Is A.I. more than the sum of its parts?

[PeroK]

The reason games like chess don't impress me much is that they are structured, boundedd and limited. It's basically a complicated math problem. Calculating strategy based on the rules is completely in-line with what I'd expect a computer to be able to do.

Even so, it did work out everything that humans have ever worked out about chess and more. To say that exhibits no intelligence whatsoever is stretching a point. Effectively that means that to play chess at a superhuman level requires no intelligence.

If you go back to the 1980s, plenty of people thought that a computer could never beat the world champion. Let alone by teaching itself how to play. That represents a massive moving of the goalposts. Now that it can do it, it doesn't really count. The same goes for LLMs. I'm sure many people would have cited their capabilities as representating intelligence. Until, of course, they can do it and then it doesn't count.

If you look at this thread. Dave said it needs to work something out for itself. I gave AlphaZero teaching itself to play chess. And you say that doesn't count.

In short, there is nothing a computer could do that would count as intelligence. Even if you have specified the criteria, once it meets those criteria, it doesn't count anymore.

 

[FactChecker]

Without a body and a set of sensors, I fail to see how one can classify any set of atoms as "intelligent".

By that criteria, a blind paraplegic would have zero intelligence.

If a computer system contains more in its memory than any human can have, and can manipulate the information in combinations better than any human can, wouldn't that imply intelligence by some reasonable standards?
I think there is a lack of appreciation of some of the AI systems and methods that have been developed. One AI form are "expert systems". For instance, a computer can be given a vast storehouse of medical indicators of ailments with associated probabilities. It can then manipulate combinations of facts to determine a most likely diagnostic. That is certainly some form of "intelligence" far beyond a four function calculator.

 

[Hill]

Doesn't it simply mean that there are different kinds of "intelligence"? Human intelligence, computer intelligence, canine intelligence, bee intelligence, octopi intelligence, etc.?

 

[jack action]

What does a machine have to do to register on your intelligence scale?

Have an imagination. Getting out of what seems logical.

For example, AlphaZero was given the rules of chess, but not any advice on how to play.

But you had to set the rules. Unbreakable boundaries that the algorithm cannot break.

Intelligent life forms discover the rules as they go along. Things that seemed a limit once are not anymore because they explore their environment, which they know nothing about; they discover it.

They do things because they can, not because they are asked to or because they need to do it.

Intelligence can't be a binary thing, because dogs, cats and crows clearly have some intelligence.

Good example. Here is a dog playing with a ball in a stairway:

​

Why does it do that? What sets in motion those atoms in this particular motion?

This one is playing with a man-made machine, and a kid even participates at some point:

​

The question is not only "Why does the dog play with the ball and the machine?" but:

• "Why does the kid play with the dog playing with the machine?"
• "Why did the adult build the machine for the dog to play with?"
• "Why am I watching a video of a dog and a kid playing a man-made machine?"

By that criteria, a blind paraplegic would have zero intelligence.

If a person has none of the five senses at birth, they will most likely be brain-dead. A brain needs interactions to develop.

If a computer system contains more in its memory than any human can have, and can manipulate the information in combinations better than any human can, wouldn't that imply intelligence by some reasonable standards?

Not for me. For example, sorting coins requires the intelligence of differentiating how coins differ. By that definition, this coin sorter is a little bit intelligent:

​

And this one is more intelligent because it is faster:

​

The reality is that the intelligence lies ahead, within the machine builder. It doesn't necessarily make the machine closer to "artificial" intelligence. Making machines bigger and faster is not a prediction of creating synergy either (the subject of this thread).

 

[FactChecker]

If a person has none of the five senses at birth, they will most likely be brain-dead. A brain needs interactions to develop.

I didn't say "at birth"

Not for me. For example, sorting coins requires the intelligence of differentiating how coins differ. By that definition, this coin sorter is a little bit intelligent:

Comparing that to some of today's AI systems is like comparing a first-grade student to Einstein.
I don't think you appreciate what some of the AI systems can do.

