AI's Learning Curve Is Steep. So Is the One Behind It.
AI doesn't just save time. Learning to use it well makes everything else faster to learn. Two curves, stacked, with my own career and whole education systems both starting to measure the lift.
Ash Youssef
· 12 min read
The first curve: getting good at AI is the hard part
AI is like a skyscraper with the fast lift only available from the fifth floor. There are stairs and escalators scattered around the foyer and lobby, and lots of advice on the best route up. Most people stop there, look around the foyer, and decide the place isn't really for them. "I have no intention of walking up five flights of stairs!" they say to themselves.
The first curve is the climb to the fifth floor. Take the stairs. It's worth it.
Working with AI properly takes deliberate effort. You learn to give it context. You learn to give it intent. You learn to teach it what your work is for, and to preload it with how you think, so the next answer matches your standards instead of an average of the internet's. Intent is the unit of value here: an AI agent will execute well only against intent that's self-contained. That's not natural for most people. Most of us live in our own heads, where context is implicit. With AI it has to be explicit.
The institutions are admitting this. Article 4 of the EU AI Act, in force since February 2025, legally requires every school, university, and any organisation that deploys AI to ensure "a sufficient level of AI literacy" among the staff who use it. The regulators conceded the point: you can't deploy without first teaching people the tool. UNESCO has published a global AI Competency Framework for students and for teachers. The OECD and the European Commission have built an AI Literacy framework that will sit alongside reading and maths in PISA 2029.
The cautionary tale of what happens when you skip this is South Korea. In March 2025 the country launched the world's most ambitious AI-in-schools rollout: an "AI Digital Textbook" for maths, English, and informatics across primary and secondary levels. By the end of one semester, the textbooks had been demoted to "supplementary material" and the teachers' union had sued the Minister of Education. The blocker wasn't the technology. 98.5% of teachers said pre-rollout training was insufficient. They had been handed a tool without being taken up the first curve.
Even Sal Khan, founder of Khan Academy and the most public bull on AI tutoring, walked back his 2023 optimism in April 2026: "For a lot of students, it was a non-event. They just didn't use it much." The tool was there. The curve wasn't climbed.
The second curve: my experience as one data point
Three years ago I was confident in maybe twelve technologies. Today I work across more than sixty. Without AI, in the same time, I might have added five or six, on a good run.
The way it actually happened in the early days was simple. I'd write something in PHP or JavaScript, ask AI to turn it into Swift or Kotlin, then study the result. That move alone expanded my world from web into iOS and Android. More importantly it taught me how different languages think. I already knew how to program. I just didn't know all the dialects.
What changed for me is the order of learning. The old way: months of theory, then practice, then projects. The new way: I build first, with AI as the experienced colleague sitting next to me, and only then do I circle back to the concepts. They land harder when I've already shipped something that uses them. It feels like a fundamentally different learning model, and it works surprisingly well.
The compounding payoff isn't financial or even practical. It's psychological. Knowing that almost no client brief, no platform, no problem really intimidates you anymore is priceless. I honestly can't imagine a better feeling for any developer than reading a brief and thinking, yeah, I can build that.
It isn't just me
The reason I'm writing this isn't to brag about a personal tally. It's that the same compounding shows up in much larger samples, often in the places that need it most.
In Edo State, Nigeria, the World Bank ran a six-week trial in mid-2024. 800 senior-secondary students used Microsoft Copilot for English, with their teachers in the room. The result: roughly 0.31 standard deviations of measured learning gain, equivalent to nearly two years of typical schooling, in six weeks. The lowest-performing students gained as much as the rest. The lift carried into unrelated end-of-year exams.
In the UK, Eedi's adaptive maths tutor ran a randomised controlled trial across 20 schools and 2,901 Year 7-8 students in the 2023-24 academic year. 10 to 15 minutes per week of use produced 2 to 4 months of additional maths progress. That's a meaningful uplift on a tiny dose.
The aggregate picture is similar. Khanmigo went from 40,000 students to 700,000 in one academic year. MagicSchool passed 6 million educators by October 2025, more than the entire US public-school teacher count. RAND's 2025 survey found 53% of US K-12 teachers used AI for school in 2024-25, up from 25% the year before. Adoption doubled in twelve months.
What I'd take from this, with the South Korean caveat firmly in mind: the second curve is real, but it sits on top of the first. The teachers and learners getting the lift are the ones who climbed the literacy curve before they touched the tool. Estonia is now rolling ChatGPT Edu to every upper-secondary student, with Education Minister Kristina Kallas framing the goal as AI that is "designed to guide thinking rather than provide ready-made answers." That's the first curve baked into the deployment.
Stefania Giannini, UNESCO's Assistant Director-General for Education, compares the shift to "the invention of writing and the printing press." That sounds melodramatic until you sit with the numbers above.
Where this leaves the rest of us
The principle isn't about code. It's about the speed at which you can move from "I don't know how to do this" to "I can do this." That speed has changed. For anyone who's prepared to climb the first curve, the second one bends the way it has never bent for a working professional before.
