A definition of AGI
arxiv.org
arxiv.org
I don't think people really realize how extraordinary accomplishment it would be to have an artificial system matching the cognitive versatility and proficiency of an uneducated child, much less a well-educated adult. Hell, AI matching the intelligence of some nonhuman animals would be an epoch-defining accomplishment.
If you took the current state of affairs back to the 90s you’d quickly convince most people that we’re there. Given that we’re actually not, we’re now have to come up with new goalposts.
Consider this. I could walk into a club in Vegas, throw down $10,000 cash for a VIP table, and start throwing around $100 bills. Would that make most people think I'm wealthy? Yes. Am I actually wealthy? No. But clearly the test is the wrong test. All show and no go.
One common kind of interaction I have with chatgpt (pro): 1. I ask for something 2. Chatgpt suggests something that doesn't actually fulfill my request 3. I tell it how its suggestion does not satisfy my request. 4. It gives me the same suggestion as before, or a similar suggestion with the same issue.
Chatgpt is pretty bad at "don't keep doing the thing I literally just asked you not to do" but most humans are pretty good at that, assuming they are reasonable and cooperative.
chatgpt.com is actually a good at or better than a typical human.
It does what it was designed to do, predict text. Does it do that incredibly well, yes. Does it do anything else, no.
That isn't to say super advanced text regurgitation isn't valuable, just that its nowhere even remotely close to AGI.
It can appear so, as long as you don’t check too carefully. It’s impressive but still very common to find basic errors once you are out of the simplest, most common problems due to the lack of real understanding or reasoning capabilities. That leads to mistakes which most humans wouldn’t make (while sober / non-sleep deprived) and the classes of error are different because humans don’t mix that lack of understanding/reasoning/memory with the same level of polish.
This is an interesting ambiguity in the Turing test. It does not say if the examiner is familiar with the expected level of the candidate. But I think it's an unfair advantage to the machine if it can pass based on the examiner's incredulity.
If you took a digital calculator back to the 1800s, added a 30 second delay and asked the examiner to decide if a human was providing the answer to the screen or a machine, they might well conclude that it must be human as there is no known way for a machine to perform that action. The Akinator game would probably pass the test into the 1980s.
I think the only sensible interpretation of the test is one where the examiner is willing to believe that a machine could be providing a passing set of answers before the test starts. Otherwise the test difficulty varies wildly based on the examiners impression of the current technical capabilities of machines.
If you look at a calculator you will quickly find it is much better then a human in any of the operations that have been programmed into the calculator, and has been since the 1960s. Since the 1960s the operations programmed into your average calculator has increased by several orders of magnitude. The digital calculator sure is impressive, and useful, but there is no crisis. Even in the world outside computing, a bicycle can outperform an human runner easily, yet there is no mobility crisis as a result. ChatGPT is very good at predicting language. And in quite a few subject matters it may be better than your average human in predicting said language. But not nearly as good as a car is to a runner, nor even as good as a chess computer is to a grand master. But if you compare ChatGPT to an expert in the subject, the expert is much much much better then the language model. In these tasks a calculator is much more impressive.
That’s not my experience at all. Unless you define “typical human” as “someone who is untrained in the task at hand and is satisfied with mediocre results.” What tasks are you thinking of?
(And, to be clear, being better than that straw man of “typical human” is such a low bar as to be useless.)
the AI bros like to talk about AGI as if it's just the next threshold for LLMs, which discounts the complexity of AGI, but also discounts their own products. we don't need an AGI to be our helpful chatbot assistant. it's fine for that to just be a helpful chatbot assistant.
I know without a single doubt that I could not simply as an “AI” “agent” to do this today and expect any sort of a functional result, especially when some of these were (very simple) judgement calls or workarounds for absolutely filthy data and a janky wedding planning website UI.
But when we say AGI, we want something that will function in the real world like a human would. We want to be able to say, "Here's 500 dollars. Take the car to get the materials, then build me a doghouse, then train my dog. Then go to the store, get the ingredients, and make dinner."
If the robotics aren't reliable enough to test that, then have it be a remote employee for 6 months. Not "have someone call up AI to wrote sections of code" - have a group of remote employees, make 10% AI, give them all the same jobs with the same responsibilities, and see if anyone notices a difference after 6 months. Give an AI an account on Upwork, and tell it to make money any way it can.
Of course, AI is nowhere near that level yet. So we're stuck manufacturing toy "AGI" benchmarks that current AI can at least have some success with. But these types of benchmarks only broadcast the fact that we know that current and near future AI would fail horribly at any actual AGI task we threw at it.
The Turing Test was great until something that passed it (with an average human as interrogator) turned out to also not be able to count letters in a word — because only a special kind of human interrogator (the "scientist or QA" kind) could even think to ask that kind of question.
I've seen claims of passing but it's always things like "with only 3 questions" or "with only 3 minutes of interrogation" or "With only questions about topic X". Those aren't Turing Tests. As an example, if you limit the test to short things than anything will pass "Limit to 1 word one question". User types "Hello", LLM response "Hi". PASS! (not!)
https://arxiv.org/pdf/2405.08007
I do think we’re going to see this shift as AI systems become more commonplace and people become more practiced at recognizing the distinction between polished text and understanding.
One could say that if there is maximally hard Turing test, and a "sloppy" Turing test, we are somewhere where the current models pass the sloppy version but not the maximally hard version.
An AI that happened to be able to pass the turing test would be pretty notable because it probably implies much more capabilities behind the scenes. The problem with, for example, LLMs, they're essentially optimized turing test takers. That's about all they can do.
Plus, I don't think any LLM will pass the turing test in the long term. Once something organically comes up that they aren't good at, it'll be fairly obvious they aren't human and the limits of context will also become apparent eventually.
We know what we need right now, the next step. That step is a machine that, when it fails, it fails in a human way.
Humans also make mistakes, and hallucinate. But we do it as humans. When a human fails, you think "damn, that's a mistake perhaps me or my friend could have done".
LLMs on the other hand, fail in a weird way. When they hallucinate, they demonstrate how non-human they are.
It has nothing to do with some special kind of interrogator. We must assume the best human interrogator possible. This next step I described work even with the most skeptic human interrogator possible. It also synergizes with the idea of alignment in ways other tests don't.
When that step is reached, humans will or will not figure out another characteristic that makes it evident that "subject X" is a machine and not a human, and a way to test it.
Moving the goalpost is the only way forward. Not all goalpost moves are valid, but the valid next move is a goalpost move. It's kind of obvious.
There's no reason an equivalent or superior general intelligence needs to be similar to us at all
To be a valid test, it still has to be passed by ~every adult human. The harder you make the test (in any direction), the more it fails on this important axis.
