eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[Marco]

When you think that you can trust the AI doing the right thing: How Medicare Advantage Plans Use Algorithms To Cut Off Care For Seniors In Need

Some near minimum wage human needing to meet his quota would do any better?

Capitalism + triage makes for a messy business.

[Fungus]

Microsoft just laid off one of its responsible AI teams (https://www.platformer.news/p/microsoft-just-laid-off-one-of-its).

They only want irresponsible teams...

[coppice]

Microsoft just laid off one of its responsible AI teams (https://www.platformer.news/p/microsoft-just-laid-off-one-of-its).

They only want irresponsible teams...

Being responsible doesn't mean them being responsible for anything good. That article seems to indicate they produced little of value.

[SiliconWizard]

When you think that you can trust the AI doing the right thing: How Medicare Advantage Plans Use Algorithms To Cut Off Care For Seniors In Need (https://science.slashdot.org/story/23/03/13/2332231/how-medicare-advantage-plans-use-algorithms-to-cut-off-care-for-seniors-in-need).

Anyway, who needs AI ethics? Microsoft just laid off one of its responsible AI teams (https://www.platformer.news/p/microsoft-just-laid-off-one-of-its).

The whole point of promoting AI on a large scale as it is now is precisely to circumvent any ethics, morals and liability.

And not just ethics, but laws as well.
With "AI", many laws are bound to change to completely flip liability on its head. Obviously IP laws are also going to be changed in a major way, as "AI" is currently pretty much based on stealing IP on a very large scale without people even knowing it or having any say in it.

[Bicurico]

A human can read many books and as a consequence learn something by training the neural network in his brain. If he then uses this knowledge to i.e. write a book, nobody would rais IP issues withe the authors of the books that were read, unless whole sentences would have been copied, which is not the case in this example.
Or, consider a painter that is inspired by other contemporary painters. This happend for any major style, from expressionism to cubism, from renaissance to abstract painting. Just go to the Louvre and see how painters inspired themselves.

With AI, the learning works identically by training a neural network. The original data is not stored in the database. If you would search the GPT model, no readable content would be found.

The only valid objection would be: has OpenAI the right or authorisation to read the books? Are the books open source or was the book purchased?
 

[vk6zgo]

When you think that you can trust the AI doing the right thing: How Medicare Advantage Plans Use Algorithms To Cut Off Care For Seniors In Need

Some near minimum wage human needing to meet his quota would do any better?

Capitalism + triage makes for a messy business.

For an example of a disastrous use of Algorithms Google "Robodebt".

This was set up because the responsible politicians & much of the appropriate Department's management were convinced that there was a lot of cheating by recipients of Unemployment Relief, so they set out to "claw payments back".

In the real world, the problem was minimal, but the system was set up, anyway, & it proceeded to send bills for outlandish amounts which the recipients were alleged to owe to the Dept.

These turned out to, in most cases be totally incorrect, & to add insult to injury, there was no real way to dispute the bills.
Eventually, it was discontinued, as it had become a scandal, & is now the subject of a Royal Commission, with the principals ducking & weaving to try to avoid responsibility.

Hardly the fault of the Algorithms, as they just did what they were set up to do by stupid or malicious (sometimes both) actors.
Some people in the Dept tried to point out it would "all end in tears", but they were steamrollered by its enthusiastic supporters.

[eti]

As sceptical as I am about those who become a bit too enamoured with “AI”, I’ll admit it was VERY helpful when I asked it various questions about LISP programming just now!

[EEVblog]

https://betterprogramming.pub/exploring-the-intersection-of-ai-and-physics-the-role-of-chatgpt-in-code-generation-dacb6fd6f96b

[Fungus]

I’ll admit it was VERY helpful when I asked it various questions about LISP programming just now!

Did it say "Dont!"?

(    )

[Neutrion]

A human can read many books and as a consequence learn something by training the neural network in his brain. If he then uses this knowledge to i.e. write a book, nobody would rais IP issues withe the authors of the books that were read, unless whole sentences would have been copied, which is not the case in this example.
Or, consider a painter that is inspired by other contemporary painters. This happend for any major style, from expressionism to cubism, from renaissance to abstract painting. Just go to the Louvre and see how painters inspired themselves.

