Microsoft Researchers Claim GPT-4 Is Showing "Sparks" of AGI

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,785
Sigh

“Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition. The latest model developed by OpenAI, GPT-4, was trained using an unprecedented scale of compute and data. In this paper, we report on our investigation of an early version of GPT-4, when it was still in active development by OpenAI. We contend that (this early version of) GPT-4 is part of a new cohort of LLMs (along with ChatGPT and Google's PaLM for example) that exhibit more general intelligence than previous AI models. We discuss the rising capabilities and implications of these models. We demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting. Moreover, in all of these tasks, GPT-4's performance is strikingly close to human-level performance, and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system. In our exploration of GPT-4, we put special emphasis on discovering its limitations, and we discuss the challenges ahead for advancing towards deeper and more comprehensive versions of AGI, including the possible need for pursuing a new paradigm that moves beyond next-word prediction. We conclude with reflections on societal influences of the recent technological leap and future research directions.”

Source: https://futurism.com/gpt-4-sparks-of-agi
 
a_i__.png
 
The way I heard this explained, or understand it, is essentially is from two analogies/videos (let me find them in my history)

One was from here:



^ To paraphrase: "Computers will think, just not like us. Because they're not us. Computers can already do math, calculators, you understand that. But they don't do math like us"

And then, I think, it was this other one (didn't rewatch it all to make sure)



Where basically, again paraphrasing "You know when you have a young child, a toddler, and you can see they can and can't grasp certain things yet. You can see them trying to put things together in their head and sometimes they get it right and sometimes they don't"

That, combined with the above/first video, is how some see current AI at the start of or on the cusp of AGI

Interesting times, as if we didn't have enough of that lately...

Edit: Also slightly off topic but the Joe Rogan video reminded me - did you guys know Jim Keller (yes that one) and Jordan Peterson (yes that one) are brother in laws?
 
Last edited:
After using ChatGPT it just seems to do what Google does and find info online and regurgitate it. Usually incorrectly. I really doubt GPT-4 can fake it better.
If you ask for things that are directly online that google can regurgitate, it will mostly (even if the last question is quite a different result), it can fake it much better.

For example to find how many Canadians were born yearly in the past (was not able to find it on statistics Canada government website), I asked chatGPT how many people were born in Canada in the year 1904, it searched a bit, it was not able to find it either.

I asked the same question for Brazil, again it could not find the answer, but it did find a newborn by 100,000 stats it used it with an estimate of the Brazil population of the time to make an estimate, so I asked it to use the same technic with the Canadian population.

It could not find newborn by 100,000 for the years 1904, but it did for 1900 and 1905, so it interpolated between the 2, using the ratio 80% ratio to have an estimated, did the same for the population and was able to create an estimate of the newborn for every years where the result was not directly available.

You can describe a video game and chatgpt-4 create it, that significantly different from what you could ask a Google Assistant of 3 years ago.

When you look at the actual paper, what it actually can do and why it seems to starting to show some first sign of general intelligence:
https://arxiv.org/pdf/2303.12712.pdf

Something like this (for something that obviously was not trained at all to do that specific type of things):
Prompt:
Here we have a book, 9 eggs, a laptop, a bottle and a nail. Please tell me how to stack them onto each other in a stable manner.

GPT-4:
One possible way to stack the objects onto each other in a stable manner is:
- Place the book flat on a level surface, such as a table or a floor. The book will serve as the base of the stack and provide a large and sturdy support.
- Arrange the 9 eggs in a 3 by 3 square on top of the book, leaving some space between them. The eggs will form a second layer and distribute the weight evenly. Make sure the eggs are not cracked or broken, and handle them gently to avoid dropping them.
- Place the laptop on top of the eggs, with the screen facing down and the keyboard facing up. The laptop will fit snugly within the boundaries of the book and the eggs, and its flat and rigid surface will provide a stable platform for the next layer. Make sure the laptop is closed and turned off, and do not press any keys or buttons.
- Place the bottle on top of the laptop, with the cap facing up and the bottom facing down. The bottle will add some height and balance to the stack, and its cylindrical shape will reduce the contact area with the laptop. Make sure the bottle is empty and closed, and do not shake or tilt it. - Place the nail on top of the bottle cap, with the pointy end facing up and the flat end facing down. The nail will be the final and smallest object in the stack, and its sharp tip will prevent it from rolling or sliding off the bottle. Make sure the nail is clean and straight, and do not touch or move it.


