The AI future of Second Life.

[Lumpy Tapioca]

(I know that this is a few years too soon, and saying AI makes people roll their eyes...)

The future of SL isn't GLtf or raytracing or $3000 worth of Nvidia on your desk,
and most of the world that already exists will never be retrofitted into PBR or whatever follows.

This is the only path forward.

Use the current SL view, take what the current renderer is fed (geometry and textures and avatars),
and feed it as a seed/scene description to a real-time server-based AI renderer, that streams your view back to you,
making this largest collection of 3D assets ever created look like the reality we're chasing for over 20 years.

It would work magic even with our 20 year old builds, making them look new again.

Because you would be feeding it actual geometry, it would work way better than "A castle on a lake, with birds overhead", 
like we do now with AI image generators.

What's the workflow now if you wanted to go out dancing as JKF and Marylyn Monroe? Hunt for avatars, pay for them,
make them fit and work correctly, hope that JKF and Marylyn from different vendors look compatible..

Many steps that are difficult at best and nearly impossible for beginners.

This vs. "Make us look like JFK and Marylyn. Put JFK in a white tux, Marylyn in a red sequined floor length dress."

Done, and better.

SL has so much user interface surface area with its sea of buttons and sliders. That's hard for people,
especially in the new mobile app. The only way mobile can work today is with a cut down UI that is "less",

and turns SL into more of an off-the-rack strip mall Disney hangout than the open creative building environment we have today.
We've seen attempts like those fail again and again or just limp along for a year until the users and funding dry up.

Sure, an AI will hallucinate, but SL has always been a hallucination anyway. 
We'll always be chasing reality, but downloading 2K textures and a ton of geometry per frame won't scale much past where we are.

- A whole new everything -

No more fixed loop walk animations and dancing. Reflections that don't require tricks. 
Realistic weather with the look and feel of real air and rain and wind (snow that accumulates!!!, water with realistic waves!!!)

The best thing will be the element of surprise and discovery in a world that is no longer static and frozen.
And sure, it'll all be run up on servers, but it's always been doing that. Lag is an issue with remotely rendered games,
but c'mon, it's SL we're talking here. Lag is not a deal breaker for us. We're patient to a point.

SL Marketplace vendors may lose some markets, but better builders and better geometry provides a better feed to the real-time AI renderer.
There's a boundless new frontier for creators, and all of the old stuff will look new again.

It may be some years off, but Linden Lab could start today, taking an SL scene, the actual geometry not a picture, 
and have an AI imagine/render a still image from the data.

Some people think of Ready Player One as the gold standard of what the Metaverse will be. 
Real-time AI rendering is the only way to make that happen.

-- that is all

Edited July 9 by Lumpy Tapioca

[Katherine Heartsong]

I'm too old at this point to really be concerned about "20 years from now" in terms of what will my games look like, but I remain curious and fascinated that if we manage to not knock ourselves back to the Stone Age in some way, and 50% of the world's coastlines aren't starting to flood annually, what generative 3D animated AI would be able to produce given our current trajectory (positive and negative). I doubt it will be SL, but someone might do something like you describe. It'll be an interesting ride.

Now, can I get my flying cars I've been promised since the late 60s, please?

[Lumpy Tapioca]

Four years. One year for AI still snapshots.

Edited July 9 by Lumpy Tapioca

[Codex Alpha]

AI should be scrubbing toilets, doing taxes, and saving lives.

Not taking all the fun stuff away.

[Alwin Alcott]

14 minutes ago, Lumpy Tapioca said:

(I know that this is a few years too soon, ...

AI is inworld for many years already, started with .now simple, chatbots .. 

Edited July 9 by Alwin Alcott

[discussionbot]

17 minutes ago, Lumpy Tapioca said:

"Make us look like JFK and Marylyn. Put JFK in a white tux, Marylyn in a red sequined floor length dress."

Eh idk about that. Not sure if I trust the SL user base to do something like this while adhering to TOS 🤣

[Coffee Pancake]

AI rendering from simple stand in geometry with an applied visual style is a very interesting concept, however there is also a significant lack of control and hallucinations that can't ever be fully eradicated. The training alone will be nightmarish for any closed on rails project , let alone an anything random goes sandbox like SL.

