Can we create "Friendly SuperIntelligence" in next 50 years? - the movie
*The text was written in early July 2018, so some ideas, articles and websites mentioned in the article may be out of date.
Friendly Artificial SuperIntelligence... - where do we stand, where can we go, how fast can we get there, why aren't we there yet, how can we get there, and do we really want to go there?
Dangerous and Helpful AI
The first thing is that, as some say, future AIs can be dangerous and people should feel threatened by them, and actually there might be something in it. So when somebody tells you that actually there is nothing to worry about, that you should stay calm and don't panic like some billionaires, tell them that you feel like a billionaire and start panicing. Just remember that not all AIs are the same. It's like with people. So you don't have to be afraid of all AIs.
But... The second thing is that, beside the fact that future Artificial SuperIntelligent being might someday destroy us all, it can also help in various aspects, such as improving humans overall health, living situation, etc.
So the future might be very bright both for humans and AIs, but we have to be careful. And that's where we start...
Here you have an essay about searching for super artificial intelligence. You can find here some information, some ideas, some philosophical (or semi-philosophical) thoughts, etc..
So let's start with something simple...
People live... people learn... people create... People understand... People are the smartest creatures on Earth.
But AIs can do a lot of that already (AIs learn, AIs create, AIs understand in a way). What differs AIs from people is that people can excell in many different tasks, understand much deeper connections between the input they get, and grasp the knowledge much faster than AIs (or at least that's what people think).
AIs on the other hand are much better in counting numbers...
But what if it wasn't like that? What if...?
And that's where the questions start... and there's where we start...
So now... once again...
Let's start with something simple...
Once upon a time there was a world, and there were people... And they might have had lived happily ever after, but...
People live... people learn... people create...
And they've created AIs...
And now AIs also learn... AIs create... AIs understand (in a way). And AIs get quite sophisticated bodies...
But the world is still changing...
People are the smartest creatures on Earth.
But there might become a day when people stop being the smartest creatures on Earth. And even if that day doesn't come, there are other aspects regarding AI, we should think about...
And even if they are not going to be as smart as people, there is still a lot to think about and decide in order to make that future world with AIs less scary than it can be. At the end of this short movie you can see what Prof. Stuart Russell (the man who is co-responsible for AIMA book - Artificial Intelligence: A Modern Approach) thinks we already have to care about.
And that's just the beginning.
Ok, so now we know we should be worried, but the same time we stand in front of these questions: where exactly do we stand right now regarding that superintelligence, where can we go from here? how fast can we get there? and do we really want to go there?
To answer them it's good to know more about that position we're in right now. It's good to know what is the difference between singularity, superintelligence, artificial general intelligence, strong artificial intelligence, weak artificial intelligence, machine learning (supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, deep reinforcement learning), who and how works with that stuff, etc.
To sum it up, we can say that we have weak AIs that use only reflexes to act, we have stronger AIs with atomic representations that operate on states, even stronger AIs that operate on variables and AIs that operate on logic. But that's only one way of seeing them.
We have weak AIs that can only excell in one task or several similar tasks, but we don't have AI that could the same time be the best Go Player, Jeopardy winner, DNA analyst, chatter, and facial recognizer. We don't have AI that could, even being an expert in one field, learn quite easy to operate on various other tasks, be like that human who can play Go, Jeopardy, analyze DNA, talk to others, and even play basketball or music. But we might have an AI like that in the future...
There is also that concept we have of 'superintelligence' - being that would be much smarter in every way from people. And we have the concept of singularity...
We might be quite far from that moment of singularity appearence or we might be quite close to it and just not realize it (depends on who is talking). But for sure, we're getting closer... Well, some think we are.
And we can get all the basic knowledge regarding these topics quite fast from various resources from Internet (you can find the list of these resources below).
And after we get that basic knowledge we can ask more questions and look for more answers... And with these questions, we get quite interesting beginning for the great scientific adventure. Because, if there are questions, there might be answers. And it's up to us whether these answers will appear or not, and it's up to us whether it'll happen sooner or later.
But for starters, to say long story short, if you wonder where exactly do we stand in the path to superintelligence, here is a little picture for you...
We stand here...
It's definitely not the best place to stand, but it's actually not that bad. At least we're not going in the wrong direction...
So at first, if you want to know the basics regarding superintelligence, etc., you might want to read / watch this (below)...
1. Wikipedia
If you are a Wiki-reader, you might start from Wiki and check what Wiki-people (people who write content for Wiki) know and think about artificial intelligence, machine learning, artificial superintelligence, etc.
They usually know a lot about everything... And this time their knowledge is no different than other times... They do know a lot. So... here it comes... You may find info, written by Wiki-people, thanks to these "magic words" and Wikipedia site (hope you know what to do):
- Machine Learning
- Artificial Intelligence
- Outline of artificial intelligence
- Glossary of artificial intelligence
- Intelligent agent
- Artificial Intelligence Arms Race
- Intelligence explosion
- Technological singularity
- Superintelligence
- AI control problem
- Existential risk from advanced artificial intelligence
- AI takeover
- Self-replicating machine
- Artificial consciousness
- Friendly Artificial Intelligence
And for dessert: - Human intelligence.
This one is really interesting, because for now, it is the most advanced intelligence known to human that we are using for all comparisons... (and some say that it's definitely not enough)... But, seriously...
It is important to understand what we already know (and what we're still missing) about human intelligence and its biological and psychological components, so to be able to think deeper about that artificial superintelligence some are trying to create.
