the bedrock. It is where the large language models themselves get built. Where the training infrastructure gets designed. Where the chips that run AI get made. Where the power plants that feed the chips get planned. AI applications sit on top of foundation models. Foundation models sit on top of compute providers like AWS and Azure. These providers sit on top of data centers, chips, and power. Every layer down, the work gets more specialized, more capital-intensive, and more awe-inspiring.

My favorite layer is the one at the very bottom. The semiconductors. The places where all of this software shows up as electrons bouncing across chips that hold trillions of transistors. Let’s look into what it actually takes to make one machine on the TSMC assembly line:

In the book Chip Wars, Chris Miller describes that, “The best approach was to shoot a tiny ball of tin measuring thirty-millionths of a meter wide moving through a vacuum at a speed of around 200 mi per hour. The tin is then struck twice with a laser, the first pulse to warm it up, the second to blast it into a plasma with a temperature of around half a million degrees, many times hotter than the surface of the Sun. This process of blasting tin is then repeated 50,000 times per second to produce EUV light in the quantities necessary to fabricate chips.” Doing this required a laser that was more powerful than any that existed. “These lasers were monuments of machining in the best German industrial tradition. Because around 80% of the energy a carbon dioxide laser produces is heat and only 20% light, extracting heat from the machine is a key challenge. Trumpf [the company] had previously devised a system of blowers with fans that turned a thousand times a second, too fast to rely on physical bearings. Instead, the company learned to use magnets so the fans floated in air, sucking heat out of the laser system without grinding against other components and imperiling reliability.” “The company proposed a laser with four components: two “seed” lasers that are low power but accurately time each pulse so that the laser can hit 50 million tin drops a second; four resonators that increase the beam's power; and ultra-accurate ‘beam transport system’ that directs the beam over 30 meters toward the tin droplet chamber; and a final focusing device to ensure the laser scores a direct hit, come on millions of times a second.”

This one stage in the semiconductor industry's manufacturing processes was so fascinating that I have to go on. As Chip Wars put it, “every step requires new innovations. Specialized gasses in the laser chamber had to be kept at constant densities. The tin droplets themselves reflected light, which threatened to shine back into the laser and interfere with the system... the company needed industrial diamonds to provide the “windows” through which the laser exited the chamber, and had to partner to develop new, ultra-pure diamonds.” “It took [them] a decade to master these challenges and produce lasers with sufficient power and reliability. Each one required exactly 457,329 component parts”. All so that one supplier could make a laser for the EUV supplier.

These two companies then had to go to a third company, Zeiss, to manufacture a mirror that was so smooth that, “if these mirrors were scaled to the size of Germany… their biggest irregularities would be a tenth of a millimeter.” Then they had to direct this light, for which they had to invent mechanics so precise that “they could be used to aim a laser to hit a golf ball as far away as the Moon.” “The result was a machine with hundreds of thousands of components that took tens of billions of dollars and several decades to develop”.

And that all is used to produce one machine on the TSMC semiconductor assembly line.

This is the sort of engineering work that is required to support the future of technology. It’s like a fractal. The more you zoom in, the more you realize there is to zoom in on, infinitely. And the further I zoom back again, the more I see that it is near infinite in its complexity.

The people who built this will not be easily replaced. The semiconductor industry currently employs more than 277,000 workers in high-paying R&D, design and manufacturing jobs. These will increasingly all be enabled by AI. And they are certainly enabling AI.

Back To The Future of Jobs

I look back at most conversations that I see people having about AI and education and they feel like they are mostly focused on teaching people how to use ChatGPT to learn traditional school subjects. AI literacy is not learning to chat with machines. It is learning to move down the stack from consumer to creator. Almost nobody is discussing how we teach young people to actually do the work across these five layers.

Step back and look at the five layers as a whole. At the top, AI literacy is wide. Anyone can do it. Everyone needs to. Prompt engineering is the new keyboard. The next layer down, the workflows, is where the typical worker of the future lives. Marketers, designers, animators, engineers, finance analysts, lawyers, doctors, teachers. The day-to-day jobs of the next two decades belong to the people who can run AI pipelines well. The third layer, the code, is where small teams of leveraged builders create the tools that everyone else uses. The fourth layer is where AI meets the physical and scientific frontier. This is where the great civilizational advances will come from. Drug discovery. New materials. Energy abundance. The fifth layer is where civilization itself rests. The semiconductors. The data centers. The reactors that power them. The compilers that translate dreams into electrons.

The deeper humans delve into these layers, the more we advance society.

The conversations I see about AI in education focus almost entirely on using AI as a tool to teach the old subjects. There is a much more important conversation almost no one is having: how do we teach people the new subjects? How do we teach them to build, not just to consume? How do we move them from being users of AI to being creators with AI?

This is the literacy that will decide the next century.