What definition of "learn" are you using?

To chomp on data using the rules of math and physics, to derive somthing new, add that "new" to the data set, iterate. Aka "learn".
The learn model for humans (in basic terms), K-12, college/university, Phd+

It can't be done with AI ML, classic catch-22, the limits boundary seperates humans from machines. AI ML shines in "evaluation" and "analysis" realms, same as what humans can do, just much much faster. AI Ml is also rooted in "positive" space, since a pic of white-noise with some blob in it is not a Cheetah.

I would still like to see the accuracy of the neural models are in classifying object type when evaluating say AI generated images. I guess if you made an AI model of say a Canon 50mm prime glass lens coupled to a cropped cmos sensor, then I suspect images made from that AI model will be hard to detect as fake. But that said, there's still limits there.

Some day, you'll be able to click the button that says "give me AI model for the Canon EOS R5 2023 model with a Canon 50mm prime lens #0570C002"
And each time you click that "give me model" button, the AI creating the AI model you get alters the model ever so slightly to accomodate the errors found in cmos image sensors and the errors in lens glass (add more error metrics as you see fit), this way every AI model downloaded for use will create non-identical images, just like real world variances between all Canon's set up that same exact way. In real physical space, if you own a Canon setup, all the images from that setup will typically all have some identifying errors that tie all the images back to that specific setup. But with the AI I mention, every image would appear to come from unique Camera's as long as you grab a newly generated AI model to use, this way no two images will ever appear to come from same "camera". Building a single AI model that in itself creates the errors needed, is still a single AI model, not as good.

 

To chomp on data using the rules of math and physics, to derive somthing new, add that "new" to the data set, iterate. Aka "learn”. The learn model for humans (in basic terms), K-12, college/university, Phd+

I don’t mean to be obtuse but I don’t understand this explanation. What data is being evaluated, what is “the data set”. How does that describe “learning” in a scholastic or academic context? I genuinely can’t follow this.

It can't be done with AI ML, classic catch-22, the limits boundary seperates humans from machines. AI ML shines in "evaluation" and "analysis" realms, same as what humans can do, just much much faster. AI Ml is also rooted in "positive" space, since a pic of white-noise with some blob in it is not a Cheetah.

How are you using the terms Artificial Intelligence and Machine Learning? What is the relevance of “negative space” to this? Why is that relevant? Is it different from human learning or are we still talking about learning at all.

Building a single AI model that in itself creates the errors needed, is still a single AI model, not as good.

It the discussion of modeling cameras a digression or is it part of the explanation of learning?

I don’t mean to be a pain, but I am having a very hard time following you and you seem to think this is an important insight. Maybe you could be a little more systematic and lay out the argument you are making without assuming the discursive “evidence” helps to understand it?

Again, I would like to get some idea where you are trying to go with this—not give you a hard time.

 

AI and ML are very abstract to most. I perhaps can't explain in text the info I am trying to convey.

What is the relevance of “negative space” to this? Why is that relevant?

As example, look at the pics in post #12. Any one image is the "data set" for the pic. You can certainly have a collection of pics of object type X, like "these are all horses", hence a larger collection of data of type X, a bigger data set. You have two observers, the human eye and the camera with AI model behind it. To humans, all of the pics that look like whitenoise are in the negative sense, they have no real object type to the human, other than we say it's general whitenoise, which is like saying "imaginary number". In essence, easily "detected" as FALSE if someone says "hey, that's a pic of a donut". On the flip side you have the positive realm, object types humans can easily ID as a specific thing, elephants, cars, rockets, frisbees, buses, pencils, etc etc. However, the AI model can falsely ID both negative and positive pics (object types).

When you get the human to believe the false AI, that's no different than some people who say a man can become pregnant.

The collision of AI ML and humans is gonna be a challenge.