Ministers from very different starting points are converging on the same message. France's Élisabeth Borne calls AI "a supporting brain" that "must not dispense with exercising one's own brain." Singapore's Chan Chun Sing frames the goal as "the Pedagogy of One." The UAE has made AI a mandatory subject from kindergarten to grade 12, with roughly a quarter of content devoted to ethics. Saudi Arabia is training 500,000 teachers and rolling AI basics to 6 million pupils. These aren't fringe positions. This is what national education looks like in 2026.
The risk is the South Korean one, but at individual scale. If the only way you ever use AI is to type a question and accept whatever lands back, the first curve never starts and the second one never arrives. The cost isn't dramatic, which is what makes it easy to miss. It's slow, compounding under-use of a capability that's available to you, against peers who took the time to learn how to brief it, push back on it, and build it into how they work.
That's the bet worth making, regardless of what you do for a living. Climb the first curve deliberately. Then watch what happens to everything else.
A generational read
There's a sub-pattern in the adoption data worth pausing on. On raw usage, the younger you are, the more you use AI at work. 83% of Gen Z, 73% of millennials, 60% of Gen X, 52% of boomers (LSE, 2025). Same direction across most surveys. By volume, the kids are out in front.
But volume isn't value, even at work. The LSE figure measures workplace use; what it doesn't measure is the kind of work AI is being put on. Gen Z's workplace AI tends to be tactical and high-frequency: drafting captions, formatting, summarising, generating content for channels. The economic density per hour of use is low. Millennials are the executors. Heavy daily users running AI through delivery work, where the intent is set by someone else and the bounded job is to ship. Gen X is disproportionately the cohort pointing AI at the decisions that actually move money: vendor choice, restructure, pricing, hiring, strategy. The Randstad and SurveyMonkey workplace reports both put Gen X in that bridge position. High enough on adoption to count, senior enough on org charts for what they use it for to compound.
Stack volume against intent and a results curve shows up. Volume drops with age. Intent rises with age. The two together get you peak useful output somewhere around Gen X. Then the curve resets at the boomer end, where volume falls off so far that high intent can't recover it.
And millennials aren't being left behind. They are carrying the day-to-day weight of AI execution and they will inherit the strategic seat over the next five to ten years. The current picture is a snapshot of who's where in their career, not who'll end up running the show.
Three readings of why Gen X is currently in the bridge position. One: career stage. Forties and fifties is when many of us run things, so AI lands on work that already moves money. Two: tech-wave history. The cohort that came of age on the PC, lived the dot-com bust, learned the cloud, deployed mobile, and survived social has adaptive muscle that pre-dates AI; AI is just the latest wave to ride. Three: a quieter possibility. Gen X tends to use AI deliberately rather than constantly. What gets reported as scepticism, from inside, can look like discrimination about what's worth handing over.
I don't think it's any one of those. Probably all three, weighted differently by person. The wider point is the same. Adaptability is the currency in the AI-onsphere, but it isn't enough on its own. Without a fixed direction, you'll spend hours circling. The people getting the most from AI right now are the ones with adaptability and a clear sense of what they're trying to do with it. Right now, that intersection sits disproportionately with Gen X. It won't stay that way for long. The generations behind us will catch up on direction the same way they already lead on volume.
There's an awkward implication in this for the teaching workforce. Most Western teachers sit squarely in the millennial and Gen X bands, the same generations the data says are best positioned to use AI for high-value work. Yet the country-by-country evidence earlier in this post says current teachers aren't equipped to teach it. Both can be true. Adapting AI to your own work is one skill. Teaching someone else how to climb the curve is a different one. The teachers personally climbing, often against the institutional grain, aren't yet a critical mass. That gap is exactly what the next question is pointing at.
Who teaches the climb?
Which leaves a question I keep coming back to. AI is new. It moves faster than any prior shift in working tools. And there's a third quality, harder to name, something about the way the tool changes shape under your hands as you use it, the way teaching it well is half the job of using it well. And, teaching that you can, quite literally with your own words magic the tool into a better shape that works for you.
The studies pulled together for this post all point in the same direction. The countries getting genuine learning lift from AI are not the ones who handed the existing teaching workforce a new tool and walked away. They are the ones who built a specialism around it. Estonia paired its national rollout with intent-led prompt design and teacher training. Singapore published an AI in Education Ethics Framework alongside the tools. Saudi Arabia is putting 500,000 teachers through mandatory training before the six million pupils touch it. South Korea, the one that skipped this step, is the cautionary tale we've already met.
I suspect we're about to see a new core discipline emerge alongside language, maths, and science. Not "computing" or "ICT": those subjects exist already and they aren't the same thing. Something that sits at the intersection of critical thinking, applied epistemology, and craft. How to brief. How to verify. How to push back. How to compound your own learning by using a tool that learns alongside you. PISA 2029 will start measuring it. The UAE has already mandated it. The UK, US, and EU are circling.
Whichever way you slice it, this is a major shift. The people teaching this well in ten years won't look much like the people teaching computing today. They'll be specialists in something that hasn't quite been named yet. And the gap between students who get them and students who don't will be the educational fault-line of the next generation.
If you're stuck on where to seriously start on your AI journey, wherever you are with it, the thing worth aligning first is intent. Align yourself and your AI work with the intent of your studies, your career, your business, your cause, whatever you're trying to build. AI works a whole lot better when what it's working with is angled by your intent. Nate B Jones writes on this better than I do; his Substack is the best ongoing read I know on it.