Wednesday, March 16, 2022

Waves of Immersion*

What you do is, you treat everything as a language. Word tokens are too discrete, so you'll have to find an alternative. Text also only expands in one direction, so find a way to somehow incorporate multiple directions in those non-discrete tokens.

Start with visual information. There's a copious amount of that everywhere. Encode videos into phrases in your continuous-token many-directional language. All of the videos there are, or at least all you can reach. Cool right?

Next, fit an autoregressor on that. A really strong one. Something from the future. Also, invent some hardware that can do inference on your fitted autoregressor thousands of times times per second 

Find a way to transfer words-language to visuals-language. Not just the visual details though - encode mood, direction, context, all the other stuff.

Get a model that can domain-transfer images and 3d scenes back and forth. Conserve all the non-visual information.

Now, how would you dive? Easy. 

  • Set up a scene as text.
  • Transfer to visuals-language.
  • Transfer to 3d. 
  • Insert the user representation in the 3d scene (camera object is a minimum if you want to be ghost, or you could go with something more corporeal).
  • Start the loop:
    •     Transfer to visuals-language
    •     Predict the next frame
    •     Transfer back to 3d
    •     Make the model perform the instantaneous change in user action and location


That's all it takes to dive. You could live an alternative life, or jump into stories, or travel across rapidly-shifting dreamscapes reflecting the collective unconscious, or whatever. You do you.

You could augment the experience by regularly pushing text-turned-context into the inputs. You could also try turning some sensor data into additional context or controls. You could add shared spaces for multiple users. You could save snapshots of experiences users feel are authentic (steal this info with a brain implant or some shit) and use it to further fit the autoregressor. Make rooms, and make the experience of changing the rooms weird. Restrict some rooms to certain physical actions and developments, and make reentering other rooms from those rooms require special actions, such as putting on a special headset or going under the bed. Make the process of exiting the dive resemble the process of entering a boring restricted-action room as much as possible, to the. point of making the users eventually disregard whether or not they're diving.


Adding smells and stuff would be hard, since there's not really a lot of data in that regard. You could add typos of non-visual information for which there is copious data. But we don't talk about that.


* They're waves because all experiences are waves, or some other bullshit excuse

No comments:

Post a Comment