With AI, the learning works identically by training a neural network. The original data is not stored in the database. If you would search the GPT model, no readable content would be found.

The only valid objection would be: has OpenAI the right or authorisation to read the books? Are the books open source or was the book purchased?

Except, that none of those human painters could produce paintings on a thousands/second scale, aviable globally.
And  although everybody learns from everybody, intellectual property rights are there for a reason.
So if you think AI should be treated like a human, just because some aspects are the same, than with this logic you could also treat a table as a dog.

[madires]

The new GPT-4 can interpret images: OpenAI Announces Chat GPT-4, an AI That Can Understand Photos (https://petapixel.com/2023/03/14/chat-gpt-4-will-let-you-turn-text-into-video-and-is-coming-next-week/)

[Axk]

Can someone with Plus try this query and see if GPT-4 gives the correct answer?
The Skype Bing bot gives the correct answer with a link to an article, the free version of Chat GPT (GPT-3.5) is producing plausible nonsense like it always did.
So I wonder if it's GPT-4 or the Bing bot is using web searches in some way:

Query: "Give an example of a real story existing in literature where a young man is seeking help from a detective to solve a mystery with his uncle and father dying under strange circumstances and where it turns out Ku Klux Klan was involved"

Correct answer: "The Five Orange Pips by Sir Arthur Conan Doyle"

[Marco]

Bing AI is known to already use GPT-4 BTW.

[Bicurico]

A human can read many books and as a consequence learn something by training the neural network in his brain. If he then uses this knowledge to i.e. write a book, nobody would rais IP issues withe the authors of the books that were read, unless whole sentences would have been copied, which is not the case in this example.
Or, consider a painter that is inspired by other contemporary painters. This happend for any major style, from expressionism to cubism, from renaissance to abstract painting. Just go to the Louvre and see how painters inspired themselves.

With AI, the learning works identically by training a neural network. The original data is not stored in the database. If you would search the GPT model, no readable content would be found.

The only valid objection would be: has OpenAI the right or authorisation to read the books? Are the books open source or was the book purchased?

Except, that none of those human painters could produce paintings on a thousands/second scale, aviable globally.
And  although everybody learns from everybody, intellectual property rights are there for a reason.
So if you think AI should be treated like a human, just because some aspects are the same, than with this logic you could also treat a table as a dog.

"Except, that none of those human painters could produce paintings on a thousands/second scale, aviable globally." --> that has exactly nothing to do with IP matters
"And  although everybody learns from everybody, intellectual property rights are there for a reason." --> IP is mainly obsolete and the IP of a book expires 75 years after the death of the author (if I recall correctly), so many books used by GPT have no IP - you can seatch Project Gutenberg to get an idea of how many books there are!
"So if you think AI should be treated like a human, just because some aspects are the same, than with this logic you could also treat a table as a dog." --> sorry, but I don't understand your example (table for dog). A neural network does EXACTLY try to replicate the way a biological brain works. It solves problems in the same way. And, incredibly, a side-effect of GPT is that it can provide insight on how the brain really works. So, yes, in my opinion it makes sense to treat content created by AI in a similar way as if it was created by a human.

This link has already been posted on this thread: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/

I really recommend everyone to read it. It is written by the creator of Wolfram Alpha and goes into length explaining how GPT works, why it works so good and what we can expect.

I would say that 2023 is the year in which mankind managed to replicate a neural network in the size remotely comparable with the one of the human brain (smaller, but in the same league).

According to Stephen Wolfram, the main current limitation of GPT and alikes is the fact that the neural network is trained and operated sequentially, while the biological brain works in parallel.

Also, the reason we get GPT in 2023 and not like 10 years earlier is because of the sheer computing power required, which was not available sooner. And, of course, the cost resulting from this requirement.

Coming back to IP:

When you buy a book (analog or digital), you "own" the book and can read it as often as you want. You can learn from it. The knowledge you gain is yours and is not linked to any contract to the IP owner of the book. Consider i.e. a physics, electronics or math book. You buy it to learn from it. The only think you are not allowed to do is to replicate the book or parts of it. You might do some quotes, but they need to reference the book.

