4 ms·
My main experience with the demo on the site is that I asked it to add a simple feature which would not even necessarily let you win the game, and it would spen
by djaro 10d ago
My main experience with the demo on the site is that I asked it to add a simple feature which would not even necessarily let you win the game, and it would spend >100k tokens in a loop of "laws broken" until maybe adding it, maybe not. After burning half a million tokens just to add an extra line of walls so that teleporting 2 squares up wouldn't let me win, my API key got rate limited and broke the loop.
On top of that, the solutions feel like patchwork. I asked it to let spacebar flip the board horizontally, and it responded by making the board completely symmetrical including 2 flag poles. At some point it just has to say "this isn't possible without breaking the laws" or think of an actual workaround, because if I was making a game, suddenly having 2 finishes would be unwanted behavior for me.
- andy12_ 10d agoI think the demo is this way simply because looking at the LLM find wacky ways of implementing features without breaking the law is fun and drives the point across. In practice I imagine you would write something like "If you don't see a clear way of implementing a feature without breaking the law, ask me for directions" in AGENTS.md