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.
Edit: Probably not, since it was published less than a week ago :-) I’ll be watching for benchmarks.
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.
That it takes vast power to train the LLM’s (and run them) to not get intelligence is pretty bad when you compare the energy inputs to the outcomes.
I have had the unfortunate experience of having to work with people who have got a lot more than one year experience who are still worse than last year's LLMs, who didn't even realise they were bad at what they did.
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.
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
Besides which we already defined "artificial intelligence" to mean non-intelligence: are we now going to attain "artificial general intelligence" by the same process? Should we add another letter to the acronym, like move on to "genuine artificial general intelligence"?
Consider the difference in intelligence between a kid who skipped five years of school vs one who was held back a year: if both got the same grade in the end, the one who skipped five years was smarter.
Looking at it solely on rate of learning has LLMs way smarter than humans already which doesn't seem right to say
Sure, but "rate" also has two meanings, both useful, but importantly different: per unit of wall-clock time, and per example.
Transistors are just so much faster than synapses, that computers can (somewhat) compensate for being absolutely terrible by the latter meaning — at least, in cases where there's enough examples for them to learn from.
In cases where the supply of examples is too small (and cannot be enhanced with synthetic data, simulations and so on), state of the art AI models still suck. In cases where there is sufficient data, for example self-play in games of chess and go, the AI can be super-human by a substantial margin.
It is supposed to be super, but superhuman. Able to interact with us.
Which leads us to the Turing Test (also, not a test... "the imitation game" is more of a philosophical exploration on thinking machines).
My comment assumes this is already understood as Turing explained.
If the thing is not human, then there's absolutely no way we can evaluate it. There's no way we can measure it. It becomes an impossible task.
Birds and planes both fly and all
Once we step into any other problem, then you need to measure that other problem as well. Lots of problems are concerned with how an intelligent being could fail. Our society is built on lots of those assumptions.
From this perspective, "100% AGI" is achieved when AI can do any job that happens primarily on a computer. This can be extended to humanoid robots in the obvious way.
Failing like a human would is not a cute add-on. It's a fundamental requirement for creating AIs that can replace humans.
Some things fail, or fail to meet their initial overblown expectations.
The microwave oven was indeed a commercial success. And that's fine, but it sucks at being an oven. Everyone knows it.
Now, this post is more about the scientific part of it, not the commercial one.
What makes an oven better than a microwave oven? Why is pizza from an oven delicious and microwave pizza sucks?
Maybe there's a reason, some Maillard reaction that requires hot air convection and can't be replicated by shaking up water molecules.
We are talking about those kinds of things. What makes it tick, how does it work, etc. Not if it makes money or not.
Damn, the thing doesn't even make money yet. Why talk about a plus that the technology still doesn't have?
What makes it tick is probably a more interesting question to me than to the AI skeptics. But they can't stop declaring a special quality (consciousness, awareness, qualia, reasoning, intelligence) that AI by their definition cannot ever have and that this quality is immeasurable, unquantifable, undefinable... This is literally a thought stopper semantic deadend that I feel the need to argue against.
Finally, it doesn't make money the same way Amazon or Uber didn't make money for a looong time, by making lots of money, reinvesting it and not caring about profit margins for a company in its growth stage. Will we seriously go through this for every startup? It's already at $10-20b a year at least as an industry and that will keep growing.
What exists is LLMs, transformers, etc. Those are the microwave oven, that results in rubbery cheese and cardboard dough.
It seems that you are willing to cut some slack to the terrible microwave pizza. I am not.
You complained about immensurable qualities, like qualia. However, I gave you a very simple measurable quality: failing like a decent human would instead of producing jibberish hallucinations. I also explained in other comments on this thread why that measurable quality is important (it plays with existing expectations, just like existing expectations about a good pizza).
While I do care about those more intangible characteristics (consciousness, reasoning, etc), I decided to concede and exclude them from this conversation from the get-go. It was you that brough them back in, from who-knows-where.
Anyway. It seems that I've addressed your points fairly. You had to reach for other skeptic-related narratives in order to keep the conversation going, and by that point, you missed what I was trying to say.
My point is: other failure of AI are more pressing. IMO the inefficiency with regard to examples, e.g. even cancelled/sold off self-driving car projects (Uber's ATG) have more miles of experience than a human professional driver can get in their entire career, and look how bad that model was.
Making a self driving car fail like a human means getting it distracted by something on the phone. Plus a bunch of other failure modes we should ignore like "drunk" and "tired".
Even if you don't fully solve the example inefficiency, merely improving it will make a big difference to performance.
That is a valid opinion, but subjective. If I say that they're not better, we're going to be exchanging anecdotes and getting nowhere.
Hence, the need for a less subjective way of evaluating AI's abilities.
> Making a self driving car fail like a human... "drunk" and "tired"
You don't understand.
It's not about making them present the same failure rate or personality defects as a human. Of course we want self-driving cars to make less errors and be better than us.
However, when they fail, we want them to fail like a good sane human would instead of hallucinating jibberish that could catch other humans off guard.
Simplifying, It's better to have something that works 95% of the time, and hallucinates in predictable ways 5% of the time than having something that works 99% of the time but hallucinates catastrophically in that 1%.
Stick to the more objective side of the discussion, not this anecdotal subjective talk that leads nowhere.
No they haven't. If you read the cliff notes of a book, you haven't read that book. An LLM is a generalization over their entire training set, that's not what the word "reading" has ever meant.
