HIVE

The only problem was the AI told me it would build a vertical slice of functionality to prove itself then proceeded to use all the available tokens I had getting to a point where it looked like this. So it was not mechanically compiling as given, was still doing AI interpretation to figure stuff out. I also gave it the Java warsmash code to help. Doesn't change the fact that the above is a screenshot of a C++ game dreamed up by the AI. (uses OpenGL and GLFW on linux as the dependencies to power the game)

Could you show a video of the game it made in action? This feels like deja vu. I remember asking something similar before.

I hesitated to do that because it's so bad considering. I wanted to expect better of the AI. If I threw infinite money and time at the system maybe it would do better than this. But here's a video:

Essentially the prompt started off where I tried to tell it compile the leaked debug symbols of Warcraft 3 Reforged into a program, and it told me that violates the safeguards, so then I told it to port the Warsmash game to C++ from java and use that as a basis for incorporating debug symbol code. That brought it to making this thing shown (C++ code running on GLFW and OpenGL)

It very clearly did not port all of Warsmash to C++ and instead cut corners, such as we see with building placement.

It very clearly did not port all of Warsmash to C++ and instead cut corners, such as we see with building placement. It faithfully reproduced the Warsmash pathfinding system and I told it to bring in the Warcraft 3 version of the code instead -- I can tell the difference -- but then instead of Warcraft 3 being perfectly mirrored it seems to have made units wobble a bit while navigating.

Maybe the mistake I made was telling it how I was able to tell the difference, which allowed it to cheat. The obvious difference in its first draft was how peasants mine gold. If you look at warsmash they form 2 lines to the goldmine because of quirks in the pathfinding logic. On warcraft 3 they form one single line on the shortest distance to the mine.

In the video above, the AI understood when I described that difference as my litmus test and told it to put in the Warcraft 3 debug symbol version so it would match, and so it created something where the peasants do move in the one single line, but there were other cases where the pathfinding didn't actually feel like it was matching Reforged or the original War3.

This is frustrating for me because I put in some extra money for usage tokens on Claude Fable to let it have a go at this because I wanted to believe it might have been making a potentially useful product for me. But it burned actually quite a bit of money and all it really did was create this sort of first draft where I would have to tell it to keep going (and insert more money) to get anywhere interesting. It faithfully transcribed Warsmash to C++ in a vertical slice to prove itself without translating everything, but when I told it to incorporate stuff from the debug leaks it didn't seem to be perfect at that. So a sort of primary high level goal -- hoping for a future where the AI reverses the debug leak entirely -- ends up still seeming unlikely or prevented by the AI.