In real world application, the development of AI based tools have moved fast in terms of sheer number of models being developed and used. What's lagging are the tools that can validate a data set of unknown origin, in which Nicholas Carlini describes some in the 2017 IEEE vid a few posts back. The neural model ID's object types, it's good until an adversary changes a few pixels, then the model falsely ID's the object type, the adversary has tricked the model, etc. The tools needed to validate the results that come from AI ML (neural) models (aka "tools") are lacking. When the dog is ID'd as a bird, we humans can see that error, but if the analysis and results are all done in digital world and they carry along through the system, well, you can now see how things can get messed up real fast.

As for the "learning" part, ask the question from the other direction, show examples of what any AI model has learned new and has put that back into the data set. "new" is like discovering (deriving) a new DNA sequence that is responsible for a specific illness. What AI model can do that?

 

AI and ML are very abstract to most. I perhaps can't explain in text the info I am trying to convey.

Thanks for taking the time to answer.

I don‘t think we are going to be able to have a meeting of the minds on this. What you are sating just doesn’t hang together for me and what I know about AI, and Machine Learning, and the nature of the model (Stable Diffusion) that produces those images.

I am not saying you are wrong because I can’t work out the case you are trying to make. So, I am just going to leave it as unresolved, at least for now. Maybe something will come up to help clarify it.

Thanks again for the extensive reply and willingness to explain.

 

Can AI / ML really "learn" anything? Technologies created by humans really can't do such, they can only use existing data sets (knowledge) to derive limited answer set. AI ML cannot solve any complex problems, like time travel, or big universe questions.

Hi,

The way it stands now it is seems to require knowing everything about everything in order to render an answer about anything that is asked of it.

The way I see it is like thinking about a TOE where:
"You can not give a precise answer about anything unless you know everything."
In other words, even the knowledge we have as humans is always going to be incomplete and possibly proven false until we know everything about the universe and it's creation.

I see AI as something like that too. That's because I think a lot of it is pure comparative analysis, which is how the AI program I worked with for a while was able to provide certain answers you would think it would not be able to. That also seems to be why it takes so much training to get it working.
The only sidestep to this I see right now is that it can have the internet as a learning tool with all of it's information, but then it also has to learn what is false as well as what is true. That still allows it to learn on it's own which is kind of nice because feeding information takes forever. I got so bored I just couldn't put up with it anymore.
It seems the only way to get it to know 'false' from 'true' on the internet is it would have to read multiple articles on the same subject and do the "Space Shuttle walk", where if it finds 2 articles that agree with each other and 1 that disagrees, it takes the first two articles as true. Likewise, if it finds 3 that agree and 2 that do not agree with those 3, it takes the 3 as true. So majority wins. That's probably the only mechanism. The problem comes up where at in the beginning, if it finds 2 that agree and 1 that doesn't and it takes those 2 as true, then later finds the 1 that did not agree suddenly takes on 2 more that agree with it, then it has to switch to the second set of articles as true. Thus it flips it's decision over time.
I find articles on the web that are not necessarily false, but their math writeup makes typographical errors, rendering the entire article almost moot. The only recourse is to try to replicate the procedure for finding the right math, which can take time. Could AI ever do that.

Hey I just thought of something that may be interesting.
How do we teach an AI object to understand different colors, like blue, green, red, yellow, etc.?
I think it would have to have an image sensor of some type so that it can look at the different pixels and their color components like the BGR bitmap files/images.

 

If some future AI becomes self-aware, intelligent and can go beyond learning to the intuition of the unknown I don't think they would want to be anything remotely human. We are likely slugs in the universe of intelligence and physical abilities.

 

DIY AI?

 

If you get to see Killer Robots on Netflix, you'll get an idea of the realms Ai and ML live in.
ML is not learning anything, it's more like very fast mapping of local sensor data into large sets of modeling data, they call it "ML". "Ai" is a classical oxymoron, just more buzz words for the era, like "cloud" and "zero trust".

 

https://www.technologyreview.com/s/610434/why-humans-learn-faster-than-ai-for-now/


It's not surprising that these brute force machine learning or "deep" learning systems have problems when there is little to signal a 'good' path from a 'bad' path. There is very little intelligence in current AI that people didn't already program into it.