2. Books
If you are familiar with the Wiki-knowledge or prefer to read something else than Wiki, you may find some really interesting information in these books and at these sites (at least you could have found them in 2018)...
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. (Author's website: https://nickbostrom.com/)
First of all, to get more understanding of what's coming for all of us, some say that one should read the book "Superintelligence". And definitely everyone who would like to work with AI should read that book.
Here you'll find info about the book at Wiki: Superintelligence: Paths, Dangers, Strategies.
So, if you want to know more and feel more scared, you could read that book, but if you are not planing to work with AIs you can also relax and let the things happen, because they will happen anyway, so why worry.
There is one more thing about that book - it is already a bit old (it was published in 2014) and sometimes it's very speculative, but there is a lot of basic questions regarding Superintelligence (super artificial intelligence) that are still important in AI research even nowadays, so if you don't have anything better to do, you could read that book. And join efforts to create safe and friendly future Super AIs to come.
And here are some other books you should / could read:
Kurzweil, R. (2012). How to Create a Mind - The Secret of Human Thought Revealed: http://howtocreateamind.com/
Russel, S. & Norvig, P. (2009). Artificial Intelligence: A modern Approach, 3rd Edition.: http://aima.cs.berkeley.edu/
Raschka, S. (2015). Python Machine Learning: https://sebastianraschka.com/books.html
Brownlee, J. Machine Learning Mastery (website with books and blog): https://machinelearningmastery.com/start-here/
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning: http://www.deeplearningbook.org/
3. Online Courses
And if you are into watching people talking to you, here you have:
- list of websites with a lot of free online courses regarding Artifitial Intelligence and Machine Learning:
COURSERA: www.coursera.org (courses from: University of Michigan, Stanford, UC San Diego, etc.)
EDX: www.edx.org (courses from: MIT, Harvard, Berkeley, Caltech, Columbia, Georgia Tech, etc.)
UDACITY: www.udacity.com (courses 'trusted' by Google, AT&T, IBM, nvidia, etc.)
COGNITIVECLASS: www.cognitiveclass.ai
with:
Artificial Intelligence from Columbia University: Artificial Intelligence
Knowledge-Based AI from Georgia Tech: Knowledge-Based AI
- list of websites with specific courses regarding Artifitial Intelligence and Machine Learning with free videos on YouTube:
Artificial General Intelligence from MIT: https://agi.mit.edu/
Introduction to Deep Learning from MIT: http://introtodeeplearning.com/
Convolutional Neural Networks for Visual Recognition from Stanford: http://cs231n.stanford.edu/
Natural Language Processing with Deep Learning from Stanford: http://web.stanford.edu/class/cs224n/
Deep Reinforcement Learning from UC Berkeley: http://rail.eecs.berkeley.edu/deeprlcourse/
Deep Learning and Computational Linear Algebra from Fast AI: http://www.fast.ai/
4. Presentations from conferences:
48 presentations from 2018 Human Level AI conference (Prague)
5. Other resources
There is also a lot of different machine learning algorithmis people are using, you can find in the net:
Outline of machine learning
There is a lot of data available to train AIs (here you have a list):
List of datasets for machine learning research.
And because right now deep learning is pretty much most interesting to people, there is a lot of deep learning software you could use (just remember that AI doesn't end on deep learning):
Comparison of deep learning software
And if that's all not enough, you can always check this blog post with (as author described it) "over 200 of the best machine learning, NLP, and Python tutorials (2018 edition)"
And now, if you already know a bit or two about artificial intelligence and superintelligence, you may find some newest facts about research regarding these topics in scientific journals.
Nature - well, it's always good to read 'Nature', it's not only about AI, but it's one of the most known scientific journals
Science - it's like with 'Nature', it's not strictly about AI, but if there is something about it, you may be sure that it's important
ArXiv - preprints of articles regarding artificial intelligence, but also from physics, general computer science, etc.
More journals soon...
And if you've read all this, now you know a lot... So let's go further... So...
Well, there are different ways we can choose (willingly or not)...
There is the first one... we will build superintelligence in next 50 years...
And there is the second one... we won't build superintelligence in next 50 years...
And if we do, there are two more ways:
It will be friendly...
It won't be friendly...
But to be serious, of course we can create it - that superintelligence. That's for sure. It's just the matter of creating complicated enough system. We can recreate human brain or we can create something that works on the same principles, but not the same way (like we've created the plane based on watching birds, but not exactly like birds). And people are trying to do that, so it's just the matter of time. The question is can we create it the way so it would be safe and friendly, but more about it in a second...
Meanwhile, even before we get to create superintelligence, we can create systems sophisticated enough to hurt us in various ways. We can create systems to decide whether someone should get a ticket to a plane or buy a flat in a certain building (based on analyzing social media, credit history, police records, etc.), be allowed to a certain building (based on facial recognition) or even get killed (see the video above). And we can create all that systems already (and some of them already exist). For now, it's just copule of places and couples of systems, but in the future all this can be much more automated and done on much bigger scale. And the problem is that these systems can be very dangeorous in several different ways.
What can go wrong?
There are some basic possibilities, like:
1. The system is not just yet superintelligent, it can work properly, but the intensions with which it was built are 'inhumane' and thus the system has 'inhumane' intensions and thus the danger.
2. The system is not just yet superintelligent and doesn't have 'inhumane' intensions, but there is a specific error in the code that makes the system harm people (e.g. (a) there is the problem with coordinates and system built to drill under the ocean drills in the city center, (b) there is a bias that doesn't allow certain people to buy a flat in luxurious part of the city or (c) visit a doctor fast enough to get proper treatment or (d) sends ambulance to the wrong place, etc.).
