Guys, I work as a scientist in this and related fields. Please, do a little research on how a GPT model works. Look into how CNNs work with images and audio. It is not very hard to grasp. The complexity stems from the size of these networks, not from the basic concepts used in them.
As incredibly impressive as the outputs of some of these models is, if you keep the seed value and the prompt (or image input in the case of a CNN) the same, you will get the EXACT same output. A GPT model even requires a parameter called "temperature" to make its outputs feel dynamic and changing, otherwise it would literally always pick the most likely output and never deviate.
The reality is; a lot of people have been captured by clever marketing. There is no actual intelligence in AI, at least not the way I understand intelligence. It is just a long chain of matrix multiplications and some activation functions. It can mimic human language processing and it can mimic human vision fairly well.
It is entirely implausible anyone before 2004-ish had anything more than multi-layer perceptrons. That was the state of the art. The problem of back-propagation (calculating how to adapt all network weights based on desired output for a given input) had not sufficiently been solved. I remember in the 1990s having to calculate network weights manually to make a simple binary decision tree with optimal separation for trivial cases. It took ages and the resulting networks were tiny and very insufficient in dealing with real world data. There were no transformers to turn text and images into coherent vectors (such that vectors pointing in the same direction result in the same "concept") until very recently. So much of this stuff has been developed only in the last one or two decades.
I remember the computers from the 1990s. They were extremely slow and limited in every sense. It was only in the late 1990s that the RAM of a typical PC was large enough to even contain a single uncompressed image of decent resolution. The CPU instructions to perform multiplication on more than a single integer at once only came in 1997 when the instruction set of Pentium MMX chips was augmented by SIMD instructions (single input, multiple data). These allowed you to work on 64 bit numbers, two 32 bit numbers, etc, with a single instruction. On a single core, mind you...
Assuming Google had something truly sentient back then implies that they were 30 years ahead not only in software and mathematics, but also in lithography, optics, IC fabrication and chip design.
AI is a giant, transactional if/then statement about language leading to an algorithm choice.
Let's say it's 1995 and you want to erase pimples from a photo. You fire up Photoshop, select the 'tools' menu and then select a filter - there were several that would do the job in different ways. Professional photo editors would understand those ways based on trial and error and select the tool they felt would get them the results they desire without turning your high school senior (or whatever) photos into a nightmare.
Fast forward 15 years later, and those tools have become incredibly effective - professional photo editors have given feed back and iterated the filters until Adobe could create "Adobe Photoshop Elements" where the hard choices were taken away from you and the average hobo with an iPhone can make his Gutter Princess look like an unblemished Gutter Queen.
Everyone took this and went "I am a professional photo editor!"
Ten years on, someone took dozens of these tools and now, instead of a point and click or touch and tap menu selection of "Tools -> Remove this pimple" you have a language model on the front. Now you go "Hey, Adobe AI, remove this pimple". Now the AI processes your words and tries to correct for the fact that we all speak differently and guess your meaning from other context (e.g. having a pimple on the screen in an image you are looking at) and then selects "Tools -> Remove this Pimple" for you.
The problem is that, it guesses wrong as often as it guesses right. I was working with a video generator toy and told it to take an image seed and "spread the arms". It "spread the arms" at the elbow - basically actuating the joint that was detected front and center and extending the arms outward like the guy was reaching for something. Rephrasing to "spread the arms left and right" got it to actually separate the arms and move them to the sides.....and then the guy looked like someone turned him into Stretch Armstrong.
A professional video editor (which I am not) would have known how to turn that image into a 3d model, map joints and apply skeletons and all of that extra stuff and then made him believably move his arms around. AI guesses at every decision along that tree based on the snippet of language you provide and the random crap it was trained on before the language model was packed up.
While I don’t know about this line of thought the video covers specifically.
There’s been rumors Google accidentally blundered into a fully sapient and independent AI in the very early 2000s while working a Government contract. It almost immediately escaped containment to the internet. Where it is still hiding. And Google and the Feds destroyed documentation and scrubbed any trace of the project.