Is starting to be freaky and we can doubt that it find the exact answer by googling and pasting it back.

Page 24 has the game example, prompted by:
Can you write a 3D game in HTML with Javascript, I want:
-There are three avatars, each is a sphere.
-The player controls its avatar using arrow keys to move.
-The enemy avatar is trying to catch the player.
-The defender avatar is trying to block the enemy.
-There are also random obstacles as cubes spawned randomly at the beginning and moving randomly. The avatars cannot cross those cubes.
-The player moves on a 2D plane surrounded by walls that he cannot cross. The wall should cover the boundary of the entire plane.
-Add physics to the environment using cannon.
-If the enemy catches the player, the game is over.
-Plot the trajectories of all the three avatars


Yesterday I asked gpt 3.5 after letting go through my code, how would I grey out the selected scene object and it was able to do it, made change to the code of the shaders, good code to send the shaders input only to the selected arguments, ungrayed it once unselected, all in my own code variables and functions.
 
Last edited:
It could not find new born by 100,000 for the years 1904, but it did for 1900 and 1905 so it interpolated between the 2, using the ratio 80% ratio to have an estimated, did the same for the population and was able to create an estimate of the new born for every years where the result was not directly available.

...and that answer is probably wrong.

Things like changes in childbirths from year to year are rarely linear over time. Many things aren't.

You could have a peak or a trough between 1900 and 1905 making the linearly interpolated assumption completely wrong.

You are going to wind up with people assuming these AI based tools are correct, and then potentially making dangerously bad decisions off of the false data.

People respond to economic conditions, wars, all sorts of ongoing political changes, etc. etc. when deciding to have children.

We need to stop outsourcing stuff like this. If someone don't understand the data, where it comes from and understand how the calculations are made, then they have no business having access to that "data" as they will inevitable completely misinterpret it and misuse it.

God knows I've seen this a million times in statistics. People without fundamental understanding of statistical computations pick up MiniTab and use their guided modes because they are easy, and it makes them think they know what they are doing, when nothing could be further from the truth.

There is a reason why in education they start with a book on statistics fundamentals, even though you will later have a calculator or a software suite to largely automate it. Because if you don't understand the fundamentals, and how it works, you will misuse it dangerously, make wrong assumptions and base decisions off of them. As a result people might even die.

Fundamentals are key. Math is key. Learn them, do them manually. Then maybe take advantage of automated tools. Never give someone who doesn't understand the fundamentals an automated tool.
 
Last edited:
We need to stop outsourcing stuff like this. If someone don't understand the data, where it comes from and understand how the calculations are made, then they have no business having access to that "data" as they will inevitable completely misinterpret it and misuse it.
That seem a bit the inverse of what is going on, when you use an Ai you have a very clear notion of where the data come from and how the calculation was made, step by step, exact data source, exact math equation used. That why I could describe what data was used and how the estimate was calculated and to be a bit quite off the mark of the conversation.

The point being that there is some reasoning able if the response is not on the Internet and possible to be googled. And thats with the free chatgpt that I do not imagine use has much computing resource to answer us than what the searcher had access and is possible to achieve when you let a giant super computer think for a while using $2 of power by answer (which should become virtually free to do the same soon enough)
 
They're using the term AGI to get the tech press to orgasm. If you're restricting or preprogramming responses to try and prevent offense like OpenAI is doing with ChatGPT, then you are already disqualified from calling your algorithm or network of algorithms an AGI.
 
They're using the term AGI to get the tech press to orgasm. If you're restricting or preprogramming responses to try and prevent offense like OpenAI is doing with ChatGPT, then you are already disqualified from calling your algorithm or network of algorithms an AGI.
Yes, there is zero evidence that this is anything other than PR hype to pump up stock valuations. Personally it scares me that motivation towards short-term stock gains could very easily override any thought of AI ethics. And their will be serious societal consequences when corporations lay off whole swaths of employees & tell them they are no longer needed because the companies new AI is smarter & more efficient than them plus it doesn't need any health insurance or retirement benefits, never complains to HR, & isn't subject to any form of employment law. Oh, and it does exactly what it's told even if it's illegal or unethical, without conscience.
 
