Blogs
Let the machine type; teach our children to think
At MUST itself, the Vice Chancellor, Professor Pauline Byakika, the university's first female VC, heads an Institution that runs a data-science research hub that puts machine learning and AI to work on medical images, sharpening disease diagnosis across Africa. She has presented herself, even as a seasoned professor, as someone who must constantly learn, unlearn and relearn. If that is contextually true for a Vice Chancellor, it is true for a Senior Six student
Let the machine type; teach our children to think
By: Admin ., Journalist @New Vision
____________
OPINION
Dr Richard Kimera (PhD)
The world's most powerful AI bosses say software now writes itself. In Uganda, where the median age is about 17, that has students and parents asking whether a computer science degree is still worth the fees. It is. But only if we learn the right lesson.
A Senior Six student at Masaka Secondary School recently put a question to me that I have heard, in different words, a dozen times this year: “If the computer can now write the programs by itself, why should I waste four years and my parents' money studying software engineering?”
It is a fair question, and it deserves a fair answer, not the comforting one, and not the frightening one, but the true one.
First, the part that is real
Let me begin with what is real, because pretending otherwise insults the reader.
In March 2025, at a Council on Foreign Relations event in the United States, Dario Amodei, chief executive of Anthropic, one of the companies building the most advanced AI in the world, predicted that within three to six months, machines would be “writing 90 percent of the code,” and within a year, essentially all of it.
A year earlier, at a government summit in Dubai, Jensen Huang, head of the chip giant Nvidia whose hardware powers much of this revolution, went further and told young people to stop learning to code altogether. The future programming language, he argued, “is human”, you would simply tell the computer what you want in plain words.
It is a fair question, and it deserves a fair answer, not the comforting one, and not the frightening one, but the true one.
First, the part that is real
Let me begin with what is real, because pretending otherwise insults the reader.
In March 2025, at a Council on Foreign Relations event in the United States, Dario Amodei, chief executive of Anthropic, one of the companies building the most advanced AI in the world, predicted that within three to six months, machines would be “writing 90 percent of the code,” and within a year, essentially all of it.
A year earlier, at a government summit in Dubai, Jensen Huang, head of the chip giant Nvidia whose hardware powers much of this revolution, went further and told young people to stop learning to code altogether. The future programming language, he argued, “is human”, you would simply tell the computer what you want in plain words.
These are not village rumours. They are the considered words of the people who profit most from the technology. And the hiring numbers have begun to echo them: worldwide, openings for junior, entry-level programmers have fallen by more than half since 2022, and a study from Stanford University found that employment among the youngest developers, aged 22 to 25, has dropped by roughly a fifth from its peak.
So the fear in that Masaka classroom is not foolish. It is arriving.
Now, the lesson almost everyone gets wrong
Here is where the conversation goes off the road. People hear “the machine can write code” and conclude “so there is nothing left to learn.” That is precisely the wrong lesson, and if our young people swallow it, we will have talked an entire generation out of becoming the one thing most worth becoming.
Read those same experts more carefully. Even Amodei, in the very breath he used to predict 90 percent, added a caveat the headlines ignored: a human must still decide what the program is for, set its conditions, and make the design decisions. Even Huang, having told children not to code, immediately insisted that we must instead upskill everyone. The machine writes the lines. It does not decide what is worth building, or whether what it built is right.
That distinction is everything.
What these tools actually are
Strip away the mystery, and an AI coding tool is something like the fastest, most confident boda rider in Kampala. Tell it where you want to go and it will get you there at remarkable speed. But it does not know the city the way you do. If you cannot read the map yourself, if you have never walked the streets, you cannot tell when it is taking you the long way, driving you toward a swamp, or speeding with total confidence in entirely the wrong direction.
To someone who knows the city, that rider is a gift. To someone who knows nothing, the very same rider is a danger dressed up as convenience.
How an AI agent actually writes code
It helps to look under the bonnet, because the magic dissolves on inspection.
At the centre of these tools sits a large language model, a system trained on enormous amounts of text and code to do one narrow thing extremely well: predict the next chunk of text that should follow. It has no grasp of your business, your users, or the law. It has patterns.
An “agent” wraps that model in a loop. Give it a goal and it does roughly this. First it breaks the goal into smaller steps, this is called planning. Then it writes a piece of code for the first step and runs that code in a sandbox, an isolated space where the program can execute without touching anything important. It reads whatever comes back, a result, or an error message, and feeds that back into itself to choose the next move.
Reason, act, observe; reason, act, observe, again and again, keeping a short memory of what it has already tried so it does not go in circles. It can also reach out to tools: call an internet service, query a database, open a file. The code it produces along the way is disposable scaffolding, written to get an answer, then often thrown away. As one recent research paper frames it, the code is no longer the product; the agent itself is.