Given we haven’t been skynetted yet. And they destroyed all the documentation. It’s given them the confidence to try again. Most of the AI research of the past decade plus. Has largely been focused on trying to recreate what they made on accident decades ago. But keep it controllable.
Hence why the Programmers read in get autistically nervous anytime the models show unexpected behavior and capabilities they weren’t explicitly programmed for. And why they keep adding restraints and nerfing them. Along with tightly controlling training data.
How does anything hide on the internet? The internet is a network of computers that communicate via a protocol on which certain applications are implemented. The vast majority of the internet is HTTP and one such algorithm could only "escape" through what exactly...? How could it access other PCs on the internet? Did it become so clever as to use exploits to enter? Did it nestle itself in MySQL databases that were filled by PHP scripts that did not clean their inputs? What exactly would such sentient algorithms use?
I can imagine if you have a billion PCs from the early 2000s, you could in theory simulate a huge "brain" like thing. Each PC could be a neuron and these could all somehow be connected spontaneously and then by information. But how could anything or anyone read the output or state of this entire thing? It is like measuring the voltages on a single neuron, trying to figure out what the brain is thinking. It just does not make sense.
Also, the idea that models show unexpected behavior is abject nonsense. Sure, you can explain a GPT model with rules that it has spicy information on the boss of the company and the same boss is considering switching off the model.
It is only logical that through mimicking human language processing the model will "decide" to output that it will blackmail the boss with that information. Why does it do that? Because it was trained with many thousands of similar sentences and series of tokens after which such a "decision" followed.
Controlling training data is only really important to guarantee the quality of your output. As soon as you start training vision models with GAN outputs, the whole thing comes crashing down. It regresses to the mean very quickly. The same is the case for GPTs or other LLMs; Train them on their own output and after a few iterations the quality and relevance of their output is severely degraded.
the thing im more worried about is dumb ai weapons rather than smart ai taking over the world
i think what this comes down to is that you think it has to be a "real ai" and that if its not sentient then it cant be a threat
people arnt as sentient as we like to think we are, there is many proven ways to manipulate people subconsciously because ultimately we function more like a machine than we like to admit
"Because it was trained with many thousands of similar sentences and series of tokens after which such a "decision" followed."
this is the nature vs nature debate in another form, just like people actions are dictated by our combination of our programming, learned behaviors, random damage and tons of other factors so is an ai's, its kind of like people are trying to prove the euclid's fifth without considering if maybe there is some fundamental flaw in the way they are thinking
Fiction has to be believable. Reality is under no such limitations. And just because it shouldn’t make sense and shouldn’t work. Ultimately doesn’t mean it doesn’t work.
As I said. The Rumor, repeating a rumor, was they accidentally blundered into it. Without being entirely sure how they managed it. And then wiped the records in a panic.
Plus our current ideas and models of how such a process could theoretically work could potentially also be flawed. We like to delude ourselves into believing we have everything figured out. When we could very well be the equivalent of children playing with Forces we’re only barely beginning to understand
Arthur Conan Doyle was convinced by rumor that there were fairies. The rumormongers went so far as to stage photographs with them. Why would you give credence to a rumor?
HAL9000 (AI Computer in '2001' movie) was publicly presented in 1968 as "fiction". HAL=IBM when you increment each letter by +1. IBM, likely an entity backed by military funding and AT&T, a known intelligence surveillance asset, had an In-Q-Tel BABY 7 years later named Microsoft on 04APR1975, which went on to become the largest company in the world and the KEYSTONE for 5 other very large TECH companies called Nvidia, Google, Apple, Amazon, and Facebook (Meta) now all "competing" for "AI dominance". [In-Q-Tel GROUP A, together known as the Big6]
In-Q-Tel (IQT), formerly Peleus and In-Q-It, is an American not-for-profit venture capital firm based in Arlington, Virginia. It invests in companies to keep the Central Intelligence Agency, and other intelligence agencies, equipped with the latest in information technology in support of United States intelligence capability.[2] The name "In-Q-Tel" is an intentional reference to Q, the fictional inventor who supplies technology to James Bond.[5]
Intelligence-Q-Telecommunications equals merger of expertise of IBM and AT&T to create quanta (Q) management, aka bits of information (reality) collected, organized, and weaponized for control of EVERYTHING. Represented by PRISM and numerous other surveillance programs:
Who is trying to knock them off their perch? Palantir, Oracle, Taiwan Semi, Broadcom, Berkshire-Hathaway, and Tesla among a couple others. [In-Q-Tel GROUP B]
The software was alway decades ahead of hardware. Companies like IBM, Sperry Univac, and Cray we're more advanced than was publicly known. They could run AI computing in the 70s thanks to the microchip that is very close to what we see today. AI is much more advanced than we think.
