Marc Andreessen on the atomization of AI

Re:Marc Andreessen on the atomization of AI by thr0waway1239 on Wednesday September 14, 2016 @01:50AM (#12494506)

I was thinking about this notion of using simulated worlds for doing training. An open source simulated world would be fantastic - to the extent that it is actually realistic. For example, maybe collect archived gameplay from some MMORPG and use it to build the simulated world. As new players play in the game, the gameplay data is collected and available for the community to use. Obviously this is not well thought out, but it would be an excellent candidate for a group like OpenAI, assuming they aren't already building something like that.

I wonder if it will be a chicken and egg problem though? Maybe you cannot even create decent gameplay without already having a trove of data, but bad gameplay will not attract more data from potential players?

[ Parent | on HN ]

Re:Marc Andreessen on the atomization of AI by dharma1 on Wednesday September 14, 2016 @04:38AM (#12495017)

https://gym.openai.com

Just needs integration with popular game engines like UE and unity3d, at the moment it mostly works with 8bit game emulators

[ Parent | on HN ]

Re:Marc Andreessen on the atomization of AI by thr0waway1239 on Wednesday September 14, 2016 @08:48AM (#12496290)

Fantastic. We should find a way to place these learning environments at random locations on a world map and interact with them via the popular VR devices :-)

[ Parent | on HN ]

Re:Marc Andreessen on the atomization of AI by outsideline on Wednesday September 14, 2016 @10:22AM (#12497101)

There are several groups that have already done this and there are several tools similar to OpenAI gym. Why is only one company mentioned when it comes to this? You do understand that a large chunk of OpenAI gym are open source software packages like physics engines, box2d?

Here's what you need to make an Ai Gym :

> Physics engine (pick your open source package)

> Simulator (pick your open source package)

> Write a thin wrapper around it

It can be done and has been done by many individuals in less than a week or two of dedicated work.

[ Parent | on HN ]

Re:Marc Andreessen on the atomization of AI by thr0waway1239 on Wednesday September 14, 2016 @12:51PM (#12498873)

Since I honestly don't know enough about this topic, it would be nice if you can mention the other companies. I just mention OpenAI because they are a prominent organization which happen to have both open and AI in their name.

Maybe I am also not being very clear, but I am actually visualizing a game which is popular enough to attract a lot of gameplay, but has gameplay elements which are actually the kinds of things which the AI problems are trying to solve. E.g. imagine capturing gameplay from a popular multiplayer drone game to improve drone navigation.

Drones apparently don't fly very well autonomously.

http://spectrum.ieee.org/automaton/robotics/drones/skydio-ca...

Now, imagine if it were possible to get open data where any researcher or group could work on this problem using a gameplay dataset. Perhaps someone with an unexpected take on the problem could make some important contribution?

[ Parent | on HN ]