That is genuinely powerful. It is also why its failures are the kind that bite. The model is straining to produce code that looks right and passes the immediate test, not code that is correct in every situation it will later meet. It can write a function that handles the example you gave it and quietly mishandles the one you did not: a customer with no middle name, a transaction at midnight on the last day of the financial year, a mobile network that drops halfway through a payment. The agent will report success. Only a human who understands the system will know where to look for the rot.
The cliff the machines fall off
This is not my hunch. A research paper posted this June to arXiv, an online archive where scientists share findings before formal peer review, by Zhenfeng Cao, a researcher in China, lays out this new “agentic” way of building software and then tests where it breaks. Drawing on a careful set of trials known as EvoClaw, it found something that should be pinned to the wall of every computer lab in East Africa.
When AI agents are handed small, neat, self-contained tasks, they succeed more than eight times out of ten. But when they are asked to do what real software actually demands, keep a living system running over time, where today's change must not break what was built last month, and where small early mistakes quietly multiply, their success collapses to below four in ten.
The bigger, messier, and more consequential the system, the more it needs a human being who genuinely understands it.
The tool can fool even the experts
You might think the danger is only for beginners. It is not. In a careful controlled trial, the research group METR found that experienced open-source developers using AI assistance were actually about 19 percent slower, even as they felt faster. Other reviews report that AI written code carries more bugs and gets rewritten more often than code thought through by a person. The tool feels quick while quietly costing you. That is the trap.
We have seen a smaller version of it before. A student who masters the calculator but never learned arithmetic cannot tell when the screen shows nonsense because a zero was mistyped. The calculator did not make him careless; skipping the foundations did. AI is now a calculator for almost everything, and the temptation to skip the foundations has never been stronger, or more expensive.
And our most important systems are not toys
What are our most consequential systems? The mobile money platforms that move a vast share of this country's money. The systems behind URA, NSSF, the hospitals, the banks, the Electoral Commission. These are not classroom exercises. They are exactly the places where a confident, plausible-looking error, the kind these tools produce most easily, can cost money, trust, or lives.
Benoni Katende, the chief technology officer at NSSF, put the real job plainly at a tech showcase in Kampala: find a genuine problem, and use technology to solve it, faster, cheaper, better. Notice what that demands. Not faster typing. Understanding deep enough to know which problem matters and whether the solution actually works.
Why this is Uganda's moment, not Uganda's threat
Here is the striking thing. Even as powerful voices abroad tell the young not to bother, young Ugandans are pouring intocomputing. Makerere College of Computing and Information Sciences is among the largest of its kind on the continent, with lecture halls built to seat over ten thousand. At Mbarara University of Science and Technology, the computer science intake is reported to have grown from around 30 students to more than 100 admitted, a pattern echoing across the country. The “AI for Uganda” initiative bluntly notes a surge in demand for these skills against a shortage of people who actually have them.
The universities are leaning in, not out. At MUST itself, the Vice Chancellor, Professor Pauline Byakika, the university's first female VC, heads an Institution that runs a data-science research hub that puts machine learning and AI to work on medical images, sharpening disease diagnosis across Africa. She has presented herself, even as a seasoned professor, as someone who must constantly learn, unlearn and relearn. If that is contextually true for a Vice Chancellor, it is true for a Senior Six student.
At Makerere, Vice Chancellor Prof Barnabas Nawangwe says AI is already transforming the university's research; Kyambogo's Vice Chancellor, Prof Eli Katunguka, observing that students have already embraced these tools, says simply: “It is something we cannot ignore.”
In March, Makerere and the national research network switched on a National AI Research Cloud to give local innovators real computing power. Academies such as Refactory have trained hundreds of developers, roughly 700 over five years, with job placement rates above 90 percent.
This hunger is correct. Uganda is among the youngest nations on earth: a median age of about 17, more than half of us under 30, and roughly 700,000 young people entering the job market every year, far too many of whom never find meaningful work. We cannot afford to misread this moment, and we cannot afford to be mere consumers of tools built somewhere else.
We have done better before. When the banks would not serve ordinary people, Ugandans helped build and spread mobile money, and the region leapfrogged ahead of far richer countries. That did not happen because people pushed buttons. It happened because people understood the technology well enough to bend it to our own problems.
So how should a young person study now?
To that student at Masaka, and to the parent weighing the fees, here is my honest advice: yes, study software engineering, IT, computer science. But study them differently from the generation before you.
Learn the foundations the machine itself is built on, how computers, data, and networks really work, and how to reason through a problem step by step, because that knowledge is what lets you judge the tool instead of merely obeying it. Use AI freely, but treat it the way a good supervisor treats a fast, clever, slightly careless intern: let it do the typing, then read every line, question it, and deliberately break it to find where it fails. Build real things from beginning to end, and do not run from the boring, difficult parts, the debugging, the fixing, the long maintenance. That frustrating work is exactly where judgment is grown, and it is exactly the work AI is worst at.
