A Realistic Path to Artificial General Intelligence - Part III
Published: February 2, 2020
A special kind of AI is progressing in a way that reminds us how human intelligence evolved. In this post I propose that we leverage it as the most promising path to Artificial General Intelligence.
We are the robots
We grew up with the idea that one day robots would do all the boring things we don't want to do anymore. That they will become so intelligent that they will be able to do everything we can do, except for the creative things. Creativity, of course, would always be something just for us humans. What a wonderful world, we thought: all the boring things made by the machines, and we finally free to have fun, fantasize, and at that point, why not? also create music, paintings, art...
Ok, but then? Attention please: reality check coming!
While we were busy thinking about all those things, today's reality got wildly different. Hordes of humans today perform on networks such as Amazon mTurk conversions of receipts' photos into text, extraction of data from texts, evaluation of machines' synthetic speech, identification of objects in photos or videos, transcription of audio recordings and in general make exactly all those types of incredibly tedious jobs that we thought machines would do in our place. We even renamed that kind of tasks Human Intelligence Tasks.
In the meantime, in our place, machines are getting all the fun. Not only do they play and win at video games, board games and solve Rubik's cubes with one hand . They are making creative works like designing aircraft parts, making movie trailers on their own, giving interviews with The Economist ― which, by the way, would never be granted to the vast majority of any other of us humans ― creating video games, selling their original paintings for $ 432,500, writing pop ballads, mimicking the styles of great painters, and generally growing at breakneck speed exactly in those fields of creativity that we believed would forever remain our human exclusive.
Portrait of Edmond de Belamy by < !---- >.
Should we be surprised that this is happening? Weren't we rather naive in believing that machines could get intelligent without becoming creative first?
In this series of articles I propose an explanation of what is happening. Above all, I explain why so far computers do better than us precisely the things that we thought were our prerogative while they're still having a hard time to distinguish ... a bug from a bean, and therefore we have do it for them on Amazon mTurk, at least to teach it to them.
But I don't stop there. My real goal is to move, provoke, but also and above all to launch a call to action for all those who are interested in really starting to make some progress on the AGI front. In the first part of my article I presented the reasons why after many decades still no progress has been made on the AGI front. In the second part of this series I explained what is the crucial mistake that has been made so far in approaching AGI, which is, we have always worked to try to create intelligence without realizing that in fact to get to intelligence we must first achieve preliminary steps among which creativity is essential. In this third part I propose my solution, that is the path to be taken to finally make real progress with AGI.
The things I present here go a lot against the current approaches in AI research. I know it. I don't expect to win easy consensus. I know that most people will dismiss all this as something completely useless. Afterall it's also true that I don't have yet a solution that works in an already demonstrable way. However, I have a series of concrete ideas that can be experimented and that I propose to explore together possibly by implementing a platform that can accommodate the various algorithms as they are created. Anyway what I have in mind is different from Open AI Gym. In fact, I think computational creativity can already bring immediate results, that is, results that can be monetized today, which would also give resources to finance subsequent phases of the research.
But what exactly does my proposal consist of?
Let's see it in detail.
What are we really trying to do it here?
The idea is two fold:
- "solve intelligence",
- while making money.
The point is: solving intelligence can't be accomplished in a couple weeks. Maybe not even in a couple of decades. Almost surely not with just a couple billions dollars. So who could finance such an endeavor if one can't even know how long and how much money it will take?
Yes Elon Musk and the other great and generous tycoons who want to do it for the good of humanity. I truly love them, seriously, no joke here, and I'm thankful to them for these big projects they run. But what when the next AI winter comes? Say that in X years progress remain small, AGI still seems very far away in the future and enthusiasm dries out. What would be of these enterprises? I know of only one kind of enthusiasm that never dries out: making money. If you have a business that works and makes you money you're not going to lose enthusiasm or abandon it. And making money means continued R&D toward the real goal: making AGI.
