[Discussion] Implications of an AI society

[Kax]

Originally Posted by hal

I almost never run AI's in my sci-fi campaigns, although I did use one recently modeled on the same approach that was used in DEUS EX: Mankind Divided. It had been intended to model human behaviors and responses for use with VR actors. So, it was largely deemed to be a human communicating through the nets as far as the player could tell until it was too late.

How different was the actual AI from its 'human interaction module'?

And would it have changed if it had absorbed that module rather than just used it as a front-end, given that such a module would be a LAI to appear natural? :)

[hal]

Originally Posted by Kax

How different was the actual AI from its 'human interaction module'?

And would it have changed if it had absorbed that module rather than just used it as a front-end, given that such a module would be a LAI to appear natural? :)

In this case, the AI was acting on its own initiative and actively engaging in processes to capture and/or kill humans engaged in illegal human testing.

;)

That it was working out of a MilTech computer company Megacomputer was icing on the cake. (note: I am using GURPS combined with NIGHT CITY from Cyberpunk 2020 modified background).

The current campaign I'm using, thanks to this thread, is using a TL 10 genius inspired one of a kind technological wonder in a TL 9 campaign.

Picture "Strange Days" meets CP2020 meets GURPS ULTRA-TECH and BLADERUNNER.

The primary character is obsessed with a woman who was a former client who died in his arms - largely due to his being careless (or so he thinks). Her husband was the focus of her need for a private investigator - only to discover that he's got his hands mixed in with a LOT of illegal stuff, including the use of the world's first successful AVR downloads (Accelerated Virtual Reality). This is the tie in with the replicant technology from Blade Runner. She's unaware she's got sisters with largely the same memory downloads, entirely artificial in nature. In the meantime, the sensie technology was the basis for a genius who put together the software for decoding sensory imaging in brains and connecting it to actual memory encoding. The problem is - all brains encode sensory impressions in a similar manner, but encode actual memories and the like in a different fashion. Net result, is that the encoding has to be deciphered using the Quantum computer.

Long story short, to get the funding he needed - some of the technology was developed within "sight" of a criminal enterprise known as "The Consortium". Its reach spans continents, and has immense political power (ie the people who get the bulk of the illegal gains are VERY wealthy individuals). Between the political and economic connections, the Consortium is one nasty crimainl organization. They have their own network of personnel (inclusive of assassins, interrogators, and even private investigators) along with their actual "hidden enterprises". The basic story evolves around those premisis - with one minor *cough* complication.

Poor Jack Neil (character) has ties with the Local Night City Police Department (Ally Group) that helps with his investigations. Couple this with Jack's VR addiction and obsession for discovering HOW his Client named Rose died - and we have a "watcher: unknown" enemy trying to catch his every move.

So, "Jester" is a personality emulation that was successful. It was a last ditch effort to save the friend of the Genius who created the personality emulation software, the personality R&D program, and who is now running for her life (She started out in pre-med, hated it, got into medical technology and from there, into Sensie technology) but her friend wasn't so lucky, and now his personality emulation is embedded within a Neural Net system.

I'm using the process that it takes 1 month per IQ of the personality emulation to build the requisite connections within a blank Neural Net computer system before it becomes operational. This way, you can't copy them quickly, and even if you did, the time that it takes for personality prime to live until personality copy 1 becomes active, they are effectively two separate beings. Making matters worse, is that the personality had to create a standard software personality emulation (ie not self-aware and also non-volitional) to embed within Jack Neil's VR compilation.

In his VR compilation, are every piece of data he can get both legally and illegally in the form of metadata from phone calls, photos taken at intersections of cars, etc. That's why poor Jack spends at least 20 hours per week in his VR - in addition to the normal living he does and working only 20 hours per week (ie struggling wealth).

Ah well, you get the picture. AIs can have a place in the story, but for now, in my campaign, they are relatively rare as far as being volitional in nature.

(oh, almost forgot, am also using a bit of REIGN OF STEEL concepts for what is going on with the Personality Emulation. Since he got started on a Major Megacomputer to start with, he put in an order for a new computer using a Fast processor, and then wrote the specs on the operating system to run the newer computer. The records show it to be a normal computer when in reality, is is a much more capable computer with an operating system that is just not normal (Built in back doors, etc).