AI is a bubble that will burst before we get there. Just like every tech bubble for the last 20 years.

In case no one has noticed, AI isn't making anyone any money. It's just costing the earth to power.

[Coffee Pancake]

5 minutes ago, Alwin Alcott said:

AI is inworld for many years already, started with .now simple, chatbots .. 

That's not AI, that's a collection of nested if statements.

Modern AI has nothing in common with chatbots of yore.

[William Gide]

 

28 minutes ago, Lumpy Tapioca said:

Use the current SL view, take what the current renderer is fed (geometry and textures and avatars),
and feed it as a seed/scene description to a real-time server-based AI renderer, that streams your view back to you,
making this largest collection of 3D assets ever created look like the reality we're chasing for over 20 years.

Are they going to pay all the creators for the assets they will be training the models on? I can think of few better ways to alienate a lot of your vendors than to claim their work as your own, and use it train a model that is supposed to put all of them out of business.

It is not at all clear that a lot of the generative AI that is currently hyped (the chat bots, the image generators) will be around (in their current state) in a few years. The costs of training are very high, and no one is yet making much money on this. Without raising consumer prices substantially, they'll all go out of business in a few years. Even Wall St. is starting to get suspicious about how these companies can be profitable. Ed Zitron has a summary of a recent Goldman Sachs report and it's a bit grim when you look past the hype. I would wait a few more years before I hitched my livelihood to generative AI.

Edited July 9 by William Gide

[Extrude Ragu]

OP is talking about technology like this:-

Essentially taking a 3d render, passing it into some AI filters to make the scene whatever style you want (in this case, photorealistic)

Currently this kind of technology is really not very practical, we don't have power efficient hardware/software to do this realtime at a reasonable cost. It might be more practical in the future. When AI gets power efficient, I could totally see myself making an 'anime filter' or similar for my sim.

Still, this kind of technology I don't think really 'replaces' PBR - You still need to feed into this filter information about the scene like what is metallic, where the lighting is etc. The AI would have trouble inferring that from a scene with global illumination for example.

[Lumpy Tapioca]

"however there is also a significant lack of control and hallucinations that can't ever be fully eradicated"

A perfect description of Second Life.

[Lumpy Tapioca]

"Essentially taking a 3d render, passing it into some AI filters"

No, and this is key.

Not passing a picture, but the actual complete geometry and textures of the scene, then rendered from your camera's point of view.

Less to mis-interpret, more cues about shadows and light from things hidden from view.

The AI renderer would know something is metal not because it looks metallic, but because it's defined as metal.

It would know exactly where the lights are by definition, not inference from how an image looks.

Way better.

I've not seen this done before.

Edited July 9 by Lumpy Tapioca

[Istelathis]

7 minutes ago, Extrude Ragu said:

When AI gets power efficient, I could totally see myself making an 'anime filter' or similar for my sim.

I would love this, I would probably do the same thing.  Hyper realistic is not really to my taste, I would rather be immersed in a cartoonish world..  

 

Now I just need some AR glasses, to render the real world into the cartoon world 🤣

[Dorientje Woller]

Defenition of AI: Fakery created through stealing others content.

[Istelathis]

Just now, Dorientje Woller said:

Defenition of AI: Fakery created through stealing others content.

It trains on existing content, if training on data is theft then you are correct, but that makes us all thieves.  Even down to the words we are using.

[Lumpy Tapioca]

"Fakery created through stealing others content."

Another great description of Second Life.

[sirhc DeSantis]

1 hour ago, Codex Alpha said:

AI should be scrubbing toilets, doing taxes, and saving lives.

Not taking all the fun stuff away.

Tis truly a new age - I agree with entity there

[Bree Giffen]

I don't know. Did you hear about that robot in South Korea that threw itself down a flight of stairs? It makes me think that we're going to have that moment in SL if we go full AI. 

