Watch the "Alpha Go" documentary. It might still be on Netflix.

quote:

Originally posted by RGRacing:

quote:

Originally posted by RichardC:
A natural blonde, dyeing her hair brunette.

Winner Winner ...

=====
STOLEN ! -- (will be sent to my smarty pants blonde daughter)

I never miss an opportunity to hit her with another blonde joke.

I'm a computer science PhD student. AI and machine learning are not my specific research area, but I've done a lot of work with them.


The most general, non-technical definition of AI that I can come up with is "software that looks at a set of inputs and makes decisions," although the definition isn't perfect.


That can mean things like looking at an image and deciding whether it's a picture of a dog or a cat, or looking at all the sensor output on a self driving car and deciding what the car should do next, or a million other things.


Machine learning specifically refers to AI systems where you "train" the system to make correct decisions by giving it many examples of inputs and the correct decisions based on those inputs.

quote:

Originally posted by slosig:
AI is not linear, and it involves the computer "learning" and making decisions based on what it has learned rather than only on what was programmed in.


Another term used lately is "machine learning". I dunno if this simply another way of saying the same thing of if "machine learning" is somehow distinct from "Artificial Intelligence".

The first part of this is actually machine learning. If you could sit down and write a really good piece of decision-making software that never needed to learn anything it would still be AI.

quote:

Originally posted by cyanide357:
Now, there are actual researchers on AI that I'm sure are legitimate. It as well as machine learning.


But my gut feeling in regards to the AI buzz-wording going around these days it just glorified rebranded skip logic / branch logic where all the scenarios/responses have to be pre-programmed.


I think most things currently in play are half-baked implementations from startups trying to get VC funding.


Or that is just the cynic in me.

This is not true at all. The current standard technique in machine learning involves constructing a general-purpose network and then training the network parameters on huge quantities of inputs paired with the desired outputs. The network can then be fed an input similar to those it was trained on, but has never seen before, and make a decision on it. The networks used are not branching decision trees.


The structure of the network that is used is critical and requires human input, and there are parameters used to tune the training process, but the training itself is completely automatic.


To return to an earlier example, you can train a network by giving it 100,000 pictures of dogs and cats, each one labeled "dog" or "cat" based on what is in the picture. Then you can give it a new picture of a dog it has never seen before and it should be able to tell you the picture is of a dog.

quote:

Originally posted by Jimineer:
20 or so years ago it was called Neural Networks but has evolved into AI and machine learning.


Machine learning is AI applied to a machine with a computer that causes the machine to perform some task based on measurements or observations by the machine using various sensors like visual, temperature, speed, and other measurable parameters.

AI is much, much more than just neural networks.


The "machine" in "machine learning" isn't some machine the computer learns to control. The "machine" IS the computer. "Machine learning" refers to computer systems where the system learns to do what you want from you giving it lots of examples of what you want it to do. ML systems are used for all kinds of things. Some of them are machine control tasks, like self driving cars. Many of them are not, like image or speech recognition, weather prediction, predictive text input, all kinds of stuff.

quote:

Originally posted by ArtieS:
My view, and it's probably not shared in the AI community, is that currently AI is best thought of as a kind of machine learning that takes the place of a degree of programming.

AI is a lot more than machine learning, but modern machine learning systems are capable of looking at larger quantities of data than humans can deal with and can infer relationships that are too complex for humans to recognize.


Modern machine learning systems do a better job at many tasks than it would be possible for humans to simply sit down and write software for.

quote:

Originally posted by BurtonRW:
As far as I’m concerned, it’s not AI until it can pass the Turing Test. Repeatedly (to account for examiner or other structural bias in the administration).


-Rob

In modern parlance, this idea is referred to as "strong" or "general" AI. People have been trying to do this for maybe 50 years and essentially no progress has been made.


The field of AI today is concerned with systems that perform specific tasks rather than general AI.

quote:

Originally posted by radioman:
I'm still lost on how AI would differ from traditional software. Let's take, as an example, the alarm system on my house. It has several doors to monitor concurrently, so that's multiple inputs. I sometimes enter through the front door, and sometimes I enter through my back door or side door.


....


That seems like an AI activity, but honestly, I could have written this code on the IBM-XT clone I had 25 years ago.


I still fail to see what's really new about AI compared to what we were doing 25 years ago. It's just that now we have faster computers and more memory, and maybe more sources of data to look for correlation with. I just don't see how AI is somehow different. --please help me out here.


Also, Is the hardware for AI different, or is it the same stuff we used 25 years ago, only a faster version with more memory and more data inputs?

Modern AI systems deal with tasks that are too complicated to sit down and write a bunch of if-then-elses and get anywhere close to solving the problem.

What you're talking about used to be called "expert systems," and that is a part of the AI field, but has fallen almost completely out of use.

On the hardware question, there is some specialized hardware out there, but AI software mostly runs on more-or-less "normal" computers. A lot of AI software makes use of the massively parallel processing power of graphics cards (modern high-end graphics cards have 1000+ very fast math-oriented processors and a bunch of memory designed to feed them as fast as possible) and some of it runs on clusters of machines that each have a bunch of graphics cards.

