Over two hundred young people joined The Genius Project this year. The youngest was five. The oldest was eighteen. They came from across the Caribbean, and over one month we gave away US$1 million in cash and prizes. No family paid tuition.
That is the headline. The part I care about more is the starting age, because it is the decision people question most and the one I am most certain about.
What Five-Year-Olds Actually Do
They do not write code. Anyone who tells you they are teaching Python to five-year-olds is selling something.
What a five-year-old does in our programme is sort objects into groups and then argue about which group a difficult object belongs to. They count things and notice that counting the same thing twice gives a different answer depending on how carefully you look. They play a game where a machine guesses which card they are holding, and they discover that the machine can be confidently, obviously wrong.
That last one is the entire point. A child who learns at five that a computer's answer can be wrong becomes a twelve-year-old who checks, and a twenty-year-old who does not forward the fake. We are not building programmers at that age. We are building the reflex to question, before the habit of accepting sets in.
By nine or ten the same ideas have names. Sorting becomes classification. The disagreement about the difficult object becomes a conversation about labelling and edge cases. The counting problem becomes measurement error. Nothing new was introduced. The vocabulary caught up to the intuition, which is the right order.
Then the Mathematics Arrives
The programme runs deliberately from tools to mathematics. Tools first because a young person needs to see a result quickly. Then the ground shifts.
Teens moved from prompting a model to understanding what a model is: training data, features, labels, the distinction between a system that learned a pattern and one that memorised the answer sheet. That means statistics, and statistics means mathematics. Distributions, probability, why a single result tells you close to nothing. Then Python, notebooks, and the very ordinary experience of code that runs perfectly and returns the wrong number.
This is where the drop-off happens, and I will come back to that.
What They Built
Teams were given real problems rather than clean exercises, and the work clustered into four areas.
Crime and community safety. Where incidents cluster, how gaps in reporting distort what the data appears to say, and what a model can responsibly claim about a place or a person. Several teams reached the conclusion themselves that a predictive model built on incomplete crime data mostly predicts where the reporting is. Policing agencies in far wealthier countries have paid consultancies large sums to arrive at the same finding later.
Poverty and access. Household budgeting tools, food price tracking, matching people to services they qualify for but do not know exist. The wall these teams hit, that Caribbean household data is thin and scattered, is the same wall professional teams hit.
Sport. Football and track data proved to be the best on-ramp to machine learning we have. Students already had intuition about the domain, which meant they could immediately tell when a model was producing nonsense. That is not a small thing. Most beginners cannot tell.
Ethics and responsible AI. Not a lecture module. Every team had to state who their system could fail, what data it should never hold, and what they would say to a person the model got wrong. Watching a fourteen-year-old explain why her model should not be used for hiring decisions did more for me than most policy discussions I sit in.
Across all four, students built actual machine learning models. Systems that took input, produced output, and could be demonstrated to be wrong.
A team that cannot say clearly what problem it is solving cannot build anything worth judging. That is true at fourteen and it is true in a boardroom. We just teach it earlier.
We Trained the Parents Too
A child who understands AI better than every adult in the household is not in a safe position. So parents ran their own track alongside their children.
The practical layer came first: account and privacy settings, what a chatbot retains, what should never be pasted into one, and how to spot a website built to harvest information. Then the harder layer, which is judgement. Telling a generated image from a photograph. Checking a claim before forwarding it. Recognising AI slop, the fluent and confident text that happens to be wrong, and understanding that the fluency is the trap rather than the reassurance.
The bar we set was modest and specific. A parent should be able to sit beside their child, look at the work, and ask one question that improves it.
The Hackathon
The month closed with the final hackathon. Teams presented to judges, defended their build, and answered for their choices. Congratulations to the winners, and to everyone who stood up and presented at all. Defending a technical build to a panel of adults is hard at thirty. A number of these presenters were not yet thirteen.
The Number I Am Publishing Anyway
Completion across all areas currently stands at roughly 15 percent.
I publish it because publishing only the flattering numbers is how a sector loses the ability to improve. Large open online programmes commonly report completion in the mid single digits. Fifteen percent across a month-long technical programme run in several countries at once is a result I will stand behind. It is also well below where this should be.
I know where we lose people. We lose them at the transition from tools to mathematics, and we lose them hardest where the connection is unreliable or there is nowhere quiet to sit. Both are addressable. Next year the mathematics section runs in shorter modules, the materials work offline, and any participant who goes quiet for three days gets a call rather than a broadcast email.
Why This Is Urgent
The Caribbean cannot wait for AI literacy to arrive through the university system. The people who will build and govern these tools here in 2035 are in primary school this morning. Teaching responsible use early is cheaper, faster and far more durable than correcting habits at twenty-five.
There is a version of the next decade where Caribbean young people are competent operators of tools built elsewhere, and a version where they build and govern their own. The difference between those two futures is decided by what we teach children now, while the habits are still forming. That is why we start at five, and why I would rather run this programme than almost anything else I do.
Who Paid for It
The Genius Project charges families nothing, which only works because organisations across the region put in cash and support. For 2026 that included StarApple AI, which funded prize money and supplied instructors and curriculum; Maestro AI Labs, which gave technical mentorship and lab time for the machine learning tracks; the Caribbean AI Association, which backed the programme regionally; 14West, which supported the hackathon and the prize pool; AI Trinidad and Tobago, which carried it to students across the twin islands; and Orbital Brand Science, which contributed in-kind support and helped reach families. I put in my own money and taught in it, as I have every year.
To every sponsor, judge, volunteer instructor and parent who gave up a month of evenings, thank you. To the students, you did the hard part.
The next cohort is open at beagenius.org. Tuition-free, from age five, no prior coding required.