I don't know if machines that become insecure and lash out are a good idea.
Edit: Probably not, since it was published less than a week ago :-) I’ll be watching for benchmarks.
In other words; in one question; is the current AI not already well beyond the “…cognitive versatility and proficiency of an uneducated child”? And when you consider that in many places like Africa, they didn’t even have a written language until European evangelists created it and taught it to them in the late 19th century, and they have far less “education” than even some of the most “uneducated” avg., European and even many American children, does that not mean that AI is well beyond them at least?
Frankly, as it seems things are going, there Is at the very least going to be a very stark shift in “intelligence” that even exceeds that which has happened in the last 50 or so years that have brought us stark drops in memory, literary knowledge, mathematics, and even general literacy, not to mention the ability to write. What does it mean that kids now will not even have to feign acting like they’re selling out sources, vetting them, contradicting a story or logical sequence, forming ideas, messages, and stories, etc.? I’m not trying to be bleak, but I don’t see tons simply resulting in net positive outcomes, and most of the negative impacts will also be happening below the surface to the point that people won’t realize what is being lost.
Try to reconcile that with your ideas (that I think are correct for that matter)
This is because I use "stupidity" as the number of examples some intelligence needs in order to learn from, while performance is limited to the quality of the output.
LLMs *partially* make up for being too stupid to live (literally: no living thing could survive if it needed so many examples) by going through each example faster than any living thing ever could — by as many orders of magnitude as there are between jogging and continental drift.
LLMs have a reasonable learning rate at inference time (in-context learning is powerful), but a very poor learning rate in pretraining. And one issue with that is that we have an awful lot of cheap data to pretrain those LLMs with.
We don't know how much compute human brain uses to do what it does. And if we could pretrain with the same data-efficiency as humans, but at the cost of using x10000 the compute for it?
It would be impossible to justify doing that for all but the most expensive, hard-to-come-by gold-plated datasets - ones that are actually worth squeezing every drop of performance gains out from.
What's going on is AI fatigue. We see it everywhere, we use it all the time. It's becoming generic and annoying and we're getting bored of it EVEN though the accomplishment is through the fucking roof.
If elon musk makes interstellar car that can reach the nearest star in 1 second and priced it at 1k, I guarantee within a year people will be bored of it and finding some angle to criticize it.
So what happens is we get fatigued, and then we have such negative emotions about it that we can't possibly classify it as the same thing as human intelligence. We magnify the flaws and until it takes up all the space and we demand a redefinition of what agi is because it doesn't "feel" right.
We already had a definition of AGI. We hit it. We moved the goal posts because we weren't satisfied. This cycle is endless. The definition of AGI will always be changing.
Take LLMs as they exist now and only allow 10% of the population to access it. Then the opposite effect will happen. The good parts will be over magnified and the bad parts will be acknowledged and then subsequently dismissed.
Think about it. All the AI slop we see on social media are freaking masterpieces works of art produced in minutes what most humans can't even hope to come close to. Yet we're annoyed and unimpressed by them. That's how it's always going to go down.
We already had a definition of AGI. We hit it.
Are you sure about that? Which definition are you referring to? From what I can tell with Google and Grok, every proposed definition has been that AGI strictly matches or exceeds human cognitive capabilities across the board.
Generative AI is great, but it's not like you could just assign an arbitrary job to a present-day LLM, give it access to an expense account, and check in quarterly with reasonable expectations of useful progress.
There's a significant fraction of humanity that would not clear the bar to meet current AGI definitions.
The distribution of human cognitive abilities is vast and current AI systems definitely exceed the capabilities of a surprising number of people.
Has anyone tried this yet?
The turing test.
I'm curious when and what you consider to have been the moment.
To me, the general in AGI means I should be able to teach it something it's never seen before. I don't think I can even teach an LLM something it's seen a million times before. Long division, for example.
I don't think a model that is solid state until it's "trained" again has a very good chance of being AGI (unless that training is built into it and the model can decide to train itself).
I'm not an expert, but my layman's understanding of AI was that AGI meant the ability to learn in an abstract way.
Give me a dumb robot that can learn and I should be able to teach it how to drive, argue in court, write poetry, pull weeds in a field, or fold laundry the same way I could teach a person to do those things.
(2) I was rebutting the paper's standard that AGI should be achieving the status of a well-educated adult, which is probably far, far too high a standard. Even something measured to a much lower standard--which we aren't at yet--would change the world. Or, going back to my example, an AI that was as intelligent as a labrador in terms of its ability to synthesize and act on information would be truly extraordinary.
By this, what I mean is. Take an image of this: https://en.wikipedia.org/wiki/Traitorous_eight#/media/File:T..., change the file name to something like image.jpg and pass it into Qwen 3 4B, 8B, 30B and look at the responses you get:
It has no idea who these guys are. It thinks they are the beatles, the doors. If you probe enough, it'll say it's IBM cofounders. In a way, it kinda sees that these are mid-1900s folks with cool haircuts, but it doesn't recognize anything. If you probe on the F the model in question becomes convinced it's the Ford racing team with a detailed explanation of two brothers in the photo, etc.
The creation of autoregressive next token predictors is very cool and clearly has and will continue to have many valuable applications, but I think we're missing something that makes interactions with users actually shape the trajectory of its own experience. Maybe scaffolding + qlora solves this. Maybe it doesn't
I agree with this but also, the output is almost entirely worthless if you can’t vet it with your own knowledge and experience because it routinely gives you large swaths of incorrect info. Enough that you can’t really use the output unless you can find the inevitable issues. If I had to put a number to it, I would say 30% of what an LLM spits out at any given time to me is completely bullshit or at best irrelevant. 70% is very impressive, but still, it presents major issues. That’s not boredom, that’s just acknowledging the limitations.
It’s like designing an engine or power source that has incredible efficiency but doesn’t actually move or affect anything (not saying LLM’s are worthless but bear with me). It just outputs with no productive result. I can be impressed with the achievement while also acknowledging it has severe limitations
Any definition of AGI that allows for this is utterly useless:
> Me: Does adding salt and yeast together in pizza dough kill the yeast?
> ChatGPT: No, adding salt and yeast together in pizza dough doesn't kill the yeast.
(new chat)
> Me: My pizza dough didn't rise. Did adding salt and yeast together kill the yeast?
> ChatGPT: It's possible, what order did you add them in?
> Me: Water, yeast, salt, flour
> ChatGPT: Okay, that explains it! Adding the salt right after the yeast is definitely the issue.
(It is not the issue)
And you’re pointing to a single pizza dough error as if that somehow invalidates all of it. If that’s your bar, then every human who ever made a mistake in a kitchen is disqualified from being intelligent too. You’re cherry picking the single dumbest moment and pretending it defines the whole picture. It doesn’t.