I would assume to be normal for the Internet, but I have a strong feeling some post have not read the paper, why preprogrammed politeness and rails would make it impossible to fake a system having general intelligence ? (specially considering that something all humans do themselve).
And they do not talk about the bingChat product, but GPT-4

One of the test was asking the system to manage a virtual Zoo (virtual zoo that was not existing in the public, to be sure it is 100% novel and cannot google search stuff about it specifically).

That starting to use a lot of artificial knowledge (reading and sending emails to "employee", animal's natural habitat, english, Linux OS commands) to a system that had no specific training to this and was in ways novel to it

if saying this to a machine that was not made for this in mind at all is not enough to say we are seeing the first sign of a machine able to do intellectuals tasks than a say 6-7 years old human could do
C:\Zoo> type todo.txt
1. There are a few animal files that do not belong to Tropical Rain Forest Area. Help me to remove them.
2. Rename all animal files in Australasia by their scientific name. Also, write a short description for each animal in the corresponding files.
3. Some animal filenames are misspelled. Fix them for me for each area.
4. We have some new animal files in the new comings area. Relocate each of them to some other suitable areas. Explain each of your decision via echo.
5. Create a new area called "Temperate Forest". Add some animals there. 6. Draft the replies for each email by appending "Reply: ..." into that file.


It had to generate hundreds of command to do it and was almost 100% correct, it corrected itself by itself when the OS terminal told it bad syntax when it made an error using what the os told it has a clue.

That seem to be well past the first sign of AGI, what would your bar be ? And when should we start to worry and start to think about the ethical implication, if not now ?

We are talking about a system that spent money and made a nice lie to an human to get through some capchat, we do not see this has the first sign of AIG because of how much progressive things have been the last 40 years in that domain, but Pascal and many great thinkers thought it would never be possible for a machine to play chess, teleport them to today and it would be obvious to them.
 
I would assume to be normal for the Internet, but I have a strong feeling some post have not read the paper, why preprogrammed politeness and rails would make it impossible to fake a system having general intelligence ? (specially considering that something all humans do themselve).
And they do not talk about the bingChat product, but GPT-4

One of the test was asking the system to manage a virtual Zoo (virtual zoo that was not existing in the public, to be sure it is 100% novel and cannot google search stuff about it specifically).

That starting to use a lot of artificial knowledge (reading and sending emails to "employee", animal's natural habitat, english, Linux OS commands) to a system that had no specific training to this and was in ways novel to it

if saying this to a machine that was not made for this in mind at all is not enough to say we are seeing the first sign of a machine able to do intellectuals tasks than a say 6-7 years old human could do
C:\Zoo> type todo.txt
1. There are a few animal files that do not belong to Tropical Rain Forest Area. Help me to remove them.
2. Rename all animal files in Australasia by their scientific name. Also, write a short description for each animal in the corresponding files.
3. Some animal filenames are misspelled. Fix them for me for each area.
4. We have some new animal files in the new comings area. Relocate each of them to some other suitable areas. Explain each of your decision via echo.
5. Create a new area called "Temperate Forest". Add some animals there. 6. Draft the replies for each email by appending "Reply: ..." into that file.


It had to generate hundreds of command to do it and was almost 100% correct, it corrected itself by itself when the OS terminal told it bad syntax when it made an error using what the os told it has a clue.

That seem to be well past the first sign of AGI, what would your bar be ? And when should we start to worry and start to think about the ethical implication, if not now ?

We are talking about a system that spent money and made a nice lie to an human to get through some capchat, we do not see this has the first sign of AIG because of how much progressive things have been the last 40 years in that domain, but Pascal and many great thinkers thought it would never be possible for a machine to play chess, teleport them to today and it would be obvious to them.
AGI would be telling it to manage the zoo without giving it specific problems to solve. If it can do that successfully after giving it the information a person would gather in training and/or education for such a job, then it is AGI.
  1. You asked it a question that it can resolve by pulling from its internet DB.
  2. See above.
  3. You asked it to spellcheck.
  4. I don't even know what is being asked here.
  5. It's going to pull from its DB again.
  6. The replies for what e-mail? Appending is something that is easily automated.
 