Leisure Suit Larry in the Land of the Lounge Lizards

After reading up on neuroscience, basically the problem is twofold:

A. A single human neuron is much smarter than an artificial one. Whereas the neurons in artificial neural networks are passive for the most part, just comparing inputs and weights to produce a response, a human neuron will make decisions on its own and can adjust how it responds to input in real time based on the inputs. Studies indicate that one biological neuron requires an entire network of artificial ones to simulate adequately.

B. The human brain (and animal brains for that matter) aren't just messes of neurons that "come together" into a system magically. Brains have their own architecture. And one of the tasks that they're optimized for is efficient memorization and problem-solving.

I also think our current models of neural networks that just take input and produce a response are too simplistic. Real neurons form entire circuits that can perform much more complex functions, such as timing. ANNs aren't capable of this, at least in the configuration they're usually used in.

 

What I have been thinking about now from time to time is just what should 'ai' be, not as much anymore about what it "is" right now mostly because it's too inferior at the time.

There are so many dimensions to the human intelligence with all kinds of different directions a thought can take. Are we using motor skills, charging batteries, solving problems, teaching, learning, etc.

I'd hate to have to try to code a program to have all the skills needed to be human.

 

I'd hate to have to try to code a program to have all the skills needed to be human.

It's not possible to do such. You can code pieces of human nature, and even those pieces won't be 100% human-like.

The "skills" you speak of are not exactly skills in AI. AI is coded at the fundamental rules level. Like math, we know all the rules of math. Like physics, we know all the rules of physics. Like language, we know a lot about languages. However, you can only code what we know, and in grand scheme of things we humans know very little. And perhaps yes, you can code the rules needed for an AI robot to go seek out the materials and processes needed to build a clone of itself, but that's still not a robot thinking for itself. Terminator T-800 bot is perhaps the closest parallel, notice the T-800 could not make itself more advanced, no self ability for gain-of-function there, etc.

 

AI in my opinion is becoming more and more integrated into all systems lately. In the last few years, I've been hearing about different innovations. Yes, there is still work to be done, but the results are getting better and better.

 

It's not possible to do such. You can code pieces of human nature, and even those pieces won't be 100% human-like.

The "skills" you speak of are not exactly skills in AI. AI is coded at the fundamental rules level. Like math, we know all the rules of math. Like physics, we know all the rules of physics. Like language, we know a lot about languages. However, you can only code what we know, and in grand scheme of things we humans know very little. And perhaps yes, you can code the rules needed for an AI robot to go seek out the materials and processes needed to build a clone of itself, but that's still not a robot thinking for itself. Terminator T-800 bot is perhaps the closest parallel, notice the T-800 could not make itself more advanced, no self ability for gain-of-function there, etc.

Hi,

Not saying you are wrong, but do we have some proof of that?
Sometimes some things are not done simply because nobody has done them yet

 

A. A single human neuron is much smarter than an artificial one

AI is as dumb as the persons programming the models.

As example, AI cannot develop new math or physics. Everything else is just permutations, it's synonymous with differential equations where you can have lots of solutions. As example, if you wanted a new SARS virus then you can have an AI model that can churn out new strains, but those new strains are confined by math and physics (the rules of SARS), so essentially nothing new being "created" by AI.

If you wanted a HAL to take over the world, it can only do so by way of a human programming in how it would do that. If you were able to program all the ways to take over the world, let's say on billion ways (each way many peta bytes big), there is no way for AI to determine which way would be the "best" way, because the metrics of "best" have to already be defined and programmed into the model, programmed by a human.

AI is also astoundingly restrcited by digital bits and ability to process, we have no way to store the amount of data & programming needed by "do anything" or "tall task" AI, nor do we have the compute power to get conclusions in a timely matter even with quantum computing speeds.

AI is currently applied in micro scale, NLP (chatbots in text and audio, etc), robots that can determine if it's on ice or rocks or sand or water, autonomous attack drones, etc. It's all very micro in application. A gazillion micro AI models will never add up to "decided on it's own take over the world".

I predict we are hundreds of years away from any AI being "dangerous" in context of "thinking for itself".