And this one is more intelligent because it is faster:

You keep mentioning speed. I'm not sure I agree.

 

[javisot]

Intelligence: is the ability to acquire, understand, and apply knowledge to reason, solve problems, and adapt to new situations. In general, it is the ability to solve problems.

I still think that a machine capable of solving all existing problems is not intelligent, no more intelligent than the person who designed that machine. The machine cannot solve the problem of creating itself.

But I'll pose it as a question in case anyone has a different opinion: Does anyone think that a machine capable of solving all existing problems is intelligent? Why?

 

[FactChecker]

Intelligence: is the ability to acquire, understand, and apply knowledge to reason, solve problems, and adapt to new situations. In general, it is the ability to solve problems.

Ok. That sounds reasonable.

I still think that a machine capable of solving all existing problems is not intelligent, no more intelligent than the person who designed that machine. The machine cannot solve the problem of creating itself.

1) Why demand "all existing problems"? Would you demand that of a human? Can't it be intelligent about specific subjects?
2) What would you say about an AI program that:
a) Contains more facts on a subject than any human has. (like expert systems that combine the knowledge of many world experts)
b) Is able to acquire more experience on a subject than any human can possibly have. (like computer chess programs that play against each other millions or times and record the results of different decisions)
c) Is able to piece more relevant facts together and reach conclusions than any human can (like logic engines that can piece together thousands of logical statements to reach conclusions)

All of this is happening now.

 

[PeroK]

The reason games like chess don't impress me much is that they are structured, boundedd and limited. It's basically a complicated math problem. Calculating strategy based on the rules is completely in-line with what I'd expect a computer to be able to do.

Out of interest, to what level do you play chess?

 

[javisot]

1) Why demand "all existing problems"? Would you demand that of a human? Can't it be intelligent about specific subjects?

Regarding point 1, the common definition of intelligence is the one I mentioned, which can generally be understood as "the ability to solve problems". When we ask about an intelligence comparable to or superior to human intelligence, we are asking about an intelligence capable of solving at least the same problems as a human.

We don't know the exact list of problems a human can solve, but that doesn't matter; let's suppose there is a machine capable of solving all existing problems. That machine would be "very intelligent" according to this definition of intelligence. (At this point, I add the reflection: "the intelligent one is the person who built that machine")

2) What would you say about an AI program that:
a) Contains more facts on a subject than any human has. (like expert systems that combine the knowledge of many world experts)
b) Is able to acquire more experience on a subject than any human can possibly have. (like computer chess programs that play against each other millions or times and record the results of different decisions)
c) Is able to piece more relevant facts together and reach conclusions than any human can (like logic engines that can piece together thousands of logical statements to reach conclusions)

All of this is happening now.

I would reply that the humans who created that AI program were very intelligent, but not the AI program itself.

 

[PeroK]

PS just to emphasise that AlphaZero was very different from conventional chess engines. In particular, it didn't calculate every move. Instead, it used pattern recognition to select "candidate" moves in each position - rather like a human would. And, of course, it decided for itself that this is how the game should be played. It wasn't told how to play the game in any shape or form. Only what was allowed.

Non chess players usually think that chess is just about calculation (thinking n moves ahead). But, thinking ahead is no use unless you can assess the end positions. Unlike what you see in the movies, not every position has a checkmate in four moves waiting to be found. For a conventional chess engine, position assessment is specified by the human designers, who specify the algorithm. This allows it to assess each position based on human knowledge of the game. And, generally, the engine looks at every move that is allowed by the rules in every position. I.e. human intelligence + computer brute force.

By contrast, AlphaZero had to figure out for itself how to assess each position and what candidate moves it should consider (i.e. not just everything that the rules allow).

