"Dice forget. Cards remember"

Very interesting video highlighting a fundamental mechanic involving dice and cards in game design.

Both are randomizers: they introduce uncertainty into a situation. But they differ in one important way. One roll of the dice never affects future rolls nor is affected by a previous one. But drawing a card from a deck changes its shape and influences its future draws, creating randomness with memory.

By shuffling the deck we clear that memory and resets the situation. Future actions are freed from the influence of the past. That introduces a very important design decision: when to reshuffle.

Related to that, in deck-building games removing cards from the deck is sometimes even better than adding them. You’re reducing the options and making your hand more predictable.

In other words: Dice are stateless; Cards are stateful. Watching the video made me think a lot about LLMs.

It’s curious that the analogy applies so directly. The models are stateless by nature but by building a harness around them, we make these probabilistic machines stateful.

They are still random-ish and pseudo-stateful, but from a practical standpoint they can remember, in a similar way to a deck of cards.

In game design, deciding when to reshuffle has a huge influence on the mechanics and style of the game. There’s not a single right answer. Shuffling more often is not necessarily better than shuffling less often. It all depends on the goals, the audience and the intended feel of the game.

There’s a similar situation with Context Management in agent design. How to spend those precious tokens is not a question with a single right answer either. The task, costs, user preferences… all need to be taken into account.

When building or using an agent, how will you decide when to shuffle the cards?

🎲🃏