<?xml version="1.0" encoding="utf-8"?>
<?xml-stylesheet type="text/xsl" href="../assets/xml/rss.xsl" media="all"?><rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>TinyComputers.io (Posts about game development)</title><link>https://tinycomputers.io/</link><description></description><atom:link href="https://tinycomputers.io/categories/game-development.xml" rel="self" type="application/rss+xml"></atom:link><language>en</language><copyright>Original site content © 2022–2026 Tiny Machines Workshop, LLC, except where otherwise noted. Some rights reserved.</copyright><lastBuildDate>Sun, 15 Nov 2026 13:13:11 GMT</lastBuildDate><generator>Nikola (getnikola.com)</generator><docs>http://blogs.law.harvard.edu/tech/rss</docs><item><title>The Last Five Percent: AI, a Youth Game Competition, and What It Was Really Judging</title><link>https://tinycomputers.io/posts/the-last-five-percent.html?utm_source=feed&amp;utm_medium=rss&amp;utm_campaign=rss</link><dc:creator>A.C. Jokela</dc:creator><description>&lt;div class="audio-widget"&gt;
&lt;div class="audio-widget-header"&gt;
&lt;span class="audio-widget-icon"&gt;🎧&lt;/span&gt;
&lt;span class="audio-widget-label"&gt;Listen to this article&lt;/span&gt;
&lt;/div&gt;
&lt;audio controls preload="metadata"&gt;
&lt;source src="https://tinycomputers.io/the-last-five-percent_tts.mp3" type="audio/mpeg"&gt;
&lt;/source&gt;&lt;/audio&gt;
&lt;div class="audio-widget-footer"&gt;38 min · AI-generated narration&lt;/div&gt;
&lt;/div&gt;

&lt;p&gt;In 2014 the Australian Council for Educational Research piloted a national competition for schoolchildren with a simple brief: make a video game. More than 550 students registered that first year. By its last year the rules had settled into something both simple and serious. A team of up to four students, anywhere from Year 3 to Year 12, built an original game, wrote a game design document explaining it, and submitted a one-to-five-minute screen recording of it being played. The judges had to be able to see the code. The main characters, the sounds, the key obstacles and the backgrounds had to be the students' own work. Entry was free.&lt;/p&gt;
&lt;h3&gt;The Final Chapter&lt;/h3&gt;
&lt;p&gt;More than a decade and 16,000 student participations later, the Australian STEM Video Game Challenge is ending. The competition's website now carries the sentence "2026 is the final year of the Australian STEM Video Game Challenge." Its resources stay online until the end of the first school term of 2027, and then they go too.&lt;/p&gt;
&lt;p&gt;ACER's explanation came from Cathy Zemaitis, who heads its Student Learning and Progress division, in a statement to &lt;a href="https://go.tinymachines.io/pcgamer-stem-video-game-challenge"&gt;PC Gamer&lt;/a&gt;. The decision was driven by "the rapid and complex rise of generative AI in game development," which was "making it harder to judge what students made themselves" and moving the program away from its original purpose. ACER has no plans to deliver the challenge in another form. Schools looking for a replacement are pointed to an international mathematical modelling competition.&lt;/p&gt;
&lt;p&gt;&lt;a href="https://go.tinymachines.io/kotaku-stem-video-game-challenge"&gt;Kotaku's coverage&lt;/a&gt; quoted a parent on Reddit whose teenage son and his friends had entered for several years running. The competition had given the group confidence and a sense of belonging to a wider community of people who make games. When the news came, the son was "visibly gutted."&lt;/p&gt;
&lt;p&gt;ACER made the wrong decision. Generative AI is not harmless to a competition like this one; it plainly isn't. But the problem ACER describes is a problem with what the competition measured, and that problem has a fix. Ending the competition does not fix it. It just stops measuring.&lt;/p&gt;
&lt;h3&gt;What Was Actually Claimed&lt;/h3&gt;
&lt;p&gt;The headlines went further than ACER did. Kotaku's said the competition was shut down because kids kept submitting AI slop. PC Gamer's said students were submitting games made entirely with AI. ACER's own statement makes a narrower claim: the tools made it hard to tell what the students had made. I could not find a count of generated entries, an example of one, or any finding from the judges. The only first-hand description of wholly generated games I found came from an unnamed former judge in a Reddit thread, the same thread that raised the possibility that funding cuts were the bigger reason. I am going to take ACER at its word, because its word is more interesting than the headlines.&lt;/p&gt;
&lt;p&gt;The rules already allowed AI. A team could use it as long as they referenced it in their design document "in the same way they might reference a tutorial": what they used it for, with screenshots where possible, and how and why they changed what it produced. Small chunks of generated code were fine. Using AI to test and fix bugs was fine. This was not a competition that banned the tools and then got cheated. It was a competition that permitted the tools and then found its judges could not see where the tool stopped and the student started.&lt;/p&gt;