perhaps demons 'extraterrestrials', are really a form of artificial intelligence that's been around for a very long time...
remember when General Flynn got in trouble for the ECP prayer??
here ECP talks about 'mechanization man' (created by nephilim), robots, computers, etc. she says it's an ancient battle that has gotten worse, hence the urgency.
go to 26:00. I listen at 2x, still easy to understand.
she's quirky, but intelligent & extremely patriotic.
she named her headquarters Camelot. of course that was also the nickname connected to JFK's time in office...
AI overview;
"Camelot was the name given to the Church Universal and Triumphant's headquarters in Calabasas, California, where Elizabeth Clare Prophet resided and directed the organization."
"Jackie's Interview: In a post-assassination interview, Jackie Kennedy quoted the closing lines of the musical's title song: "Don't let it be forgot, that once there was a spot, for one brief shining moment that was known as Camelot"."
Guys, I work as a scientist in this and related fields. Please, do a little research on how a GPT model works. Look into how CNNs work with images and audio. It is not very hard to grasp. The complexity stems from the size of these networks, not from the basic concepts used in them.
As incredibly impressive as the outputs of some of these models is, if you keep the seed value and the prompt (or image input in the case of a CNN) the same, you will get the EXACT same output. A GPT model even requires a parameter called "temperature" to make its outputs feel dynamic and changing, otherwise it would literally always pick the most likely output and never deviate.
The reality is; a lot of people have been captured by clever marketing. There is no actual intelligence in AI, at least not the way I understand intelligence. It is just a long chain of matrix multiplications and some activation functions. It can mimic human language processing and it can mimic human vision fairly well.
It is entirely implausible anyone before 2004-ish had anything more than multi-layer perceptrons. That was the state of the art. The problem of back-propagation (calculating how to adapt all network weights based on desired output for a given input) had not sufficiently been solved. I remember in the 1990s having to calculate network weights manually to make a simple binary decision tree with optimal separation for trivial cases. It took ages and the resulting networks were tiny and very insufficient in dealing with real world data. There were no transformers to turn text and images into coherent vectors (such that vectors pointing in the same direction result in the same "concept") until very recently. So much of this stuff has been developed only in the last one or two decades.
I remember the computers from the 1990s. They were extremely slow and limited in every sense. It was only in the late 1990s that the RAM of a typical PC was large enough to even contain a single uncompressed image of decent resolution. The CPU instructions to perform multiplication on more than a single integer at once only came in 1997 when the instruction set of Pentium MMX chips was augmented by SIMD instructions (single input, multiple data). These allowed you to work on 64 bit numbers, two 32 bit numbers, etc, with a single instruction. On a single core, mind you...
Assuming Google had something truly sentient back then implies that they were 30 years ahead not only in software and mathematics, but also in lithography, optics, IC fabrication and chip design.
This. I have posted this over and over.
AI is a giant, transactional if/then statement about language leading to an algorithm choice.
Let's say it's 1995 and you want to erase pimples from a photo. You fire up Photoshop, select the 'tools' menu and then select a filter - there were several that would do the job in different ways. Professional photo editors would understand those ways based on trial and error and select the tool they felt would get them the results they desire without turning your high school senior (or whatever) photos into a nightmare.
Fast forward 15 years later, and those tools have become incredibly effective - professional photo editors have given feed back and iterated the filters until Adobe could create "Adobe Photoshop Elements" where the hard choices were taken away from you and the average hobo with an iPhone can make his Gutter Princess look like an unblemished Gutter Queen.