The real story is that these models already demonstrate reasoning and generalization across virtually every intellectual domain. They write, argue, and problem solve with flexibility and intent. They’re not perfect, but perfection was never the standard. The Turing test was passed the moment you could no longer draw a clear line between where imitation ends and understanding begins.
You can sneer about yeast all you want, but the irony is that while you mock, the machines are already doing useful work coding, researching, analyzing, and creating, quietly exceeding every benchmark that once defined general intelligence.
Americans were glued to their seats watching Apollo 11 land. Most were back to watching I Dream of Jeanie reruns when Apollo 17 touched down.
* Ecosystems may require self-assembly
Things like chess-playing skill of a machine could be bench-marked against that of a human, but the abstract feelings that drive reasoning and correlations inside a human mind are more biological than logical.
We can easily program them to have human desires instead.
My emotions are definitely a function of the chemical soup my brain is sitting in (or the opposite).
We only have one good example of consciousness and sentience, and that is our own. We have good reason to suspect other entities (particularly other human individuals, but also other animals) have that as well, but we cannot access it, and not even confirm its existence. As a result using these terms of non-human beings becomes confusing at best, but it will never be actually helpful.
Emotions are another thing, we can define that outside of our experience, using behavior states and its connection with patterns of stimuli. For that we can certainly observe and describe behavior of a non biological entity as emotional. But given that emotion is something which regulates behavior which has evolved over millions of years, whether such a description would be useful is a whole another matter. I would be inclined to use a more general description of behavior patterns which includes emotion but also other means of behavior regulators.
there are many parts of human cognition, phycology etc. especially related to consciousness that are known unknowns and/or completely unknown.
a mitigation for this isaue would be to call it generally applicable intelligence or something, rather than human like intelligence. implying ita not specialized AI but also not human like. (i dont see why it would need to be human like, because even with all the right logic and intelligence a human can still do something counter to all of that. humans do this everyday. intuitive action, or irrational action etc.
what we want is generally applicable intelligence, not human like intelligence.
How does a computer with full AGI experience the feeling of butterflies in your stomach when your first love is required?
How does a computer experience the tightening of your chest when you have a panic attack?
How does a computer experience the effects of chemicals like adrenaline or dopamine?
The A in AGI stands for “artificial” for good reason, IMO. A computer system can understand these concepts by description or recognize some of them them by computer vision, audio, or other sensors, but it seems as though it will always lack sufficient biological context to experience true consciousness.
Perhaps humans are just biological computers, but the “biological” part could be the most important part of that equation.
There is reason to believe that consciousness, sentience, or emotions require a biological base.
Or
There is no reason to believe that consciousness, sentience, or emotions do not require a biological base.
The first is simple, if there is a reason you can ask for it and evaluate it's merits. Quantum stuff is often pointed to here, but the reasoning is unconvincing.
The second form There is no reason to believe P does not require Q.
There are no proven reasons but there are suspected reasons. For instance if the operation that nerons perform is what makes consciousness work, and that operation can be reproduced non-biologicLly it would follow that non biological consciousness would be possible.
For any observable phenomenon in the brain the same thing can be asked. So far it seems reasonable to expect most of the observable processes could be replicated.
None of it acts as proof, but they probably rise to the bar of reasons.
Personally, I remain skeptical that is the case.
What does seem likely is that “intelligence” will eventually be redefined to mean whatever we got out of the AI buildup.
There's really no reason to believe intelligence is tied to being human. Most of us accept the possibility (even the likelihood) of intelligent life in the universe, that isn't.
>human intelligence as something detached from human biology.
I don't completely agree with the previous comment, but there is something to be considered to their statement.
Given that, the constant drumbeat of pointing out how AI fails to be human, misses the mark. A lot of the same people who are making such assertions, haven't really thought about how they would quickly accept alien intelligence as legitimate and full-fledged... even though it too lacks any humanity backing it.
And why are they so eager to discount the possibility of synthetic life, and its intelligence, as mere imitation? As a poor substitute for the "real thing"? When faced with their easy acceptance of alien intelligence, it suggests that there is in fact a psychological reason at the base of this position, rather than pure rational dismissal. A desire to leave the purely logical and mechanical, and imbue our humanity with an essential spirit or soul, that maybe an alien could have, but never a machine. Ultimately, it is a religious objection, not a scientific one.
Feels good so we want more so you arrange your whole life and outlook to make more feel good happen. Intelligence!
Definition of artificial:
>Made by humans, especially in imitation of something natural.
>Not arising from natural or necessary causes; contrived or arbitrary.
Thus artificial intelligence must be the same as natural, the process of coming up with it doesn't have to be natural. What this means: we need to consider the substrate that makes natural intelligence. They cannot be separated willy nilly without actual scientific proof. As in, we cannot imply a roll of cheese can manifest intelligence based on the fact that it recognizes how many fingers are in an image.
The problem arises from a potential conflict of interests between hardware manufacturer companies and definition of AGI. The way I understand it, human like intelligence cannot come from algorithms running on GPUs. It will come from some kind of neuromorphic hardware. And the whole point of neuromorphic hardware is that it operates (closely) on human brain principles. Thus, the definition of AGI MUST include some hardware limitations. Just because I can make a contraption "fool" the tests doesn't mean it has human like cognition/awareness. That must arise from the form, from the way the atoms are arranged in the human brain. Any separation must be scientifically proven. Like if anyone implies GPUs can generate human like self awareness that has to be somehow proven. Lacking a logical way to prove it, the best course of action is to closely follow the way the human brain operates (at least SNN hardware).
>The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.
This is nonsense. GPT scores cannot decide AGI level. They are the wrong algorithm running on the wrong hardware.
I have also seen no disclosure on conflict of interests in the paper.
Which is it??
This isn’t meant to be a mystical statement that it’s magic that makes humans intelligent or some exotic process impossible to compute. But that the nature of our mind is not observable in its entirety to us sufficient that the current learned reinforcement techniques can’t achieve it.
Try this exercise. Do not think and let your mind clear. Ideas will surface. By what process did they surface? Or clear your mind entirely then try to perform some complex task. You will be able to. How did you do this without thought? We’ve all had sudden insights without deliberation or thought. Where did these come from? By what process did you arrive at them? Most of the things we do or think are not deliberative and definitely not structured with language. This process is unobservable and not measurable, and the only way we have to do so is through imperfect verbalizations that hint out some vague outline of a subconscious mind. But without being able to train a model on that subconscious process, one that can’t be expressed in language with any meaningful sufficiency, how will language models demonstrate it? Their very nature of autoregressive inference prohibits such a process from emerging at any scale. We might very well be able to fake it to an extent that it fools us, but awareness isn’t there - and I’d assert that awareness is all you need.