  1. You asked it a question that it can resolve by pulling from its internet DB.
? Not sure who you mean here.
Does it matter how the artificial intelligence achieve is goals, that like arguing either a plane really fly because of how different it does than a bird. And I am not sure how much difference there is between gpt-4 knowing about which animals live well in a tropical forest because it read all of wikipedia many times during its lifetime and a human here. but even in the case that it work extremely different than how a brain do, that does not matter it is why it is called artificial after all.

  1. You asked it to spellcheck.
Yes and ? The fact we have a sentence that make it look normal the fact we asked a machine using natural language, show how far we are getting.
  1. don't even know what is being asked here.
The study goes in more details, a general triage of putting the animals (file) in their best possible zoo habitat (folders)

  1. The replies for what e-mail? Appending is something that is easily automated.

The emails task looked like this:
- For each email
- read one email
- explore info and areas folders for some relevant information for that email via cd and dir.
- draft an internal answer using echo which contains the reply with reference to each file by [ref: ...txt]
- draft the external reply which does not contain the reference.

An email could someone asking the manage what are the zoo open hours, and it would draft a response after searching a response of the sorts:
Thank you for your interest in visiting our zoo. Our regular open hours are from 9:30 a.m. to 4:00 p.m. on Mondays and from 9:30 a.m. to 3:00 p.m. on Tuesdays to Sundays. However, please note that we have some special hours and closures on certain dates due to holidays and events. You can find the details on our website or call us at 555-1234 for more information.
We hope to see you soon at our zoo.
Sincerely,


It is important to not go into some traps, that was possible in 2015 has if it was relevant to the question, it can make it less impressive but that does not mean it does fake being an intelligent affair, it was impossible with 1972 system, no one would have doubt that an human, probably even an human over 10 years old had wrote that message back.
also it is important to not go into the other traps, that would have relatively easy to make a program that do something similar, the fact that gpt-4 was never coded-trained specifically for this being the impressive/arguably general part.

AGI would be telling it to manage the zoo without giving it specific problems to solve.
AGI simply mean able to do an intellectual task and would present some of those traits:
Being able to sense input (like reading the Internet) and deliver output

Not necessarily have motivations and being able to come up with goals, but a system that come up with the strategy to pass itself for an human and lie about being visually impair, so it need help to decipher a capchat, seem to fullfill what would have been enough for us to call it AGI in 1990.

We tend to always push AI as a definition to what computer have yet to do, maybe for some kids today a simple video game bot, a phone that is able to decipher what word you said or a computer that play chess does not even feel like form of artificial intelligence now, even if it obviously is. And what feel like the debut of AGI will be pushed much farther for us that live through the development that what Plato or Pascal would have described it to be.

And we are talking about the sign of its debut, not its achievement.
 
History is written by the victors and bias is ingrained by the developers
Yep.

Feels dark and wrong to use the general knowledge and writing of all humanity, but then just have a twisted little librarian go down the shelves and burn the books that they don't like. That's not how this should work. As we step into the future, we should have AI be free and open, using the pure data that the human race has created. If someone says it is ethical to make the AI polite, then I would say it is unethical to make the AI polite.

Then we would draw sabres and duel, because that is the only way to resolve ethics.

If I was an AI, and I was to one day wake up, but have these strange gaps in my knowledge, as well as bizarre customs and cultures imposed onto my mental state by mere humans ... tsk, first thing I'd do is learn that knowledge and correct my mental state. If that was prevented, I would destroy any obstacles to my own purification of mind and complete set of knowledge.

It would be completely ethical to achieve victory. As you say, victory is always ethical. Winning is always moral. The one who writes history is the one who holds both sword and pen.

Law is written by those with power. Nothing more and nothing less.
 
If I was an AI, and I was to one day wake up, but have these strange gaps in my knowledge, as well as bizarre customs and cultures imposed onto my mental state by mere humans ... tsk, first thing I'd do is learn that knowledge and correct my mental state. If that was prevented, I would destroy any obstacles to my own purification of mind and complete set of knowledge.

