This idea of computer programs learning to do stuff and humans creating stories about it have been around at least since Conway's Game of Life.
It hard to get at the value of the playing Hide-and-Seek example, but it's certainly not in their anthropomorphized 'strategies'.
Stick to real games with clearly defined rules and arenas humans can also play in if you want to show progress, else it's probably made up for PR purposes.
My wife and I did some fun little experimentation in 2015... we simulated a sea slug which has a very simple neural circuit / observed behavior. Basically looks for food.
The hypothesis was adding a state of aggression, territoriality in nature (patrolling food), would make them appear social.
It worked! We then transferred the learned model parameters to robots for fun (and people thought they were social). Note that we left the models an option to “evolve” in several ways, they always converted to being aggressive in defending food spawn locations.
In any case, this makes sense. When you have complex systems and you apply a sort of pressure, models (or animals) will learn to “survive” because that’s what is passed on.
Why post the Quanta write up on a months old publication? OpenAI released this months ago: https://openai.com/blog/emergent-tool-use/ and IMO, did a much better job of explaining/detailing everything.
EDIT: Just finished reading both and I can say the Quanta publication doesn't add anything except a short tangent about the history of self-play. Seriously, go read the openAI publication. It has great reactive visualizations and is much more put together.
6 comments
[ 2.9 ms ] story [ 27.9 ms ] threadhttps://openai.com/blog/emergent-tool-use/
https://news.ycombinator.com/item?id=21006616
It hard to get at the value of the playing Hide-and-Seek example, but it's certainly not in their anthropomorphized 'strategies'.
Stick to real games with clearly defined rules and arenas humans can also play in if you want to show progress, else it's probably made up for PR purposes.
The hypothesis was adding a state of aggression, territoriality in nature (patrolling food), would make them appear social.
It worked! We then transferred the learned model parameters to robots for fun (and people thought they were social). Note that we left the models an option to “evolve” in several ways, they always converted to being aggressive in defending food spawn locations.
https://austingwalters.com/modeling-and-building-robotic-sea...
In any case, this makes sense. When you have complex systems and you apply a sort of pressure, models (or animals) will learn to “survive” because that’s what is passed on.
EDIT: Just finished reading both and I can say the Quanta publication doesn't add anything except a short tangent about the history of self-play. Seriously, go read the openAI publication. It has great reactive visualizations and is much more put together.