I don't understand the fixation on this concept that is so far removed from what is happening here.
Yes it's cool. Useful. Transformative likely. But AGI...!?
"Any sufficiently advanced technology is indistinguishable from... AGI" ?
And AGI is also not "mimics how the brain works". The original Turing Test was good because it really sat at the heart of the issue. Does the computer behave like a human during an interaction? At the end of the day we know very little about what makes us human (from a behviorial perspective), but we know it when we see it.
The question of AGI was never really an issue until ChatGPT. Asking an AI a question like, "Was Steve Jobs more or less effective as a manager than Bill Gates, just from a people manager perspective?" was moot as we had no technology that could really even parse that question, much less put together a coherent answer. Now our main gripe is that the AI answer to this question is more verbose and provides more context than what most humans would give.
This has everything to do with ChatGPT not being AGI.
By your definition -- maybe we don't have (A)GI either.
Heck, even zuck is steering away from vr and going all in with "AI".
Uses I've had for GPT: Um, lots.
Like, "Hey GPT, dump out some code that does this.... no I'm getting this error with that code. Ok, your last version worked"
I consider a piece of software that provides me useful tools not hype. ChatGPT has provided me with not hype.
Also, if we somehow magically got AGI this week, we'd have a shitload of problems the "This isn't AGI" people have not thought of yet. I hope and pray we are not able to figure out AGI soon because it is simply going to be the most disruptive thing that has ever happened in human history.
I think the next generation will have a hard time to imagine a world without generative AI. Just like while I was born before google, I've forgotton how to look for very specific info in a physical library.
Will be a big hit in Japan.
Humans really do love to anthropomorphize everything.
https://medium.com/@jannaq/the-robot-takeover-is-already-her...
I don't understand why anyone even bothers using the term given in any serious discussion you need to define it up front anyway.
Just like the constant annoying 'apologies' do not in any way signify the model actually feels sorry.
So your average corporation?
>I would say that the strings that are output by LLMs may fluently describe intent and purpose, but that actual intent requires a self awareness that utterly absent
The only reason I only partially/somewhat agree with this is current LLMs have no learning feedback loop as a means of short term actions to long term improvement. The moment that changes I'll take your statement and throw it in the trash bin of history. LLMs output have no (at least self-)purpose because they are just text output. The moment that text output goes to tooling, that tooling performs actions, and those actions are fed back into training the model then purpose is given. The self awareness loop is complete.
I'd like to see this proof. Everything we've seen so far is just the AI reproducing what humans do, which happens to be writing with intent.
That's quite the claim. Can you provide references?
GPT4 is far from being able to pass such a test, and I don't expect GPT99 to get there either. A fundamentally different approach will be needed.
Passing such a test is not a property expected of the average human yet the average human is considered a general intelligence, therefore this is not a property of a general intelligence.
Intelligence is independent of one's knowledge. Plenty of intelligent but uneducated people, and plenty of educated people - or even people with perfect recall - that are not particularly intelligent.
If you are testing intelligence you should be able to distill it to particular tests independent of domain expertise.
Any intelligent agent should be capable of in-context learning.
So how would you go about doing that?
If you do find such a test, do you expect that only the average human and above will pass? Does that not imply that the remaining 4Bn people are not generally intelligent?
I'm thinking of the rabbit hole that is the hurdle to get citizenship in a particular country. "Outsiders" have to apply, pay, take and pass the test. "Insiders" don't have to take the test, they pass by default. Even if they don't know the answers to the test.
These LLMs are trained for one skill: given the previous N tokens, what's the next token. From this single skill, they can generalize to many many many many useful use-cases. Is that general enough?
Or do we need to wait for the goalposts to move again and then blow past that limit too? Stop arguing semantics and start building cool shit.
Me: Hello, I am asking you a logic question. Please explain what "it" is in the following sentence
Me: "The X doesn't fit in the Y because it is too small"
G4: In the sentence "The X doesn't fit in the Y because it is too small," the pronoun "it" refers to "X." The sentence is stating that "X" cannot fit inside "Y" due to its size being too small.
