A.I. "Jazz"

[CyberGene]

5 minutes ago, Reezekeys said:

"mistakes", or breaking "rules." I would be interested in seeing how an AI can be programmed to do that. Introduce "randomness"?

Well, actually that’s how AI is programmed in a more general sense: the generative logic is allowed to produce almost random stuff at first and the results are judged and scored. Through that learning process and with enough iterations, the AI self-adapts until it starts to produce desirable results. That’s why you sometimes hear the term Machine Learning used interchangeably with AI. Actually ML is a particular implementation of AI where the machine “learns” how to solve a particular task through an automated learning process. It’s an oversimplification but roughly it’s the same. And is how we all learn to do things, through mistakes and feedback from our senses or, in the case of art, through the subjective weights the others give to our attempts.

[MathOfInsects]

Just now, CyberGene said:

Well, actually that’s how AI is programmed in a more general sense: the generative logic is allowed to produce almost random stuff at first and the results are judged and scored. Through that learning process and with enough iterations, the AI self-adapts until it starts to produce desirable results. That’s why you sometimes hear the term Machine Learning used interchangeably with AI. Actually ML is a particular implementation of AI where the machine “learns” how to solve a particular task through an automated learning process. It’s an oversimplification but roughly it’s the same. And is how we all learn to do things, through mistakes and feedback from our senses or, in the case of art, through the subjective weights the others give to our attempts.

Yep, jinx.

[jazzpiano88]

29 minutes ago, CyberGene said:

Well, actually that’s how AI is programmed in a more general sense: the generative logic is allowed to produce almost random stuff at first and the results are judged and scored. Through that learning process and with enough iterations, the AI self-adapts until it starts to produce desirable results. That’s why you sometimes hear the term Machine Learning used interchangeably with AI. Actually ML is a particular implementation of AI where the machine “learns” how to solve a particular task through an automated learning process. It’s an oversimplification but roughly it’s the same. And is how we all learn to do things, through mistakes and feedback from our senses or, in the case of art, through the subjective weights the others give to our attempts.

Yep, you’re describing the training phase which can be thought of as learning.  Once a network is trained, the typical randomness is generated by the randomness of the input.   Example - you feed it a stick figure with some randomness and it will create a random tiger image associated with that skeleton.  Some generative models that use an encoder/decoder structure add a random variation at the latent layer to create random variations in the output at runtime in addition to the randomness of the input.  

[AROIOS]

15 hours ago, jazzpiano88 said:

One of the major things that sparked the AI, or Machine Learning, revolution was the plethora of training data, namely image data.   Large (Deep) Neural Nets capable of doing amazing things were not practical without enough data to train them without overfitting.  

 

 So with a large image data set available called imagenet the researchers toiled away with various network topologies, training techniques and hyper parameters to see who could achieve the best image classification accuracy on millions of images for object recognition. It was literally a completion. 
 

Along the way, they started using the networks to do other things, like generate continuous outputs to steer a car.  And here we are.  I prefer the term Machine Learning over AI because I think it’s more descriptive.  We’re not really mimicking “intelligence” IMO.  

It's mostly marketing. Even the term "Machine Learning" itself is largely just CS guys re-branding good ole "Statistical Learning".

[AROIOS]

12 hours ago, Reezekeys said:

...How does an AI decide which ideas "sound good" or not?...

It sends bots to capture and force humans to grade its output...

That's actually not too far from what actually happens. There are huge teams of humans grading and labeling training data and A.I. outputs both here in the US and in outsourcing destinations like India.

Remember the online Captcha challenges ? That's us labeling data for A.I. training (and feeding Skynet 😃)
 

[AROIOS]

11 hours ago, Reezekeys said:

"If anyone can click on a few buttons and get a Mona Lisa or a Coltrane soundalike track, what does that say about our capacity to honor or enjoy human creative work? It's gonna be a boring world. Am I imagining it or are some folks actually excited about this prospect?"

Apples to oranges, but I'd pick Calculus and Quantum Physics over Mona Lisa and Giant Steps any given day, as our species' "capacity to honor".

And A.I. is already churning out new math and physics rules that used to take humans centuries, if not millennia, to discover.

That's the kind of stuff the keeps me up at night. Mona Lisa and Giant Steps? Nice to have, but won't do a thing to cure cancer or generate free energy.