Hi,
Not saying you are wrong, but do we have some proof of that?
Sometimes some things are not done simply because nobody has done them yet

Proof? Sure. You see any AI model that is building new tech for rocket motors. The answer is no, we know this from deduction logic, because if AI could do it then we would be 1,000 or 10,000yrs into the future on new rocket motor technology.

 

AI is as dumb as the persons programming the models.

As example, AI cannot develop new math or physics. Everything else is just permutations, it's synonymous with differential equations where you can have lots of solutions. As example, if you wanted a new SARS virus then you can have an AI model that can churn out new strains, but those new strains are confined by math and physics (the rules of SARS), so essentially nothing new being "created" by AI.

If you wanted a HAL to take over the world, it can only do so by way of a human programming in how it would do that. If you were able to program all the ways to take over the world, let's say on billion ways (each way many peta bytes big), there is no way for AI to determine which way would be the "best" way, because the metrics of "best" have to already be defined and programmed into the model, programmed by a human.

AI is also astoundingly restrcited by digital bits and ability to process, we have no way to store the amount of data & programming needed by "do anything" or "tall task" AI, nor do we have the compute power to get conclusions in a timely matter even with quantum computing speeds.

AI is currently applied in micro scale, NLP (chatbots in text and audio, etc), robots that can determine if it's on ice or rocks or sand or water, autonomous attack drones, etc. It's all very micro in application. A gazillion micro AI models will never add up to "decided on it's own take over the world".

I predict we are hundreds of years away from any AI being "dangerous" in context of "thinking for itself".



Proof? Sure. You see any AI model that is building new tech for rocket motors. The answer is no, we know this from deduction logic, because if AI could do it then we would be 1,000 or 10,000yrs into the future on new rocket motor technology.

Hi,

To be succinct, that's not really a proof. That's one example that may not have worked, yet. I'm not saying it will be true some day, but it's just one example that does not work right now. Things are changing, and they are changing pretty fast. Once we can start working at the quantum level for everything we will have an entirely new world. The limits are being pushed with what we thought could never be possible.

There does seem to be some unexpected results from 'ai' models, but I can't be sure yet if they are true or not because of the way false info is presented all over the web. I'll be keeping an eye open though.

Have you ever heard of an evolutionary math solver? That kind of program can actually solve for physical functions like the circumference of a circle from its diameter (that's just a simple example). It may not just 'decide' to create a function for that, but it can actually come up with the variables like pi*diameter, eventually.

How did Elon Musk come up with his new electric motor design after years of having almost the same technology. That was probably a mix of computer and human work though.

 

Yes, AI learns (can see patterns that humans can't and quickly make misleading correlations from those patterns) faster today. It can tell if you drink beer or eat refried beans from a knee x-ray.

https://thedebrief.org/how-ai-used-...nd-bean-eating-habits-and-why-its-concerning/

But, can it predict the number of people that mix refried beans in their beer and drink it from looking at a skull x-ray?

 

Yes, AI learns (can see patterns that humans can't and quickly make misleading correlations from those patterns) faster today. It can tell if you drink beer or eat refried beans from a knee x-ray.

It does not learn anything new, it can only match data to a model.

Things are 1st modeled, which of course already needs tons and tons of data to build an accurate model that also carries precision. This is the ML part. You then feed that model untrained data (data not used to build the model). AI then can figure out if the data supplied falls within that model or some variance of.

Again, AI learns nothing new, it's a fancy probability engine. Input some "unknown" data, react to that data accordingly. I quote "unknown" because it's not likely an AI model of roads will be getting input data about the location of stars in the sky, etc.

AI in autonomous vehicles can fail very fast, just go paint a curving white line from the middle of the two lanes, start on a dash and then curve it out to the road apron in about 500ft length, see how the AI vehicle handles that. AI in autonomous has many things modeled, the vehicles can "drive themseleves" fairly "ok", but they will encounter numerous unforseen circumstances, and in some of those cases people will die.

I also note, mobile (cars, drones, etc) AI anything is very restricted, because 1) not enough data storage, and 2) not enough mobile compute power.