If AI "reads" the book it does learn from it, too. It does NOT store its contents, though. It will, instead, train the underlying neural networks. What gets stored are parameters. These are abstract and you wouldn't be able to reconstruct the original book from them. No copyright violation in sight.

The only debate is: could an author (of a book, website or image) refuse to allow his IP to be processed by AI?

Where does AI start? What about printed text to audio readers? Could those be used?

What if you train the AI by reading the book to it? Would that be acceptable?

Can AI watch all free TV channels and learn from the broadcasted contents? They are free to air, so why not put a "bot" to watch the contents and learn from it?

All of this is far more complex and my bet is that it is virtually impossible to decide what to do.

Imagine portable super computers, the size of a mobile phone: the student takes it with him to his Uni lectures to train "his" pet AI...

I have to stop writing here, because I am starting to get into SCI-FI mode.

But to finish it off: look at the technological complexity involved until humaity got GPT-4.0. And still, it is only a fraction of what a human being can do. Consider that a human is a self-replicating machine! Now tell me that life on earth started by casual proteins mingling in oceans and through evolution led to humans. This in just a few hundred or thousand million years? Hard to believe.

Regards,
Vitor

[bob808]

The crazy starts around 10 minute mark

[Marco]

According to Stephen Wolfram, the main current limitation of GPT and alikes is the fact that the neural network is trained and operated sequentially, while the biological brain works in parallel.

I think that mischaracterises the article.

I think the core problem they point out is this :

ChatGPT doesn’t internally “have loops” or “recompute on data”. And that inevitably limits its computational capability

Lets say we work through a Gardner Puzzle, we form and reject hypothesis using internal dialogue and/or external scratchpad and only then formulate the solution. The language model has to solve it inside it's hidden layers in one go AND repeat all that work a couple dozen times (the output history of characters doesn't help much, it can't really work iteratively through that). Well it can try any way, it's going to fail on most puzzles not in the training set.

There's some interesting research which tries to replicate hypothesis based problem solving, but nothing as successful as just making the streaming predictors larger. It's amazing what GPT can accomplish with the level of redundant processing it must do.

[Bicurico]

You are correct - I forgot about the loops. But lack of parallel processing is what slows down current AI models.

[coppice]

You are correct - I forgot about the loops. But lack of parallel processing is what slows down current AI models.

You mean a well organised team can finish a job faster than a lone worker? Who knew?

[SiliconWizard]

Bing AI is known to already use GPT-4 BTW.

Of course, that's why MS invested billions of USD in it, as I mentioned earlier.

[Neutrion]

"Except, that none of those human painters could produce paintings on a thousands/second scale, aviable globally." --> that has exactly nothing to do with IP matters
"And  although everybody learns from everybody, intellectual property rights are there for a reason." --> IP is mainly obsolete and the IP of a book expires 75 years after the death of the author (if I recall correctly), so many books used by GPT have no IP - you can seatch Project Gutenberg to get an idea of how many books there are!
"So if you think AI should be treated like a human, just because some aspects are the same, than with this logic you could also treat a table as a dog."

--> sorry, but I don't understand your example (table for dog). A neural network does EXACTLY try to replicate the way a biological brain works. It solves problems in the same way. And, incredibly, a side-effect of GPT is that it can provide insight on how the brain really works. So, yes, in my opinion it makes sense to treat content created by AI in a similar way as if it was created by a human.

A dining table and a dog have one thing in common. That does not mean, you ought to treat them as equal.
You are finding similarities withing the workings of an AI system with humans and conclude, that because there are these similarities we should treat them as similar, and comlpetely exclude the whole lpethora of issues around.
I don't want to be off but this is the way how communism was created. They ignored a lot of fine levels of realities which still were important, but they somehow forgot to calculate with it.

Coming back to IP:

When you buy a book (analog or digital), you "own" the book and can read it as often as you want. You can learn from it. The knowledge you gain is yours and is not linked to any contract to the IP owner of the book. Consider i.e. a physics, electronics or math book. You buy it to learn from it. The only think you are not allowed to do is to replicate the book or parts of it. You might do some quotes, but they need to reference the book.