Although chess is indeed a limited sphere in which to test a neural network, the way AlphaZero played chess marked a significant advance in computer thinking. In this limited sphere it was, as many leading players admitted, like seeing the game played by a superior intelligence. Indeed, if the algorithms had been specified by humans, it would not have been so effective - as its neural network ultimately saw deeper into the game than any human had.

And that is a step towards independent intelligence. Even if in a limited sphere.

 

[javisot]

I remember the anecdote of the second game between Kasparov and Deep Blue. At a certain point in the game, Kasparov got angry because he thought IBM was cheating. Kasparov thought that playing against Deep Blue meant playing against a machine limited to calculating the most probable next move, but Deep Blue's move didn't seem to follow that premise; it seemed to be playing a deeper game, as if it were being controlled by humans, according to Kasparov at the time.

Kasparov simply wasn't perfect, and his criticism wasn't justified.

But I agree with russ-watters; chess is a very limited and restricted game (unlike, for example, Go). It's true that chess is often considered in the Turing test, but I personally believe it's a game limited enough that it doesn't allow for arbitrarily precise discrimination between humans and machines.




(My highest online elo is 2132 in 1 minute games and 2256 in classic games)

 

[jack action]

I didn't say "at birth"

Well, of course, a person with a fully developed brain becoming blind and paraplegic won't drop to "zero intelligence" overnight; especially knowing it still has other senses.

Comparing that to some of today's AI systems is like comparing a first-grade student to Einstein.

That is my point: Einstein is very intelligent, and a first-grade student is a little less. Therefore, if you consider a computer intelligent for accomplishing a human task, you should also consider a mechanical machine being intelligent to some level to accomplish a similar task, in a simpler form.

You keep mentioning speed. I'm not sure I agree.

I might be wrong, but I feel you are impressed at a machine that can solve problems almost instantly, while a human would take years to solve (analyzing millions of patterns, for example, or simply reading books to find information vs reading a database for a program). In such a scenario, the machine isn't smarter than the human, just faster.

2) What would you say about an AI program that:
a) Contains more facts on a subject than any human has. (like expert systems that combine the knowledge of many world experts)

Does having more knowledge make one more intelligent? Again, if we assume man-made objects are intelligent, that would mean a dictionary (a book) is more intelligent than any human.

b) Is able to acquire more experience on a subject than any human can possibly have.

Then you should consider a web crawler downloading all visited websites having some form of intelligence, because it accumulates knowledge in its database. The AI program is not intelligent because it finds what you expect it to find, given the mathematical tools you gave it. Otherwise, you would consider a simple algorithm that can locate a black dot on an image with a white background, also having some form of intelligence. Yet, it is just a tool that basically locates a "0" in a sea of "1"s.

c) Is able to piece more relevant facts together and reach conclusions than any human can

Again, this is just a tool that does the work commanded by the human; one that the human could do if given enough time (which might take many generations of humans for a complex problem).

I still fail to see how one can say that AI is "more intelligent" than humans, or on its way to it, unless one considers doing the work faster as a major contribution to what characterizes intelligence.

By contrast, AlphaZero had to figure out for itself how to assess each position and what candidate moves it should consider (i.e. not just everything that the rules allow).

So you think a computer with an algorithm that uses the most commonly used passwords to crack a human-made password is a "smarter" machine than one using a brute force attack? It is only more efficient, which only proves that the programmer is smarter, not the machine.

The fact that the machine gathers the data as it plays the game is still just another tool used to eliminate the process of having humans filling a database, which would give the same results. It just saves time, with the disadvantage of no one knows exactly what is in the database. That doesn't make the machine more intelligent.

Indeed, if the algorithms had been specified by humans, it would not have been so effective - as its neural network ultimately saw deeper into the game than any human had.

And that is a step towards independent intelligence. Even if in a limited sphere.