A warning especially for those heading into computing at universities like MUST: it has never been easier to let the machine write your coursework and your code, and many students now quietly do exactly that. It is the surest way to walk away with a certificate and none of the judgment the certificate is supposed to prove. Do the hard problem yourself first, by hand, until you understand it, then let AI critique your answer, stretch it, and show you a better way. Used in that order, the tool sharpens you. Used in reverse, it hollows you out.
Aim, in the end, not to be the quickest typist of code but to become the person who decides what should be built and whether it is right. Cao's paper calls that person the “intent architect.” Whatever we name it, it is the role the machines cannot fill.
The choice in front of us
The countries and the young people who win the coming years will not be the ones who used AI the most. They will be the ones who understood it the most. East Africa can raise a generation of button pushers, helpless the moment the tool stumbles, or a generation of builders who direct these machines at our own problems, on our own terms.
The tools are now cheap and everywhere. Understanding is not. That is the thing still worth paying school fees for.
The machines can type. Let us make sure our children can still think.
The writers is Assistant Lecturer, Faculty of Computing and Informatics, Mbarara Univeristy of Science and Technology.
So the fear in that Masaka classroom is not foolish. It is arriving.
Now, the lesson almost everyone gets wrong
Here is where the conversation goes off the road. People hear “the machine can write code” and conclude “so there is nothing left to learn.” That is precisely the wrong lesson, and if our young people swallow it, we will have talked an entire generation out of becoming the one thing most worth becoming.
Read those same experts more carefully. Even Amodei, in the very breath he used to predict 90 percent, added a caveat the headlines ignored: a human must still decide what the program is for, set its conditions, and make the design decisions. Even Huang, having told children not to code, immediately insisted that we must instead upskill everyone. The machine writes the lines. It does not decide what is worth building, or whether what it built is right.
That distinction is everything.
What these tools actually are
Strip away the mystery, and an AI coding tool is something like the fastest, most confident boda rider in Kampala. Tell it where you want to go and it will get you there at remarkable speed. But it does not know the city the way you do. If you cannot read the map yourself, if you have never walked the streets, you cannot tell when it is taking you the long way, driving you toward a swamp, or speeding with total confidence in entirely the wrong direction.
To someone who knows the city, that rider is a gift. To someone who knows nothing, the very same rider is a danger dressed up as convenience.
How an AI agent actually writes code
It helps to look under the bonnet, because the magic dissolves on inspection.
At the centre of these tools sits a large language model, a system trained on enormous amounts of text and code to do one narrow thing extremely well: predict the next chunk of text that should follow. It has no grasp of your business, your users, or the law. It has patterns.
An “agent” wraps that model in a loop. Give it a goal and it does roughly this. First it breaks the goal into smaller steps, this is called planning. Then it writes a piece of code for the first step and runs that code in a sandbox, an isolated space where the program can execute without touching anything important. It reads whatever comes back, a result, or an error message, and feeds that back into itself to choose the next move.
Reason, act, observe; reason, act, observe, again and again, keeping a short memory of what it has already tried so it does not go in circles. It can also reach out to tools: call an internet service, query a database, open a file. The code it produces along the way is disposable scaffolding, written to get an answer, then often thrown away. As one recent research paper frames it, the code is no longer the product; the agent itself is.
That is genuinely powerful. It is also why its failures are the kind that bite. The model is straining to produce code that looks right and passes the immediate test, not code that is correct in every situation it will later meet. It can write a function that handles the example you gave it and quietly mishandles the one you did not: a customer with no middle name, a transaction at midnight on the last day of the financial year, a mobile network that drops halfway through a payment. The agent will report success. Only a human who understands the system will know where to look for the rot.
The cliff the machines fall off
This is not my hunch. A research paper posted this June to arXiv, an online archive where scientists share findings before formal peer review, by Zhenfeng Cao, a researcher in China, lays out this new “agentic” way of building software and then tests where it breaks. Drawing on a careful set of trials known as EvoClaw, it found something that should be pinned to the wall of every computer lab in East Africa.
When AI agents are handed small, neat, self-contained tasks, they succeed more than eight times out of ten. But when they are asked to do what real software actually demands, keep a living system running over time, where today's change must not break what was built last month, and where small early mistakes quietly multiply, their success collapses to below four in ten.
The bigger, messier, and more consequential the system, the more it needs a human being who genuinely understands it.
The tool can fool even the experts
You might think the danger is only for beginners. It is not. In a careful controlled trial, the research group METR found that experienced open-source developers using AI assistance were actually about 19 percent slower, even as they felt faster. Other reviews report that AI written code carries more bugs and gets rewritten more often than code thought through by a person. The tool feels quick while quietly costing you. That is the trap.