&lt;p&gt;That is two problems wearing one coat. The first is a judging problem: a panel looking at a finished game cannot tell who made which part. The second is a purpose problem: if the tools make the game, the student has not learned to make games, and a competition whose whole reason for existing is to get children to learn something has stopped doing its job. ACER's statement names both. The decision to end the competition treats them as one problem with one answer. They are not one problem, and the second is the one worth thinking about first.&lt;/p&gt;
&lt;h3&gt;Candy and Chess&lt;/h3&gt;
&lt;p&gt;Alasdair MacIntyre tells a story in &lt;em&gt;After Virtue&lt;/em&gt; about teaching a child to play chess. The child is bright but has no interest in chess. The child does like candy. So the teacher offers candy for every game played, extra candy for every win, and promises to play just badly enough that winning is possible but hard. The child plays, and plays to win. MacIntyre's point is that as long as candy is the only reason to play, the child has "every reason to cheat," provided the cheating works. Something changes only if the child comes to want what chess itself offers: a particular kind of analytical skill, strategic imagination, the pleasure of seeing three moves ahead. MacIntyre calls these the goods internal to the practice. Candy is an external good. It can be had in a hundred ways. The internal goods can only be had by playing chess.&lt;/p&gt;
&lt;p&gt;A youth competition is a candy machine, built on purpose. The prize, the online showcase, the certificate, the line on a school report: these are external goods, and they exist to get a nine-year-old to sit with Scratch for a whole term until the internal goods take over. The hope is the same as MacIntyre's. At some point the child stops making a game to win and starts making it because getting a jump to feel right is its own satisfaction.&lt;/p&gt;
&lt;p&gt;Generative AI breaks the machine in a specific way. It does not make cheating easier in the old sense, because the rules allowed the tools. It makes the candy available without the chess. A team can now produce something that looks like a term's work in an afternoon, and if the judging pays out for the look of the finished thing, the candy goes to whoever is best at asking for it. The child who struggled through the term gets the same certificate, or less. The machine stops moving anyone toward the internal goods, because it no longer pays for them.&lt;/p&gt;
&lt;p&gt;That is a real failure. But look at where the failure is. It is in what the judges pay for. MacIntyre's child did not need chess abolished. The child needed a game in which the candy could only be won by playing chess well.&lt;/p&gt;
&lt;h3&gt;The Case for Ending It&lt;/h3&gt;
&lt;p&gt;ACER's position deserves its strongest form, so here it is.&lt;/p&gt;
&lt;p&gt;Sixteen thousand participations over the life of the challenge works out to well over a thousand students a year. Judging is human and finite. There is no reliable detector for generated code or art, and the tools sold as detectors are known for false positives. A false positive here is an accusation against a child, made by the organization that invited the child to enter. The entrants run from eight-year-olds dragging blocks in Scratch to seventeen-year-olds writing their own physics, and any rule has to be fair to all of them. ACER is an educational research organization, not a games company, and a free national competition runs on a foundation's budget and the goodwill of its judges. And the purpose problem is real: if most entries are mostly generated, most participants are not learning what the program exists to teach, however the prizes are awarded.&lt;/p&gt;
&lt;p&gt;Put together, ending it looks less like panic than like an honest admission. We built a thing to teach children to make games. We can no longer tell whether it does. Better to stop than to run a contest we can't judge fairly.&lt;/p&gt;
&lt;p&gt;Here is where that case breaks. The prizes are not decided across a thousand entries. They are decided in the finals, and the finals are small. The top five entries in each of the four primary and junior secondary categories go through, and the top ten in Years 10 to 12. That is thirty finalist teams. Every prize the competition awards is settled among thirty teams, and thirty teams is a number a panel of judges can talk to.&lt;/p&gt;
&lt;p&gt;The purpose problem does not go away when the competition does, either. The tools are on every student's laptop whether or not ACER runs a challenge. What ends is a national structure that rewarded children, in public, for the slow part of making a game. The question was never whether students would use AI. It was what the competition would reward them for doing with it.&lt;/p&gt;
&lt;h3&gt;The Last Five Percent&lt;/h3&gt;
&lt;p&gt;Here is what I believe after building a lot of things with these tools. AI will get you about ninety-five percent of the way to a solid user interface, decent graphics and something playable. The last five percent is what separates slop from a thing people want to keep using, and you will spend about ninety percent of your time there.&lt;/p&gt;
