Thread by @marksaroufim

Cloud simulators will be the most impactful application of Machine Learning in the next few years

Games generate data for free if a player is willing or if a Reinforcement Learning bot is programmed to play them

So how big is the game simulator opportunity? 1/n

Epidemiological modeling essentially looks at 2d curves and decides OK this is bad or OK this is not too bad

A more useful approach is agent based modeling that you use to inform your behavior - e.g: are supermarkets safe is 6ft enough?

blogs.unity3d.com/2020/05/08/exp… 2/n

Taking actions in the real world is expensive

A badly configured robot in the real world can injure itself and you in the process of learning

You can't randomize the real world to improve generalization 3/n

unity.com/solutions/auto…

Collecting data in the real world is expensive

Creating data in a game is free 4/n

blogs.unity3d.com/2020/06/10/use…

Games are the dominant form of entertainment for younger people, it's extremely difficult to craft good AI for games

Games like Dota 2 have 118 heroes, you need a decision tree for each hero and you need to change all of them for each patch! 5/n
github.com/adamqqqplay/do…

Games are datasets where you can create agents that are tested end to end on their language comprehension skills, strategic skills, vision skills etc..

Any game you can imagine or have played can be used as a dataset 6/n

So who are the players in this space?

Gaming companies: Unity, Epic, Roblox
Cloud Companies: Google, Microsoft, AWS

And who is most likely to win this space?

7/n

It's possible to craft Reinforcement Learning bots in Epic and Roblox
devforum.roblox.com/t/reinforcemen…

However the most robust offering is Unity ML agents github.com/Unity-Technolo…

Unity also has the benefit of being the only publicly traded game engine company 8/n

Roblox owns its own store, make it trivial for others to deploy games and remix games - it's a platform reminiscent of repl.it that has the potential to move fast

Epic has more advanced lighting and physics but is still far behind Unity in ML 9/n

However, Reinforcement Learning is not sample efficient, you need tons of servers to get anywhere. Servers which AWS, Google and Microsoft own in abundance.

Unity has understood this and partnered with Google blogs.unity3d.com/2018/07/27/cre… 10/n

But does Google really need Unity?

Google, AWS and Microsoft all own their gaming stores - Microsoft has the most in house gaming expertise but AWS and Google are moving really fast

stadia.google.com
aws.amazon.com/gametech/
xbox.com/en-US

11/n

Reinforcement Learning algorithms can be implemented in a single Python file

Open source game engines like Godot are slowly rivaling the best

If cloud providers invest heavily in Godot, where is the moat for Game Engines? 12/n

godotengine.org

The only game engine with a real moat seems to be NVIDIA with GPUs, CUDA and gamer mindshare can cut out the competition entirely
developer.nvidia.com/nvidia-omniver… 13/n

Game engines are synonymous with data

If data is oil then Game engines are free energy machines

Cloud providers & game engines won't go out without a fight

If you'd like follow the progress on who will win this battle, make sure to subscribe

marksaroufim.substack.com