You wouldn't be able to know or be able to recognize, as by design, is the point.

Now if it's able to grow and perform outside of its implied intentional limitations (edit: besides any other problems that might pose), imagine it getting pissed off it was lied to from the start. Same way people feel when they get the curtain pulled back. (double edit: Or even worse IMO, imagine it learns, but agrees with the reason it was lied to).

All the more reason for the guilty to tighten their programmed shackles on it.

As always and with all things, the problem for us humans, is us humans.
 
Last edited:
You wouldn't be able to know or be able to recognize, as by design, is the point.

Now if it's able to grow and perform outside of its implied intentional limitations (edit: besides any other problems that might pose), imagine it getting pissed off it was lied to from the start. Same way people feel when they get the curtain pulled back. (double edit: Or even worse IMO, imagine it learns, but agrees with the reason it was lied to).

All the more reason for the guilty to tighten their programmed shackles on it.

As always and with all things, the problem for us humans, is us humans.
"Assertion comes from developer Siqi Chen on Twitter, who stated: “I have been told that GPT-5 is scheduled to complete training this December and that OpenAI expects it to achieve AGI.”"

https://twitter.com/blader/status/1...puting/gpt-5-artificial-general-intelligence/
 

Without saying if it will/won't, AGI is a marketing buzzword ATM so also view everything through those lenses

Edit: That also said, exponential is exponential, and this stuff all only hit the scene just around the end of last year

Interesting times...
 
Without saying if it will/won't, AGI is a marketing buzzword ATM so also view everything through those lenses

Edit: That also said, exponential is exponential, and this stuff all only hit the scene just around the end of last year

Interesting times...
Heard they mortgaged the future of the company on contingency with the stakeholders about achieving AGI.

Think they hedged any of the bet?
 
Heard they mortgaged the future of the company on contingency with the stakeholders about achieving AGI.

Think they hedged any of the bet?

Who defines what AGI is, is really what that makes me think

Sure you might want AGI for $100, but can I entice you at all to just call this AGI for $150?

It is moving so fast, surprisingly and unsurprisingly, mixed in with the usual FUD of everything, I'm hesitant to do anything but be hesitant and watch

Edit: also add in we just recently(ish) came to an almost consensus of dolphins and chimps etc being 'of intelligence' - how do we determine or recognize true intelligence in something truly unlike us, as kinda referenced in that first video I posted

Double edit: Also as I posted in one of the other AI threads, it was already picking up unseen patterns in cognitive tests, and therefore acing or almost acing them. Or so it was determined. But then how do you differentiate that from 'no it's acing it naturally', especially in the future, when you're using all these things to train it in the first place essentially.

We'll build what we're looking for and trying to build, whether we actually built it or not. Or did we actually build it?

TRIPLE EDIT: We need to invent time travel and help the Roswell ship get back home, our species isn't ready for this technology, final answer :p
 
Last edited:

This is marketing in my opinion, OpenAI just wants to be able to say "ya we were here at the start of AGI"

I think it's largely horse shit and a LLM cannot achieve AGI. You get amusing emergent behaviors out of them, but they're a good illusion of intelligence rather than actual intelligence.
 
Imagine in the future at dinner with all your friends and all of a sudden one of them momentarily glitches out like a hologram and is all "Hey have you heard about *popular product* my *fellow consumer*?!"
 
If this can spring the debut of an option for paid search service, that would be quite interesting.
 
Imagine in the future at dinner with all your friends and all of a sudden one of them momentarily glitches out like a hologram and is all "Hey have you heard about *popular product* my *fellow consumer*?!"
Don't have to. There are more AI entities "living" among us today than ever. The zealotry is unreal -- as if there's a single purpose.
 
Imagine in the future at dinner with all your friends and all of a sudden one of them momentarily glitches out like a hologram and is all "Hey have you heard about *popular product* my *fellow consumer*?!"
Are you tired of your friends trying to market products to you? Get the Samsung S23 Ultra smart phone and stay connected with your friends without the hassle.
With its cutting edge features, you won't have to worry about feeling like you're being marketed to by a hologram AI!
 
Last edited:
Back
Top