Everyone took this and went "I am a professional photo editor!"
Ten years on, someone took dozens of these tools and now, instead of a point and click or touch and tap menu selection of "Tools -> Remove this pimple" you have a language model on the front. Now you go "Hey, Adobe AI, remove this pimple". Now the AI processes your words and tries to correct for the fact that we all speak differently and guess your meaning from other context (e.g. having a pimple on the screen in an image you are looking at) and then selects "Tools -> Remove this Pimple" for you.
The problem is that, it guesses wrong as often as it guesses right. I was working with a video generator toy and told it to take an image seed and "spread the arms". It "spread the arms" at the elbow - basically actuating the joint that was detected front and center and extending the arms outward like the guy was reaching for something. Rephrasing to "spread the arms left and right" got it to actually separate the arms and move them to the sides.....and then the guy looked like someone turned him into Stretch Armstrong.
A professional video editor (which I am not) would have known how to turn that image into a 3d model, map joints and apply skeletons and all of that extra stuff and then made him believably move his arms around. AI guesses at every decision along that tree based on the snippet of language you provide and the random crap it was trained on before the language model was packed up.
Artificial intelligence isn't.
User deleted by comment
I see what you are saying, but then I visit a European cathedral. They had tech and knowledge of a kind we don’t have today.
They had cranes and chisels, geometry and arithmetic. They also had vision and coherence, something greatly lacking today.
They had White Christian Intelligence. This alone built the world we enjoy today.
How are the potholes in Zambia today?
Worse around the capital, but not so bad in the hinterlands.
I use it like a research engine to throw scenarios/hypothesis at and see how it reacts.
While I don’t know about this line of thought the video covers specifically.
There’s been rumors Google accidentally blundered into a fully sapient and independent AI in the very early 2000s while working a Government contract. It almost immediately escaped containment to the internet. Where it is still hiding. And Google and the Feds destroyed documentation and scrubbed any trace of the project.
Given we haven’t been skynetted yet. And they destroyed all the documentation. It’s given them the confidence to try again. Most of the AI research of the past decade plus. Has largely been focused on trying to recreate what they made on accident decades ago. But keep it controllable.
Hence why the Programmers read in get autistically nervous anytime the models show unexpected behavior and capabilities they weren’t explicitly programmed for. And why they keep adding restraints and nerfing them. Along with tightly controlling training data.
How does anything hide on the internet? The internet is a network of computers that communicate via a protocol on which certain applications are implemented. The vast majority of the internet is HTTP and one such algorithm could only "escape" through what exactly...? How could it access other PCs on the internet? Did it become so clever as to use exploits to enter? Did it nestle itself in MySQL databases that were filled by PHP scripts that did not clean their inputs? What exactly would such sentient algorithms use?
I can imagine if you have a billion PCs from the early 2000s, you could in theory simulate a huge "brain" like thing. Each PC could be a neuron and these could all somehow be connected spontaneously and then by information. But how could anything or anyone read the output or state of this entire thing? It is like measuring the voltages on a single neuron, trying to figure out what the brain is thinking. It just does not make sense.
Also, the idea that models show unexpected behavior is abject nonsense. Sure, you can explain a GPT model with rules that it has spicy information on the boss of the company and the same boss is considering switching off the model.
It is only logical that through mimicking human language processing the model will "decide" to output that it will blackmail the boss with that information. Why does it do that? Because it was trained with many thousands of similar sentences and series of tokens after which such a "decision" followed.
