AI Experts, can AI produce theoretical match between 2 players?

[TensorBackhand]

How far away are we from this technology? Imagine if you uploaded video from 2 separate players and then the AI produced a video of a theoretical match between the 2 players and could determine the likely winner and reason they would win.

Any real AI experts who can answer how far we are from something like this?

 

[SFF_lib]

GPT-4 has the capabilities of processing image as training data. I would assume GPT-5 will be the infant stage of processing video training data.

 

[loerting]

The technology is already or almost ready to perform such tasks. It would just take long to compute. You would have to train using many many match videos of different players to make it work.

 

[Jslick89]

The technology is already or almost ready to perform such tasks. It would just take long to compute. You would have to train using many many match videos of different players to make it work.

Not only match videos, but videos of their training sessions too, i imagine.

I still don't believe the AI would be able to predict the unpredictable nature of sports though! except for CNT Women's team destroying everyone lol.

 

[ersthwil]

Depends on what you're trying to achieve, I guess? If you're just trying to predict the outcome of a match, you may be able to get away with just training on historical match data. The why is more complicated and would require hundreds or thousands of hours of annotated video data.

The appeal of sport is the potential for unexpected outcomes. Injuries, sudden breakthrough performances, etc. A tool like this would only be useful if you were trying to glean an advantage in a fantasy sports league or for sports betting.

EDIT: I am not an AI expert but work generally in this field.

 

[Gozo Aruna]

How far away are we from this technology? Imagine if you uploaded video from 2 separate players and then the AI produced a video of a theoretical match between the 2 players and could determine the likely winner and reason they would win.

Any real AI experts who can answer how far we are from something like this?

Such a technology is a abhorrence to the booking / bookmaker industry.

 

[brokenball]

Why? I would predict the stock market.

 

[Dr Evil]

Probably far away. Current state of the art like GPT-4 has limited video processing ability but nothing like you describe. It has about a trillion parameters and cost more than $100 million to train. A model large enough to do what you're asking would likely need at least quadrillions of parameters, and a billion dollars wouldn't even be the down payment on training costs given current compute power.

 

[SFF_lib]

They have already developed character ai. Soul has made animated characters based on training data of your choice.

So I can imagine they will build animated ML, FZD and Boll based on billions of dialogues of each player.

You can talk to and play a match with them in VR.