# Has it been theoretically proven or dis-proven that a machine can generally think and communicate on its own?

I am not looking for proof of AI (too broad, I know) concepts trying to solve a specific business problem, instead for a proof (or dis-proof) or a model (or a source of the same) where a machine is capable of

• General thinking and decision making.
• Collecting data as it see fit (use any channel at their disposal).
• Communicate as it see fit (with whom, how, where and when).
• Develop and evolve an agenda.

Note - I wrote an article more than 5 years ago - why a software cannot be intelligent. Since then I have not read a convincing argument against it.

Edit

Intelligence is the ability to take a decision in the unseen, unfamiliar, and unvisited circumstances. So, my definition of Artificial intelligence is - a machine doing essentially the same (doing, not replicating or simulating).

I am looking for a proof or disproof that machine is capable of doing things (written above), not the following

1. Instances where machine as simulated a certain aspect or property of intelligence.

2. Machine replicating a certain human behavior to achieve a goal.

• Read your article and by your definition of intelligence, I think many humans would also be classified as not being intelligent...while I would probably agree to that in regards to many humans, I think for practical reasons we'd have to count humans as intelligent. I also disagree with regards to WHY and WHAT. Computers are far superior to even humans at specifically identifying the WHY's and WHAT's for taking certain actions. Also, humans may program neural nets but the decisions made are purely determined by the computer and its experiences, IOW its training data. Exactly like humans. – Dunk Jan 16 '18 at 22:50
• Can you clarify what you mean by "AI". Is it the traditional use of "strong-ai" (as in "superintelligence" or "artificial general intelligence")? Intelligence is a spectrum, and algorithms are quite capable of all of the aspects of thinking and communicating, just at a current performance below human-level, except in special cases of "strong-narrow AI" (see AlphaGo). – DukeZhou Jan 17 '18 at 1:27
• @DukeZhou Thanks for your comment. I have just updated my question to clarify exactly what I am looking for. Please let me know if I need to clarify further. – gurvinder372 Jan 17 '18 at 6:24
• In the article you say "Software cannot take a decision on its own, it’s your hardwired logic programmed into the code that make your software do what it does." I think you could make software do this with running nlp on the keywords of a scripting language that takes in a problem as string and attempts to output a script that solves the problem and maybe a classification model to decide whether it wants to execute a script or not (implying it has a method to execute scripts on its own). – nickw Sep 23 '19 at 15:41

We have not been able to create a truly intelligent AI yet, according to your definition. So we have no real life proof-of-concept that shows that it actually works.

But based on the current research, there is no known property of the human brain that cannot be modeled in software/hardware. We do not understand the human brain enough yet - and most likely lack the computational power and required storage - to create an artificial copy. But as long as we do not identify a property of the human brain that cannot be translated into a piece of software/hardware, it is most likely possible. I honestly cannot even imagine such a property within the bounds of physics that would create such a barrier.

Of course this is no formal proof, but without a clearly identified show-stopper there is no reason to believe that it is not possible to create a thinking machine.

• Thanks for your answer. I have just updated my question to clarify exactly what I am looking for. Please let me know if I need to clarify further. – gurvinder372 Jan 17 '18 at 6:22
• @gurvinder372, If you don't narrow the sense of "the unseen, unfamiliar, and unvisited circumstances", then no free lunch theorem allows to say that there can't be any intelligence more powerful than a blind search. Notice that the theorem doesn't restrict an algorithm to be Turing computable. Humans aren't more intelligent than a blind search by your definition too. – red75prime Oct 17 '18 at 3:36

We don't even know how to prove that humans think (or not) yet. Or maybe it would be better to say that we don't really know what thinking is. In either case, there doesn't seem to be much reason to think (heh) that a "machine thinking" needs to be the same as a "human thinking". So no, I don't think it's been formally proven that a machine cannot in principle be said to "think".

Roger Penrose wrote a long and somewhat controversial book explaining why machines can't think - The Emperor's New Mind. But I'm not aware of anybody who considers his argument definitive. Actually, I don't know of anybody who takes it very seriously at all these days.

In the end, I don't think it matters. We humans choose (or seem to choose) to think that we think, but external behavior is generally what we care about, not subjective internal mental states. So if a computer acts like it can think, does it really matter if it's "really" thinking or not?

• Thanks for the link, looks interesting. Could you please add some quotes from the book, just to complete the answer? I don't think defining thinking - the process of considering or reasoning about something - is a concern, but I agree to execute the same in machine will require us to understand how it is done by our brain, which will lead to a meta problem since we will have to describe the process of thinking by thinking only. – gurvinder372 Jan 18 '18 at 6:36

General thinking seems as the toughest requirement to me. An AI looks at a new problem, tries to remember the most similar problem with a known solution and apply a similar solution to the new problem. It might try to guess how big mistake it has made and try to modify some parameters of the solution randomly (pseudo-randomly unless given a random or chaotic enough input such as a single wild animal behaviour or the weather).

A neural network given a hardware comparable in performance and storage capacity to the human brain and an insane quantity of input (e.g. the whole Internet including all attachable input devices) could probably be trained to mimic human general thinking believably enough. Some thinkers suggested that a human brain is basically just a very advanced machine.

My summer jobs made agree - as a postman, I performed better when I let my fingers find the mailbox not searching for it, and so did I as cashier, when my fingers would type in the correct code for the bakery products, fruits or vegetables (no barcode) without me even noticing what the goods were sometimes.

There might be some dis-proof based on an assumption that an AI is coded by humans using formal logic known to have some theoretical limits, which might be what you said in "Does software evolve?" in your article (nice job, by the way).

It almost becomes a religious question once you allow machines / algorithms to "breed" and "compete" as wildlife animals - would a long but finite time in an environment rich in resources and analog inputs be enough for an AI to evolve providing a finite set of environment parameters (gravitational constant etc.)? Reminds me of Just six numbers by Martin Rees.

• For your first paragraph: True randomness at a machine level is a solved problem. It is not solvable in software, but is in hardware, and devices that generate completely unpredictable numbers can be bought cheaply. There is no known limitation on these devices that would require an AI to use wild animal behaviour or weather patterns in preference. In practice, it is a similar concept - monitoring a "true random" environment to extract entropy. Just it turns out that environment can be very small and integrate very cleanly with a machine. – Neil Slater Jan 16 '18 at 15:24
• @Neil Slater how do determine randomness entropy is a measure of the chaos or order of a section of our environment if you determine the section to be chaotic it only is from the limits of that observation and maybe be part of larger section that maybe observed to be ordered and not random. The production of random numbers from that hardware can only be judged from the amount of time you observe the production and is only a section similarly a larger selection may suggest a negation of the content of it. – Bobs Jan 17 '18 at 0:48
• Thanks for your answer. I have just updated my question to clarify exactly what I am looking for. Please let me know if I need to clarify further. – gurvinder372 Jan 17 '18 at 6:22
• – DukeZhou Jan 17 '18 at 17:34
• @Neil Slater I don't think the problem of true randomness has been solved in hardware – DuttaA Mar 31 '18 at 16:07