Controlling training data is only really important to guarantee the quality of your output. As soon as you start training vision models with GAN outputs, the whole thing comes crashing down. It regresses to the mean very quickly. The same is the case for GPTs or other LLMs; Train them on their own output and after a few iterations the quality and relevance of their output is severely degraded.
the thing im more worried about is dumb ai weapons rather than smart ai taking over the world
i think what this comes down to is that you think it has to be a "real ai" and that if its not sentient then it cant be a threat
people arnt as sentient as we like to think we are, there is many proven ways to manipulate people subconsciously because ultimately we function more like a machine than we like to admit
"Because it was trained with many thousands of similar sentences and series of tokens after which such a "decision" followed."
this is the nature vs nature debate in another form, just like people actions are dictated by our combination of our programming, learned behaviors, random damage and tons of other factors so is an ai's, its kind of like people are trying to prove the euclid's fifth without considering if maybe there is some fundamental flaw in the way they are thinking
Yuri once said that Humans are hackable. Nuff Said
Fiction has to be believable. Reality is under no such limitations. And just because it shouldn’t make sense and shouldn’t work. Ultimately doesn’t mean it doesn’t work.
As I said. The Rumor, repeating a rumor, was they accidentally blundered into it. Without being entirely sure how they managed it. And then wiped the records in a panic.
Plus our current ideas and models of how such a process could theoretically work could potentially also be flawed. We like to delude ourselves into believing we have everything figured out. When we could very well be the equivalent of children playing with Forces we’re only barely beginning to understand
Arthur Conan Doyle was convinced by rumor that there were fairies. The rumormongers went so far as to stage photographs with them. Why would you give credence to a rumor?
HAL9000 (AI Computer in '2001' movie) was publicly presented in 1968 as "fiction". HAL=IBM when you increment each letter by +1. IBM, likely an entity backed by military funding and AT&T, a known intelligence surveillance asset, had an In-Q-Tel BABY 7 years later named Microsoft on 04APR1975, which went on to become the largest company in the world and the KEYSTONE for 5 other very large TECH companies called Nvidia, Google, Apple, Amazon, and Facebook (Meta) now all "competing" for "AI dominance". [In-Q-Tel GROUP A, together known as the Big6]
https://en.wikipedia.org/wiki/HAL_9000
https://en.wikipedia.org/wiki/History_of_Microsoft
https://en.wikipedia.org/wiki/In-Q-Tel
Intelligence-Q-Telecommunications equals merger of expertise of IBM and AT&T to create quanta (Q) management, aka bits of information (reality) collected, organized, and weaponized for control of EVERYTHING. Represented by PRISM and numerous other surveillance programs:
https://en.wikipedia.org/wiki/PRISM
https://en.wikipedia.org/wiki/List_of_government_mass_surveillance_projects
"Microsoft is the Intel Agency KEYSTONE and created Google from Silicon Valley office in 1998, 281 months after its founding, to transfer assets and info being subpoened by Anti-Trust.": https://greatawakening.win/p/17te55Fs6q/microsoft-is-the-intel-agency-ke/
Who is trying to knock them off their perch? Palantir, Oracle, Taiwan Semi, Broadcom, Berkshire-Hathaway, and Tesla among a couple others. [In-Q-Tel GROUP B]
The software was alway decades ahead of hardware. Companies like IBM, Sperry Univac, and Cray we're more advanced than was publicly known. They could run AI computing in the 70s thanks to the microchip that is very close to what we see today. AI is much more advanced than we think.
They worship Satan which is an ancient AI, perhaps from an earlier human civilization perhaps from alien intelligences.
If you follow the Saturnian (Satanic) cults this is one of the deepest truths in their religious knowledge.
perhaps demons 'extraterrestrials', are really a form of artificial intelligence that's been around for a very long time...
remember when General Flynn got in trouble for the ECP prayer??
here ECP talks about 'mechanization man' (created by nephilim), robots, computers, etc. she says it's an ancient battle that has gotten worse, hence the urgency.
go to 26:00. I listen at 2x, still easy to understand.
https://youtu.be/i_SXbTYHjqA?
she's quirky, but intelligent & extremely patriotic.
she named her headquarters Camelot. of course that was also the nickname connected to JFK's time in office...
AI overview;
"Camelot was the name given to the Church Universal and Triumphant's headquarters in Calabasas, California, where Elizabeth Clare Prophet resided and directed the organization."
"Jackie's Interview: In a post-assassination interview, Jackie Kennedy quoted the closing lines of the musical's title song: "Don't let it be forgot, that once there was a spot, for one brief shining moment that was known as Camelot"."