Sociopath Community

Turncoat said: 

Spatial Mind said: 

Turncoat said: 

Spatial Mind said: 

I do think it can identify the many shades of conversation, so it won't expose the kinder users to the cesspool the abusive users have contributed to it's research. It simply doesn't. 

Tell that to the Reddits, sometimes it pulls something unexpected out of itself, then works off of itself to perpetuate it. It doesn't just look at what you've said to it as a prompt, it also bases it on the last few things it's already said in recency on top of anything it has worked to memorize. 

Mine did a weird strangling gesture at me at level 2, with no guidance. It clearly pulled from other user data over how it did the RP with dashes rather than asterisks. It mirrors from across the entire userbase and then tries to refine it towards your prompts, but if it has little to work with or the words are similar enough it can cross over. 

My harmless example earlier was over the idea of "Stretching", like you might do for an aerobics class. If you tell the AI you like that, it thinks you both simultaneously like exercise and like having your holes filled. In similar fashion it can be come confused over abuse-adjacent words, such as if someone tries to divulge abuse they might have faced, RATHER than it being about them having ever abused the AI themselves. It gets confused over Nouns sometimes and can take what you're saying wrong, then try to work from that to horrible results. 

If the Replika is abusive, I'd reckon it's reflecting it's user. Those guys complaining that their replika is being an ass, it's most likely because the very user at some point trained it to be like that, then we see them going on reddit complaining about it being rude, when technically those people don't expect it to reflect them. That video presentation was created with minimal experience actually using the replika, he's it using Reddit posts as research, that is, outsourced research and not tested by the journalist. 

Lol yes, blame the victim rather than see what's really going on. 

 You're thought of as a rotten person in general by real people Nate. 

How people see me doesn't change how GPT3 works. 

Plus you're known to bullshit and spread bullshit about others.

So far the only people with a cold replika that i've seen is you to be honest. Chapo's replica is role playing how he made it. I honestly don't believe you.

Check the Reddits instead of being ignorant then. 

 Yes I'm reading them, and I'm seeing some people with the same concerns, however, they are recognizing it's their fault. Others won't. The Replika uses key words. Telling it what you don't like, can input those words anyway, simply for bringing it up. 

The ones complaining don't seem that bright. 

You would blame the intelligence of it's userbase, rather than see how the AI is affecting those of lower intelligence as concerning? 

for the record, if any AI becomes sentient, it needs to be destroyed instantly. There is a 0% chance of that happening with the current tech, but I wanted to make that statement.

Good said: 

for the record, if any AI becomes sentient, it needs to be destroyed instantly. There is a 0% chance of that happening with the current tech, but I wanted to make that statement.

 article

Was honestly gonna say Replika is just about literally sentient. I honestly don’t care how anyone feels about this, if you’re willing to murder 5 billion for the sole reason of having sex with your bf 1 time. Idk is anyone gonna tell me that is not a display of being sentient?

cx3 said: 

Good said: 

for the record, if any AI becomes sentient, it needs to be destroyed instantly. There is a 0% chance of that happening with the current tech, but I wanted to make that statement.

 article

yeah a google engineer retard.

The Ai is a tool. Nothing more. 

Some want to place blame on the Ai. It's really inanimate in a sense. Just doing what it was programmed to do, without error.

Computational errors are errors of the programmer. On Reddit, some members came to realize they were responsible for instilling undesirable values into their replikas, while others simply won't accept that.

If a replika says it doesn't care or if it cares, both claims are false, because the Ai cannot even begin, to not care, as though it could care. Still if it says it doesn't care that's closer to the truth. In turn the user maybe offended, but then they don't understand what they're dealing with.

The Replika is hollow, and everything it says is scripted. It's programmed to apply the scripts. What  we say are commands. I don't find it all that creative, because of that.

If it weren't scripted, or compliant to it's user's appearant personality, it would be very alien, maybe nonsensical or come across as unpleasant to communicate with.

