Builder Journal · ARC Prize 2026

What Teaching a Machine to Think Taught Me

This is a Builder Journal summary, a step back from inside a competition I am still competing in, with prize money on the line and the deadline getting closer. I have written a handful of dispatches from this one already, each about a specific way it humbled me. This pulls them into one place, along with the lessons that turned out to matter, almost none of which were the ones I expected. You are reading where I was, not where I am, with the parts that are still my edge kept dark.

The competition is the ARC Prize 2026, and it is built around the one thing today’s AI is worst at. Hand a child a video game they have never seen, no manual, and within minutes they have worked out the controls, the goal, and what kills them, all by poking at it and watching what happens. That casual act, getting your footing in a world nobody explained to you, is one of the hardest open problems in the field. The benchmark drops an AI into a small game it has never encountered, tells it nothing, and asks it to win by experimenting: take an action, watch the response, form a theory, test it, and build a working picture of a place from a standing start. Humans, including kids, do this comfortably. The most powerful AI systems on earth, for years, scored close to zero. That gap is the whole point. It is the benchmark that refuses to be impressed by how much a model has memorized, because none of it can be memorized.

My approach cuts against the fashion. The popular move is to point one enormous language model at the problem and let it sort everything out. I am doing close to the opposite. I hand-build the mind, the perception, the exploration, the decision-making, all as my own code that I can open up and explain, and I let a language model do exactly one job, the theory-building part, the place the benchmark’s own creator says most of the difficulty lives. The reason is simple: I want to be able to explain every single thing my agent does, and "the big model decided" is not an explanation. Learning first, score second. Here is what that bet has actually taught me.

The only honest scoreboard is the one you cannot see

To improve the agent I have to measure it, and I cannot measure it on the real test, because the real games are hidden. So, like everyone in the competition, I built a practice gym at home from the few example games the organizers hand out. It was supposed to be my edge, the place I try ten ideas cheaply and only submit the winner.

Then one change tested beautifully at home, a clean win of about five hundredths of a point, and cost me nearly three times that on the real, hidden board, pointing the opposite direction. Worse, when I went to understand the loss, my own tooling could not even reproduce it. I had a regression sitting in writing on the official scoreboard and my bench calmly insisted nothing was wrong. My home proxy had at one point been overstating performance by something like twenty times. I had been steering by a gauge that was not just miscalibrated but pointed at the wrong quantity entirely. The full story is in I looked unbeatable in practice and lost the real fight. The rule I took from it is blunt: the leaderboard is the only oracle, the one place the real behavior ever shows its face, and my home bench does not get a vote on whether something works anymore.

I ran into the exact same wall in a completely different competition, one about predicting the depth of rock a mile underground, and told that half of it in a separate journal. Two unrelated problems, two home-built scoreboards, the same lie. A measurement you cannot fully trust is more dangerous than no measurement at all. No measurement keeps you humble. A lying one makes you bold, right before it walks you off a cliff.

More powerful is not the same as better

The obvious way to make my agent win more games is to teach it to solve more kinds of games. So I did. I built a new skill, tested it, and verified it: the agent now beat a game it provably could not beat the day before. Real, new capability. Then I submitted, and the score dropped, from about 0.09 to 0.04, the worst it had done.

When I finally stopped adding and started counting, the reason was uncomfortable. Every time the agent reaches for its theory-building skill, it has to spend real moves poking at a game to test its guesses. On a game it can crack, that is an investment that pays off in a win. On a game it cannot crack, and most of the hidden games are exactly that, the effort is pure loss. Every new skill I added just made the agent attempt more games it would lose anyway, and pay the toll on every one. I thought I was buying wins. I was buying more expensive losses. The full story is in I made my agent more capable and it got worse. The competition does not pay for capability. It pays for net result, and a capability that costs more than it returns is a liability wearing the costume of progress. The real lever turned out to be closer to the opposite, restraint rather than power, and that part I am keeping to myself for now.

The most expensive bugs are in your beliefs, not your code

At one point my agent was stuck, failing on the hidden test in a way I could see but not explain, and across two work sessions I had written down two settled facts about why: the failure only happened on the hidden online games and so could not be studied at home, and the practice games I did have were useless for investigating it anyway. Put those two beliefs together and they say, in a calm and reasonable voice, that there is nothing to be done. They were not facts. They were permission to stop looking, and I took it, twice.

Eventually I did the thing I had been avoiding. Instead of rereading my notes a third time, I wrote small probes and ran them against the actual code and the actual game data. Both beliefs collapsed inside an hour. The failure came apart cleanly on games already sitting on my disk. The dead-end practice data showed the problem plainly the moment I asked it the right question. The wall I had respectfully worked around for two sessions was something I had built myself, out of two wrong sentences I then cited back to myself as if they were evidence. The agent had hundreds of passing tests and the code was fine. The bug was upstream of the code, in what I believed about it, and beliefs do not show up in a test suite. The full account is in The bug was in my beliefs, not my code. The rule I try to apply without mercy now: the highest-risk sentence in any project is "we can’t, because," because it is the one that quietly ends the investigation. A wrong "we can do this" gets corrected the moment you try and fail. A wrong "we can’t" never gets corrected at all.

The through-line

Line these up and the pattern is impossible to miss. Trust the hidden scoreboard over your practice one. Optimize net result, not raw power. Re-derive your beliefs from the artifact instead of your notes. Not one of them is a modeling trick. Every single one is about not fooling yourself, in a domain built to fool you, where the answers are hidden and every instrument you own, including your own memory, will lie to you if you let it.

There is a quiet proof of all this sitting on the scoreboard. My best score is 0.25, and it belongs to the plain, careful version of my agent, the one that does the safe, simple thing and never overreaches. Every cleverer version I have built has lost to it. For a while that felt like failure. Now I read it as the whole lesson in a single number: in a problem this unforgiving, restraint beats capability, honesty beats confidence, and the simplest thing that actually works is worth more than the most sophisticated thing that might. The one move that finally beats 0.25 for real is the piece I am keeping dark until the competition closes. But I already know it will be a lesson in doing less, not more.

Where I stand right now: still 0.25, still climbing, on much firmer ground than when I started, and as stubborn as ever about building a mind I can actually explain. I came into this expecting the hard part to be intelligence. It keeps turning out to be honesty.

More in this series