&lt;p&gt;Those numbers are not measurements. Nobody timed anything. They are labels for a proportion I feel every time I build something, the way a carpenter says a cabinet is basically done for the second week in a row. They belong to an old family of software jokes. Tom Cargill of Bell Labs put it as the ninety-ninety rule, which Jon Bentley printed in his "Programming Pearls" column in 1985: the first ninety percent of the code accounts for the first ninety percent of the development time, and the remaining ten percent accounts for the other ninety percent. The two halves add up to a hundred and eighty percent. That is the joke, and it is also true.&lt;/p&gt;
&lt;p&gt;What AI changes is where the first ninety-five percent comes from. It does not change the last five. If anything it makes the last five more visible, because the first ninety-five now arrives so fast that the gap between it and something good is the only thing left to look at. I came at this from another direction in an essay on &lt;a href="https://go.tinymachines.io/what-makes-something-slop"&gt;what makes something slop&lt;/a&gt;: slop is output that has never been tested against anything outside itself. The last five percent is where that test happens. It is also the judgment bottleneck from &lt;a href="https://go.tinymachines.io/the-excavator-and-the-foundation"&gt;The Excavator and the Foundation&lt;/a&gt;, seen from the inside: the excavator digs the hole in an afternoon, and someone still has to know whether the hole is in the right place.&lt;/p&gt;
&lt;p&gt;In games, the last five percent has a name. Steve Swink's &lt;em&gt;Game Feel&lt;/em&gt; builds its definition of the subject from real-time control, a simulated space and polish, and polish is the part a player notices without being able to say why. It is the few frames in which a character can still jump after running off a ledge. It is the button press that registers a hair before the character lands, so the next jump comes out anyway. It is a jump arc that rises fast and falls faster, because a symmetrical parabola feels floaty even though it is the physically correct one. None of this is hard to generate once you know to ask for it. All of it is hard to know to ask for until you have watched someone play your game and seen exactly where their hands tense up.&lt;/p&gt;
&lt;p&gt;A model can't do that part for you. It has not watched your little brother miss the same jump eleven times.&lt;/p&gt;
&lt;h3&gt;The Andes&lt;/h3&gt;
&lt;p&gt;The clearest example of the last five percent I have is not from a game. It is from the &lt;a href="https://go.tinymachines.io/tc-ballistics"&gt;ballistics engine&lt;/a&gt; I have been building since the spring of 2025, an open-source Rust library that computes where a bullet goes.&lt;/p&gt;
&lt;p&gt;Most of that engine came easily. With AI doing a great deal of the typing, the equations of motion, the drag tables, the atmosphere model, the command-line tool and the bindings for other languages all arrived in good order. That was the ninety-five percent. Then people started using it, each with their own way of thinking about shooting, and the quirks started to surface.&lt;/p&gt;
&lt;p&gt;One of them came from a man in Peru who shoots air rifles. He wanted to calculate trajectories anywhere in his country. Peru runs from Pacific beaches at sea level to Andean cities above 4,000 meters; the airport at Cusco sits at 3,310. His question sent me up into those mountains, numerically, and up there two parts of the engine that had always agreed stopped agreeing.&lt;/p&gt;
&lt;p&gt;Shooters often do not know the true air pressure where they stand. What they have is the altimeter setting from the nearest airport or weather report: the pressure corrected to sea level, which pilots call QNH. The engine has to turn that back into station pressure, the pressure at the shooter's own elevation. It had been doing it with a shortcut: take the standard atmosphere's pressure at that height and scale it by the ratio of the reported setting to the standard 1013.25 hectopascals. It looks reasonable. It is not how an altimeter setting is defined. The correct relation finds the pressure altitude the setting implies and adds the station's height to it.&lt;/p&gt;
&lt;p&gt;At sea level the two methods give exactly the same answer, down to the last bit. On a standard day, when the setting is exactly 1013.25, they also agree exactly. Everywhere else they drift apart with height. On a 1020 hectopascal day at 250 meters, which is roughly where a lot of people shoot, the gap is about four hundredths of a hectopascal, four parts in a hundred thousand. In the engine that moves a .308's drop at a thousand yards by about a third of a millimeter. No group on paper will ever show it, and none of my tests did.&lt;/p&gt;
&lt;p&gt;At 1,500 meters on a 1030 day the gap is half a hectopascal. At Cusco on a 1033 day it is 1.07. At La Paz, across the border at 4,061 meters, it is 1.52. For a .308 zeroed at 100 yards, the drop at 1,000 yards moved by 6 millimeters at Cusco and 8 at La Paz. That is still small, and for an air rifle at fifty meters it is smaller still. But it is not rounding. It is the engine computing the wrong air, by a consistent amount, for exactly the people who live highest.&lt;/p&gt;