If AI "reads" the book it does learn from it, too. It does NOT store its contents, though. It will, instead, train the underlying neural networks. What gets stored are parameters. These are abstract and you wouldn't be able to reconstruct the original book from them. No copyright violation in sight.

The only debate is: could an author (of a book, website or image) refuse to allow his IP to be processed by AI?

Where does AI start? What about printed text to audio readers? Could those be used?

What if you train the AI by reading the book to it? Would that be acceptable?

Can AI watch all free TV channels and learn from the broadcasted contents? They are free to air, so why not put a "bot" to watch the contents and learn from it?

All of this is far more complex and my bet is that it is virtually impossible to decide what to do.

Imagine portable super computers, the size of a mobile phone: the student takes it with him to his Uni lectures to train "his" pet AI...

I have to stop writing here, because I am starting to get into SCI-FI mode.

But to finish it off: look at the technological complexity involved until humaity got GPT-4.0. And still, it is only a fraction of what a human being can do. Consider that a human is a self-replicating machine! Now tell me that life on earth started by casual proteins mingling in oceans and through evolution led to humans. This in just a few hundred or thousand million years? Hard to believe.

Regards,
Vitor

Here you listed the issues a bit. AI is not a human and before we would elaborate its possible effect on societies, and have at least theoretical legal and ethical and financial rules for its working , we just unleash it to wait and see as if we had nothing to loose.

Training AI for free: This is also an issue which would have to be discussed first, before getting to this point.
But most of the people who think about it agree, that a general AI which is smarter than humans will not be able to controlled by humans.
But with the training that is what is happening, that scientist are working towards this goal.
And even the not so smart AI as discussed here, can do a lot of trouble. Why should we let this advancement going on without any barriers, and ethical, democratical questions asessed first?
Just to help these poor little companies? Why? Because they promise? Well I can also promise you a lot of things if you pay me a lot of money.

Why cant't we access to the really old informations, as to how exactly build a nuclear bomb? Or a bio weapon? Why is it tightly controlled? It is just a tool, isn't it, so it's not bad.....

[Bicurico]

AI, no matter how powerfull, will not by itself, constitue a threat to human kind. It is the company that owns the AI that might be a threat.

And such kind of threats already exist, here are some examples:

1) Google: when you do a search, the results showing are filtered by parameters setup by Google for their own benefit. If Google wants, content is hidden from users while other content is forced upon users. Try to do a Google search for a product review. The first page will be hits trying to sell you the product, instead of links to actual reviews. That is obviously for ad profit, as it would be difficult to get it that wrong.
2) Facebook: the company controls the whole network of people using it and even those not using it! Indiviual profiles are made and exploited. There is a movie that visualizes the control of Facebook.
3) Any modern smartphone will track your geolocation. Google at least sends you a monthly report where you can see (like themselves) where you have been on any moment.
...

In a modern society, we are full monitored (digital money, toll boxes, smartphones, digital receipts, etc.). If you as an idividual start to worry goverment members, they can easily eliminate your credibility based on the huge data that exists.

So, while I think that AI, GPT, etc., on their own are just a tool that is very helpful, it can be used to further control human kind.

You can check it for yourself: ChatGPT is biased (though responses change over time). While it refuses to produce text for some political ideologies, it happily replies on other political ideologies. That is on purpose.

But I do see the benefit of quickly learning something or getting help when programming, etc.

[Neutrion]

The crazy starts around 10 minute mark

Interesting!!
But anyway this was anticipated years before.  Everybody who thought about it knew that sooner or later these things will start to happen. And still basically we did not do anything to prepare for it in the last years.
(But having instead interesting debates about what is a woman and what is man, and how should we call them.)
So the basic approach was: But just experiment with it, its so interesting, until we can't see destruction, we can not be sure that it will be so bad.

I actually started to wonder what will come first. The mass panic when people start to realize what this is, or the job and income losses on a mass scale affecting the not so top level intellectual jobs, and the consequences of that.

I also watched this:

To watch a smart guys talking about positive solutions.
Well this is some tech optimistic Stanford professor and still. 
His idea: Well we have to teach children to think creatively, to work with the AI. 
And than? When the AI will become more creative what happens  than?

Next step: Try to teach the childrens properly to pray to the AI to not to be eradicated, and to get some extra scores and food...

