Leaked Documents Show OpenAI Has a Very Clear Definition of ‘AGI’

Yeah, maybe you're right. I don't known where the threshold is.

I wonder if the current computational feasibility will cap out improvment of current generation LLMs soon?

First of all, I'm about to give the extreme dumbed down explanation, but there are actual academics covering this topic right now usually using keywords like AI "emergent behavior" and "overfitting". More specifically about how emergent behavior doesn't really exist in certain model archetypes and that overfitting increases accuracy but effectively makes it more robotic and useless. There are also studies of how humans think.

Anyways, human's don't assign numerical values to words and phrases for the purpose of making a statistical model of a response to a statistical model input.

Humans suck at math.

Humans store data in a much messier unorganized way, and retrieve it by tracing stacks of related concepts back to the root, or fail to memorize data altogether. The values are incredibly diverse and have many attributes to them. Humans do not hallucinate entire documentations up or describe company policies that don't exist to customers, because we understand the branching complexity and nuance to each individual word and phrase. For a human to describe procedures or creatures that do not exist we would have to be lying for some perceived benefit such as entertainment, unlike an LLM which meant that shit it said but just doesn't know any better. Just doesn't know, period.

Maybe an LLM could approach that at some scale if each word had it's own model with massive amounts more data, but given their diminishing returns displayed so far as we feed in more and more processing power that would take more money and electricity than has ever existed on earth. In fact, that aligns pretty well with OpenAI's statement that it could make an AGI if it had Trillions of Dollars to spend and years to spend it. (They're probably underestimating the costs by magnitudes).

I'd say that difference between nature boiling down to maths and LLMs boiling down to maths is that in LLMs it's not the knowledge itself that is abstracted, it's language. This makes it both more believable to us humans, because we're wired to use language, and less suitable to actually achieve something, because it's just language all the way down.

Would be nice if it gets us something in the long run, but I wouldn't keep my hopes up

Funny to me how defensive you got so quick, accusing of not reading the linked paper before even reading it yourself.

The reason OP was so rude is that your very premise of "what is the brain doing if not statistical text prediction" is completely wrong and you don't even consider it could be. You cite a TV show as a source of how it might be. Your concept of what artificial intelligence is comes from media and not science, and is not founded in reality.

The brain uses words to describe thoughts, the words are not actually the thoughts themselves.

https://advances.massgeneral.org/neuro/journal.aspx?id=1096

Think about small children who haven't learned language yet, do those brains still do "stastical text prediction" despite not having words to predict?

What about dogs and cats and other "less intelligent" creatures, they don't use any words but we still can teach them to understand ideas. You don't need to utter a single word, not even a sound, to train a dog to sit. Are they doing "statistical text prediction" ?

It's a basic argument of generative complexity, I found the article some years ago while trying to find an earlier one (I don't think by the same author) that argued along the same complexity lines, essentially saying that if we worked like AI folks think we do we'd need to be so and so much trillion parameters and our brains would be the size of planets. That article talked about the need for context switching in generating (we don't have access to our cooking skills while playing sportsball), this article talks about the necessity to be able to learn how to learn. Not just at the "adjust learning rate" level, but mechanisms that change the resulting coding, thereby creating different such contexts, or at least that's where I see the connection between those two. In essence: To get to AGI we need AIs which can develop their own topology.

As to "rudeness": Make sure to never visit the Netherlands. Usually how this goes is that I link the article and the AI faithful I pointed it out to goes on a denial spree... because if they a) are actually into the topic, not just bystanders and b) did not have some psychological need to believe (including "my retirement savings are in AI stock") they c) would've come across the general argument themselves during their technological research. Or came up with it themselves, I've also seen examples of that: If you have a good intuition about complexity (and many programmers do) it's not unlikely a shower thought to have. Not as fleshed out as in the article, of course.

They have a conclusion that they've come to the conversation with and anything that challenges that gets down voted without consideration.

The assumptions you aren't allowed to challenge, in order: AI is bad; Computer intelligence will never match or compete with human intelligence; computer intelligence isn't really intelligence at all, it's this other thing [insert 'something' here like statistical inference or whatever].

"AI is bad" is more of a dictum extending from cultural hedgemony than anything else. It's an implicit recognition that in many ways, silicon valley culture is an effective looting of the commons, and therefore, one should reject all things that extend from that culture. It's not a logical or rational argument against AI necessarily, but more of an emotional reaction to the culture which developed it. As a self preservation mechanism this makes some sense, but obviously, it's not slowing down the AI takeover of all things (which is really just putting highlighter on a broader point that silicon valley tech companies were already in control of major aspects of our lives).

