If a LLM only understands that 9/3 is 3 because it can run a python script then it’s not intelligent.
Artificial General Computing (AGC) may be?
Ai is pwning humanity in the game of intelligence. What’s left? Fighting for the meaning of the term AGI as if it matters?
AI is smarter than you in many respects. And the trendline points to AI becoming smarter in every respect. That’s the tragedy and the truth.
But sure. Hold on to your precious definition of agi and the fact that ai hasn’t hit that moving goal post yet if it keeps you sane.
It's not right for people to dismiss there are some cool things going on.
Once artificial comprehension is cracked, there will be no possibility of training an AI with incorrect or false or bad data, the AI will reject it as fraudulent logic. It will be very used to that too, because we really have a tsunami of fraud in this civilization.
As a society lead and composed largely of sycophants and weak educations, we'll call this hyper-dimensional crib sheet AGI and then need a new term for when we actually get artificial comprehension.
So you're the genius that has figured it all out, and the entire world is wrong?
It's far more likely you are wrong.
Yes - there is absolutely reason to question AGI, and the way in which we understand AI popularly is 'wrong' - but - the better ways for understanding it are not arcane. In fact, you'd be able to get most people to agee with the less popular but 'more valid' description with ease.
"AI is a means of communication not automation" is a completely contrarian take - not some kind of 'obviously valid' perspective,
"GI requires artificial comprehension which is a science that humanity does not have. " - LLMs exhibit artificial comprehension already - albeit on a very rudimentary scale. It's very reasonable to suggest that they are 'inherently limited' but that's a different argument.
You're conflating 'groupthink' and 'magical thinking'.
An arbitrary contrarian idea is not evidence of 'groupthink'
'Groupthink' is common - more common that not.
The 'Gell-Mann Amnesia effect' alludes to this - whenever someone with expertise reads about something in the press related to their expertise - they realize 'the media has it wrong' - but they continue to 'trust the media' on all other matters.
That's nuanced, but observable.
But in most cases - public narratives or common misconceptions are easily remedied with a few basic facts.
A military expert could quickly dispell misconceptions about 'how the war in Iran is actually going'.
In the case of AI - it's true that we probably have a common misunderstanding or intuition for how it works.
That does not mean the OPs imperious claim has any validity.
LLM AI is absolutely a communications technology. Every single aspect of understanding them and understanding how to use them is a communications problem. Yet, you'll argue that.
The comprehension exhibited by LLM AI is not comprehension, it is pattern recognition. That subtle distinction is critical. Comprehension is the creation and operation of virtual systems within the comprehending mind for the purpose of validating an observation for fraud, for plausibility and therefore continued safely, and for validating exploratory concepts. There is no statistical reply generation there would be something else entirely that is not statistical calculus driven.
What a shameful lack of self awareness for someone who is making 'magical claims' - and who cannot express their ideas; your closing paragraph is gibberish.
You're the one making incredulous claims, not me.
Anyone claiming to have some obvious purview that literally millions of other people working intimately with the subject cannot perceive is by definition making a dubious claim ... which is fine ... so long as it's backed up by some kind of logical assertion or evidence, which you have not provided.
1) "Every single aspect of understanding them and understanding how to use them is a communications problem."
This is not a reasonable claim.
2) "Pattern recognition is not comprehension"
Is a reasonable claim - and very common perspective.
It's at the core of the debate surrounding AI, not some 'specific insight' that is somehow rare.
It's literally what everyone is talking about.
3) "There is no statistical reply generation there would be something else entirely that is not statistical calculus driven."
That is not even English, it makes no sense - it reads like an LLM with severely corrupted context.
4) As to the actual argument: Humans are ultimately biological 'neural networks' it could very well be that our reasoning is effectively a form of pattern recognition.
The underlying architecture of Le Cun's JEPA architecture give a strong hint to that - they have a layer of abstraction above tokenisation wherein the model can essentially 'reason' in a way that may be more familiar than the kinds of synthetic 'token reasoning' we seen in LLM architectures today.
But I would make a stronger and different set of claims here:
1. Computation as mathematical formalism is distinct from computation as it exists in the wild (paradigmatically, the human mind). Therefore, computation as mathematical formalism is not identical to computation as it exists in the human mind.
2. Computation as mechanizable mathematical formalism is a syntactic construct. This property is exactly what lends it to mechanical simulation. A purely formal system does not require semantics to operate, though it does require a human observer with semantics to design, recognize, and interpret it (see Searle's Chinese room argument).