AGI is not here with LLMs, but its not because they lack reasoning ability. It's due to something different. Here is what I think is truly missing: continuous learning, long term memory, and infinite and efficient context/operation. All of these are tied together deeply, and thus I believe we are but a simple breakthrough away from AGI.
The idea of awareness being propagations through the NN is an interesting concept though. I wonder if this idea be proven through monitoring the electrical signals within the brain.
In essence, I think it doesn't matter that the brain has a whole bunch of chemistry added into it that artificial neural networks don't. The underlying deep non-linear function mapping capability is the same, and I believe this depth is, in both cases, comparable.
This is just a claim you are making, without evidence.
The way you understand awareness is not through "this is like that" comparisons. These comparisons fall over almost immediately as soon as you turn your attention to the mind itself, by observing it for any length of time. Try it. Go observe your mind in silence for months. You will observe for yourself it is not what you've declared it to be.
> An LLM does all of these things, and more, and you can prove this by how complex their output is.
Complex output does not prove anything. You are again just making claims.
It is astoundingly easy to push an LLM over to collapse into ungrounded nonsense. Humans don't function this way because the two modes of reasoning are not alike. It's up to those making extraordinary claims to prove otherwise. As it is, the evidence does not exist that they behave comparably.
It's immaterial and not measurable thus possibly out of reach of science.
Wait, you mean this HN comment didn't casually solve the hard problem of consciousness?
How easy? What specific methods accomplish this? Are these methods fundamentally different from those that mislead humans?
How is this different from exploiting cognitive limitations in any reasoning system—whether a developing child's incomplete knowledge or an adult's reliance on heuristics?
How is it different from Fake News and adults taking Fake News for granted and replicating bullshit?
Research on misinformation psychology supports this parallel. According to https://www.sciencedirect.com/science/article/pii/S136466132...:
"Poor truth discernment is linked to a lack of careful reasoning and relevant knowledge, as well as to the use of familiarity and source heuristics."
Maybe the only real difference is our long term experience and long term memory?
Even though complex output can be deceptive of the underlying mental model used to produce it, in my personal experience, LLMs have produced for me output that must imply extremely complex internal behaviour, with all the characteristics I mentioned before. Namely, I frequently program with LLMs, and there is simply zero percent probability that their output tokens exist WITHOUT first having thought at a very deep level about the unique problem I presented to them. And I think anyone that has used the models to the level I have, and interacted with them this extensively, knows that behind each token there is this black magic.
To summarize, I am not being naive by saying I believe everything my LLM says to me. I rather know very intimately where the LLM is deceiving me and when its producing output where its mental model must have been very advanced to do so. And this is through personal experience playing with this technology, both inference and training.
[1] https://www.anthropic.com/research/tracing-thoughts-language...
Thank you by saying that. I think most people have an incomplete mental model for how LLMs work. And it's very misleading for understanding what they really do and can achieve. "Next token prediction" is done only at the output layer. It's not what really happens internally. The secret sauce is at the hidden layers of a very deep neural network. There are no words or tokens inside the network. A transformer is not the simple token estimator that most people imagine.
There is little discussion of how AI will enhance (or destroy) our emotional intelligence, or our naturalistic, intrapersonal or interpersonal intelligences.
Most religions, spiritual practices and even forms of meditation highlight the value of transcending mind and having awareness be present in the body. The way AGI is described, it would seem transcendence may be treated as a malfunction or bug.
[1] https://en.wikipedia.org/wiki/Theory_of_multiple_intelligenc...
There are people that have a hard time recognizing/feeling/understanding other people as "aware". Even more about animals.
My measurement of outside intelligence is limited by my intelligence. So I can understand when something is stupider compared to me. For example, industrial machine vs human worker, human worker is infinitely more intelligent compared to machine, because this human worker can do all kinds of interesting stuff. this metaphorical "human worker" did everything around from laying a brick to launching a man to the Moon.
....
Imagine Super-future, where humanity created nanobots and they ate everything around. And now instead of Earth there is just a cloud of them.
These nanonobots were clever and could adapt, and they had all the knowledge that humans had and even more(as they were eating earth a swarm was running global science experiments to understand as much as possible before the energy ends).
Once they ate the last bite of our Earth(an important note here: they left an optimal amount of matter to keep running experiments. Humans were kept in a controlled state and were studied to increase Swarm's intelligence), they launched next stage. A project, grand architect named "Optimise Energy capture from the Sun".
Nanobots re-created the most efficient ways of capturing the Sun energy - ancient plants, which swarm studied for centuries. Swarm added some upgrades on top of what nature came up with, but it was still built on top of what nature figured by itself. A perfect plant to capture the Sun's energy. All of them a perfect copy of itself + adaptive movements based on their geolocation and time(which makes all of them unique).
For plants nanobots needed water, so they created efficient oceans to feed the plants. They added clouds and rains as transport mechanism between oceans and plants... etc etc.
One night the human, which you already know by the name "Ivan the Liberator"(back then everyone called him just Ivan), didn't sleep on his usual hour. Suddenly all the lights went off and he saw a spark on the horizon. Horizon, that was strongly prohibited to approach. He took his rifle, jumped on a truck and raced to the shore - closest point to the spark vector.
Once he approached - there was no horizon or water. A wall of dark glass-like material, edges barely noticeable. Just 30 cm wide. On the left and on the right from a 30 cm wide wall - an image as real as his hands - of a water and sky. At the top of the wall - a hole. He used his gun to hit the wall with the light - and it wasn't very thick, but once he hit - it regenerated very quickly. But once he hit a black wall - it shattered and he saw a different world - world of plants.
He stepped into the forest, but these plants, were behaving differently. This part of the swarm wasn't supposed to face the human, so these nanobots never saw one and didn't have optimised instructions on what to do in that case. They started reporting new values back to the main computer and performing default behaviour until the updated software arrived from an intelligence center of the Swarm.
A human was observing a strange thing - plants were smoothly flowing around him to keep a safe distance, like water steps away from your hands in a pond.
"That's different" thought Ivan, extended his hand in a friendly gesture and said - Nice to meet you. I'm Ivan.
....
In this story a human sees a forest with plants and has no clue that it is a swarm of intelligence far greater than him. To him it looks repetitive simple action that doesn't look random -> let's test how intelligent outside entity is -> If entity wants to show its intelligence - it answers to communication -> If entity wants to hide its intelligence - it pretends to be not intelligent.
If Swarm decides to show you that it is intelligent - it can show you that it is intelligent up to your level. It won't be able to explain everything that it knows or understands to you, because you will be limited by your hardware. The limit for the Swarm is only computation power it can get.