No, it is not. The fact that a single human brain cannot analyze the information in a given period of time, especially knowing there is also an energy limit that must be respected, doesn't mean the machine doing the job is "intelligent", even less that it has "independent intelligence". It just means that you can concentrate more power within a single machine to analyze a problem. It is not smarter, it is not smart, it is only efficient. Like using a powerful crane to lift heavy loads instead of having a single human making multiple trips or coordinating multiple humans to do the same work in the same period of time.

 

[FactChecker]

Well, of course, a person with a fully developed brain becoming blind and paraplegic won't drop to "zero intelligence" overnight; especially knowing it still has other senses.

It's similar with AI programs. You can load in a lot of smarts. Once that is done, you can ask it questions (A terminal is a "sensor". I could argue that reading memory cards is also "sensing" what is on them.)

That is my point: Einstein is very intelligent, and a first-grade student is a little less. Therefore, if you consider a computer intelligent for accomplishing a human task, you should also consider a mechanical machine being intelligent to some level to accomplish a similar task, in a simpler form.

Ok. You are the one denying that it has any intelligence.

Again, this is just a tool that does the work commanded by the human; one that the human could do if given enough time (which might take many generations of humans for a complex problem).

One of the complaints about complicated, deep neural networks is that it is virtually impossible to determine how they reach their conclusions. There are optimization problems that are virtually impossible for a human to solve, no matter how much time you give him.

I still fail to see how one can say that AI is "more intelligent" than humans, or on its way to it, unless one considers doing the work faster as a major contribution to what characterizes intelligence.


So you think a computer with an algorithm that uses the most commonly used passwords to crack a human-made password is a "smarter" machine than one using a brute force attack? It is only more efficient, which only proves that the programmer is smarter, not the machine.

I disagree. It is possible for a mediocre programmer to write a program that can solve problems the programmer could not touch.

The fact that the machine gathers the data as it plays the game is still just another tool used to eliminate the process of having humans filling a database, which would give the same results. It just saves time, with the disadvantage of no one knows exactly what is in the database. That doesn't make the machine more intelligent.

The point is that the computer can experience much more and retain the lessons learned. That is, it is smarter.

or coordinating multiple humans to do the same work in the same period of time.

"multiple humans" is sort of changing the goalpost. But even there, the limitations of communication and retaining experiences can make a group of humans less intelligent than a machine.

 

[FactChecker]

I would reply that the humans who created that AI program were very intelligent, but not the AI program itself.

A mediocre programmer can write AI programs that can solve problems which the programmer can't.
Some AI programs can look at data and examples and retain a library of conclusions and lessons learned that are far beyond the programmer's capability. I call that "intelligence".

 

[javisot]

A mediocre programmer can write AI programs that can solve problems which the programmer can't.
Some AI programs can look at data and examples and retain a library of conclusions and lessons learned that are far beyond the programmer's capability. I call that "intelligence".

You can also use a calculator to perform a calculation. Where do you think the intelligence lies in that case?

 

[FactChecker]

You can also use a calculator to perform a calculation. Where do you think the intelligence lies in that case?

When you say that AI has no intelligence, you have to address my examples, not bring up your own. I would not categorize a light switch as "having intelligence", but that is not the issue.

 

[javisot]

When you say that AI has no intelligence, you have to address my examples, not bring up your own. I would not categorize a light switch as "having intelligence", but that is not the issue.

How does your example differ from mine? In your example, you talk about an AI program capable of solving a problem that a specific programmer cannot. You can also use a calculator to perform a calculation you don't know how to do. In neither case do I think the intelligence resides in the AI program/calculator.

 

[FactChecker]

How does your example differ from mine? In your example, you talk about an AI program capable of solving a problem that a specific programmer cannot. You can also use a calculator to perform a calculation you don't know how to do. In neither case do I think the intelligence resides in the AI program/calculator.

Some AI programs can look at data and examples to draw conclusions and tendencies, then store them for later use. Some can do that beyond the capability of any human on Earth. I would count that as a form of "intelligence".