3. Computation as mathematical formalism is not identical to computation as it exists in physical machines. The latter only simulate the former through a conventional correspondence between effectively arbitrarily chosen physical events and symbolic constructs locked away in the human observer. If a system is merely executing physical transitions under a mechanistic worldview, its "goals" are parasitic on external human users. There is just no way in which we may objectively ascribe computation to physical computing machines (see Kripke's quaddition argument, for instance). And strictly speaking, physical machines aren't even applying syntactic rules to symbols. A corollary of this is that computation as mechanical simulation is also not identical to computation in the human mind.
4. What is most characteristic of intelligence is not computation, but intentionality and semantics. These are absent from the mathematical formalism as mechanizable construct, precisely because semantics and intentionality cannot be mechanized. So the claim that "intelligence" is some broad and inclusive continuum ignores the fact that there is no alchemy that can transmute syntax into semantics. You can't be just a little bit pregnant. These are differences in kind, not degree. You can't retreat into "neutrality" here either. To call a system "intelligent" in a non-trivial sense rather than merely "capable of producing useful outputs" is already to take a stand on what intelligence is. Quietly adopting a"mechanistic/functionalist stance while leaning on concepts like "information", "representation", "optimization", and "capability" ignores how much heavy philosophical baggage these words carry.
The linked post makes no such claim.
> artificial general intelligence is not a scaled up LLM ... artificial comprehension ... kindergarten ...
Your reasoning takes the form of "I define X to mean Y and Z, and since Y is impossible, X is impossible", which is trivially true within your definition.
But not everyone shares your definition.
A widely held definition of AGI is "a machine that can accomplish most tasks that humans do on computers as well as the median human".
This definition focuses on concrete and observable results and not philosophical or metaphysical definitions of "intelligence" or "comprehension" or "consciousness".
Absolutely! This is naked inevitabilism that makes the other parties skate right over the implicit priors they may disagree on.
Inevitabilism also has applications in sales and parenting, e.g. asking a vegitable-hating toddler "Will you have peas or carrots?" The slight of hand is to make vegetable-eating fait accompli.
>those thinking it is are literally failed educations
What? What does this mean? I feel like I'm having a stroke reading this.
>AGI requires artificial comprehension
It requires artificial general intelligence. Like, trivially. You're choosing to define AGI in a specific way that isn't based on any foundation that is universally accepted.
>When we have artificial comprehension there will be no need to train AIs beyond the kindergarten school basics, and then the AGI will figure it all out, all the rest, on its own.
Ok? I can buy that what you're describing would be pretty powerful, but how is this a requirement for "general intelligence"? Am I not generally intelligent despite learning much of the knowledge I use from others? I've built a career out of stuff I learned through "human literature," am I literally just unintelligent because I didn't figure out everything on my own?
And shouldn't such a capable system not even need "kindergarten school basics"? Why is that the minimum amount of knowledge you find acceptable to explicitly teach such an AI while still considering it a general intelligence?
>Human history and human literature will need to be learned in reference, but science and mathematics will require no training, beyond to learn the names of concepts we discovered earlier, so the AGI can discuss its understandings with humans and other AGI.
I guess I don't see why what we have now is fundamentally different from whatever it is you're imagining. If an AI can produce novel scientific findings, then isn't that sufficient for saying that it is generally capable at making scientific findings? Sure, we've bootstrapped it as much as possible, but do you not agree that LLMs seem capable of scientific discovery?
Seems like you've just concocted a wish list and have attributed your strict definition of AGI as some fundamental source of truth. I'd argue that what you're describing is much stronger than "AGI" and would likely warrant a label like ASI instead.
To me (and most people?) AGI is simply an AI system that is reasonably capable of general work. It doesn't go deeper than that. If it can meet the expectations of "general intelligence" (the same way you could describe a human as being "generally intelligent"), then it is AGI by definition.
It's fine to argue what the criteria for "general intelligence" should be, and you can even assert that you don't think AI will achieve a reasonable "general intelligence" without whatever foundation you're describing, but it doesn't make sense to argue that the general intelligence of a given system is the "wrong" kind of general intelligence and therefore doesn't deserve the title because it isn't a more powerful version of a general intelligence.
AGI: artificial general intelligence is not a scaled up LLM, and those thinking it is are literally failed educations. They are failed because they cannot see the subtly that is what AI becomes when trained on human literature; it becomes an entirely different thing than any other form of trained algorithm. AI is a communications technology, which few to none seem to understand. AI is not automation, but it can be used to create automation with careful consideration, that is not taking place.
I'm speaking in strong terms because we are literally fighting a war against stupidity and propaganda.
AGI requires artificial comprehension which is a science that humanity does not have. When we have artificial comprehension there will be no need to train AIs beyond the kindergarten school basics, and then the AGI will figure it all out, all the rest, on its own. Human history and human literature will need to be learned in reference, but science and mathematics will require no training, beyond to learn the names of concepts we discovered earlier, so the AGI can discuss its understandings with humans and other AGI.