Skip to content
Trending storyDeveloping

Design Arena: K3.1 Pro games finish more often than K3 games

3 articles1 sourcesince Nov 2Last article 1h ago ·

Overview

AISummary of 3 articles

Design Arena reports that K3.1 Pro's generated games are more often finishable than K3's.

It says 62% of K3.1 Pro's games can be won or finished, against 40% for K3, though both models' games start at similar rates (94% and 91%). The gap holds even for prompts that never ask for a win, level, goal or mission: 53% of K3.1 Pro's games can still be finished, against 28% for K3. Among users who prompted both models, the figures are 73% and 30%.

Design Arena also reports that, within each model's own games, a finishable game earns about 83 rating points more in head-to-head votes than one that cannot be finished, with opponents held at their leaderboard ratings.

Written by AI from the articles below · updated Nov 2, 1:09 PM ET

Check the sources:

Developments

3 developments

  1. Nov 2, 12:58 PM ET · 1 article
    Design Arena finds finishable games worth about 83 rating points more in head-to-head votes
    Design Arena: Design Arena finds finishable games win about 83 rating points more in votes
  2. Nov 2, 12:58 PM ET · 1 article
    Design Arena reports K3.1 Pro games finish more often than K3 in prompt-level comparison
    Design Arena: And K3.1 Pro builds the whole game, not just the part users described.
  3. Nov 2, 12:58 PM ET · 1 article
    Design Arena reports K3.1 Pro's games finish more often than K3's
    Design Arena: That thinking shows up in the games.

Article timeline

The articles in this story. Times are ET.

Nov 2
  1. DesignArenaOfficial
    Design Arena finds finishable games win about 83 rating points more in votes

    AIDesign Arena says that within each model's own games, a finishable game is worth about 83 rating points more in head-to-head votes than one that cannot be finished, with opponents held at their leaderboard ratings.

    Video from @DesignArena's post
  2. DesignArenaOfficial
    And K3.1 Pro builds the whole game, not just the part users described.

    AISeven in ten prompts never ask for a win, a level, a goal, or a mission. On those, 53% of K3.1 Pro's games can still be finished, against 28% of K3's. When the prompt does ask for an ending, it's 84% against 71%. It isn't down to who was prompting, either. Among the users who prompted both models, 73% of K3.1 Pro's games could be finished, against 30% of K3's.

    Video from @DesignArena's post
  3. DesignArenaOfficial
    That thinking shows up in the games.

    AIDesign Arena says K3.1 Pro starts 94% of its games and completes 62% of them as playable, against 91% and 40% for K3. The post says the gap shows up in games that can be won or finished.

    Video from @DesignArena's post

Heat Trend

Not enough continuous observations to show a trend yet.