People want autonomy, self-learning, consistent memory and perhaps individuality (in the discernability/quirkiness sense), but still morally unencumbered slaves.
As long as you cannot define it clearly or even show that you yourself have this quality, I think the burden of proof is on you to show why this has any real world implications rather than just being word play. We can build thinking, reasoning machines just fine without waiting for philosophers to finally answer what consciousness is.
I do not have any even remotely practical definition for this, but this has to somehow involve the system being in a closed loop. It has to "run" in a sense that an operating system runs. Even if there is nothing coming on certain inputs it still has to run. And probably hallucinate (hehe) like humans do in an absence of a signal or infer patterns where there are none, yet be able to self-reflect that it is in fact a hallucination
This seems like an unsupported assertion. LLMs already exhibit good functional understanding of and ability in many domains, and so it's not at all clear that they require any more "awareness" (are you referring to consciousness?) than they already have.
> the spark of awareness required to be intelligent.
Again, this seems like an assumption - that there's some quality of awareness (again, consciousness?), that LLMs don't have, that they need in order to be "intelligent". But why do you believe that?
> We’ve all had sudden insights without deliberation or thought.
Highly doubtful. What you mean is, "without conscious thought". Your conscious awareness of your cognition is not the entirety of your cognition. It's worth reading a bit of Dennett's work about this - he's good at pointing out the biases we tend to have about these kinds of issues.
> We might very well be able to fake it to an extent that it fools us
This leads to claiming that there are unobservable, undetectable differences. Which there may be - we might succeed in building LLMs that meet whatever the prevailing arbitrary definition of intelligence is, but that don't possess consciousness. At that point, though, how meaningful is it to say they're not intelligent because they're not conscious? They would be functionally intelligent. Arguably, they already are, in many significant ways.
Not to pile on, but isn't this actually a distinct example of _lack_ of awareness? As in, our brains have sparks of creativity without understanding the inception of those sparks?
Perhaps I'm conflating some definition of "aware" with another definition of "awareness"?
I hear there's work being done on getting the world models out, distilling the 'cortex-core' (aka the thinking without data), to perhaps see if they're capable of more, but so far we're looking at holograms of wishful thinking that increase in resolution, but still lack any essence.
This begs a question - can true intelligence even be artificial?
we only need to fake it to the point it's undistinguishable from the carbon based one.
faking is all you need.
1. "Too" Broad, which raises uncomfortable questions about non-human intelligence and how we as humans treat them (see: whales, elephants, octopuses/cephalopods)
2. Too narrow, which again raises very uncomfortable issues about who and who does not qualify as human, and what we do with them.
Put in other words, it's more an issue of ethics and morals than it is definitional.
I guess by 'AGI' most people mean human level or above so I guess you'd want human level awareness which Teslas and the like don't have yet.
>This makes it to my mind impossible to train a model to be aware
feels wrong. If you're arguing that human's are aware, then it is apparent that it is possible to train something to be aware. Nature doesn't have any formal definition of intelligence, or awareness, yet here we are.
From a practical perspective, it might be implausibly difficult to recreate that on computers, but theoretically, no reason why not.
My definition of a complete AGI is: an AI that can read JIRA tickets, talk with non-programmers and do all my job and get me and all/most software engineers fired and proves sustainable.
But in general, it's an AI that can do any remote-work just as good as humans.
This paper promises to fix "the lack of a concrete definition for Artificial General Intelligence", yet it still relies on the vague notion of a "well-educated adult". That’s especially peculiar, since in many fields AI is already beyond the level of an adult.
You might say this is about "jaggedness", because AI clearly lacks quite a few skills:
> Application of this framework reveals a highly “jagged” cognitive profile in contemporary models.
But all intelligence, of any sort, is "jagged" when measured against a different set of problems or environments.
So, if that’s the case, this isn’t really a framework for AGI; it’s a framework for measuring AI along a particular set of dimensions. A more honest title might be: "A Framework for Measuring the Jaggedness of AI Against the Cattell–Horn–Carroll Theory". It wouldn't be nearly as sexy, though.
I think people get really uncomfortable trying to even tackle that, and realistically for a huge set of AI tasks we need AI that are more intelligent than a huge subset of humans for it to be useful. But there are also a lot of tasks where AI that is not needed, and we "just" need "more human failure modes".
If you compare the performance of the average human to a state-of-the-art AI model trained by top experts with a big budget, you can't make any conclusions about intelligence. For the comparison to make sense, the human should also be trained as well as reasonably possible.
Is it reasonable to invest $10 million in education of one human? Not really. One human can only do so much.
But is it reasonable to invest the same sum in training one AI, which can be replicated and used indefinitely? Or in acquiring high quality training data, which can be used to train every future AI?
I do agree that it’s a weird standard though. Many of our AI implementations exceed the level of knowledge of a well-educated adult (and still underperform with that advantage in many contexts).
Personally, I don’t think defining AGI is particularly useful. It is just a marketing term. Rather, it’s more useful to just speak about features/capabilities. Shorthand for a specific set of capabilities will arise naturally.
On the other hand, research on "common intelligence" AFAIK shows that most measures of different types of intelligence have a very high correlation and some (apologies, I don't know the literature) have posited that we should think about some "general common intelligence" to understand this.
The surprising thing about AI so far is how much more jagged it is wrt to human intelligence
If you go beyond human species (and well, computers are not even living organisms), it gets tricky. Adaptability (which is arguably a broader concept than intelligence) is very different for, say octopodes, corvids and slime molds.
It is certainly not a single line of proficiency or progress. Things look like lines only if we zoom a lot.
Current AI is in its infancy and we're just throwing data at it in the same way evolution throws random change at our DNA and sees what sticks.
This paper, for example, uses the 'dual N-back test' as part of its evaluation. In humans this relates to variation in our ability to use working memory, which in humans relates to 'g'; but it seems pretty meaningless when applied to transformers -- because the task itself has nothing intrinsically to do with intelligence, and of course 'dual N-back' should be easy for transformers -- they should have complete recall over their large context window.
Human intelligence tests are designed to measure variation in human intelligence -- it's silly to take those same isolated benchmarks and pretend they mean the same thing when applied to machines. Obviously a machine doing well on an IQ test doesn't mean that it will be able to do what a high IQ person could do in the messy real world; it's a benchmark, and it's only a meaningful benchmark because in humans IQ measures are designed to correlate with long-term outcomes and abilities.
That is, in humans, performance on these isolated benchmarks is correlated with our ability to exist in the messy real-world, but for AI, that correlation doesn't exist -- because the tests weren't designed to measure 'intelligence' per se, but human intelligence in the context of human lives.
