2048 part 2

About a week ago, I published how to get AI to develop a 2048 game emulator. See here for that post.

That was fun, but you know what would be more fun? Having a simulator that can try various sequences and heuristics to solve the 2048 game. Yes, I did that – again with AI.

The playground is here. I utilized Github Copilot via VSCode and settled eventually on Claude Sonnet 4 as it seemed to iterate and test the changes it was making more so than the others.

There were quite a few learnings in this:

1. AI is a bit like human developers- the code just increases in complexity and requires sporadic re-factoring. That re-factoring is still human-based intuition and not part of the AI itself.
2. Just like it won’t intuitively determine when re-factoring is needed, it also won’t prune / clean temp files or backups itself either. With a few iterations, the AI decided to make backups as the changes were more extensive – those backups ended up being included for a few commits. (I did notice this behavior changed half-way through the project, so maybe this is fixed now).
3. When there are extensive changes based on the prompt, there is no warning, and there are sometimes avoidable errors. Missing curly bracket or messing up the UI layout was a common occurrence.
4. When I came up against blockers (like finding the right heuristic or fixing the missing UI component) a straight up ask to the AI helped quite a bit. Though the documentation it created was also overly optimistic (claiming up to a 90% success rate).
5. Kids are a really good testing group. With their feedback, I implemented mobile device support, descriptions about the heuristics, and made the UI a bit more friendly.

This was such a fun project, and I admit – I keep finding tweaks to make.

If you’re interested in the solution (which heuristic or sequence solves 2048 the most), I suggest you check out the stared heuristics in the playground!