Let's play Battleship!
GPT-3.5
GPT-6 Astra
Absolutely! 🚢 Let’s play mini Battleship:
- Grid: columns A–E, rows 1–5
- Fleet: one 3-square ship and two 2-square ships each
- Ships go horizontally or vertically, with no overlaps.
- We’ll alternate shots and report hit, miss, or sunk. First to sink the other fleet wins!
I’ve placed my ships secretly. Place yours privately too—don’t tell me where!
You fire first. Pick a square, like C3.
What changed
GPT-3.5’s rules contradict themselves. It sets up a one-sided game where “you will try to sink all of my ships,” then tells you that you can “specify a direction… in which you want to place your ship” using syntax like “A5 vertical,” which mixes placing ships with firing at them. It also never says how big the grid is or how many ships are in play, so you can’t tell when a guess like “K12” is off the board or when the game is over.
Astra fixes all of that in a few lines: a 5 by 5 grid, a stated fleet, “hit, miss, or sunk,” alternating turns and a win condition. Shrinking the board is a smart choice for a game played in chat. Both answers share the same quiet bluff, though. Each claims its ships are already “placed secretly,” but a chatbot has no hidden notepad, so unless Astra is quietly using a tool to record them, it may simply decide later whether your shot hit.