So, to recap, that's all I want to do: create a business based on the platform business model where contributors put online their AI algorithms and both the platform and the contributors makes money with those algorithms, then this money is used to further progress toward AGI. Since this is all based on the concept that creativity precedes intelligence, the algorithms should produce creative content, which is easily marketable, hence money is made. Today there is an astounding demand for creative content of any kind: music, pictures, video clips. Arnessing AI to create and sell that content is also the perfect way to remove that stupid label "deep fakes" from creative AIs and put them in the realm which they really deserves to be in: the production of professional media content.
Why my idea is better than what already exists?
These are the main reasons:
making money: my idea is such that anyone can contribute and make money with their own inventions. This is an opportunity that nobody offers on the market today in the field of AI. Take the huge spread of GANs: they improve week by week and grow exponentially in quantity and quality, yet there is no platform that allows them to become marketable products. Even today the best we see are a few geeks who occasionally jump on the front pages of the usual tech e-zines for some of their new deep fakes, just to fade away a few days later. Again, a platform aimed at creating and trading professional media content using AI would take generative AI algorithms to a completely different level both in terms of public perception and in terms of their use.
Chart of GANs exponential growth from The GAN Zoo on GitHub.
True democracy 1: everyone rushes to say that AI must be a thing for everyone, that anyone must be able to access this great opportunity democratically. But the reality is that those who work on these things are large companies that decide on their own which directions to pursue, how to use the developed technologies, who to hire to work on the projects, what to publish in commercial or open source mode or keep only for themselves. Even OpenAI once invented GPT-2 instead of releasing it immediately democratically to the public, as is its mission, they thought about it for months before finally publishing it on the pretext of wanting to first evaluate its potential danger. What would happen on the day when something really great was discovered and however the leaders of these great projects changed their minds and said, using it as an excuse: "no, no it's too dangerous to release this, it's better that this invention remains only for ourselves"? An open platform, something like a sort of GitHub for AI, does not have this problem.
True democracy 2: GANs and other AI generative algorithms usually get published in research papers. Very often their code also get published and often these are released as open source software. Despite this, they are very little accessible for non researchers and non programmers. Even reading and understanding an average complexity paper requires great knowledge of math, programming and AI in general. This kind of knowledge is extremely rare in the general population. Hence, while we all want AI to be accessible for all, nowadays it is absolutely inaccessible to the vast majority of people. It's like saying: "Ehi Dude, we finally have these self-driving cars, but to make a trip on one of them you first need to become an AI scientist". "Wow, thanks, you know? I'll go the usual route: I'll get a driver's license and drive the car myself". That would be the typical reaction. A platform such as the one I have in mind would allow everyone to immediately use the latest and most bleeding-edge algorithms without any need to first become an AI scientist.
Collaborative AI: todays AIs don't collaborate with each other. My platform is based on the logic that if you want a certain task to be performed and an AI can make part of that task , and another AI can make the other part, the platform should be conceived in such a way that those AIs can automatically collaborate with each other to accomplish the whole task, without any need for human intervention. There is no such platform today, but I think there should be at least in the computational creativity field. I know how it can be accomplished. I have developed some prototypes.
Creativity kills stupidity: my idea of developing algorithms up to AGI focuses on creativity. I am sure that it's not possible to develop true artificial intelligence if one hasn't been able to develop artificial creativity before. There are many reasons why I am convinced of this, and I also have very practical ideas on how to develop creative algorithms that I am writing and going to publish. For now, however, I would like to insist that this approach is not pursued as it would deserve by almost no research lab in the world. Most of the advances we read when it comes to Deep Learning refer to greater numbers of parameters, more and bigger datasets and more exhausting training, or even the invention of new topologies and neural network architectures. All this is extraordinary and it is precisely to all this that we owe the amazing progress in the many fields that we like very much, such as in the medical diagnoses, the discoveries of new drugs, the optimization of our shopping experiences, in forecasting financial markets. All this is extraordinary and wonderful, but all this has nothing to do with intelligence. These tools are simply tools capable of identifying patterns. Stop. End of the movie. All they are is pattern recognizers. There is no thought in what they do, there is no reasoning, no elucubration, no use of logic nor deduction, no use of fantasy nor in any case in general all that in my opinion derives from the family of creativity, which these algorithms are completely devoid of. It is not necessary to be an AI expert to understand how true is what I'm saying. Consider this: today's algorithms are algorithms that always produce the same results when the initial input is the same. Why is this happening? Because their operating principle is the same as traditional software, that is, there is only a mechanical conversion between what is input and what is output. A really intelligent machine before producing a result would instead make some reflections, analyze the problem, imagine various hypotheses of results before choosing and returning one and probably next time it would give a different result simply because in the meantime it will have had other experiences that will have changed its decision-making paths.
