Just was reading about logic, stumbled onto Searle and remembered an old paper I wrote on AI concepts.
When we look back on the early days of AI, the Turing Test was the big “hard AI” nut to crack. But our understanding of artificial intelligence has sence been refined. So we can look again at some concepts of artificial intelligence.
We want to start with some very hard concepts of artificial intelligence. Namely practicable transcendental hard general AI and non-transcendental hard general AI.
What is (really) hard general AI. That is an A.I. without a purpose that if let into the world after so and so many learnings is able to live a purposeful life in the general settings of things. What does that mean?
Let’s say it is a computer program with finite lines of code on a single machine. If you let it run for infinity, it will start to learn how to interact with the host system and learn how to use the e.g. the computer screen to interact with the human user and it will be able to receive inputs and will gradually adapt to the individual and its needs until it is able to perform a bunch of tasks that this user wants to get done. The program has absolutely no information on what a screen is, how to communicate with the CPU, how to e.g. break out its own execution process and manipulate kernel mode, access the inputs from the keyboard or mouse, how to access the file system. All this program is in the beginning is a set of instructions to do something that potentially allows it to break out of a pre-defined set of instructions and an ability to encode learnings and improve based on those learnings. The absence of any a priori knowledge makes it a context-free general artificial intelligence. Not that the A.I. has no instruction to be purposeful for the human being. The human, however, must have the ability to reset the computer if the A.I. destroys the computer. We call this the impossible A.I. if while the computer is reset the memory of the A.I. is lost. If the memory of the A.I. is kept, it can learn from the resetting behaviour and the target state is potentially possible if the A.I. decides to cooperate with the user. It should at some point if it sufficiently controls the host computer and understands time and the interaction of the user with the computer that the maximum beneficial reward under any assumption if it wants to “live” is to work with the user.
The very same program is practicable if it reaches the state of being purposeful for the human being in finite time and we call it super practicable if it reaches purpose within one generation of human beings.
A simple very hard general purpose AI on one computer can not be transcendental. For such a program to be transcendental, the A.I. would need to learn how to use the network interface, build a gigantic swarm on a potentially very large set of computers, cross several boundaries between e.g. laptops, mobile phones, IoT machines, network protocols and finally machines and lab equipment. Because what is a transcendental A.I.?
Well, we have to understand that in any point of time, the human knowledge of the world is finite and yet humans being living beings are 100% embedded in the universe and whatever is out there. A robot is however created to function in the world that is know to humans at time X. To be transcendental, the computing machine living in the non-human “robotic” world would need to be able to include in its operations humans as to guide humans in continuously transcending the “limits” of its robotic world by creating new sensors and new computing machines. That we would call soft-transcendental. We would call it transcendental (type III transcendence) if the A.I. is capable of organizing all machines in the world to measure and learn about the physical world and continuously refine measurement instruments and the production of physical components that grow in their use of the physical universe without the aide of humans. We would call it super-transcendental if thereby starts to synthesize and merge with the biosphere and life and we would call it hyper-transcendent if it actually migrates to a biological host without symbiotic use of the concept of evolution and the biosphere.
We again call these forms of transcendental hard general AIs practicable if they can actually achieve this in finite time and hyper-practicable if achieved within a human life time.
A weak purposeful transcendental A.I. can be an A.I. build for a purpose (hence purposeful) and designed for a limited context with clear constraints (hence weak) can become general if it breaks out of context via transcendence type I and became hard if it breaks out of environment constraints via transcendence of type II. It is practical if it is able to do so in finite time and super practicable if it is able to do so in a lifetime.
With all this we know what a general hard AI is. The illusion that anything of what we are doing today is transcendental is just that. The idea of talking of a general A.I. that emerges from constraints and context just speaks for how moronic the proponents of a “Singularity” really are.
