Resolves positively if there is an AI which can succeed at a wide variety of computer games (eg shooters, strategy games, flight simulators). Its programmers can have a short amount of time (days, not months) to connect it to the game. It doesn't get a chance to practice, and has to play at least as well as an amateur human who also hasn't gotten a chance to practice (this might be very badly) and improve at a rate not too far off from the rate at which the amateur human improves (one OOM is fine, just not millions of times slower).
As long as it can do this over 50% of the time, it's okay if there are a few games it can't learn.
I spoke a bit about this market in a comment on another market:
https://manifold.markets/ScottAlexander/in-2028-will-an-ai-be-able-to-gener#xauyrw6cY45QGI3yxks7
Can it practice similar games? When a human is playing a video game for the first time, they're usually drawing on experiences from playing other video games, and video games take advantage of this by making many of the inputs/outputs similar across games (left stick usually moves you around, right stick makes you look around, etc.)
I think this would require human level generality. Current video game AIs are this good because they get to play for so long they basically explore the whole state space of the game - not literally the whole state space, but for each possible state, the AI experienced analogous states during the training phase.
When it comes to games like puzzle games or strategy games, it seems to me that this requires having the same ability as humans to generalize from little experience, and is nearly equivalent to predicting human-level AGIs before 2028.
one actionable consequence of this opinion is that I think the divergence between these two questions is wildly larger than it should be. I will be taking this bet. https://manifold.markets/ScottAlexander/by-2028-will-there-be-a-visible-bre
This is essentially my view. An agent that passes this test would be dramatically more economically useful than almost any form of historical automation, since the conditions are so general
what kind of learning during the game would be ok?
e.g.
- self learning after the game has already started?
- using a training-algorithm (e.g. backprop) to update an intemediate modell (e.g. the reward function)?
for instance: with the world-modell approach [1], the agent learns about vision, than it learns predictions about the world, and then it learns strategy in that imagined world. The world-model is first created by observing the word, so no self-training here. The control-model is created by self-learning on the world-model, but not on the real game. So that would would be self-learning, but not on the game, but instead inside its own mind (kinda like a human-player simulates the game world in his head to come of with strategies before trying them out in the real game)
in this example: If an AI would train a world-model (V and M Model in the paper) by observing games, but then use a pre-trained* C-Model, how would that resolve this market?
* pre-trained on other games, not the specific on in question.
[1] https://worldmodels.github.io/
Related markets
Related markets