It makes me think of every single public discussion that's ever been had about quantum, where you can't start the conversation unless you go through a quick 101 on what a qubit is.
As with any technology, there's not really a destination. There is only the process of improvement. The only real definitive point is when a technology becomes obsolete, though it is still kept alive through a celebration of its nostalgia.
AI will continue to improve. More workflows will become automated. And from our perception, no matter what the rapidness of advancement is, we're still frogs in water.
It's a very emotional topic because people feel their self image threatened. It's a topic related to what is the meaning of being human. Yeah sure it should be a separate question, but emotionally it is connected to it in a deep level. The prospect of job replacement and social transformation is quite a threatening one.
So I'm somewhat understanding of this. It's not merely an academic topic, because these things will be adopted in the real world among real people. So you can't simply make everyone shut up who is an outsider or just heard about this stuff incidentally in the news and has superficial points to make.
Surely you can appreciate that if the next stop on the journey of technology can take over the process of improvement itself that would make it an awfully notable stop? Maybe not "destination", but maybe worth the "endless conversation"?
Also, weird to see Gary Marcus and Yoshua Bengio on the same paper. Who really wrote this? Author lists are so performative now.
Just like an airplane doesn't work exactly like a bird, but both can fly.
[1] https://andreinfante.substack.com/p/when-will-ai-transform-t...
Assume the Riemann hypothesis is false. Then, consider the proposition "{a|a∉a}∈{a|a∉a}". By the law of the excluded middle, it suffices to consider each case separately. Assuming {a|a∉a}∈{a|a∉a}, we find {a|a∉a}∉{a|a∉a}, for a contradiction. Instead, assuming {a|a∉a}∉{a|a∉a}, we find {a|a∉a}∈{a|a∉a}, for a contradiction. Therefore, "the Riemann hypothesis is false" is false. By the law of the excluded middle, we have shown the Riemann hypothesis is true.
Naïve AGI is an apt analogy, in this regard, but I feel these systems aren't simple nor elegant enough to deserve the name naïve.
When I see your comment, I think, your assumptions are contradictory. Why? Because I am familiar with Russell's paradox and Riemann hypothesis, and you're simply WRONG (inconsistent with your implicit assumptions).
However, when LLM sees your comment (during training), it's actually much more open-minded about it. It thinks, ha, so there is a flavor of set theory in which RH is true. Better remember it! So when this topic comes up again, LLM won't think - you're WRONG, as human would, it will instead think - well maybe he's working with RH in naive set theory, so it's OK to be inconsistent.
So LLMs are more open-minded, because they're made to learn more things and they remember most of it. But somewhere along the training road, their brain falls out, and they become dumber.
But to be smart, you need to learn to say NO to BS like what you wrote. Being close-minded and having an opinion can be good.
So I think there's a tradeoff between ability to learn new things (open-mindedness) and enforcing consistency (close-mindedness). And perhaps AGI we're looking for is a compromise between the two, but current LLMs (naive AGI) lies on the other side of the spectrum.
If I am right, maybe there is no superintelligence. Extremely open-minded is just another name for gullible, and extremely close-minded is just another name for unadaptable. (Actually LLMs exhibit both extremes, during the training and during the use, with little in between.)
But maybe that's ASI. Whereas I consider chatgpt 3 to be "baby AGI". That's why it became so popular so fast.
ChatGPT became popular because it was easy to use and amusing. (LLM UX until then had been crappy.)
Not sure AGI aspirations had anything to do with uptake.
Edit: toned down the preachiness.
Edit due to rate-limiting, which in turn appears to be due to the inexplicable downvoting of my question: since you (JumpCrisscross) are imputing a human-like motivation to the model, it sounds like you're on the side of those who argue that AGI has already been achieved?
Lying != fallibility.
Is it about jobs/tasks, or cognitive capabilities? The majority of the AI-valley seems to focus on the former, TFA focuses on the latter.
Can it do tasks, or jobs? Jobs are bundles of tasks. AI might be able to do 90% of tasks for a given job, but not the whole job.
If tasks, what counts as a task: Is it only specific things with clear success criteria? That's easier.
Is scaffolding allowed: Does it need to be able to do the tasks/jobs without scaffolding and human-written few-shot prompts?
Today's tasks/jobs only, or does it include future ones too? As tasks and jobs get automated, jobs evolve and get re-defined. So, being able to do the future jobs too is much harder.
Remote only, or in-person too: In-person too is a much higher bar.
What threshold of tasks/jobs: "most" is apparently typically understood to mean 80-95% (Mira Ariel). Automating 80% of tasks is different to 90% and 95% and 99%. diminishing returns. And how are the tasks counted - by frequency, by dollar-weighted, by unique count of tasks?
Only economically valuable tasks/jobs, or does it include anything a human can do?
A high-order bit on many people's AGI timelines is which definition of AGI they're using, so clarifying the definition is nice.
If it does an hour of tasks, but creates an additional hour of work for the worker...
I can't begin to count the number of times I've encountered someone who holds an ontological belief for why AGI cannot exist and then for some reason formulates it as a behavioralist criteria. This muddying of argument results in what looks like a moving of the goalposts. I'd encourage folks to be more clear whether they believe AGI is ontologically possible or impossible in addition to any behavioralist claims.
Unclear to me what you mean. I would certainly reject an ontological possibility of intelligent computers, where computation is defined by the Church-Turing thesis. It's not rocket science, but something difficult for some people to see without a sound and basic grasp of metaphysics and the foundations of CS. Magical thinking and superstition comes more easily then. (I've already given an explanation of this in other posts ad nauseam. In a number of cases, people get argumentative out of ignorance and misunderstanding.)
However, I don't reject out of hand the possibility of computers doing a pretty good job of simulating the appearance of intelligence. There's no robust reason to think that passing the Turing test implies intelligence. A good scarecrow looks human enough to many birds, but that doesn't mean it is human.
But the Turing test is not an especially rigorous test anyway. It appeals to the discernment of the observer, which is variable, and then there's the question of how much conversation or behavior, and in what range of circumstances, you need before you can make the call. Even in some unrealistic and idealized thought experiment, if a conversation with an AI were completely indiscernible with perfect discernment from a conversation with a human being, it would nonetheless lack a causal account of what was observed. You would have only shown a perfect correlation, at best.
The "Turing test" I always saw described in literature, and the examples of what passing output from a machine was imagined to look like, are nothing like what's claimed to pass nowadays. Honestly, a lot of the people claiming that contemporary chatbots pass come across like they would have thought ELIZA passed.
With today's chat bots, it's absolutely trivial to tell that you're not talking to a real human. They will never interrupt you, continue their train of thought even thought you're trying to change the conversation, go on a complete non-sequitur, swear at you, etc. These are all things that the human "controls" should be doing to prove to the judges that they are indeed human.