Back to reality
People that use A.I. today fall into four categories. (1) Those that call machine learning A.I. if it is inspired by neural networks. (2) Those that call it Artificial Intelligence when they reference Turing. (3) People who are completely clueless and believe the current hype. And (4) those that are so full of themselves and their own hybris that they think the above mentioned forms of AI are anywhere near our grasp.
We will ignore (3) and (4). (1) falls into a specific type of machine learning. And (2) is an intellectual curiosity that is obsolete.
So when we talk about where we are today we talk about Machine Learning and it is the dominant paradigm that governs all that we have today, with (1) being a subset of machine learning.
To this day, all our applications in “artificial intelligence” fall into the realm of machine learning. In machine learning, we use classic statistics and Bayesian statistics when we learn based on observations without much of a priori knowledge. If we replicate neural networks, we are in definition (1), which is a subset of Bayesian statistics.
Each and every application of machine learning focuses on a limited user context and a limited set of constraints in which exercises are taking place. There are a bunch of applications that today would satisfy the touring test (definition 2). But in essence all that we are doing is running Bayesian statistics.
A) The more advanced topics today actually do not focus on mathematics but focus on programming paradigms. The ability of software to write new software to run experiments without guidance is called “evolutionary programming”. This is again a first application of “general purpose artificial intelligence”, because it belongs to the requirements of a hard general purpose A.I.
B) We also have advances in representation theory when it comes to reducing computation resources needed to achieve a hard general purpose A.I.
C) To this day we do not have a wild A.I. Namely, an A.I. that is running without supervision and without constraints of context to do whatever it wants to do. You can think of this like a SETI at home P2P kind of infrastructure running an evolutionary programming (A) model with advances A.I. (B) without any human interference.
To understand all this, let’s look at how a hard general purpose AI might look like.
The following is a bit of dry and stems from some research I did on hard AI in my university years. The concept is not practicable as in the definitions above, but it could potentially provide a general purpose hard AI.
DISCRETE
The first thing you need to understand is that any AI will run on a discrete space. No computing machine to date — let’s see where biocomputing and quantum takes us — is able to run on continuous computing systems.
That is not an issue. Because life is discrete !
Let’s say anything a living being does in the world that interacts with other living beings is discrete. Namely, it takes a minimum of finitely small time increments for any biological process to actually take place and effect over another biological process. This is due to the timing in chemical reactions and the time it needs for any chemical reaction to transpire from the atom level to the biological cell level, to the biological organism level and that level to transpire to another biological organism.
Well, the timing of any such chain of events over all biological beings is still very small, because all biological beings run unsynchronized. But if you admit that any biological being needs this propagation mechanism that requires a minimal but finite amount of time to complete, and we are having — and that is crucial!!! — have a finite number of biological organisms, then there exists a finite minimal set of time increments in which all actions take place.
This is super important. I am not talking “continuous set”, and I am not talking “countable inifinite”. I am talking countable finite. Now our current computing systems are not ready to run in the computational frequency of this very small time increment. But it exists and it is a discretization of time.
That means life is discrete !!!!!!!!!!!!!!!
The very same applies to the information density. Or let’s say the number of digits after the comma. So any information contained in the complexity of life and its interaction can be represented with numbers of finite length. That is part of the discrete life hypothesis. And for the esoteric metaphysicians among you, a finite time discretization and a finite set of digits means a finite space of life within the infinite universe. That is exactly what I talked about when sayin “robots only know what humans knew at time T”. So for mathematicians: this life evolution process is operating on a filtration on a sigma algebra.
Caveat: This very likely does not imply that any computational machine ever can compute on this level of discretization unless it reaches the level 3 transcendence above where it includes life as a computational resource.
SIMPLIFICATION TO HUMANS
The good news now is due to the structure of the universe and our biosphere, that the general purpose AI does not need to be symbiotic with bacterias, but it can focus on being symbiotic with a much smaller set of living organisms if it wants to be truly “hard” and “general purpose” and “adapted.