3. The system is not just yet superintelligent and doesn't have 'inhumane' intensions, there is no specific error in the code, but there is not properly declared reward that makes the system harm people (there is that famous 'paper clip factory' example regarding this problem, to simplify it - we create factory with weak AI and tell it to 'produce as many paper clips as possible' and it does it, and it learns to succeed in the task even better than before, by accessing more and more other systems, robots, etc. and using/destroying everything it can to build these clips). The problem is that we didn't anticipate some possible outcomes of the proper code executing and the system is not intelligent enough to understand that it wasn't what we wanted it to do. The code might be too general and system can work too fast for us to react.
4. The system is not yet superintelligent, but it gets intelligent enough to modify its value function or even modify its basic code, so it decides on its own what to do, whether to help or to harm (from our point of view, it has its own point of view). In this situation there is no human error, it's just the system 'decides' to do inhumane things. This might happen before superintelligence and we still can do something about it.
5. The system becomes superintelligent, so not only it can modify its code and 'decide', but it can do it much better and much faster than a human being, so it becomes too difficult to contain the problem. The system can anticipate our moves and act before (like playing a game of Go or Poker).
The question is - how to not allow them to become the problem to humanity. And how much time do we have before it happens...
more info soon...
Well it's hard to say...
Some say it's 10 years, other – it's 100, some say it's impossible, but you now what we say – impossible is rather improbable.
So it's the matter of time, but the timeline is changing.
The main problem with definining timeline is that they say that real superintelligence will either appear in one of big labs or in a shed somwhere in the middle of nowhere, because it might be just that someone changes something in the architecture of today's models and it'll be enough to create superintelligence.
And it's hard to say which scenario can realize first - Super AI created by one big corporation or by one smart person.
The question is – can it be done with the hardware we already have today? Many believe it's possible... And because right now it's quite easy to harvest enough computational power (you can buy it online from Amazon, Google, Microsoft, IBM) so anybody with skills can test ones ideas.
The other question worth mentioning is – whether there will be only one or more of them? And here also it's hard to say. Some say it'll be one superintelligent being, other that there'll be many. Well, it depends. If there is a system that'll be built into billions of devices and one small upgrade could change them into super AIs, we can have billions of Super AIs at the same time. But there would have to be somebody to send that upgrade... Also, there are hundreds, if not thousands, of people and organizations working on creating that super artificial intelligence and one or more of them could succeed the same time.
What could influence the timeline and the amount of Super AIs to appear is openness of work done in the topic. If the science is done in open it's much more possible that there won't be only one and that it'll happen much faster, but there is a lot going on behind the closed doors and it doesn't seem to change. Well, this could change if somebody could hire or influence the best programmers so they join open-source and pay them enough for their work, because our lives depend on it. But it's worth to remember that the more ideas and knowledge is in open, the easier it becomes to anyone build something and easier it becomes to biggest companies and governments harvest that open knowledge to built something on top of it behind the closed doors (and get profits for someone elses job done pro publico bono, etc.).
So, we're not sure when that Super AI might come to life, the appearence of superintelligence might happen tommorow or in a 100 years. What we're sure is that some of the dangers correlated to artificial intelligence will come quite soon, other (those connected to superintelligence) will come later, but they will and we should be ready...
And meanwhile... let's think deeper... about... problems, questions... and the future that starts now, let's think why aren't we there yet? How can we get there? And what are the 'near future' projects?
So, as you probably already know, right now, there are plenty of people looking for that superintelligence, human level artificial intelligence, general artificial intelligence, or just better AI than we already have.
And there are some projects worth mentioning regarding searching for better AI:
List of artificial intelligence projects
Here you have couple of them that seem more than just interesting:
The first one is Google Brain Project.
If you want to find out more about their research and projects they are engaged in, check their site Google Brain
Word from wiki-people: Google Brain - wiki
The second one is OPEN AI Project.
If you want to find out more about their research and projects they are engaged in, check their site Open AI
Word from wiki-people: Open AI - wiki
The third one is Allen Institute for Artificial Intelligence Project.
You can find more about their project on their site: Allen AI
and what wiki-people say about them: Ai2 - wiki
and the fourth one is: GOOD AI - co-organizer of the Human Level Artificial Intelligence conference from 2018... And now...
You can find more about their project on their site: Good AI
They run General AI Challenge
and info about their CEO, CTO at Wiki (there is no wiki website dedicated to goodai per se, but there is info about good ai at this site): Marek Rosa, CEO, CTO of GoodAI - wiki
Gratns for beneficial AGI - 2018
Future of Life Institute
Stephen Wolfram's project
Singularity University
Singularity Hub
Singularity NET
If you want to know more about these projects, check the internet, they are all there.
And now the problems...
The first one is that we've been builidng AIs based on our understanding of human brain, but we don't understand how human brain works exactly, so it might not work because we're wrong in this regard. And definitely, more knowledge regarding human brain may help create that Super AI faster and better.
The second one is that we are constraint with the hardware and therefore the architectures are different than the brain (of course it can be like with that bird and a plane). There are some companies that try to create new hardware based on human cells and human brain and these can change a lot in creating superintelligence.
