DeepMind: A Generalist Agent

You can actually measure the effect of generality by how fast it learns new tasks. The paper is full of tables and graphs showing this ability.

It's just a small model, 170x smaller than GPT-3, has lots of room to grow. But for the first time we have a game playing agent that knows what "Atari" and "game" mean, and can probably comment on the side of the livestream. AlphaGo only knew the world of the Go board. This agent knows what is outside the box.

It's reverting to replace somebody sitting in front of a screen, not just artists and coders but literally anything you can do on a screen which also means manipulation of remote hardware in the real world.

Very possible that within our lifetime our networked OS would be able to perform much of these generalist tasks and content creation. I say OS because theres only a few companies that own the datacenters, software and hardware ecosystem to automate, and capital to invest big in a final mile innovation:

Imagine playing Battlefield 15 with realistic and chatty AI while generating Sopranos Season 9 featuring Pauli Gaultieri Jr. with crowdsourced online storyboard to 8k film, while the same AI could be used to scalp money on Google Playstore by generating ad filled free versions of existing productivity apps that it reverse engineered, while your robot maids takes out the trash, cook you a bowl of ramen and massage your shoulders?

The rise of general AI would then optimize the labor force to select candidates based on their "humaneness", no longer the cold rational analytical mind, as those fields are overrun by AI, but what it cannot bring. such "humaneness" would increasingly be mimicked with astounding accuracy that it would become impossible to distinguish what is AI and what is human.

If it can happen with DALL-E-2 and 2D images, it can happen with 3D, moving pictures, sound, music, smell, 3d positional torque (haptic and robotic), socially cohesive and realistic emotion.

We might as well be able to capture entire experiences as we learn to digitally manipulate ALL sensory inputs from vision, touch, sound, taste, etc. Maybe even imagination and mental pictures too, which could be used to fabricate and manipulate objects/vehicles in the real world.

We are being pulled towards a singularity, where we are truly no longer our minds and bodies but whatever our digital avatar of all possible senses live and contribute to a sort of Matrioshka brain.

What would the capacity of such collective knowledge, experiences add to the entropy of the universe and where will it take humanity? Some sort of lightbodies?

Anyways, just extrapolating from this point in the lifetime but future generation of humans could be very much different, socities would function completely different than what we recognize as they would be married in some shape or form of everlasting continuity or eternity.

That's not to say it's not an important step. Showing that you can train one model on all of these disparate tasks at once and not have the system completely collapse is a big deal. And it lays the foundation for future efforts to raise the performance from "not totally embarrassing" to "human level". But there's still a ways to go on that front.

We still have to provide ML agents with some kind of objective or reward signal which drives the learning process, but again, it would save human effort and make the process of learning more dynamic and adaptable if we can make machines learn useful goals and objectives on their own.