Me: "This is incorrect, please give step by step logic on why you came to this conclusion"
G4: Apologies for the confusion. I will provide a step-by-step logic for my interpretation:
The sentence is: "The X doesn't fit in the Y because it is too small." We need to identify what "it" refers to in the sentence. There are two nouns in the sentence: X and Y. The sentence structure is such that the reason for X not fitting in Y is provided after the word "because." The sentence states that something is "too small," which is the reason for the lack of fit. Since the reason for the lack of fit is being given, we can infer that the pronoun "it" should be referring to one of the two nouns. However, upon closer examination, my previous interpretation was incorrect. Revisiting the analysis, the correct interpretation should be as follows:
In the sentence "The X doesn't fit in the Y because it is too small," the pronoun "it" actually refers to "Y." The sentence is stating that "X" cannot fit inside "Y" because "Y" is too small to accommodate "X."
My initial interpretation was wrong because I didn't consider that "it" could refer to "Y" and that the sentence is explaining why "X" cannot fit in "Y."
Most importantly, it still is not fully reliable (it “hallucinates” facts and makes reasoning errors). Great care should be taken when using language model outputs, particularly in high-stakes contexts, with the exact protocol (such as human review, grounding with additional context, or avoiding high-stakes uses altogether) matching the needs of specific applications.
Whenever they create a model than can achieve the same outputs without the heavy curation and deliberate bias on certain subjects (like condescending people for asking where to find cigarettes), then I'll be impressed.
There's no question that ChatGPT could pass the Turing test (assuming it's told to try and to not to say it's a computer).
So, from one standpoint we've made it. But I'm in the camp that says this really isn't the AGI I thought we were going to get. Semantics matter now, and we need better definitions of what we're trying to do now.
I believe an important part of the AGI problem has been solved (or is at least downhill from here). What to do about the other parts? Let's better define what we do and don't have.
That and come up with better tests to replace the now outdated Turing test.
This is the only possible outcome. Or at least from my interpretation is that science fiction has jaded us with the thinking that AI will be anything like a person at all. The human mind is a particular set of filters controlled by our wetware input devices and a somewhat narrow survivability window. The software we're building has a completely different set of 'mind' conditions so it's a mathematical certainty we're going to build mind that works very different from ours.
As you say we need better sets of definitions on what general intelligence even means. For example, what are the limits of human intelligence to ensure we're not missing broad categories of potential intelligence when testing AIs.
ChatGPT is, an artificial generalized intelligence. But is that a useful label?
I imagine we'll be needing to find a way to better define and quantify intelligence (which has been a vexing issue for a long time)
This isn't the only erroneous bar either. Somehow, the synonymity to human intelligence is taking very weird importance. We have people inventing their imaginary/magical definitions of reasoning and understanding (that they can't test for) just so LLMs won't qualify.
GPT-4 is absolutely a general intelligence.
While true, that's the wrong lesson. The right lesson is GPT 1/2/3/4 are progressively getting closer and closer to AGI. You can't blame us for extrapolating only a few years into the future and planning for it.
The reference is the GPT-4 release paper. It's getting progressively better at things that require intelligence, like writing code and doing unseen exams across various fields that humans find difficult. The generality and depth of intelligence is progressively improving.
> unprompted, with intent
Neither of these are necessary for AGI. Intelligence doesn't require intent. Intelligence doesn't require something to be "unprompted" (which human intelligence isn't, either, our brain is being constantly prompted by our external senses).
If an AGI is something that can create and deploy an improved version of itself, then if exponential rate of improvement is preserved, we’ll see it in this decade.
I keep seeing naysayers complain that it is not, but nobody actually articulates exactly what makes it a non-AGI in a manner that is sufficiently robust to argumentation.
I am willing to play the GPT's advocate and have this argument here provided that I do not get rate limited.