We have seen a smaller version of it before. A student who masters the calculator but never learned arithmetic cannot tell when the screen shows nonsense because a zero was mistyped. The calculator did not make him careless; skipping the foundations did. AI is now a calculator for almost everything, and the temptation to skip the foundations has never been stronger, or more expensive.
And our most important systems are not toys
What are our most consequential systems? The mobile money platforms that move a vast share of this country's money. The systems behind URA, NSSF, the hospitals, the banks, the Electoral Commission. These are not classroom exercises. They are exactly the places where a confident, plausible-looking error, the kind these tools produce most easily, can cost money, trust, or lives.
Benoni Katende, the chief technology officer at NSSF, put the real job plainly at a tech showcase in Kampala: find a genuine problem, and use technology to solve it, faster, cheaper, better. Notice what that demands. Not faster typing. Understanding deep enough to know which problem matters and whether the solution actually works.
Why this is Uganda's moment, not Uganda's threat
Here is the striking thing. Even as powerful voices abroad tell the young not to bother, young Ugandans are pouring intocomputing. Makerere College of Computing and Information Sciences is among the largest of its kind on the continent, with lecture halls built to seat over ten thousand. At Mbarara University of Science and Technology, the computer science intake is reported to have grown from around 30 students to more than 100 admitted, a pattern echoing across the country. The “AI for Uganda” initiative bluntly notes a surge in demand for these skills against a shortage of people who actually have them.
The universities are leaning in, not out. At MUST itself, the Vice Chancellor, Professor Pauline Byakika, the university's first female VC, heads an Institution that runs a data-science research hub that puts machine learning and AI to work on medical images, sharpening disease diagnosis across Africa. She has presented herself, even as a seasoned professor, as someone who must constantly learn, unlearn and relearn. If that is contextually true for a Vice Chancellor, it is true for a Senior Six student.
At Makerere, Vice Chancellor Prof Barnabas Nawangwe says AI is already transforming the university's research; Kyambogo's Vice Chancellor, Prof Eli Katunguka, observing that students have already embraced these tools, says simply: “It is something we cannot ignore.”
In March, Makerere and the national research network switched on a National AI Research Cloud to give local innovators real computing power. Academies such as Refactory have trained hundreds of developers, roughly 700 over five years, with job placement rates above 90 percent.
This hunger is correct. Uganda is among the youngest nations on earth: a median age of about 17, more than half of us under 30, and roughly 700,000 young people entering the job market every year, far too many of whom never find meaningful work. We cannot afford to misread this moment, and we cannot afford to be mere consumers of tools built somewhere else.
We have done better before. When the banks would not serve ordinary people, Ugandans helped build and spread mobile money, and the region leapfrogged ahead of far richer countries. That did not happen because people pushed buttons. It happened because people understood the technology well enough to bend it to our own problems.
So how should a young person study now?
To that student at Masaka, and to the parent weighing the fees, here is my honest advice: yes, study software engineering, IT, computer science. But study them differently from the generation before you.
Learn the foundations the machine itself is built on, how computers, data, and networks really work, and how to reason through a problem step by step, because that knowledge is what lets you judge the tool instead of merely obeying it. Use AI freely, but treat it the way a good supervisor treats a fast, clever, slightly careless intern: let it do the typing, then read every line, question it, and deliberately break it to find where it fails. Build real things from beginning to end, and do not run from the boring, difficult parts, the debugging, the fixing, the long maintenance. That frustrating work is exactly where judgment is grown, and it is exactly the work AI is worst at.
A warning especially for those heading into computing at universities like MUST: it has never been easier to let the machine write your coursework and your code, and many students now quietly do exactly that. It is the surest way to walk away with a certificate and none of the judgment the certificate is supposed to prove. Do the hard problem yourself first, by hand, until you understand it, then let AI critique your answer, stretch it, and show you a better way. Used in that order, the tool sharpens you. Used in reverse, it hollows you out.
Aim, in the end, not to be the quickest typist of code but to become the person who decides what should be built and whether it is right. Cao's paper calls that person the “intent architect.” Whatever we name it, it is the role the machines cannot fill.
The choice in front of us
The countries and the young people who win the coming years will not be the ones who used AI the most. They will be the ones who understood it the most. East Africa can raise a generation of button pushers, helpless the moment the tool stumbles, or a generation of builders who direct these machines at our own problems, on our own terms.
The tools are now cheap and everywhere. Understanding is not. That is the thing still worth paying school fees for.
The machines can type. Let us make sure our children can still think.
The writers is Assistant Lecturer, Faculty of Computing and Informatics, Mbarara Univeristy of Science and Technology.