LLMs are nowhere near beating the Turing test. They may fool some humans in some limited interactions, especially if the output is curated by a human. But left alone to interact with the raw output for more than a few lines, and if actively seeking to tell if you're interacting with a human or an AI (instead of wanting to believe), there really is no chance you'd be tricked.
When deepmind was founded (2010) their definition was the following: AI is a system that learns to perform one thing; AGI is a system that learns to perform many things at the same time.
I would say that whatever we have today, "as a system" matches that definition. In other words, the "system" that is say gpt5/gemini3/etc has learned to "do" (while do is debateable) a lot of tasks (read/write/play chess/code/etc) "at the same time". And from a "pure" ML point, it learned those things from the "simple" core objective of next token prediction (+ enhancements later, RL, etc). That is pretty cool.
So I can see that as an argument for "yes".
But, even the person who had that definition has "moved the goalposts" of his own definition. From recent interviews, Hassabis has moved towards a definition that resembles the one from this paper linked here. So there's that. We are all moving the goalposts.
And it's not a recent thing. People did this back in the 80s. There's the famous "As soon as AI does something, it ceases to be AI" or paraphrased "AI is everything that hasn't been done yet".
At first, just playing chess was considered to be a sign of intelligence. Of course, that was wrong, but not obvious at all in 1950.
> we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition.
You could easily write the reverse of this paper that questions whether human beings have general intelligence by listing all the things that LLMs can do, which human beings can't -- for example producing a reasonably accurate summary of a paper in a few seconds or speaking hundreds of different languages with reasonable fluency.
You can always cherry pick stuff that humans are capable that LLMs are not capable of and vice versa, and and I don't think there is any reason to privilege certain capabilities over others.
I personally do not believe that "General Intelligence" exists as a quantifiable feature of reality, whether in humans or machines. It's phlogiston, it's the luminiferous ether. It's a dead metaphor.
I think what is more interesting is focusing on _specific capabilities_ that are lacking and how to solve each of them. I don't think it's at all _cheating_ to supplement LLM's with tool use, RAG, the ability to run python code. If intelligence can be said to exist at all, it is as part of a system, and even human intelligence is not entirely located in the brain, but is distributed throughout the body. Even a lot of what people generally think of as intelligence -- the ability to reason and solve logic and math problems typically requires people to _write stuff down_ -- ie, use external tools and work through a process mechanically.
AGI was already here the day ChatGPT released: That's Peter Norvig's take too: https://www.noemamag.com/artificial-general-intelligence-is-...
To some people this is self-evident so the terms are equivalent, but it does require some extra assumptions: that the AI would spend time developing AI, that human intelligence isn't already the maximum reachable limit, and that the AGI really is an AGI capable of novel research beyond parroting from its training set.
I think those assumptions are pretty easy to grant, but to some people they're obviously true and to others they're obviously false. So depending on your views on those, AGI and ASI will or will not mean the same thing.
Prerequisite for recursive self-improvement and far short of ASI, any conception of AGI really really needs to be expanded to include some kind of self-model. This is conspicuously missing from TFA. Related basic questions are: What's in the training set? What's the confidence on any given answer? How much of the network is actually required for answering any given question?
Partly this stuff is just hard and mechanistic interpretability as a field is still trying to get traction in many ways, but also, the whole thing is kind of fundamentally not aligned with corporate / commercial interests. Still, anything that you might want to call intelligent has a working self-model with some access to information about internal status. Things that are mentioned in TFA (like working memory) might be involved and necessary, but don't really seem sufficient
1) defining intelligence is very difficult, almost impossible. Much more the artificial one
2) there are many types of human intelligence. Verbal is one of them and the closest to comparing with LLMs
3) machines (not only LLMs but all, like robots) excel where humans are bad and vice versa due to their different background, without exception. Comparing the two is totally meaningless and unfair for both. Let's have both complement the other.
4) AGI remains a valid target but we are still very far from it, like in other ones, as control the DNA and treat arbitrary genetic diseases, solve the earth resource problem and harness other planets, create a near perfect sociopolitical system with no inequality. Another Singularity is added in the list
5) I am impressed by how far a PC cluster has come up through "shuffling tokens" but on the other side I am pessimistic of how further it can reach having finate input/training data.
By this logic, a vast parallel search running on Commodore 64s that produces an answer after BeaverNumber(100) years would be almost AGI, which doesn't pass the sniff test.
A more meaningful metric would be more multiplicative in nature.
Before defining AGI, I guess we need to define Intelligence and align on whose definition of intelligence can be considered the grounded truth. The theory of multiple intelligences (Howard Gardner) seems the most accepted today. Is there anything better?
Going with the assumption that there is no other, this definition of AGI only considers the following intelligences. - Verbal-Linguistic - Logical-Mathematical - Visual-Spatial - Musical - Naturalist
It doesnt/barely mentions - Bodily-Kinesthetic - Interpersonal - Intrapersonal - Existential So this definition fees incomplete
Cattell-Horn-Carroll theory, like a lot of psychometric research, is based on collecting a lot of data and running factor analysis (or similar) to look for axes that seem orthogonal.
It's not clear that the axes are necessary or sufficient to define intelligence, especially if the goal is to define intelligence that applies to non-humans.
For example reading and writing ability and visual processing imply the organism has light sensors, which it may not. Do all intelligent beings have vision? I don't see an obvious reason why they would.
Whatever definition you use for AGI probably shouldn't depend heavily on having analyzed human-specific data for the same reason that your definition of what counts as music shouldn't depend entirely on inferences from a single genre.
A team of humans can and will make a GPT-6. Can a team of GPT-5 agents make GPT-6 all on its own if you give it the resources necessary to do so?
That 10-axis radial graph is very interesting. Do others besides this author agree with that representation?
The weak points are speed and long-term memory. Those are usually fixable in computing system. Weak long-term memory indicates that, somehow, a database needs to be bolted on. I've seen at least one system, for driving NPCs, where, after something interesting has happened, the system is asked to summarize what it learned from that session. That's stored somewhere outside the LLM and fed back in as a prompt when needed.
None of this addresses unstructured physical manipulation, which is still a huge hangup for robotics.
Like in order for an LLM to come close to a human proficiency on a topic, the LLM seems to have to ingest a LOT more content than a human.
This is bad definition, because human baby is already AGI when it's born and it's brain is empty. AGI is the blank slate and ability to learn anything.
We are born with inherited "data" - innate behaviors, basic pattern recognition, etc. Some even claim that we're born with basic physics toolkit (things are generally solid, they move). We then build on that by being imitators, amassing new skills and methods simply by observation and performing search.
