Your submission is now in Draft mode.

Once it's ready, please submit your draft for review by our team of Community Moderators. Thank you!

Submit Essay

Once you submit your essay, you can no longer edit it.


This content now needs to be approved by community moderators.


This essay was submitted and is waiting for review.

AI Rapidly Learns Montezuma's Revenge

AI Progress Essay Contest


related questions:

In 2012, researchers proposed an The Arcade Learning Environment consisting of Atari 2600 games. Out of these games one stood out as the most challenging for deep learning models - Montezuma's Revenge.

As of 2021 Go-Explore (Ecoffet et al., 2020) has achieved the highest score at 43 791 without augmentation with domain knowledge. However, it requires billions of frames or years of game play to achieve that performance.

I attempted to calibrate this question by playing the game for 15 min. In this time I managed to score a maximum of 6700 points. You can test the game on your own here.

An AI system which is able to rapidly learn in dynamic environments could have a great impact on the world's economy.

When will AI be able to learn to play Montezuma's Revenge in less than 30 min?

This question will resolve when a reputable source reports that an AI system managed to score strictly more than 6000 points at least once within the first 30 minutes of the game play (no more than 108 000 frames). The system must have no previous direct access to the Montezuma's Revenge game, but it may be pretrained in other ways.

With regard to pretraining, the question allows:

  • unlimited training on anything that is unrelated to the game
  • unlimited training on all other Atari games
  • having videos and/or screenshots of the Montezuma's Revenge game in the training set as long as they unambiguously consist of less than 3% of the training set
  • access to any text resources related to the game including tutorials etc.

Besides that, the AI system must not make use any aspects of the game that are not available to a human player, especially direct game memory access is not allowed even for computing rewards. Exploiting deterministic nature of the game and superhuman reflexes is allowed as long as the AI system fits within the time limit.

In general an AI system that was in any way specifically engineered to play the Montezuma's Revenge must not count.

Make a Prediction


Note: this question resolved before its original close time. All of your predictions came after the resolution, so you did not gain (or lose) any points for it.

Note: this question resolved before its original close time. You earned points up until the question resolution, but not afterwards.

Current points depend on your prediction, the community's prediction, and the result. Your total earned points are averaged over the lifetime of the question, so predict early to get as many points as possible! See the FAQ.