An agent that playtests games from screenshots.
Here it plays Freedoom and my phone game Purrfect Fit. It reads the map, works out the drag itself, and replays every bug twice before it counts.
Built on Meta's Muse models for the Meta Global AI Developer Hackathon @AIatMeta
An agent that playtests games from screenshots.
Here it plays Freedoom and my phone game Purrfect Fit. It reads the map, works out the drag itself, and replays every bug twice before it counts.
Built on Meta's Muse models for the Meta Global AI Developer Hackathon @AIatMeta
Fair question. That's why the replay is there. The agent saying it found a bug doesn't count. The same inputs get played again on a fresh copy of the game, twice, and the game has to do the same wrong thing both times. If it doesn't, the report gets thrown out. So what gets filed is the game misbehaving, not the AI imagining it. It's a second pass for solo devs, not a replacement for real players.
Ha, yeah the account has had a few lives. The replay is holding up well. On my test game with 5 planted bugs it found all 5, and every report replayed the same way twice. On a game it had never seen it got 3 of 5, and one false report still slipped through, so it's not perfect yet.
@Survivalguy9@AIatMeta A one to ten scale for a "how many" question, plus "you're" for "your". You really should have had an AI look at that before you hit post.
@blakecasimir@AIatMeta Haven't tried it myself, this one only plays and tests games, it doesn't make them. My guess is code is the closest since it's just text. Maps seem harder because they have to actually play well. Would be fun to point this agent at whatever a local model makes and see what breaks
An agent that playtests games from screenshots.
Here it plays Freedoom and my phone game Purrfect Fit. It reads the map, works out the drag itself, and replays every bug twice before it counts.
Built on Meta's Muse models for the Meta Global AI Developer Hackathon @AIatMeta
@AIatMeta The harder test was a game it had never played. It found 3 of 5 planted bugs, and 1 false report got through. It got fooled once more on the clean build of that game. So it is useful, not perfect.
@AIatMeta How it did: on the game I tuned it on, it found 5 of 5 planted bugs with 0 false reports. On the same game with no bugs, it won in 47 steps and reported nothing.