I've worked in the tech sector for over 20 years and AI scares the hell out of me. Computer systems make wonderful servants which can accomplish great things under the control of their owner(s). But AI has opened the door to computer systems becoming the decision makers, which we are learning to rely on more heavily throughout almost every attribute of our lives. In a broad sense, machine learning simply provides the continual data stream and analysis to support AI evolution and growth.

AI can have an actual sensory feedback to make decisions based on newly learned "skills" by creating and storing new code.

Joe Rogan, if you're familiar with his podcast, just had on his set Lex Fridman who is an MIT scientist who apparently works with AI. I haven't listened to this episode yet, #1292, but more than likely he'll delve into artificial intelligence. Could be interesting.

quote:

Originally posted by radioman:

I'm still lost on how AI would differ from traditional software. Let's take, as an example, the alarm system on my house. It has several doors to monitor concurrently, so that's multiple inputs. I sometimes enter through the front door, and sometimes I enter through my back door or side door.

Maladat mentioned it already but only at the back end and just a mention. But it deserves expansion.

The key that distinguishes computer code is the ability to do "If then" statements or some variation of it such as If then else if else if etc. Or For Case A, do 1, Case B, do 2, Case C, do 3.

Great computer code captures all the possible iterations and also all the possible exceptions. But these have to be coded. And most error traps are simply "if it's not any of the acceptable values, do this" which usually is just an error code or to exit the program.

AI would have the attribute of "fuzzy thinking" where it can handle something that's not quite right but not quite wrong either. It would have a process to determine its proper response to the thing that is not quite right and not quite wrong, i.e. something it was not originally coded for, and then to codify the proper response for future similar situations. Thus, it "learns."

Link to original video: https://youtu.be/15PK38MUEPM

42

0-0

^^^^^^

This ( the video above ) is great!

Very informative - thanks Side_shot.

This message has been edited. Last edited by: HighZonie,

People mean stuff like machine learning, although that doesn't even mimic actual intelligence yet. (Turing test.)

Maybe it will someday, but not yet.

Is it just a difference of degree, between what a computer can do now and what would actually appear intelligent, or is it something quantitatively different?

This CGP Grey video is pretty good for machine learning:

https://youtu.be/R9OHn5ZF4Uo

I'm on my phone so I can't embed.

My understanding has been that AI means that the device has the ability to learn from its mistakes

I'm reading , Ian McEwan's 2019 book Machines Like Me, and they mention the trolley problem - using it with self driving cars. A great read, with AI.

I've never done any type of programming with the kind of tasking AI and the like is designed to handle, the programming I do is designed to handle very specific tasks efficiently enough that it can scale where required. Regarding AI them, as our resident PhD candidate has stated, my interpretation would be that AI programming is less deterministic, by quite a large degree, to the typical type of programming most programmers do to accomplish very well designed tasks, such as banking software for example, or manipulating large datasets in extremely and predictably specific ways. So my uneducated interpretation is that programming AI may share many characteristics of "regular" programming, but the desired result is much more generalized, for example like a turing test, which by definition is not the goal compared to a decision tree for a banking transaction, which would completely fail if it were not 100% deterministic and perfectly and accurately predictable.

My buddy at work says that someday you may be able to buy a wife and some children that are human robots that could pass any turing type test you could throw at "them". Even with that, I still can't see how said machines could be self aware, even though no one can define that either. Funny but could be true some day.

The subject of AI is one we touched in the industry study I used to lead.

The best explanation I've come across covers 3 levels of AI.
ANI - Artificial Narrow Intelligence - a computer program that can accomplish one task better than a human - example Google Maps. Can figure out locations and driving directions better than a human. But that's all it can do. Can't make plumbing recommendations.

AGI - Artificial General Intelligence - an advanced computer program that can do most tasks as well as a human - this is the holy grail of AI when programs can accomplish virtually an task as well as a human. At this point the program is capable of improving itself and it is a very short trip to the next level. Guestimates on when AGI will happen range from decades to never depending on computer speeds and whether such a program can even exist.

ASI - Artificial Super Intelligence - a computer system able to complete any task at a speed and level as far beyond human capability as we are above an ant.

The below link it a great primer I used with my students and I actually had the author on a conference call with my class one year.


https://waitbutwhy.com/2015/01...ce-revolution-1.html

quote:

Originally posted by Rey HRH:

Maladat mentioned it already but only at the back end and just a mention. But it deserves expansion.

The key that distinguishes computer code is the ability to do "If then" statements or some variation of it such as If then else if else if etc. Or For Case A, do 1, Case B, do 2, Case C, do 3.

Great computer code captures all the possible iterations and also all the possible exceptions. But these have to be coded. And most error traps are simply "if it's not any of the acceptable values, do this" which usually is just an error code or to exit the program.

AI would have the attribute of "fuzzy thinking" where it can handle something that's not quite right but not quite wrong either. It would have a process to determine its proper response to the thing that is not quite right and not quite wrong, i.e. something it was not originally coded for, and then to codify the proper response for future similar situations. Thus, it "learns."

Ok, this is starting to make sense. I can start to see how the learning part works: Making code for situations which were not originally coded.

Great thread. I've learned an enormous amount of information.