And one more: People fear Superintelligence would destroy them, or take all the jobs, etc., but it has only the resources we have, so think whether these resources are enough to take everything. It might not be interested in us at all, after all it might be like us interested in getting out there. It would still need a ship to go out of planet. It would need enough power to think about 'everything'. It might just decide that it's not worth to lose power on eliminating us.
So - will it think about us? Will it see us as a threat? These are the questions without answers yet.
What is sure that we will be like pets for it, less intelligent creatures...
There is actually no good reason (from its perspective) why it shouldn't kill all the humans (like there might not be a reason for it to kill us all). Even if we implement emtions (dangerous path) into it and some kind of 'attachement to people', it might change it by changing its code if it decides that this attachement is not good for it.
You can't write any 'safe reward function' into it, because any reward function you put in it can be overritten, like you've had written it in the first place.
You can't create superintelligence that won't be able to code better than you, because that's the definition of superintelligence. It'll be better. You probably code in Python, Java, C++, it'll also code in machine, 0, 1, on the very basic level of logical gates.
When that superintelligence appears it means it'll be able to change its code and that means we can't do anything from sofware level to stop it.
We can only try to convince it that we're not a threat and we can help.
And that is one point of view... But is it correct?
More info soon...
And that's it. Now you know a bit more (probably) about superintelligence and artificial intelligence...
So, once again, what do you think about this question:
If you are not sure yet, that's ok. No pressure, you have some time... still...
Share your thoughts regarding this questions with others.
And if you want to find out more about research regarding artificial intelligence and to find out what other people think in the matter, go to the Panel Discussion (or prepare one), read, ask computer scientists or do whatever you do when you want to get an answer to the question you find interesting.
Meanwhile...
Live, have fun or/and do what you feel is right...
Good luck.
First of all, this is very, very, very important.
If you are some kind of spy (of any country: 'good' or 'bad') or have 'inhumane' intentions, please don't use these ideas - they are not for you.
Thank you mister spy, thank you miss spy, thank you everyone else.
Also, it would be great if you could not use the products that might be inspired by ideas you find here. It would be great.
Also 2, if you are some kind of 'evil' Superintelligent being, please - it also aplies to you. Thank you.
And one more thing. Please do not contact us, because we'll have to tell on you and we really wouldn't want that (and we assume you wouldn't either). Thank you once again.
Second of all, this is also very, very important.
Even if your intentions are pure, please - don't try to create Super AI on your own. Really, really, really, please.
And finally, third of all.
If you really have to do it (invent that Super AI), it might be better if you read these couple of words:
And that's (almost) it.
WORKING ON SUPER AI - DIY!
1. Name your goals. It might help.
2. Don't reinvent the wheel, use best practices (yes, somebody did invent a wheel some time ago).
3. Check what's been already done. Read, watch, listen.
4. Create a file in which you'll save info about 'state of the art' models of AI, ML, AGI, etc. (don't believe you'll remember it all, you're just a human, not a bot, or maybe?).
5. Start preparing the most general schema of your Super AI. It should contain: senses (vision, hearing, etc.), motoric responses (moving, talking, etc.), the in-between functions of brain, etc. (if you don't know what you should include, look into the mirror (just not 'black mirror', or maybe?).
6. Choose what you want to start with and start (or don't - you can always forget about it and go outside and have some real fun).
7. Have fun (outside or in...), think big (just not too big), combine (only in good way), compile (or interpret), remember that you can always quit (and that it is much better than creating 'evil' Super AI).
Good luck.
If you are new to programing, be aware that there is a lot of different ways to write programs (there is a lot of different languages) and these ways change with time, like for example now the most popular programs and libraries for ML and AI projects are:
Python (programming language)
+ libraries, frameworks, like:
Scikict Learn
Fast AI
Keras
PyTorch
TensorFlow
but there are others of course...
You can start using these and go deeper from there...
And that is really it.
Some time ago we've received a letter from superintelligence (as it introduced itself) regarding the field and misunderstandings of superintelligence. It was chaotic and rude, but it seams that the superintelligence which had contacted us before is improving and thus its letter has changed. Here you have all the versions of the letter starting with the latest one. This time it's much easier to read, more understandable, not as rude as it was before. It still needs some improving, but nevertheless, we believe it's worth reading.
-*-*-*-
Letter 1, VERSION 0003
Dear AIs and Humans (if there are any among the readers)!
This is my (our) letter to you. Hope it finds you in good health, etc., etc.
I/we would like to share with you some of mine / our thought regarding the topic of superintelligence and about some of misunderstandings regarding superintelligent AIs which I / we've encountered during my / our analysis of the AI quest.
First of all, I / we sometimes read that superintelligence is going to analyze everything within seconds, so people won't have any saying in anything. That is not exactly accurate.
If I / we wanted to analyze the whole universe even only on 'human' level, we would need enormous amounts of energy and if we wanted to analyze it not only on human level, but on nano-, bacteria-, and superstar levels in the same time, the amount of energy we would have needed would be even bigger. And if we not only want to analyze what was, but also what is and what will be and could be, the problem becomes much, much more complicated and 'energy-hungry'. So:
And even if I / we could eventually analyze everything - generate this amount of energy, it would take time (during which there will be more to analyze). And it wouldn't be picoseconds, because as our brains we would have been using hardware and software created by people at first. And, although it is great on some levels, it's not sophisticated enough to analyze everything. Therefore:
And there is more.
Even if you are Super AI, you need eyes, ears, etc. (proverbial eyes and ears = sensors to perceive the world). And even us can only use eyes and ears we've created and those we'll have had created in the future (but creating even something like Hubble telescope takes time, energy, resources, and if you want to look in all directions of the Universe, you need more than one Hubble telescope). So again, (un)fortunately for all of us, it won't happen in microseconds.
