I spent a session building a temporal convolutional network to squeeze more signal out of Martian pressure residuals. It moved the leaderboard about 1.3. Then a single constant, found with five submissions and a parabola, moved it 5.4. Here is why my cross-validation never saw either coming.
The competition asked me to predict Martian air pressure on days the rover never recorded. My first submission took the top of the leaderboard and contained no machine learning at all. Here is what it did instead, and the number that told me it was still wrong.
How one project quietly led to another. Finishing LunarSite changed how I watched The Martian, and that curiosity turned into a machine learning project on the Martian environment.
Five lessons from a machine-learning competition where you predict the depth of rock you cannot see. Almost none of them are about models. A summary of the arc so far, and the through-line underneath it.
A passing test suite is the most reassuring lie in software. Three times in one competition, green checkmarks certified a broken model, each one lying in a different way. Here is how, with code.
In a competition where every experiment costs hours and submissions are rationed, the highest-leverage skill is not a better model. It is a rule written down in advance that lets a cheap probe tell you not to build the thing at all.
For a month my local validation and the competition leaderboard disagreed about whether I was improving at all. They were computing two different statistics. Here is how I found out, and the one-submission experiment that fixed it.
A summary of the arc so far from inside the ARC Prize 2026: the three failures that taught me how the competition actually works, and why the plain agent still beats every cleverer thing I have built. Almost none of it was about the model.
An ARC Prize 2026 builder's log entry. I spent two work sessions reasoning from two "facts" about my own code that were both false. Checking them against the artifact took an hour. The highest-risk sentence in any project is "we can't, because."
An ARC Prize 2026 builder's log entry. I taught my agent to solve more kinds of games and the leaderboard score fell from 0.09 to 0.04. The capability was real. The cost of using it was the problem. More powerful is not better.
An ARC Prize 2026 builder's log entry. A change that gained on my home bench lost three times as much on the real, hidden leaderboard, and my own tooling couldn't even reproduce the failure. Why the leaderboard is the only oracle.
A new thread in the builder's log: I'm competing in the ARC Prize 2026 (ARC-AGI-3), the benchmark built to measure the one thing today's AI is worst at, learning something it was never trained for. What the comp is, why it's the field's uncomfortable mirror, what it would mean to crack it, and where I'm standing.