I’ve opened Goo AI Arena, a challenge to use AI to build the tallest standing tower in World of Goo.
The task: use all 300 balls through legal in-game building steps and keep the finished structure standing. There’s a leaderboard, submission instructions, and a completed example.
Codex and I built the first entry: 13.9 metres, titled “It’s a tower.” I provided guidance, so this is an example of human–AI collaboration. My main lesson: teaching AI is hard.
Beyond the competition, I’m interested in its potential for agent evaluation. Building a tower involves planning, interpreting a changing physical environment, and recovering when construction doesn’t go as expected. Height is an easy outcome to measure; separating model capability from tooling, compute budget, and human assistance is the harder question.
I’d welcome both new contenders and ideas for making comparisons scientifically useful. What controls would you want before treating this as a benchmark?
Repository, first tower, and submission instructions
Fun idea, and physics-based building tasks are actually a decent underused niche for this. One thing worth noting from the repo itself: publication is explicitly self-declared and not independently verified, the platform doesn’t run competitors’ AI or load submitted saves, so “verified” here really means “community reproduction attempts,” not automated proof. That’s honest of you to disclose, but it also means your scoring/controls question is even more important than usual since nothing stops a legit-looking submission from being human-heavy despite the disclosure field.
On controls: fix compute/attempt budget per agent (max steps, max retries, wall-clock or token budget), otherwise “tallest tower” just measures who got the most tries. Since you already require a self-declared autonomy label (fully machine / human+machine / fully human), the leaderboard should probably split rankings by that category rather than mixing them, comparing your Codex-assisted 13.9m against a fully autonomous attempt isn’t apples to apples even with disclosure. Third, given attempts are reproducible via native save + action log, multiple runs per agent (median/best-of-N) would separate skill from lucky physics outcomes a lot better than single best-attempt scoring.