[Neutrion]

AI, no matter how powerfull, will not by itself, constitue a threat to human kind. It is the company that owns the AI that might be a threat.

How do you know that? Why it can not be a threat by itself?
But even only if the company would be a threat, do we have anything now to mitigate it?

And just because we have already some problems we shuld willingly create more?

In a modern society, we are full monitored (digital money, toll boxes, smartphones, digital receipts, etc.). If you as an idividual start to worry goverment members, they can easily eliminate your credibility based on the huge data that exists.

So, while I think that AI, GPT, etc., on their own are just a tool that is very helpful, it can be used to further control human kind.

You can check it for yourself: ChatGPT is biased (though responses change over time). While it refuses to produce text for some political ideologies, it happily replies on other political ideologies. That is on purpose.

But I do see the benefit of quickly learning something or getting help when programming, etc.

In a modern society...: Well go and live in China for a while to check what happens if these things get into the wrong hands.  The full monitoring is something you don't give a damn about until the shit hits the fan.
One reason more to learn history.

If AI develops in a pace as it seems, you soon won't have any reason to learn anything. Especially not programming.
But to have a full scale disaster it is enough if third of the people are losing their jobs for 2 years. Or one.

[xrunner]

If AI develops in a pace as it seems, you soon won't have any reason to learn anything. Especially not programming.
But to have a full scale disaster it is enough if third of the people are losing their jobs for 2 years. Or one.

What do you say to a kid who asks "Why do I need to learn anything when I can ask ChatGPT (or whatever it's called eventually)? I'll always have a smart phone or my friend will, so why do I need to take history, math, biology, or anything else when I have it all right here?" Think about newborn babies, when they get older there will have never been a time (to them) where this technology didn't exist. They will be different human beings in certain ways than any of us. What will they do different than we did when they grow up?

We're now at GPT 4. I watched a good video on it last night (linked below), and it's impressive. As far as jobs, it's going to hit certain classes of jobs hard. Any kind of receptionist job can easily be replaced by GPT 4. You would use voice to text and text to voice capability. You can give it a "system message" and steer it to what it's supposed to be acting like - such as "receptionist". You give it access to your company information as needed. This could be phone numbers, help with equipment manuals, hours of operation, anything else people can ask about on the phone, and it will handle it 24/7/365 with no complaints, no sick days, no calling in sick.

Why would you hire an unreliable human when you can eliminate them? You don't even need to outsource this to other countries anymore. So even as impressive as GPT 4 is, what is to come in the next months and years? It's going to radically change our world so get ready to get shaken.

[Bicurico]

Call center and front desk receptionist will indeed be replaced by AI.
So what? Disruptive technologies have done this to many trades in the past. Anyone remembers Kodak and analogue photography?

If you are not willing to allow this, then you cannot allow technological progress.

While many jobs will be replaced, though, new ones will be created. And the loss might be compensated by the population decrease in western society.

Why should students learn? For the same reason you learn math when you could just use your CAS enabled calculator. Or why you learn how to read and write, when you could just as well use text reader or autocorrect. Why learn foreign languages if you can use Deepl?

For me these are tools and you need to learn how to use them. Like FEA software. Anyone can run a simulation, but it needs a trained engineer to interpret the results, to make sure the input and output are correct and what to do based on the results.

If you don't know how to program, you won't build any application using ChatGPT. But if you are learning how to code, ChatGPT will be a great tutor. And if you already know how to code, it will help you find a solution when you are stuck or lazy. It works great for that.

Again, the danger is with who owns the AI. Imagine a government uses AI to automatically monitor and interpret what people are doing. What if you are flagged due to an error? How can you clear your name?

Your comparison with China is a good example: it is not the technology that is bad. It is the abuse you can use it for.

On the other hand, could an AI run government be finally without corruption and lobbying? Wouldn't that be a better government, taking decisions to the actual benefit of the people, planning in long terms and not being worried about reelection?

Could we have different AI motors running in parallel? We could elect the one we like the most.

I am really unable to further discuss this, as we cannot really anticipate what the future of AI will bring.

What really worries me is that the world has private companies that have more funds than whole countries.

Regards,
Vitor