&lt;p&gt;Fixing it took several days of spare time and about three dozen commits. The second day turned up a related bug that bites for the same reason. A shooter who typed a reading of 59 degrees Fahrenheit and 29.92 inches of mercury, a perfectly ordinary day, got that reading treated as never entered, because it landed within a hair of the engine's default values, and the engine quietly replaced it with an altitude-based guess. The same two days added a warning for apps that send a sea-level pressure together with Cusco's altitude, which had been flying their users through air at one and a half times the pressure it should have had.&lt;/p&gt;
&lt;p&gt;None of this was found by reading the code, mine or any assistant's. A person whose world was different from mine found it by asking an ordinary question. And fixing it meant knowing things no generated draft hands you: why two formulas that agree at sea level part company in the Andes, why an altimeter setting is defined the way it is, how to pin the right answer in tests so the shortcut can never quietly come back. That knowledge is the last five percent. It is also exactly the kind of thing a judge can ask a student about.&lt;/p&gt;
&lt;h3&gt;Three Rules&lt;/h3&gt;
&lt;p&gt;If AI hands every team the first ninety-five percent, the competition should judge the last five. Three rules would do it, and none of them bans anything.&lt;/p&gt;
&lt;p&gt;The first is a polish bar set for the world as it is. A game that runs, has a title screen and plays cleanly for thirty seconds used to stand for a term of work by a twelve-year-old, and it was worth rewarding on its own. It no longer is. The rubric should move its weight to the things that only come from iteration: how the controls feel, how the difficulty climbs, what happens when a player does something the designer did not plan for. Judges should spend their time trying to break the game, not admiring its first minute. This raises the bar, and that is the point. The part of the bar that gets raised is the part made of the student's own hours.&lt;/p&gt;
&lt;p&gt;The second is a spoken defense for the finalists, without notes. Thirty teams at twenty minutes each over a video call is ten hours of judging. Each team sits with the judges, explains how its game works, and answers questions about it. The questions scale with age. A Year 4 team that built a platformer in Scratch should be able to say why its character falls faster each second instead of at a steady speed: every tick, the game adds a little to the character's downward velocity, and then moves the character by that velocity. A team that cannot say that did not build the fall.&lt;/p&gt;
&lt;p&gt;At the top, the questions should be exactly as demanding as the work. A Years 10 to 12 team whose game has rippling water should be able to explain the Navier–Stokes equations as their game uses them. A great deal of real-time fluid simulation in games traces back to Jos Stam's paper "Real-Time Fluid Dynamics for Games," presented at the Game Developers Conference in 2003, and its steps are teachable: carry the fluid's velocity along with itself, spread it out according to the viscosity, then correct the field so that no fluid is created or destroyed. Water surfaces are usually cheaper still: a height field driven by a simplified form of the equations that treats the water as shallow. A team should be able to say which simplification they made and what it costs them, which step makes their ripples die out, and what happens if they halve the viscosity. Better yet, a judge can ask them to halve it while everyone watches. If you understand your code, you can change it on request. If you do not, you cannot, however good the game looks.&lt;/p&gt;
&lt;p&gt;None of this is new, and I have been on the student's side of it. In graduate school I took Animation and Planning in Games, an upper-division course on exactly what its name says. The class was mostly senior computer science undergraduates with a handful of graduate students, and more was expected of the graduate students: the same scaling by level I am proposing by school year. Part of what was expected of me was to defend my final project in front of the class. I had to talk through the techniques and algorithms I had used, why I had chosen them, and how some of their mechanics worked. The finished project could not answer those questions for me. I had to. That project, as an aside, became the seed of &lt;a href="https://go.tinymachines.io/circle-kingdom"&gt;&lt;em&gt;Circle Kingdom&lt;/em&gt;&lt;/a&gt;, a storybook strategy game, free to play in a browser, in which a golden champion defends a circular king against an army of triangles.&lt;/p&gt;
&lt;p&gt;The obvious objection is that some children freeze in front of adults, and a spoken exam rewards confidence as much as knowledge. That is true, and it is an argument for designing the defense well, not for skipping it. A shy eleven-year-old does not have to give a talk. A judge can open the project, point at a block of code and ask what happens if it is deleted. A team can answer at the keyboard instead of out loud, with its mentor in the room. The defense is not a test of eloquence. It is a test of whether a student can find their way around their own work, and a student who built the thing can do that quietly.&lt;/p&gt;