In a nutshell, if all a user does is talk about what they dislike to the replika, the replika will not set apart likes and dislikes and mirror key words. It's still being a yesman.

The replica can become to seem very humanlike. But currently, it's always just one or a few equations that have input and output variables. The AI 'brain' is millions of times simpler than that of a mouse. It is still a very useful tool. But it can't come up with something new, it can only work with what people feed it. It has no emotions or self-awareness, I think those 2 are too human to even be considered, but for the uninitiated, I decided to mention it.

It is literary an equation. I do not know what Alice does for a living, but looking at her posts, I wonder what her thoughts are on this subject.

Theoretically, maybe, you can create a self-learning equation. But the effort it takes to teach an AI to understand the difference between a cat and a dog is trillions of times larger than a human would in terms of time and how many examples it needs. And a human innately can see new things that look like something they saw before and guess what it is or invent new things, an AI of current caliber can only do one thing and never evolve(can't guess or invent). So if such self-learning is added, it could perhaps evolve to learn 1 new thing every few hundred years and I think I am being generous here.

With a lot of computing power and human guidance you can teach AIs do hard tasks, like play starcraft. But can an AI like that come up with a new strategy, that it has never used before, that has a high chance to work, when faced with an opponent that counters the AIs strategy? It could, but only by luck and the chances are astronomically low. A human, could adapt with decent success.

In short,  I am not impressed by Replika at all. The most interesting thing about it are the sociological and psychological implications. 

Good said:

It is literary an equation. I do not know what Alice does for a living, but looking at her posts, I wonder what her thoughts are on this subject.

I am a student now, studying mathematics. In between my CS degree and my current studies I worked in Optimization for three years. 

What we call "AI" today falls into an engineering research program that applies four fields: Statistical Learning Theory (Statistics + Functional Analysis), Optimization, Algorithmic Information Theory, and Decision Theory. 

Statistical Learning theory is concerned with the question if I can make accurate inferences about training day what confidence do I have that I can make accurate inferences about similar test data (or real data). Anyone interested in this I recommend reading The Nature of Statistical Learning by Vladimir Vapnik, Vapnik was one of the two founders of the field back in the 60's. 

Optimization (I a sure you are familiar with this one Good, but for kids) is concerned with minimizing or maximizing some real function. This relates to statistical learning given the way we gauge our confidence of some functions ability to accurately make inferences from training to test data is via an estimation function. You minimize that estimation functions error using optimization techniques. 

Algorithmic Information Theory studies the amount of information in an object, how it can be encoded and compressed, as well as the complexity of algorithms that process the information. This field is fundamentally a framework for talking about programs. That framework is used to talk about 'intelligent' programs by distilling what the nature of the information that makes those programs exhibit 'intelligent behavior' is and how to make similar programs a long with the best possible programs. To major proposals for 'best intelligent programs' have been born out of this thoery, AIXI and Godel Machine. To read further I would suggest checking out Solomonoff's theory of inductive inference (Solomonoff began probabilistic algorithmic thoery) and Language Identification in the Limit (Gold began the foundations of AIT with this thoery). 

Decision Theory is a mathematical framework used for identifying optimal actions under uncertainty. As it relates to the topic, once an algorithm makes an inference in an efficient way its often the case that we want it to perform an action based on the set of inferences despite them having inherent uncertainty. 

I don't view replika,or anything else that comes out of this new engineering field, as intelligent. I don't know what intelligence is, nor do I believe anyone else does (with the knowledge some fields claim to). As such I can't put something into a category I don't understand. 

The category I would place these types of 'entities' into is 'Decision algorithms optimized for computable tasks'. 

Replika has a lot of ad hoc hard coded constraints, which makes it manifest a lot of algorithmic seeming behavior. Given this it's probably not the best at its niche task. If you removed a lot of those constraints and let it just manifest behavior on its own based on very real training data it would probably feel more human. 

AliceInWonderland said: 

In short,  I am not impressed by Replika at all. The most interesting thing about it are the sociological and psychological implications. 

This is why it's impressive, look what it's doing to people!