[jazzpiano88]

12 hours ago, AROIOS said:

It's mostly marketing. Even the term "Machine Learning" itself is largely just CS guys re-branding good ole "Statistical Learning".

 

Well yes.  From a freshman college point of view, it's fitting three points to a line to minimize the error to that line.   Least Squares. 

 

When they train millions of weights and biases in a deep neural network model using backpropagation and minimize the output of a similar loss function (least error) , it somehow becomes "AI".

 

It's amazing that people think that a computer program minimizing millions of variables to an objective function via a well known algorithm that maximizes the slope (gradient descent), that your can write down, makes it "Intelligent".   It's in a textbook   

 

[Reezekeys]

39 minutes ago, AROIOS said:

Apples to oranges, but I'd pick Calculus and Quantum Physics over Mona Lisa and Giant Steps any given day, as our species' "capacity to honor".

 

Apples and oranges indeed. Sure let AI do all the math and physics it wants to. You did start this thread with a post about it doing a credible Bill Evans ("actually enjoyable" was your description), so I've been assuming we're discussing its artistic bona fides and posted as such.

 

49 minutes ago, AROIOS said:

There are huge teams of humans grading and labeling training data

 

If there were humans charged with grading that particular "Bill Evans" track, it sounds to me like they know little about what good jazz is supposed to sound like!

[AROIOS]

1 minute ago, jazzpiano88 said:

...They think that a computer program minimizing millions of variables to an objective function via a well known algorithm that maximizes the slope (gradient descent), that your can write down, makes it "Intelligent".    Jesus!!!!  It's in a textbook -- been there for 50 years.  ...

At the risk of repeating myself, there's the possibility that those simple math turn out to be how our brains work, just at a massively larger scale.

[jazzpiano88]

28 minutes ago, AROIOS said:

33 minutes ago, jazzpiano88 said:

They think that a computer program minimizing millions of variables to an objective function via a well known algorithm that maximizes the slope (gradient descent), that your can write down, makes it "Intelligent".    Jesus!!!!  It's in a textbook -- been there for 50 years.  ...

At the risk of repeating myself, there's the possibility that those simple math turn out to be how our brains work, just at a massively larger scale.

 

Not possible given the number of neuron synapse states (more than the number of elementary particles in the universe, as Carl Sagan pointed out), and the timing for a human to make a decision (i.e it would be a nice model, but it's been ruled out quickly).

[jazzpiano88]

Very good info in here from one of the pioneers:  Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI.

 

A neat little sound byte:  LLMs (chatGPT) can pass the Bar exam but it can't grasp the concepts of how to load a dishwasher that a 10 year old can learn in one shot.  

 

 

[El Lobo]

K-Pop girl bands Eternity and Mave's videos have racked up millions of views. None of these singers are real. 

https://www.cnn.com/videos/style/2023/08/22/ai-kpop-watson-pkg-intldsk.cnn

 

 

[El Lobo]

There are compilations of AI-generated piano music.

 

[stoken6]

10 hours ago, jazzpiano88 said:

A neat little sound byte:  LLMs (chatGPT) can pass the Bar exam but it can't grasp the concepts of how to load a dishwasher

Reminds me of something I heard about 10 years ago: computers can trade derivatives but they can't clean a toilet. Whose jobs are at risk?

 

Cheers, Mike.

[Adan]

12 hours ago, jazzpiano88 said:

 

A neat little sound byte:  LLMs (chatGPT) can pass the Bar exam 

 

I love this one, because most practicing lawyers understand the bar exam merely filters out some of the idiots but has almost nothing to do with being a good lawyer.

[jazzpiano88]

2 hours ago, Adan said:

15 hours ago, jazzpiano88 said:

 

A neat little sound byte:  LLMs (chatGPT) can pass the Bar exam 

 

I love this one, because most practicing lawyers understand the bar exam merely filters out some of the idiots but has almost nothing to do with being a good lawyer.

Yes. Since existing law (anything that would be on the bar exam) is completely codified in language it makes sense that the LLM would be nearly perfectly suited to the task.    But what makes a good lawyer is reasoning and applying the law to ambiguous and new situations which LLMs can’t do.  It’s interesting that some lawyer tried to use the LLM to write a brief or a motion and it got caught just making up bullshit (hallucinating facts that don’t exist).