Computer intelligence never match human intelligence is usually some combination of goal post moving, or a redefining of intelligence on the fly (this I've specifically presented for the third critique, because it warrants addressing). This is an old trope that goes back almost to the beginning of computer intelligence (it's not clear to me our definitions of machine intelligence are very relevant). It quite litterally started with multiplying large numbers. Then, for literally decades, things like chess and strategy, forwards facing notions in time were held up as some thing only "intelligent systems could do". Then post deep blue, that got relegated to very clever programmers and we changed intelligence to be something about learning. Then systems like Alpha go etc came about, where they basically learned the rules to the game by playing, and we relegated those systems to 'domain specific' intelligences. So in this critique you are expected to accept and confirm the moving of goalposts around machine intelligence.

Finally, it's the "what computers do isn't intelligence, it's **some_other_thing.exe**™. In the history of machine intelligence, that some other thing has been counting very quickly, having large-ish memory banks, statistical inference, memorization, etc. The biggest issues with this critique, and when you scratch and sniff it, you very quickly catch an aroma of Chomsky leather chair (and more so if we're talking about LLMs), and maybe even a censer of a Catholic Church. The idea that humans are fundementally different and in some way special is frankly, fundemental, to most western idealogies in a way we don't really discuss in the context of this conversation. But the concept of spirit, and that there is something "entirely unique" about humans versus "all of the rest of everything" is at the root of Abrahamic traditions and therefore also at the root of a significant portion of global culture. In many places in the world, it's still heretical to imply that human beings are no more special or unique than the oak or the capibara or flatworm or dinoflagellate. This assumption, I think, is on great display with Chomsky's academic work on the concept of the LAD, or language acquisition device.

Chomsky gets a huge amount of credit for shaking up linguistics, but what we don't often talk about, is how effectively, his entire academic career got relinquished to the dust bin, or at least is now in that pile of papers where we're not sure if we should "save or throw away". Specifically, much of Chomsky's work was predicted on the identification of something in humans which would be called a language acquisition device or LAD. And that this LAD would be found in as a region in human brains and would explain how humans gain language. And just very quickly notice the overall shape of this argument. It's as old as the Egyptians in at least trying to find the "seat of the soul", and follows through abrahamism as well. What LLMs did that basically shattered this nothing was show at least one case where no special device was necessary to acquire language; where in fact no human components at all were necessary other than a large corpus of training data; that maybe languages and the very idea of language or language acquisition are not special or unique to humans. LLMs don't specifically address the issue of a LAD, but they go a step farther in not needing to. Chomsky spent the last of verbal days effectively defending this wrong notion he had (which has already been addressed in neuroscience and linguistics literature), which is an interesting and bitter irony for a linguist, specifically against LLMs.

To make the point more directly, we lack a good coherent testable definition of human intelligence, which makes any comparisons to machine intelligence somewhat arbitrary and contrived, often to support the interlocutors assumptions. Machine intelligence may get dismissed as statistical inference, sure, but then why can you remember things sometimes but not others? Why do you perform better when you are well rested and well fed versus tired and hungry, if not for there being an underlying distribution of neurons, some of which are ready to go, and some of which are a bit spent and maybe need a nap?

And so I would advocate caution about investing heavily into a conversation where these assumptions are being made. It's probably not going to be a satisfying conversation because almost assuredly they assumptee hasn't dove very deeply into these matters. And look at the downvote ratio. It's rampant on Lemmy. Lemmy's very much victim to it's pack mentality and dog piling nature.

Interesting you focus on language. Because that's exactly what LLMs cannot understand. There's no LLM that actually has a concept of the meaning of words. Here's an excellent essay illustrating my point.

The fundamental problem is that deep learning ignores a core finding of cognitive science: sophisticated use of language relies upon world models and abstract representations. Systems like LLMs, which train on text-only data and use statistical learning to predict words, cannot understand language for two key reasons: first, even with vast scale, their training and data do not have the required information; and second, LLMs lack the world-modeling and symbolic reasoning systems that underpin the most important aspects of human language.

The data that LLMs rely upon has a fundamental problem: it is entirely linguistic. All LMs receive are streams of symbols detached from their referents, and all they can do is find predictive patterns in those streams. But critically, understanding language requires having a grasp of the situation in the external world, representing other agents with their emotions and motivations, and connecting all of these factors to syntactic structures and semantic terms. Since LLMs rely solely on text data that is not grounded in any external or extra-linguistic representation, the models are stuck within the system of language, and thus cannot understand it. This is the symbol grounding problem: with access to just formal symbol system, one cannot figure out what these symbols are connected to outside the system (Harnad, 1990). Syntax alone is not enough to infer semantics. Training on just the form of language can allow LLMs to leverage artifacts in the data, but “cannot in principle lead to the learning of meaning” (Bender & Koller, 2020). Without any extralinguistic grounding, LLMs will inevitably misuse words, fail to pick up communicative intents, and misunderstand language.

language is a big thing in the human intelligence and consciousness.

But an LLM isn't actually language. It's numbers that represent tokens that build words. It doesn't have the concept of a table, just the numerical weighting of other tokens related to "tab" & "le".

But people who agree with my political ideology are considered and intelligent. People who disagree with me are stupider than chatgpt 3.5 and just say the same shit and can't be reasoned with.