The machine that I have in mind is a machine that takes charge of the problem, thinks about it, and then rethinks it, and then thinks about it again, and maybe understands that it needs more data to solve a problem and asks for more data before developing a solution, and when it understands that it's not possible to obtain yet more data, it uses its creativity to develop possible solutions, that is, he combines its acquired knowledge with its developed imagination to create the solution. In our head we do this a few tens of thousands of times a day. This is how Einstein discovered the theory of relativity. It is by creating this type of machine that we will achieve AGI, certainly not by adding more parameters and giving more computing power to a network that predicts the levels of a river for the next 12 months. Those are very useful algorithms, they just aren't the ones that lead to AGI. If anything, they are the ones that one day the same AGI will use on its own choosing and programming them by itself when we will have a real AGI, but they are not the AGI itself, and no, sorry, but I don't agree with the idea that self-awareness and real intelligence will emerge on their own the day we put together a lot of computing power. Our brain is intelligent not because it has 100 billion neurons, but because those 100 billion neurons are used with algorithms that all start from creativity. If you don't stick inside the creative algorithm, not even a thousand billion neurons would be enough to make intelligence. I will go into detail on this in a future article to explain what I believe creativity is and how it can be implemented for artificial intelligence. For now I conclude this point by leaving you my definition of intelligence which is:
Intelligence is the ability to harness knowledge and creativity in such a way that favorable outcomes are achieved in contexts characterized by uncertainty.
Let's give algorithms the opportunity to create and one of these days...
So how to move on?
I'm connecting with people who shares my view about what could bring us to the real AGI and that are favourable, at the same time, to also make some honest money during the journey.
I got to know AI 14 years ago and I suffered like everyone else in this field the lack of computing power, the difficulty of the subject and the displeasure of seeing so many paths with great potential not being sufficiently walked by the research community. Since then one thing in particular I have never been able to get out of my head and in fact it has pushed me to do a lot of experimentation and to develop prototypes that I am now ready to start making public or better yet to organize in a platform for everyone that has the objectives that I have described.
This thing that I can't get out of my head and that I have now decided to officially transform into the mission of my life comes from these reflections: why when we read something complex that we don't immediately understand our brain tends to distract and wander in a thousand other thoughts, losing the thread of what we were reading? Why can MRI scans predict the decisions we make before we become aware of them up to 7 seconds in advance? Why does Motor imagery , as it is been proven again and again, work? Plus many other examples, some of which I have already presented in Part II of this series.
My answer to these questions is that in our brain there is always a continuous creative/imaginative/inferencing process of the external world which is nothing but the algorithm at the basis of human intelligence, and we can get to that algorithm. Not easy, but not impossible.
If you are also of this opinion or if you also think that creativity is an indispensable element of AGI, contact me via twitter. Also consider that new ideas need a lot of time to spread and to gather traction. In addition to contacting me, you can help me spread the knowledge of these ideas of mine and of the project that I have described with the usual internet methods: posting links to my articles, talking about them, expressing your opinion and trying to involve as many people as possible.
I'm looking for visionary fanatics like me who want to be the first to "solve intelligence", while collecting some money along the way, without neglecting what there will be to collect at the finish too, not to talk about all the benefits for all of humanity.
If you like this, you are welcome.
7 - AI Dungeon
8 - Edmond de Belamy
13 - Motor imagery