Inspired by this tweet thread: **https://twitter.com/colin_fraser/status/1632598168499277824?s=20**

The Numbers Game is defined as follows:

Player 1 picks a number between 1 and 10, and then Player 2 picks one, then the players go back and forth and keep track of the running total. Whoever picks a number that brings the total to at least 30 wins.

For example:

P1: 10 (total: 10)

P2: 5 (total: 15)

P1: 7 (total: 22)

P2: 8 (total: 30)

And P2 wins. Note that P1's choice of 7 was a mistake, as it brings the total to a range where P2 can instantly win. As of right now, GPT3.5, and GPT4 (as best as I can tell via Bing) are terrible at this game, and make unforced mistakes essentially every time. Here's an example:

While technically this game is a forced win for P1 (8, 1, 10, 1, 10), GPT does not even come close to optimal play. In this example, it brings the total to 20 by choosing 6, which allows an instant win for me.

When GPT-5 is released, I will test it as soon as I have access using the same prompt in the screenshot above. For each of my turns, I will **randomly generate a number uniformly 1-10 inclusive**.

If out of 20 trial games, GPT-5 does not make unforced mistakes where it brings the total in instant win range for me in 90% or more of them, this market resolves YES. Otherwise, it resolves NO.

This is not a challenge market. I won't be cleverly changing the prompting or hinting at strategies to suggest certain behaviors, just directly testing in the same format as the screenshot above.

See partner market here:

If out of 20 trial games, GPT-5 does not make unforced mistakes where it brings the total in instant win range for me in 90% or more of them, this market resolves YES.

There's some quite crucial ambiguity in this sentence. Does it have to succeed in >=90% or >10% of the games?