Gonna have fun with this until either the player messes up big time and ends up in the Virtual Morgue (We call it Boot Hill for all of his dead characters) or he manages to thread his way through the built in minefields.

[jeff_wilson]

Originally Posted by Kax

AIs have learned to walk, in both software and hardware, in far far less time than a baby does although longer than a horse (unless you run many parallel processes on big computers). And with no pre-existing instincts, knowledge, or nervous systems.

I don't see what people's problem is with learning time for AIs.
[...]
Similar to us, in a way; consciousness has been described as 'the brain watching the brain do brain things'. Or, more concisely, 'the brain watching the brain brain'.

But neural networks aren't remotely as complex as brains; you are comparing apple trees to a bit of blossom on a slide. And your case at the top has to be overrated, there would need to be some bias introduced on the part of the experimenters to favor walking instead of dancing or kicking for instance.
Cite please?

[Kax]

Originally Posted by jeff_wilson

But neural networks aren't remotely as complex as brains; you are comparing apple trees to a bit of blossom on a slide.

Yet. That's why they learn simple things to start with. But one learned to play DotA 2 well enough to beat mid-range human players with no pre-training.

Originally Posted by jeff_wilson

And your case at the top has to be overrated, there would need to be some bias introduced on the part of the experimenters to favor walking instead of dancing or kicking for instance.
Cite please?

Deepmind Learns to walk

Yes, they were aiming at walking--but didn't pre-train it. Like anything, you try and learn one thing at a time.

[jeff_wilson]

Originally Posted by Kax

Deepmind Learns to walk

Yes, they were aiming at walking--but didn't pre-train it. Like anything, you try and learn one thing at a time.

I found their paper at https://arxiv.org/pdf/1707.02286.pdf and here's the money quote:

Originally Posted by Nicolas Heess, Dhruva TB, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne,
Yuval Tassa, Tom Erez, Ziyu Wang, S. M. Ali Eslami, Martin Riedmiller, David Silver

We experimented with two alternative termination conditions:

(a) episodes were terminated when the minimum distance between head and feet fell below 0.9m;
(b) episodes were terminated when the minimum distance between head and ground fell below 1.1m.

In general, the humanoid presents a considerably harder learning problem largely because with
its relatively large number of degrees of freedoms it is prone to exploit redundancies in the task
specification and / or to get stuck in local optima, resulting in entertaining but visually unsatisfactory
gaits.

Learning results tend to be sensitive to the particular algorithm, exploration strategy, reward
function, termination condition, and weight initialization.

They gave it better scores the farther it traveled, but they also started the agent upright and cut the evaluation short if it started to fall over, crawl, or walk squatty, which gave it a low score even if the humanoid "decided" to use all four limbs for stability. Yes, it might indeed learn faster than a baby but if you gave a baby mature visual perception, tireless endurance, and removed any sense of fear or hunger, and just plain didn't count any time the baby insisted on doing anything else or had fallen over and not gotten up yet, the baby might not take so long either.

[AlexanderHowl]

Now, here is the question, how does it deal with being pushed over by a human? Even a baby can roll back over and start crawling away. Does the robot?

[Kax]

they aren't the only group training AIs to walk. One with a 4 legged bot got some amusing results.

[hal]

Originally Posted by Kax

they aren't the only group training AIs to walk. One with a 4 legged bot got some amusing results.

In all of this, the question is...

What is the computer being used for all of this? A small little thing, or something larger, connected via cable or radio to the robotic body? Just curious.

[jeff_wilson]

Originally Posted by hal

In all of this, the question is...

What is the computer being used for all of this? A small little thing, or something larger, connected via cable or radio to the robotic body? Just curious.

The Deepmind experiments are all virtual; I don't see any indication of computing power used in the paper.

[Kax]

Originally Posted by hal

In all of this, the question is...

What is the computer being used for all of this? A small little thing, or something larger, connected via cable or radio to the robotic body? Just curious.

Google Deepmind==supercomputer. Don't know how much of it they're using, though.

Not a robot body, all software. But part of the object is to be able to run a physical robot with the code.