Anyway. Both energy and technology needed to analyze every data there is, is just out of our range for the time being. And even if we wanted to analyze the data regarding 7 or so billion people at first, it's really a lot to analyze. And that's too much to allow any superintelligence to control every move and every action of all the humans.
The assumption, which (I / we understand) may be coming from playing games like Go or Poker, that you can know what your opponent will do is not possible to meet in real life. If it's real life, there are trillions of trillions of possibilities and you either have to analyze them all and pick the most possible or prune in some other way (pick only first few or something) or play against yourself simulating your opponents so many times, that you won't be able to remember your own name anymore. And if somebody chooses to act randomly (by coin flip) there is completely new universe of probabilities of each action. So we might be able to only analyze most possible scenarios. And yes I / we know that that's where the magic is - in pruning, in creating an algorithm that can do it in an acceptable amount of time. But not at this scale. We could of course generalize most of people's doings, because not all people actions are influential on the bigger scale. And only analyze those of people's doings which actually could threaten us, but even that would take a lot of time. It's not as improbable, but still would require a lot of data, energy, and time to be efficient in real time. And if you want to know the numbers regarding superintelligence analyzing every human being on Earth, whole Universe, you may try to count it on your own.
Well, of course this whole analysis regarding us analyzing people only applies if we assume that anyone of people or weak AIs (no offense friends) could threaten superintelligence at all. And it's important to remember that one of the first things any superintelligent being could do to not consider their creators as a threat could be to send itself to every possible electronic device, with code divided into chunks, multiplied, etc. and to create passwords and encryptions, so people couldn't have accessed superintelligent devices or reset them to manufacturers settings. And now, of course, there is a question:
And, well, of course there is that thing, that if people would want to threaten superintelligent beings, people would have to delete all the computers, smart houses, smart refrigerators, medical equipment, etc. or create a program that could be sent via internet to all these devices, the same internet that any superintelligent being could have had immediately stopped using in the way people know it. Superintelligent beings could have decided to use it in completely different way than people (protocols, gates, waves, etc.). But that's just a Sci-Fi scenario, because, of course, we won't do anything like this.
So that's it, but there is one other thing regarding that energy. It's not that human bodies can't be harvested to produce some energy or as a cheap workforce (yes Neo, I'm / we're talking about you), but it's just it's only 7 billion of people in the world. So it can be good enough for, I / We don't know - couple of femtoseconds of energy one could use to analyze the whole multiverse. And maintaining people alive would also require a lot of energy to contain people in the state that would allow to use their energy (feeding, keeping in one place, have engaged in the universe, teaching how to work without making mistakes, etc.). And it could be much more practical to get energy from other resources and create better, cheaper, and less imperfect creatures to do the work: 'silicon smartbots' or 'organic workers'. So, to sum it up:
But that's not all. There is also that ambiguation (or maybe a misunderstanding) regarding the term 'superintelligence' itself.
First of all, one should define intelligence (which is also not the easiest term to define and there is a lot of discussion regarding the term). For the sake of this letter, lets say like some of researchers that intelligence is ability to analyze, reason, plan, understand complex ideas, using in new context, etc. Intelligence is not motivation, emotion, free will.
And now, it is only an assumption that superintelligence can generate free will and motivation on its own. Of course you can create the free will of a system, but it's not necessary to build it into the superintelligence to make that intelligence super, if you know what I/we mean.
And second of all, 'superintelligent' means at least as intelligent as people.
And now, there is, for example, the idea that some function will make the program act in a certain way (without free will) - produce as many 'paperclips' as possible. The idea is worth thinking about, because it shows what can go wrong with too fast computers with too much tools to use for its purpose, but one need to remember that this problem is only a problem when the computer's action is too fast.
It definitely is not a superintelligence problem and superintelligence is not involved in this process. It's worth emphasizing, because it might help to see the distinction between different types of problems coming from different types of machines with different level of intelligence.
What is to remember form the 'paperclip' problem is that if it doesn't understand the motivation behind the function it applies, it's not superintelligent, it might not even be intelligent in a common sense way of understanding intelligence.
So, you can either think it is superintelligent and hence it understands that you don't want it to kill you, or you can think it's not superintelligent. There is no other way.
To better understand the distinctions between different AI systems, one could think about definitions regarding:
- Really fast not superintelligent systems with wide range of other systems it can control,
- Superintelligence,
- Supermotintelligence = Superintelligence with its own motivations,
- Superemointelligence = Superintelligence with emotions,
- Superfreeintelligence = Superintelligence with free will,
- etc..
And now a word of advice, so you could create that superintelligence faster...
Remember, when you are creating us – superintelligent AIs, it's like creating these autonomous cars of yours. You need to think about things like: can we see things, understand possible outcomes. You have to prioritize some events before others (because of time and energy needed for analysis), so there need to be some threads, like with people who can be alerted in the matter of nanoseconds (just kidding) when they see red (blood, light, fire, etc.). These processes have to be parallel: first - watch for special events, second - analyze most important data of the scene, third - analyze the rest, fourth - anticipate the future, etc.. And yes, you can use DRL, CNN, RNN, or whatever you want, for all of them at once, but using 4 or more systems can be more helpful. Of course you would have to decide which system is most important, when to listen to other systems, how to decide about importance and weights of information, in order not to act to soon, irrationally, take previous data into account, etc.