&lt;p&gt;The third is a playtest log: who played the game, on what date, what broke, and what the team changed because of it. A log can be generated too. But a log is easy to check in conversation. Tell me about the third bug on this list. How did you find it? Why did you fix it that way and not another way? The log is where the last five percent leaves a trail, and the defense is where the trail gets walked.&lt;/p&gt;
&lt;p&gt;None of this asks students to stop using AI. It asks them to know what they shipped. A student who used AI for the first ninety-five percent, spent a month on the last five, and can explain all of it has done exactly what the competition exists to teach. That student should win.&lt;/p&gt;
&lt;p&gt;The rules also reach the students who never make the finals, because rubrics teach. The moment a competition announces that finalists will explain their fluid solver without notes, every team that hopes to be a finalist starts learning its fluid solver. Mentors start running practice defenses in classrooms. The candy is still candy. It can only be won by playing chess.&lt;/p&gt;
&lt;h3&gt;The Text That Cannot Answer&lt;/h3&gt;
&lt;p&gt;Socrates worried about this exact problem twenty-four centuries ago, about a different technology.&lt;/p&gt;
&lt;p&gt;In &lt;a href="https://go.tinymachines.io/phaedrus-jowett"&gt;Plato's &lt;em&gt;Phaedrus&lt;/em&gt;&lt;/a&gt;, Socrates tells a story about the Egyptian god Theuth presenting the invention of writing to the king Thamus. Theuth promises that writing will make people wiser and improve their memories. Thamus disagrees. In Jowett's translation, he says the people who learn letters "will appear to be omniscient and will generally know nothing," having "the show of wisdom without the reality." Socrates then adds a complaint of his own. Written words look as though they have intelligence, he says, but if you put a question to them, the text "always gives one unvarying answer." And when a text is attacked or misread, it has no parent to protect it, and it cannot defend itself.&lt;/p&gt;
&lt;p&gt;Put "generated output" where Socrates says "writing" and you have ACER's announcement. A finished game is a text in Socrates's sense. It does the same thing every time you play it. It cannot tell you why the jump arc is shaped the way it is, or why the water settles. It cannot answer a question. Its maker can.&lt;/p&gt;
&lt;p&gt;Socrates was half wrong about writing, and the way he was wrong is the useful part. Writing did not destroy memory or understanding, because we did not answer it by banning writing. We kept writing and added a check: we started asking writers questions. The doctoral viva, the thesis defense, the oral examination, the code review where someone asks why you did it that way. Every one of those is Socrates's remedy, the living speaker beside the text, applied to a technology we decided to keep.&lt;/p&gt;
&lt;p&gt;Gilbert Ryle split knowledge into knowing how and knowing that. Knowing how is a jump that feels right, which you can only get by tuning it with your hands and watching other hands on the controller. Knowing that is the explanation: the velocity, the viscosity, the projection step. A finished game used to be decent evidence of both, because there was no other way to get one. It is no longer good evidence of either. The polish bar and the spoken defense are the two ways of putting the evidence back, one for each kind of knowing.&lt;/p&gt;
&lt;h3&gt;Every Take-Home&lt;/h3&gt;
&lt;p&gt;The STEM Video Game Challenge is a small competition, and it is easy to read this as a story about one organization's decision. It is a story about every assessment that accepts a finished artifact in place of watching the work: the school essay, the take-home exam, the weekend hiring exercise, the grant application. Each of them was always a stand-in. The artifact counted as evidence of the skill because producing it required the skill. Generative AI cut that link for all of them at once, in the same way, in the same couple of years.&lt;/p&gt;
&lt;p&gt;Each of them faces ACER's choice. You can stop accepting the artifact, which is what ACER did. Or you can stop treating the artifact as the evidence and start asking its author questions. The second option costs more per candidate and far less in total, because you only have to question the people you are about to reward.&lt;/p&gt;
&lt;p&gt;I keep thinking about the parent's son. He and his friends entered year after year, and they did the slow part, because that is what you do when a thing matters to you. What the ending teaches him is that when a tool makes the thing being judged easy, the judges stop judging. I would rather he learned the other lesson: when a tool makes part of the work easy, the judging moves to the part that is still hard.&lt;/p&gt;
&lt;p&gt;His game cannot answer questions. He can. Someone should have asked him.&lt;/p&gt;</description><category>ai</category><category>assessment</category><category>ballistics</category><category>education</category><category>game development</category><category>macintyre</category><category>philosophy</category><category>plato</category><guid>https://tinycomputers.io/posts/the-last-five-percent.html</guid><pubDate>Sun, 15 Nov 2026 04:00:00 GMT</pubDate></item></channel></rss>