 

[russ_watters]

Out of interest, to what level do you play chess?

Poorly and infrequently.

I'm aware that winning at chess by calculation alone is difficult and improving efficiency means the computers need to use strategy. And that it didn't take long for computers to be better at calculating (because of course) than humans, but it took longer for them to develop and execute strategy to consistently beat grand masters.

That represents a massive moving of the goalposts.

Heisenberg's Goalposts are impossible to pin down for both sides on this issue. That's the key problem for all of these AI debates. Neither side can define it well, much less can the sides agree on the definition and criteria. Both sides have accidentally described a spreadsheet or even an analog thermostat by their criteria.

In short, there is nothing a computer could do that would count as intelligence. Even if you have specified the criteria, once it meets those criteria, it doesn't count anymore.

There may be people here who have decided they won't ever accept the existence of AI, but I'm not one of them. And I think the issue of defining the criteria isn't any easier for you than for me.

But I think what would impress me is AI successfully completing mundane tasks that involve potentially unpredictable interactions with real humans. Interactions without guard rails and completely inviolable rules. It's tough to describe the criteria though, because why not provide rules even for easy/mundane tasks?

Is running a drive-through harder than beating Gary Kasparov at chess? Then why can most 16 year olds do it as their first job but AI can't do it well enough to avoid being fired despite being designed and supported by many millions of dollars and teams of the world's best engineers? And I think the answer is clearly the rules/guard rails make chess easier or rather better suited for a computer.

Where it will become hard to judge is when the engineers have written enough rules ("No, you may not have 500 cups of water.") to patch a lot of the holes. Will that mean it's smart enough to know that that's an unreasonable request or is it just saying no because an engineer told it to? That's a tough question to answer, but I also recognize that once it gets to that point for certain tasks, the answer won't matter very much anymore.

What would REALLY impress me though is when AI takes over scientific research, either loosely directed or not at all. For now it doesn't seem like we're very close to a layperson with a dream and a ChatGPT account unifying physics. For non-moderators: we get ChatGPT fever-dream personal theories FREQUENTLY.

In this limited sphere it was, as many leading players admitted, like seeing the game played by a superior intelligence.

Sure, but Gary Kasparov said that of Deep Blue 30 years ago. That, again, is the key problem: it looks like "intelligence", but is it?

 

[PeterDonis]

Moderator's note: Thread moved to General Discussion as it is not based on any actual data or research on specific AI models but is just a general discussion about AI.

 

[PeterDonis]

Given that neuroscience cannot definitively explain where consciousness originates - whether as an emergent property of the brain or something transcendent - is it be possible for complex chatbots to develop behaviours that can’t be predicted by simply analyzing the sum of their parts?

If we don't know the answer for our brains, I don't see how we could usefully answer the question about computers (whether running chatbot programs or anything else) either.

 

[PeterDonis]

Without a body and a set of sensors, I fail to see how one can classify any set of atoms as "intelligent".

Without some kind of reasonable concrete definition of what you mean by "intelligence", I fail to see how we can even usefully discuss the subject at all.

 

[Filip Larsen]

Indeed, if the algorithms had been specified by humans, it would not have been so effective - as its neural network ultimately saw deeper into the game than any human had.

Exactly. An often used measure of intelligence is the ability to detect and apply patterns, and training of neural networks are a process intended to achieve just that. It should be no surprise that when scaled up this process is able to pick up patterns we in human interactions would consider an indicator for intelligence.

Perhaps some of the disagreement or confusion in PF discussions on AI intelligence stems from the difference in adaptability between the training process and the deployment (prediction) process, with the first process providing what we could call fluid or adaptive intelligence and the later mostly just statically applying that "knowledge".

For instance, no one would claim that a NN system trained to steer a car is intelligent by itself (i.e. it is statically doing its designed job), but the training process that arrived at that particular network may indeed have found some very "clever" combinations of sensor data that no human designer had noticed before.