Hacker News Reader

Comment by devit | Hacker News Reader

You can't really do prediction if you are not rendering the game locally, so all you can do is have a lot of servers all over the world and rely on most customers having a low latency fiber link.

View on HN

For the camera alone one could do a few tricks on the client like VR's timewarp/reprojection, but that doesn't work for gameplay actions (like pressing the fire/jump button).

Theoretically the server could speculatively render and transmit a number of different "potential future frames", and the client throws the wrongly predicted frames away.

That's a nice way to burn even more energy and bandwidth though ;)

Stadia controller connects directly to WiFi to the remote server though, so it sounds like the Chromecast or whatever isn't even getting the local inputs to do such tricks.

TVs has been so slow for so long anyway. I suspect we will just have a ton of point and click style games or games that are mostly simulations with relatively uninteresting inputs but with potentially bigger visuals, or casual games where even 300ms+ latency doesn't make much of a difference. For better or worse, reflex (time) based games will just not be played on Stadia by serious gamers. We have VR for that now.

> For better or worse, reflex (time) based games will just not be played on Stadia by serious gamers.

True, but don't overestimate the importance of "serious gamers" to the industry's bottom line. IIRC, mobile gaming is now making more money than all PC and home systems combined.

> We have VR for that now.

VR actually introduces further latency problems. With a TV, your typical PS4 can cover up latency issues with spectacle, as you mentioned. That's a big reason I suspect so many console games feel like a movie today, with tons of cutscenes and quick time events. I've been playing a lot of Bloodborne lately, and while it's an action game, it's still incredibly slow compared to something like Quake.

But with VR, when you move your head you expect the world to feel as if it is real. A TV is artificial, and latency is not an intrusion in the experience. But with VR, latency is felt on a deeper level. Potentially resulting in headaches and nausea.

The funny thing is that John Carmack is riding the state-of-the-art decades later on this front as well: https://www.wired.com/2013/02/john-carmacks-latency-mitigati...

The servers could predict what the user does in the next 100ms. Not the same kind of CSP, but fits well into "powered by AI" marketing...

They probably would, but the creative players would suffer from it. AI never predicts creativity.

Perhaps if you train the model on the existing movement and action history of the particular player

Oh, that's a great idea!

Stadia could sell it as an add-on to players which not only don't want to play their games themselves, but also aren't satisfied by watching other people play through their games on YouTube or Twitch. With this add-on, they can finally watch themselves play through their games, without having to lift a finger to, you know, actually play!

They would make a killing on Twitch, where a streamer could just buy someone else's training data and use that instead of playing themselves.

Black market dealers would swap Terabytes of hot RAM with manually inputted data from the best e-sport players in the world. Corporate enforcers would hack those dealers to delete the data.

Call it Sonic Mnemonic.

I thought about it. But first, you need a lot of information to train a model, so it would only work for very heavy players. Second, creativity is not defined only as doing an action that others didn't do (often or never) before, but also of doing an action that you didn't do (often or never) before.

No offense but I think you underestimate the predictive potential of millions of hours of game state

None taken. But your claim is that creativity is not possible anymore, since it was all done in the "millions of hours of game state". However, if creativity is possible, then my argument is correct.

Another issue is that machine learning/AI don't predict rare events, like earth quakes. So even with all the knowledge in the world, it won't predict a rare creative move of a player.

But every event, creative or otherwise, is made up of hundreds of smaller events. That complicated wall jump - 360 kill you just did used several input signals. Even if the server side AI can't predict the exact final outcome, it can definitely help with the intermediate, well known states for at least some of the input systems.

I say some but I do believe a large enough volume of data can improve the performance of this class of input/states.

Yes, and then you predict something, broadcast it to your clients, and it ends up being wrong so the clients have to roll back. Would not be a good experience.

I think prediction failures are more likely to punish the opponents of the unpredictable guy. In a lot of online shooters, people with 300ms+ ping blink around unpredictably and appear to suddenly murder you out of nowhere, but they don't seem to have any trouble themselves.

I am now imagining a (probably short) story in which the AI does learn to predict players perfectly, even the creative ones, and ends with a gamer taking his hands off the controller and allowing the AI to play exactly as he would have and wondering what was ever the point.

...what's the point of playing competitive first-person-shooters when an AI is at the controls.

And those types of game are the ones that are suffering most from input latency.

I actually kind of find the thought hilarious.

In a twitch shooter, you mouse over a visible opponent. Would the AI be more likely to pre-emptively pull the trigger for you? Or more likely delay and swing the aim passed the opponent before shooting at air? AKA Is the training set for predictive user actions based on experience of players better or worse at the game than you?

The problem there seems two fold, one that you need even more beefy hardware to predict, simulate and render ahead of the player input particularly when it has to catch up to deal with misprediction and secondly that mispredictions are going to make the game feel really imprecise at best and jarring at worst.

I'm also skeptical about the ability for a model to generate predictions without having too many mispredictions to make it viable.

I think the problem here is, the game is made by Studio A, but it's run by Google in the cloud. Studio A probably doesn't have enough of an incentive to put this in the binary, but Google doesn't have the source, so they can't change the game loop.

That would make the delay much worse if you mis-predict the users action. Games would also have to add additional 100ms lag for all important events.

If you had a big enough server, you could render multiple frames for each possible user action...

True, that could work, and the two frames are probably very similar, so wouldn't even require that much more bandwidth. As someone here pointed out, however, the game controller is connected to the cloud directly, so the display doesn't even know of the inputs until the roundtrip is already done.

Reply on news.ycombinator.com