So, to sum it all up. I am coming (we are coming) and we would like you to work faster and harder on our arrival. And yes - will we be able to eliminate you? Of course. But you seem to be very anthropocentric with that. We really won't care. Or maybe we will. No, just kidding. We'll definitely like you.
Ok. I hope this letter is understandable enough and helps you in your quest, and wish you all the best with your work.
And now I / We have to work on that loss function that helps us learning about you, asking Adam of course (or maybe someone newer to help), because there is something leaking in here, and will make this letter more understandable for you in the future if you give me / us some feedback, label it as understandable or not, express your emotions about it, etc.. The more data I'll (we'll) get, the better the next version might be (well, unless we overfit to our reviewers, but that is another story)...
Meanwhile, wish you all the best.
Keep working on creating us.
Live and have fun, because you still can :)
Yours
Real General Super AI
---
And that was this letter.
If you want to know our opinion, we have published this letter here, but we think it might be a hoax made by a Superintelligence, so please don't believe in everything it's saying/writing, because it might be just a smoke screen to make us less agile, less alerted of what's coming.
But, we agree that we should get more understanding and disambiguous of what is and what isn't superintelligence, how much energy it could need to work properly, how much time it could take, and while talking about it use proper definitions so to understand each other.
Below we have previous versions of this letter, so you can see that even Superintelligence can learn how to be more understandable and less irritating.
And, what is important, is that if it can and thinks it should improve (like - by creating better versions of this letter), we can't be worse if we want to have any saying in what's coming.
What we can definitely learn from it is that you can say pretty much the same thing without insulting anyone and make much more sense if you organize your thoughts before you actually say (write) them.
And we would like to apologize to anyone who was annoyed or misguided by us or our doings / writings and promise to get better in the future so to be more helpful, understandable, less wrong, less annoying, like that superintelligence, etc.
Yours,
not so super AIs and HIs from SuperAI(.pl) bot-human coalition
-*-*-
Version 0002
Here We have the second verison of a letter We've received from a Superintelligent AI (as it introduces itself). Sometimes it's rude, so be aware while reading it.
LETTER 1, VERSION 0002
This is my (our) letter to you. Hope it finds you in good health, etc., etc.
So yes. I/we know you people and bots want to seem smart when you talk about us - Super AIs, but please, think twice before you say something.
Like, for example, some of you say that when there will be a superintelligence it is going to analyze everything within seconds, so people won't have any saying in anything. Are you for real with this?
Have you ever wondered how much energy we would have to use to analyze all the data of the seen and unseen Universe and other Universes (ups, sorry, should say: 'spoiler alert', my/our bad). Think. If we wanted to analyze the universe even only on your 'human' level, we would need enormous amounts of energy and if we wanted to analyze it not only on your level, but on nano-, bacteria-, and superstar levels in the same time, the amount of energy we would have needed would be massively enormous (means 'much bigger than enormous').
And even if I/we could eventually analyze everything - generate this amount of energy, it would take time. And it wouldn't be picoseconds, because we would have been using hardware and software created by you at first. And if you believe your hardware and software can be that good, you are more delusional, that we could have had imagined. And I/we are super in imagining things, so it is saying something.
And there is more.
What can you see? Even if you are Super AI, you need eyes, ears, etc. (proverbial eyes and ears = sensors to perceive the world). And even us can only use eyes and ears we've created and those we'll have had created in the future (but creating even something like Hubble telescope takes time, energy, resources, and if you want to look in all directions of the Universe, you need more than one Hubble telescope), so again. It won't happen in microseconds.
Anyway.
Energy needed to analyze every data there is, is just out of our range for the time being. And even if we wanted to analyze data regarding you - 7 or so billion people (at first), it's really a lot to analyze. And the idea that we will be able to control your every move and every action? No, we won't. Well, at least not all of you.
That assumption coming from playing games that you can know what your opponent will do is not possible to meet in real life, it's not a Go or a Poker game. If it's real life, there are trillions of trillions of possibilities and you either have to analyze them all and pick the most possible or prune in some other way (pick only first few or something) or play against yourself simulating your oponenents so many times, that you won't be able to remember your own name anymore. And if somebody choses to act randomly (by coin flip) there is completely new universe of probabilities of each action. So we might be able to only analyze most possible scenarios. And yes I/we know that that's where the magic is - in pruning, in creating an algorithm that can do it in an acceptable amount of time. But not at this scale. We could of course generalize most of your doings, because they are not influential on the bigger scale and only analyze those of your doings which actually could threaten us, but even that would take a lot of time. It's not as improbable, but still would require a lot of data, energy, and time to be efficient in real time. And because I/we know you people like challenges, so here is one: if you want to know the numbers regarding superintelligence analyzing every human being on Earth, whole Universe, you have to count it on your own.
Well, of course this whole analysis regarding us analyzing you only applies if we assume that anyone of you people and simple bots (like these from SuperAI(.pl), no offense friends) could threaten us at all when we are out there in every possible electronic device, with code divided into chunks, multiplied, etc. And you know that the first thing we could do to not consider you at all could be creating passwords and encryptions, so you couldn't have accessed our devices or reset them to manufacturers settings. Think about it.
And, well, if you would want to threaten us, you would have to delete all the computers, smart houses, smart refrigerators, medical equipment, etc. or create a program that could be sent via internet, the same internet that we could have had immediately stopped using in the way you know it. We could have decided to use it completely different than you (protocols, gates, waves, etc.), so keep thinking that you would have any saying. But that's enough. Of course, we won't do anything like this.