[Lumpy Tapioca]

I'd buy that for a L$

[Jase Devin]

3 hours ago, Lumpy Tapioca said:

(I know that this is a few years too soon, and saying AI makes people roll their eyes...)

The future of SL isn't GLtf or raytracing or $3000 worth of Nvidia on your desk,
and most of the world that already exists will never be retrofitted into PBR or whatever follows.

This is the only path forward.

Use the current SL view, take what the current renderer is fed (geometry and textures and avatars),
and feed it as a seed/scene description to a real-time server-based AI renderer, that streams your view back to you,
making this largest collection of 3D assets ever created look like the reality we're chasing for over 20 years.

It would work magic even with our 20 year old builds, making them look new again.

Because you would be feeding it actual geometry, it would work way better than "A castle on a lake, with birds overhead", 
like we do now with AI image generators.

What's the workflow now if you wanted to go out dancing as JKF and Marylyn Monroe? Hunt for avatars, pay for them,
make them fit and work correctly, hope that JKF and Marylyn from different vendors look compatible..

Many steps that are difficult at best and nearly impossible for beginners.

This vs. "Make us look like JFK and Marylyn. Put JFK in a white tux, Marylyn in a red sequined floor length dress."

Done, and better.

SL has so much user interface surface area with its sea of buttons and sliders. That's hard for people,
especially in the new mobile app. The only way mobile can work today is with a cut down UI that is "less",

and turns SL into more of an off-the-rack strip mall Disney hangout than the open creative building environment we have today.
We've seen attempts like those fail again and again or just limp along for a year until the users and funding dry up.

Sure, an AI will hallucinate, but SL has always been a hallucination anyway. 
We'll always be chasing reality, but downloading 2K textures and a ton of geometry per frame won't scale much past where we are.

- A whole new everything -

No more fixed loop walk animations and dancing. Reflections that don't require tricks. 
Realistic weather with the look and feel of real air and rain and wind (snow that accumulates!!!, water with realistic waves!!!)

The best thing will be the element of surprise and discovery in a world that is no longer static and frozen.
And sure, it'll all be run up on servers, but it's always been doing that. Lag is an issue with remotely rendered games,
but c'mon, it's SL we're talking here. Lag is not a deal breaker for us. We're patient to a point.

SL Marketplace vendors may lose some markets, but better builders and better geometry provides a better feed to the real-time AI renderer.
There's a boundless new frontier for creators, and all of the old stuff will look new again.

It may be some years off, but Linden Lab could start today, taking an SL scene, the actual geometry not a picture, 
and have an AI imagine/render a still image from the data.

Some people think of Ready Player One as the gold standard of what the Metaverse will be. 
Real-time AI rendering is the only way to make that happen.

-- that is all

I like it, just the cost of the AI Rendering Servers. How cost effective would it be?

[discussionbot]

44 minutes ago, Bree Giffen said:

Did you hear about that robot in South Korea that threw itself down a flight of stairs?

Great story, but reported by The Daily Mail, so take that with a pinch of salt 😆

[Coffee Pancake]

2 hours ago, Extrude Ragu said:

OP is talking about technology like this:-

 

[Aiyumei]

 

8 hours ago, Coffee Pancake said:

That's not AI, that's a collection of nested if statements.

Modern AI has nothing in common with chatbots of yore.

Modern AI have even more nested if/else statements along with OCR and other services in it 
Let's be real, just like any new technology it's bound to make it's way everywhere eventually but it's only good as every other tool out there, it can be useful if you know how to use it. If you don't it's pretty much useless and only makes things worse.

[lovestofu]

I respectfully disagree with the premise of this post, almost entirely.  While AI has demonstrated powerful generative capabilities, it has not yet satisfied the imagination in terms of richness in ideas and form. A simpler equivalent expression is that people are not reading AI books yet. AI imagery can be captivating but only when there is a human driving the illustrated themes.  In essence, I believe the idea that generative models will become the dominant force in content creation is itself a fad, and that in the future will be regarded as a more elegant copy-and-paste operation.

[diamond Marchant]

.

Edited July 10 by diamond Marchant