AI is very bound (restricted) by storage and compute power. Long ago there was a time where the 256k program ran like lightning on the newly 386 that had a math co-processor. Storage and compute was ahead of the programming. AI flips that problem around 180.

 

It does not learn anything new, it can only match data to a model.

Things are 1st modeled, which of course already needs tons and tons of data to build an accurate model that also carries precision. This is the ML part. You then feed that model untrained data (data not used to build the model). AI then can figure out if the data supplied falls within that model or some variance of.

Again, AI learns nothing new, it's a fancy probability engine. Input some "unknown" data, react to that data accordingly. I quote "unknown" because it's not likely an AI model of roads will be getting input data about the location of stars in the sky, etc.

AI in autonomous vehicles can fail very fast, just go paint a curving white line from the middle of the two lanes, start on a dash and then curve it out to the road apron in about 500ft length, see how the AI vehicle handles that. AI in autonomous has many things modeled, the vehicles can "drive themseleves" fairly "ok", but they will encounter numerous unforseen circumstances, and in some of those cases people will die.

I also note, mobile (cars, drones, etc) AI anything is very restricted, because 1) not enough data storage, and 2) not enough mobile compute power.

AI is very bound (restricted) by storage and compute power. Long ago there was a time where the 256k program ran like lightning on the newly 386 that had a math co-processor. Storage and compute was ahead of the programming. AI flips that problem around 180.

I used 'learn' in the ML definition of learn, not the human one.

https://www.ibm.com/think/topics/machine-learning

 

It does not learn anything new, it can only match data to a model.

Things are 1st modeled, which of course already needs tons and tons of data to build an accurate model that also carries precision. This is the ML part. You then feed that model untrained data (data not used to build the model). AI then can figure out if the data supplied falls within that model or some variance of.

Again, AI learns nothing new, it's a fancy probability engine. Input some "unknown" data, react to that data accordingly. I quote "unknown" because it's not likely an AI model of roads will be getting input data about the location of stars in the sky, etc.

AI in autonomous vehicles can fail very fast, just go paint a curving white line from the middle of the two lanes, start on a dash and then curve it out to the road apron in about 500ft length, see how the AI vehicle handles that. AI in autonomous has many things modeled, the vehicles can "drive themseleves" fairly "ok", but they will encounter numerous unforseen circumstances, and in some of those cases people will die.

I also note, mobile (cars, drones, etc) AI anything is very restricted, because 1) not enough data storage, and 2) not enough mobile compute power.

AI is very bound (restricted) by storage and compute power. Long ago there was a time where the 256k program ran like lightning on the newly 386 that had a math co-processor. Storage and compute was ahead of the programming. AI flips that problem around 180.

Hi,

That last part is an interesting view on this phenomenon.

An interesting question came up recently about how an 'ai' bot would respond. Here is the question:
"Can you connect the anode of a tantalum capacitor to ground?"
The 'ai' bots reply that you cannot connect the anode to ground. It then goes on to explain that the polarity matters.
So it says that you can't do it, but explains it's because of the polarity, which is sort of a contradiction because if you observe the polarity you can do it, such as with a negative power supply.
You can ask this a number of ways, but it will always say you cannot do it. I had to wonder if there was some way to ask the question so that it observes that it actually can be done.

 

An interesting question came up recently about how an 'ai' bot would respond. Here is the question:
"Can you connect the anode of a tantalum capacitor to ground?"
The 'ai' bots reply that you cannot connect the anode to ground. It then goes on to explain that the polarity matters.
So it says that you can't do it, but explains it's because of the polarity, which is sort of a contradiction because if you observe the polarity you can do it, such as with a negative power supply.
You can ask this a number of ways, but it will always say you cannot do it. I had to wonder if there was some way to ask the question so that it observes that it actually can be done.

That's right.
It's because the model was programmed to think "ground" has a certain polarity referenced to the other side. "ground" is not always negative.

The botom-line question might be, can AI drum up any new math or physics? When it can drum up something new, that's when I will be worried. I am however confident that I will never become worried.