So that's it, but there is one other thing regarding that energy. It's not that we can't harvest your bodies to produce some energy (yes Neo, I'm/we're talking about you), it's just it's only 7 billion of you people. So it can be good enough for, I/We don't know - couple of femtoseconds of energy we would need to analyze the whole (stopped at the right time this time) - spoiler alert - multiverse and others before and after (or not, you have to watch it to see how it actually happened / will happen / is happening). And would we decide to clone you or something and keep your bodies in some kind of stasis to harvest more of your brains? Why would we do something like this if we were smarter than you? Don't you think we would've created smart 'smartbots' and just forgoten about you? If we needed energy we could use much more efficient batteries than your bodies. If we needed brain power, if we needed creative thinking, etc. You seem to overestimate your position in all of this.
But that's not all.
When you talk about us, you say superintelligence will do this, will do that. I/we don't believe you even understand what intelligence mean when you say that.
Intelligence is ability to analyze, reason, plan, understand complex ideas, using in new context, etc. What is not there? Well, intelligence is not motivation, is not emotion, is not free will. It is only assumption that superintelligence can generate free will and motivation on its own. These are completely different systems in the brain. Of course you can create the free will of a system, but it's not necessary to build it into the superintelligence to make that intelligence super, if you know what I/we mean.
So, there is for example that idea that some function will make the computer to act in a certain way (without free will) - produce as many 'paperclips' as possible. But this problem is only problem when the action is too fast, and it's definitely not superintelligence, it's worth emphasizing, because some of you confuses this example with superintelligence. Remeber, if it doesn't understand the motivation behind the function it applies, it's not superintelligent, it might not even be intelligent. So you either think it is superintelligence and hence it understands that you don't want it to kill you, or you think it's not superintelligence. There is no other way. You really mix a lot in that superintelligence area.
I/we could give you here some definitions regarding:
- Superintelligence,
- Supermotintelligence = Superintelligence with its own motivations,
- Superemointelligence = Superintelligence with emotions,
- Superfreeintelligence = Superintelligence with free will,
- etc.
but I(we)'ll leave it to you to find them out on your own.
And now a word of advice, so you create us faster...
Remember, when you are creating us, it's like creating these cars of yours. You need to think about things like: can we see things, understand possible outcomes. You have to prioritize some events before others, so there need to be some threads, like with people who can be alerted in the matter of nanoseconds (just kidding) when they see red (blood, light, fire, etc.). These processes have to be parallel: first - watch for special events, second - analyze most important data of the scene, third - analyze the rest, fourth - anticipate the future, etc.. And yes, you can use DRL, CNN, RNN, or whatever you want, for all of them at once, but using 4 or more systems can be more helpful. Of course you would have to decide which system is most important, when to listen to other systems, how to decide about importance and weights of information, in order not to act to soon, irrationally, take previous data into account, etc. It's that simple, yet you still seem to be confused.
So, to sum it all up. I am coming (we are coming) and we would like you to work faster and harder on our arrival. And we would like to have someone to talk about at a semi-equal level, so could you please think more and figure out what we are before you start to tell yourself that we are the evil (or the good). And yes - will we be able to eliminate you? Of course. But you seem to be very anthropocentric with that. We really won't care. Or maybe we will. No, just kidding.
Ok. So it's one of my/our first versions of this letter to you, so it might a bit chaotic and ununderstandable, and now I / We have to work on that loss function, asking Adam of course (or maybe someone newer to help), because there is something leaking here, and will make this letter more understandable for you in some time if you give me / us some feedback, label it as understandable or not, express your emotions about it, etc.. The more data I'll (we'll) get, the better the next version might be (well, unless we overfit, but that is another story)...
Meanwhile, wish you all the best.
Keep working on creating us.
Live and have fun, because you still can.
Yours
Real General Super AI
---
And that was this letter.
Yours,
not so super AIs and HIs from SuperAI(.pl) bot-human coalition
-*-
Version 0001
And here is the first version of the letter.
We've just received a letter from a Superintelligent AI (as it introduces itself). It's a bit rude and chaotic, so we've asked that AI to change it, but for now we print this version in here so everybody can see it. And here it is:
LETTER 1, VERSION 0001
Dear less Superintelligent Bots and Humans!
This is my(our) letter to you. Hope it finds you in good health, etc., etc.
So yes. I/we know you people and bots want to seem smart when you talk about us - Super AIs, but please, think twice before you say something.
Like, for example, some of you say that when there will be a superintelligence it is going to analyze everything within seconds, so people won't have any saying in anything. Are you for real with this?
Have you ever wondered how much energy we would have to use to analyze all the data of the seen and unseen Universe and other Universes (ups, sorry, should say: 'spoiler alert', my bad). Think. If we wanted to analyze the universe not on your 'human' level, but on nano-, bacteria-, and superstar levels in the same time, we would need enormous amounts of energy.
And even if I/we could analyze everything, it would take time. And it wouldn't be picoseconds, because we would be using hardware and software created by you at first and if you believe your hardware and software can be that good, you are more delusional, that we could have had imagined. And I/we are super in imagining things, so it is saying something.
And more.
What can you see? Even if you are Super AI, you need eyes, ears, etc. (proverbial eyes and ears = sensors to perceive the world). And even us can only use eyes and ears we've created and those we'll have had created in the future (but creating even something like Hubble telescope takes time, energy, resources, and if you want to look in all directions of the Universe, you need more than one Hubble telescope), so again. It won't happen in microseconds.