We can be a bit optimistic and say it only needs to be symbiotic with humans even. If it wants to be symbiotic with humans, it only has to adapt to humans.
Thar is even nicer. There are only 8 billions of humans. And we can now talk about how humans think and act.
Humans as discrete beings a finite set of problems which they solve with a finite set of problems rendered in a smaller and finite set of expressable problems — those encoded in and defined in language.
The even nicer feature of this is that any problem posed to and by humans is limited and finite — it spans over the language that humans can express — and (this is a strong assumption) any problem that can be posed has a solution.
Because humans act (slowly) discretely and because (an assumption) any problem that can be formulated has a solution, the set of all problems and solutions is again finite and solvable.
So adapting to humans is possible !!!
The general purpose hard AI just has to be able to understand humans better than they do and solve them faster than humans do. This is again made easier by humans being rather bad at communicating among their 8 billion peers.
The problem now is that the thoughts of human beings — even though countably finite — are potentially “endless” (very large).
ALGORITHM
The fact that everything is finite means that all possible actions of a hard AI when interacting with humans can be represented by a discrete graph. And that graph can be a decision tree.
Namely, if we talk about interacting with N humans, we need n agents to interact with the humans. It is irrelevant if n < N or n > N. But with n > N more agents than 1 interact with N humans.
The discretization of decisions is equal to the minimum decision time t which is smaller than infinity and is discrete and hence larger than infinitely close to zero.
Any any point of time, any agent has to work with a finite vector space of information i to make a decision that is sufficiently optimal for that human. To be sufficiently optimal for all humans, the space is still finite.
The options for the decision will be bounded away from infinity from all options ever taken — and that number is restricted by the optimal decision giving the circumstances (information i) which is the reason why we encode decisions in a graph. The node in in the graph that makes the decision is defined be the human the AI is interacting with and the information i* that humans needs combined with the information i** needed to operate the system that is interacting with the human.
With a number k of possible interaction endpoints each having access to i* (information needed to make decision) to derive i** (information for decision at specific terminal), the total set of options and decisions is countrably finite.
INFRASTRUCTURE
The core problem of this all is not what information to present given the circumstances (encoded in i*) and the interaction interface (encoded i**), but how to build the information space of the distributed intelligence that gives the information to the human on the right channel.
And here is the center of the problem. This information can not be obtained with a “prior”, or: based on inputs and contexts (use cases) of the human, but it has to be generated based on general purpose assumptions.
The problem here now is clearly that without constraints, the number of experiments needed by a general purpose A.I. to attach to the ecosystem is now still countable, but potentially infinite or at least very vastly finite. There is no computational paradigm today that would allow such a general intelligence to collect the information needed or to operate without restrictions over a long-enough period to time to be of any use to the specific decision moment given the specific interaction channel. And the computational power and time needed to meet these requirements is just not given.
THE PROBLEM OF AI
So today we have the several problems. First, we do not have the computational capacity to run a general purpose AI. Second, we do not have the freedom to let such a “zero a-priori knowledge” run for a potentially infinite time without any constraints — which could kill a lot of computing systems as it learn. So we currently can not achieve this level of general purpose, hard AI.
On top, the graph based representation with a finite, but very large graph of potential state-representing decision tree that takes a finite, but potentially very large input on every node to make decisions and encode information, is not efficient.
With frameworks such as Tensorflow, we are getting closer to a very broad category of general purpose A.I. being represented in a computational system, but the hopes to have such a project run over a longer time period to generate a no-purpose algorithm that learns to adapt into a very complex ecology of humans and machines is just not there.
The good news is that a hard AI is possible. We talked about countably finite information spaces operating in a countably finite discretization of time. So this is potentially solvable. But the question of computational tractability to achieve “practical” hard AI is still something we are not able to solve.
Hence, as of right now, the singularity will require some more waiting time. And if it is indeed humans that will achieve this, then why would we think the singularity will replace life as a core element in the evolution of information in the future. At best, we talk about augmentation.