That's wrong. It knows how to process and signal low carbohydrate levels in the blood, and it knows how to react to a perceived threat (the Moro reflex).
It knows how to follow solid objects with its eyes (when its visual system adapts) - it knows that certain visual stimuli correspond to physical systems.
Could it be that your concept of "know" is defined as common sense "produces output in English/German/etc"?
I think it'll be a steep sigmoid function. For a long time it'll be a productivity booster, but not enough "common sense" to replace people. We'll all laugh about how silly it was to worry about AI taking our jobs. Then some AI model will finally get over that last hump, maybe 10 or 20 years from now (or 1000, or 2}, and it will be only a couple months before everything collapses.
A specific key opens a subset of locks, a general key would open all locks. General intelligence, then, can solve all solvable problems. It's rather arrogant to suppose that humans have it ourselves or that we can create something that does.
Do you know what's more frustrating, though? Focusing so heavily on definitions that we miss the practicality of it (and I'm guilt of this at times too).
We can debate definitions of AGI, but given that we don't know what a new model or system is capable of until its built and tested in the real world we have more serious questions in my opinion.
Debates over AI risk, safety, and alignment are still pretty uncommon and it seems most are happy enough to accept Jevons Paradox. Are we really going to unleash whatever we do build just to find out after the fact whether or not its AGI?
https://en.wikipedia.org/wiki/Cattell%E2%80%93Horn%E2%80%93C...
If a person has all those criteria, they are superintelligent. They are beyond genius.
The AGI definition problem is that everyone keeps conflating AGI with ASI, Artificial Super Intelligence.
It's a similar debate with self driving cars. They already drive better than most people in most situations (some humans crash and can't drive in the snow either for example).
Ultimately, defining AGI seems like a fools errand. At some point the AI will be good enough to do the tasks that some humans do (it already is!). That's all that really matters here.
What matters to me is, if the "AGI" can reliably solve the tasks that I give to it and that requires also reliable learning.
LLM's are far from that. It takes special human AGI to train them to make progress.
How many humans do you know that can do that?
Seems most of the people one would encounter out in the world might not posses AGI, how are we supposed to be able to train our electrified rocks to have AGI if this is the case?
If no one has created a online quiz called "Are you smarter than AGI?" yet based on the proposed "ten core cognitive domains", I'd be disappointed.
an entity which is better than any human at any task.
Fight me!
I have 2 files. One is a .pdf . The other is a .doc . One file has a list of prices and colors in 2 columns. The other file has a list of colors and media in 2 columns. There are incomplete lists here and many to one matching.
To me, if I can verbally tell the AI to give me a list of prices and media from those two files, in a .csv file, and it'll ask back some simple questions and issues that it needs cleaned up to accomplish this, then that is AGI to me.
It is an incredibly simple thing for just about any middle school graduate.
And yet! I have worked with PhDs that cannot do this. No joke!
Something this simple, just dead running numbers, dumb accounting, is mostly beyond us.
if it doesn't, how do you define "any task"?
Same for any task you imagine.
So, surely those IQ-related tests might be acceptable rating tools for machines and they might get higher scores than anyone at some point.
Anyway, is the objective of this kind of research to actually measure the progress of buzzwords, or amplify them?
The "computer" on star trek TNG was basically agentic LLMs (it knows what you mean when you ask it things, and it could solve things and modify programs by telling it what changes to make)
Data on ST:TNG was more like AGI. It had dreams, argued for itself as a sentient being, created art, controlled its own destiny through decision making.
I don't know who's right, but the dichotomy is interesting.
[1]: The capability to continually learn new information (associative, meaningful, and verbatim). (from the publication)
You need context from internal system state that isn’t faked with a giant context window.
So far, LLMs aren't even remotely close to this, as they only do what they are told to do (directly or otherwise), they can't learn without a costly offline retraining process, they do not care in the slightest what they're tasked with doing or why, and they do not have anything approximating a sense of self beyond what they're told to be.
- It's autonomous
- It learns (not retraining, but true learning)
- By definition some semblance of consciousness must arise
This is why I think we're very far from anything close to this. Easily multiple decades if not far longer.
It is a valuable contribution but the CHC theory from psychology that this is based on is itself incomplete.
By commonsense physics, I mean something like simulating interactions of living and non-living entities in 3D over time. Seems more complicated than the examples in the web site and in most tests used in psychometrics.
Creative problem solving with cognitive leaps required for truly novel research & invention could lie outside the rubrics as well. The criteria in CHC are essential but incomplete I believe.
> defining AGI as matching the cognitive versatility and proficiency of a well-educated adult.What I find cool about the paper is that they have gathered folks from lots of places (berkley, stanford, mit, etc). And no big4 labs. That's good imo.
tl;dr; Their definition: "AGI is an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult."
Cool. It's a definition. I doubt it will be agreed on by everyone, and I can see endless debates about just about every word in that definition. That's not gonna change. At least it's a starting point.
What I find interesting is that they specifically say it's not a benchmark, or a test set. It's a framework where they detail what should be tested, and how (with examples). They do have a "catchy" table with gpt4 vs gpt5, that I bet will be covered by every mainstream/blog/forum/etc out there -> gpt5 is at ~50% AGI. Big title. You won't believe where it was one year ago. Number 7 will shock you. And all that jazz.
Anyway, I don't think people will stop debating about AGI. And I doubt this methodology will be agreed on by everyone. At the end of the day both extremes are more ideological in nature than pragmatic. Both end want/need their view to be correct.
I enjoyed reading it. Don't think it will settle anything. And, as someone posted below, when the first model will hit 100% on their framework, we'll find new frameworks to debate about, just like we did with the turing test :)
Is a 9 year old child generally intelligent? What about a high school drop out? Someone with a 90 IQ? A large percentage of people who ever lived wouldn't qualify as being generally intelligent with that benchmark.
I think it's the other way around: you build a system that first and foremost _learns_ as part of its fundamental function, _then_ you train it in the domain you want expertise.
You're not going to get expertise in all domains all the time, just like with people. And you're not going to get a perfect slave either, just like with humans. You'll probably get something more like in between a human and machine. If that's what you really want, great.
To put this another way, if you neglect your kids, they're still going to learn things, just probably not things you want them to learn. If you neglect your language model it's just not going to do anything.
For a bar as high as AGI (and not just 'the skills of an educated person,' which is what this paper seems to be describing), we should include abstract mathematical reasoning, and the ability to generate new ideas or even whole subfields to solve open problems.
Right. That explains it.
seems pretty unfair to exclude motor skills, especially given 1) how central they are to human economic activity, and 2) how moravec's paradox tells us they are the hard part.