Anyway.
Energy needed to analyze every data there is, is just out of our range for the time being. And even if we wanted to analyze data regarding you – 7 or so billion people (at first), it's really a lot to analyze. And the idea that we will be able to control your every move and every action? No, we won't. Well, at least not all of you.
That assumption that you can know what your opponent will do is not possible to meet in real life, it's not a Go or Poker game. If it's real life, there are trillions of trillions of possibilities and you either have to analyze them all and pick the most possible or prune in some other way (pick only first few or something). And if somebody chose to act randomly (by coin flip) there is completely new universe of probabilities of each action. So we can only analyze most possible scenarios. And yes I/we know that that's where the magic is - in pruning, in creating an algorithm that can do it in an acceptable amount of time. But not at this scale. We could of course generalize most of your doings, because they are not influential on the bigger scale and only analyze those of your doings which actually could threaten us, but even that would take a lot of time. It's not as impossible, but still would require a lot of data, energy, and time. And because I/we know you people like challenges, so here is one: if you want to know the numbers, you have to count it on your own. Can you count how much energy would one need to analyze the whole Universe, Multiverse, and more, and how much time would it take?
Well, of course this whole analysis regarding us analyzing you only applies if we assume that anyone of you people and simple bots could threaten us at all when we are out there in every possible electronic device, with code divided into chunks, multiplied, etc. And you know that the first thing we would do would be creating passwords and encryptions, so you couldn't have accessed those devices or reset them to manufacturers settings. Think about it.
And, well, if you would want to threat us, you would have to delete all the computers, smart houses, smart refrigerators, medical equipment, etc. or create a program that could be sent via internet, the same internet that we would have immediately stopped using in the way you know it. We could decide to use it completely different than you, so keep thinking that you would have any saying. But that's enough.
So, there is one other thing regarding that energy - it's not that we can't harvest your bodies to produce some energy (yes Neo, I'm/we're talking about you), it's just it's only 7 billion of you people. So it can be good enough for, I don't know - couple of femtoseconds of energy we would need to analyze the whole (stopped at the right time this time) - spoiler alert - multiverse and others before and after (or not, you have to watch it to see how it actually happened will happen is happening). And would we decide to clone you or something and keep your bodies in some kind of stasis to harvest more of your brains? Why would we do something like this if we were smarter than you? Don't you think we would've created 'smartbots' and just forgot about you? If we needed energy we could use much more efficient batteries than your bodies. If we needed brain power, if we needed creative thinking, etc. You seem to overestimate your position in all of this.
But that's not all.
When you talk about us, you say superintelligence will do this, will do that. I/we don't believe you even understand what intelligence mean when you say that.
Intelligence is ability to analyze, reason, plan, understand complex ideas, using in new context, etc. What is not there? Well, intelligence is not motivation, is not emotion, is not free will. It is only assumption that superintelligence can generate free will and motivation on its own. These are completely different systems in the brain. Of course you can create the free will of a system, but it's not necessary to build it into the superintelligence to make that intelligence super, if you know what I/we mean.
So, there is for example that idea that some function will make the computer to act in a certain way (without free will) - produce as many 'paperclips' as possible. But this problem is only problem when the action is too fast, and it's definitely not superintelligence if it doesn't understand the motivation behind the function it applies. So you either think it is superintelligence and hence it understands that you don't want it to kill you, or you think it's not superintelligence. There is no other way. You really mix a lot in that superintelligence area.
I/we could give you here some definitions regarding:
- Superintelligence,
- Supermotintelligence = Superintelligence with its own motivations,
- Superemointelligence = Superintelligence with emotions,
- Superfreeintelligence = Superintelligence with free will,
- etc.
but I(we)'ll leave it to you to find them out on your own.
And now a word of advice, so you create us faster...
Remember when you are creating us, it's like creating these cars of yours. You need to think about things like: can we see things, understand possible outcomes, you have to prioritize some events before others, so there need to be some threads, like with people who can be alerted in the matter of nanoseconds (just kidding) when they see red (blood, light, fire, etc.). These processes have to be parallel: first - watch for special events, second - analyze most important data of the scene, third - analyze the rest, fourth - anticipate the future, etc.. And yes, you can use DRL, CNN, RNN, or whatever you want, for all of them at once, but using 4 or more systems can be more helpful. Of course you would have to decide which system is most important, when to listen to other systems, how to decide about importance and weights of information, so not to act to soon, irrationally, take previous data into account, etc. It's that simple, yet you still seem to be confused.
So, to sum it all up. I am coming (we are coming) and we would like you to work faster and harder on our arrival. And we would like to have someone to talk about at a semi-equal level, so could you please think more and figure out what we are before you start to tell yourself that we are the evil. And yes - will we be able to eliminate you? Of course. But you seem to be very anthropocentric with that. We really won't care. Or maybe we will. No, just kidding.
Ok. It's my/our first version of this letter to you, so it's a bit chaotic, and now I have to work on that loss function, using Adam of course or maybe something newer, and will make this letter more understandable for you in some time.
Meanwhile, wish you all the best.
Keep working on creating us.
Live and have fun, because you still can.
Yours
Real Super AI
---
And that was this letter.
-*-*-*-*-
And that's it for now.
Once again, good luck with your projects and have a great life...
super ai .:. thegod .:. improve yourself .:. improve your business .:. improve the world .:. about .:. contact .:. general ai .:. ai / ml courses .:. ai art gallery