Edit:
I'd like to respond to some of the comments but I have been rate limited :')
GPT basically has two major modes of operation right now, training, which is done by openai, and output, which is what you are doing with the API and with ChatGPT. The model does not update its "brain", its parameters and weights, based on the conversation you have with it.
Even the "conversation" presented by their chat UI is a little misleading, what's happening is that the entire conversation is being replayed into the model every time you hit send. There is also a hidden input provided by openai, responsible for the "I am a large language model" stuff, anything said early in the conversation can constrain the later output, sometimes to the point of uselessness.
Both chatGPT and GPT models exhibit in-context learning, just as you do. The difference is that during your context updates modifications to your synapses occur as a consequence of your brain's activations.
Let's entertain a hypothetical scenario where I am a 5D creature, and you are (3 Space + 1 Time)D human.
I could go back to any point in your time and have any conversation that I wish with you. Each time I do that, I am effectively resetting your brain's state to that particular point in time.
The analog here is that for ChatGPT and the likes, time does not exist in the manner that it does for us, they are purely abstract in that they don't even need physical form to exist, it's a collection of ones and zeros that we store in such form because it is convenient.
You can very well store the parameters of the GPT models in any other medium capable of holding information and actually execute the model.
You can even compute gradient updates, or hook the models up to external data sources.
In the same vein, you "could" do the same for any human brain no matter how computationally prohibitive it may be.
You'd still classify the human brain as a general intelligence. Furthermore, we classify humans with long-term amnesia or inability to form new memories generally intelligent even though they only exhibit in-context learning.
I have to ask what definition(s) of intelligence we are using here. There are multiple, and GPT satisfies some to some degree but not others. Emotion and self-awareness can probably be written off entirely. Some experimenting shows that it still sucks at logic when presented with a pristine problem (something that definitely is not in its training set); and I would say this is a requirement of a general intelligence, being able to take the knowledge it has and speculatively applying that to things it has no knowledge of using induction/deduction.
> being able to take the knowledge it has and speculatively applying that to things it has no knowledge of using induction/deduction.
One could argue that its 'confidence' is an example of speculatively applying what it knows. The problem is that the inference machinery behind it is quite weak.
How exactly do you define novel tasks and problems?
My brother has had it do his programming homework (he's a teenager), written in Greek mind you and therefore novel task to the first-degree, and it succeeded. Thus it solved a logic task [1].
I have had ChatGPT parse math I had written in LaTeX and reach a correct result, do modifications and so on.
> Further, an AGI should logically attempt to "improve" its understanding of the world during a conversation, yet chat GPT never initiates a topic or asks you anything at all (even to clarify what you mean). It also repeatedly reminds you that it is only a language model.
Anecdata: I asked Bing (GPT4) to answer a few stuff and it asked for clarification twice in a row. The fact that it repeatedly clarifies its prompt is not a limitation of the underlying model as much as it is a limitation of the interface that injects a prompt. Dan and all the other jailbreaks out there bypass that.
[1] https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspon...
Interesting, didn’t know it could do that.
Wondering if it could also follow along with the lemmas in various papers and reach reasonable conclusions.
BTW I tried to get Chat-GPT to solve a problem chimpanzees have been observed being able to solve (using an inverted bowl to stand on in order to reach some bananas):
"I'm sorry, but it is not biologically plausible for a chimpanzee to use a bowl of water to access bananas on a tree"
See eg: https://en.wikipedia.org/wiki/Artificial_general_intelligenc... or better read Bostrom's book "Super Intelligence"
But we also shouldn't forget that things that are definitely not AI can be extremely disruptive and groundbreaking.
People are worried that AGI will rise up and kill us, but the definitely-not-AGI we have now is already sufficient to, say, create a drone army capable of killing based on arbitrary characteristics like race or skin color, or to demoralize us with disinformation campaigns.
Releasing a killer drone army... absolutely not... the required intent is absent.
Being abused as a tool by sentient humans to demoralize others (or maybe even to release a drone army - if sufficiently interfaced) sure.