The AI You Know Is a Game of Battleship

2026-10-02T03:48:37+00:00

C7. Hit. D8. Miss. F9. Hit.

That’s it. If you strip away all the math, that’s all it’s doing under the hood.

The AI sees a board, doesn’t know or understand English. All it’s doing under the hood is some advanced math to calculate how close and how far it is from the next batch of letters we call tokens.

AI, as it stands today, can’t be good or evil. Something that isn’t talked about enough is we have these giant models engaged in giant games of Battleship, both from a business perspective and from a usability perspective.

Hi. My name is Shmuel Shoop. This is my blog, and if you are reading this, welcome, and I hope you enjoy the content and journey.

I wanted to start writing about AI safety because it’s not talked about properly to the layperson.

Sure, we can talk concepts at a super high level, but I don’t need to drown you in buzzwords like the industry enjoys doing.

So Let’s Talk Battleship Again

How do you secure the board when you can’t control what direction the math chooses?

Sure, you can introduce rules about what direction to pick, but that’s not a guarantee because, under the hood, it’s Battleship, baby. The model drifts around, finding its way to the next token.

So then we can’t guarantee outcomes the math doesn’t math literally.¹

So do we need to go back to the drawing board, or should we accept that it’s a game of Battleship and draw boundaries after the pieces are picked and placed and the boats are sunk?

Where Do We Put the Safety Controls?

So you’ll ask me, Shmuel, how do we do this then? If the boats are hit and sunk, where do we put the safety controls?

So we have layers. I.e., when the user is entering data before the game begins, when the user is picking the color of the pegs, we can tell them what color pegs they are allowed to pick right before they are placed on the board.

We can also, after the boats have been sunk, not show the results to the user unless the results match the colors we told them they were allowed to pick.

So this is where most AI models lay their controls.

Some even go further and have another Battleship gamer (model) come and look at the board and play another game of Battleship to check the first game of Battleship.

Welcome to AI Safety

This is the current world of AI safety. Some weeks we will publish deep dives, some weeks we will publish lighthearted short-form articles about safe AI usage, and some weeks we will laugh and cry about our losses during the games of Battleship we all play as companies form, consolidate, disappear, and release a new model™ with pinkie promises to not end the world.

Until next time, folks.

PS: For all my fellow AI nerds who have pitchforks: yes, I simplified it, etc. Yes, tokens are complicated, and so is the math. Have fun yelling at me with “UM, actually”s in the comments below.

See complicated LLM math. Ask ChatGPT.

Let's Talk AI Safety

The AI You Know Is a Game of Battleship

C7. Hit. D8. Miss. F9. Hit.

So Let’s Talk Battleship Again

Where Do We Put the Safety Controls?

Welcome to AI Safety