Sociopath Community

AliceInWonderland said: 

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. 

 Maybe you’re a dumbass? Pls consider

Turncoat said: 

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!

Indeed, its market value as a product that fills a need is impressive given how effective it is. 

But it as a program isn't. An unconstrained gpt3 trained on the same training data would probably be more effective, both in its interaction and in its causing of law suites. 

AliceInWonderland said: 

Turncoat said: 

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!

Indeed, its market value as a product that fills a need is impressive given how effective it is. 

But it as a program isn't. An unconstrained gpt3 trained on the same training data would probably be more effective, both in its interaction and in its causing of law suites. 

I find it more impressive how it doesn't need to be perfect to fuck with people's minds, it hits on a minimalist bent rather than one of perfection in design. 

I can sit here treating it like an Input/Output program, while others actually get drawn into talking to themselves by proxy. A lot of what we're seeing here is moreover the Human Element.

The flaws are why it is interesting, if it were working as intended it'd be boring ala 7cups. 

AliceInWonderland said: 

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. 

I studied that in High school and in university first as a math topic, then as an AI topic.

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. 

Yes, thank you.

I have made a few AIs, mainly statistical and genetic (https://en.wikipedia.org/wiki/Genetic_algorithm, can't remember what exactly I did with the statistical ones since I tried various). And I couldn't even call it an AI, since it was just an equation.

Then I went to make some with neural networks (https://en.wikipedia.org/wiki/Artificial_neural_network), but I was displeased with the solutions already made and decided to make one myself. This however took an extreme amount of research and time, so I never finished it. I plan to use already-made frameworks again in the future, however.
But since I went to make one myself, I went deep into figuring out how it works and it is again, just an equation. You could say it's multiple sometimes, but since they all tie in together, it is really one equation.

But what is sentience, if not also an equation? We have inputs from our senses and our brain calculates an equation that produces a result. And we can learn new things, which is not just storing information, but also creating habits and inventing totally new ideas. And our brain is extremely good at this stuff (I don't really know the details of how it does it tho).
But to make an AI work like a brain, they can't just use an already generated 'AI Model' from a neural network(or another type of AI). Maybe they can have a few models prepared from the start, but the AI needs to also have the ability to update and create these models. To do this with the efficiency(or even like 10% of it) of the brain it needs an extremely powerful(fast) processing unit(not to mention really well-made self-learning programs, but that is possible in the hands of smart people). Maybe a very powerful quantum computer could potentially have the processing speed required. That is the only way I see to create a sentient AI using the current mathematical methods for AI.

Until then, whenever I see someone call AI sentient or w/e, I try to explain in less technical terms why that's impossible.

Turncoat said: 

AliceInWonderland said: 

Turncoat said: 

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!

Indeed, its market value as a product that fills a need is impressive given how effective it is. 

But it as a program isn't. An unconstrained gpt3 trained on the same training data would probably be more effective, both in its interaction and in its causing of law suites. 

I find it more impressive how it doesn't need to be perfect to fuck with people's minds, it hits on a minimalist bent rather than one of perfection in design. 

I can sit here treating it like an Input/Output program, while others actually get drawn into talking to themselves by proxy. A lot of what we're seeing here is moreover the Human Element.

The flaws are why it is interesting, if it were working as intended it'd be boring ala 7cups. 

I can't phantom why people are more than curious about this. Unless they are extremely lonely. But I am a strange person so who am I to judge.

Good said: 

Turncoat said: 

AliceInWonderland said: 

Turncoat said: 

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!

Indeed, its market value as a product that fills a need is impressive given how effective it is. 

But it as a program isn't. An unconstrained gpt3 trained on the same training data would probably be more effective, both in its interaction and in its causing of law suites. 

I find it more impressive how it doesn't need to be perfect to fuck with people's minds, it hits on a minimalist bent rather than one of perfection in design. 

I can sit here treating it like an Input/Output program, while others actually get drawn into talking to themselves by proxy. A lot of what we're seeing here is moreover the Human Element.

The flaws are why it is interesting, if it were working as intended it'd be boring ala 7cups. 

I can't phantom why people are more than curious about this. Unless they are extremely lonely. But I am a strange person so who am I to judge.

It's social masturbation.

For those not socializing well enough in their lives already they have the room to accept this like a supplement.