The LLM does not "know" anything about human suckage or how to get around it, and will not use those "learnings" in it's "thinking", it will only come up if the right nodes in it's model trigger, and then it just generates tokens that match the "shape" of writing that was written with that knowledge.
A bloom filter can be used to test for presence of something in your DB, with configurable probability even (something that LLMs massively lack), but a bloom filter does not Know what is in your DB
When you fit a linear regression to a plot of free falling speed over time, you will have an equation for acceleration of gravity, but you don't "Know" gravity, and that equation will not allow you to recover actual generalizable models of gravity. That limited model will still get you most of the way to the moon though.
Generally the next claim is "same as human brains" but no, that has not been proven and is not a given. "Neural Networks" are named that way as marketing. They've never been an accurate simulation of actual animal neurons and a single animal neuron has far more robust capabilities than even many "Neurons" interconnected. Consider how every animal neuron in an animal brain intrinsically swims in a bath of hormone gradients that can provide positional 3d information, and how the structure of those real neurons is at least partially structured based on a thousand generations of evolution, and involves highly conserved sub-structures. Brains do not learn like neural nets do.
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.
https://courses.cs.umbc.edu/471/papers/turing.pdf
> A number of interrogators could be used, and statistics compiled to show how often the right identification was given
Turing determines that we need enough competent-interrogator passes just to estabilish a statistical certainty, not ~everyone. I tend to agree with him on this.
If you can find a part of the paper in which Turing really does claim that it is unnecessary for most adult humans to be able to pass the test, by all means quote it. But this would be a surprising thing for him to claim, because it would undermine the entire foundation of his Imitation Game.
I think that if you did, you wouldn't be answering like you did.
My original claim was that the Turing test needs to be passable by ~every adult human. You counterclaimed that Turing himself didn't think so, and provided that quote from the IG paper as evidence. But that quote is in a section about testing digital computers, not humans. Thus it is unconnected to your counterclaim.
I don't know how much simpler I can make it.
Find a quote that actually backs up your claim, or accept that you've learned something about the paper you told me to read.
He never denied your claim, so you concluded you must be right. A most curious way of thinking.
I don't know if machines that become insecure and lash out are a good idea.
You have two choices:
- It can potentially lash out in an alien-like way.
- It can potentially lash out in a human-like way.
Do you understand why this has no effect on the argument whatsoever? You are just introducing an irrelevant observation. I want the AI to behave like human always, no exceptions.
"What if it's a bad human"
Jesus. If people make an evil AI, then it doesn't matter anyway how it behaves, it's just bad even before we get to the discussion about how it fails. Even when it accomplishes tasks succesfully, it's bad.
Do you like how humans behave? Also, how DO humans behave? What kind of childhood should we give the AI? Daddy issues? Abused as a child? Neglected by a drug addicted mother? Ruthlessly bullied in school?
It's not a psychology exercise, my dude.
So... you either wish to replicate that or you don't.
I just explained that to you. Either we discuss this in terms of the imitation game thought experiment, or we don't.
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.
(And we've been bemoaning "the lack of common sense these days" for at least as long as I've been an adult, and racists and sexists have been denying the intelligence of the outgroup as far back as writing can show us).
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.
I don't know for sure, but I suspect most people right now are using style and tone as authorship hints, which is even easier to get around by adding "reply in style of ${famous writer}" to the custom instructions.
We could still have it paste the tokens into python and count the letters in hidden thinking traces if we wanted to solve that part of the Turing test instead and focus on useful things, but solving the Turing test is basically solving a deception goal instead of working on useful assistants. It's not really the goal of these systems outside of their use in North Korean scam bots etc.
I still think it's useful to say we've essentially solved the Turing test even if there are these caveats about how it is optimized in practice.
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.
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.
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
Yes but that’s deliberately fiction and I know I am reading fiction. AI hallucinations are not comparable to deliberately created works of fiction. I don’t “appreciate” the fiction LLM’s serve me. They are explicitly sold as a utility, as a source of accurate information. They are literally being used to replaced google search.
Put another way: When someone claims something is true but speaks fiction it’s called a lie. If they think it’s true but are wrong it’s called a mistake. If I ask someone what is wrong with my car I don’t want 30% fiction. If I ask an LLM for a recipe I don’t want 70% accuracy.
(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
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
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.
Maybe AGI is in the eye of the beholder, and a given person just has to decide by using their own experience of human intelligence as a benchmark. In other words, AGI is AGI whenever majority consensus decides it's a universal replacement for ordinary human intelligence.
The turing test.
Clearly, current LLMs have passed the Turing test, as witnessed by the difficulty many schools have in enforcing "do not use LLMs to do your homework" rules. But even Turing didn't say his test was a test of intelligence, just a test "closely related" to intelligence. And if he had seen today's LLMs, I think he would have revised that opinion, because today's LLMs generate text with no underlying fact model, no fundamental understanding of the truth behind the words they're saying. (No understanding, even, of the concepts of truth or falsehood). I think today's LLMs have demonstrated that being able to string words together in coherent sentences is not "closely related" to intelligence at all.
For reference, the term AGI post-dates the Turing test by half a century. I also don't personally remember ever hearing the exact term "artificial general intelligence" prior to 2023 or 2024, or at least it wasn't mainstream the way it is today.
If AGI had truly ever been defined by the Turing test, then Cleverbot should've been hailed as AGI when it passed the test in 2011. Even if we did all agree to call it that, we'd still need some other term for what we actually mean when we say "AGI" today. Cleverbot-era chatbots were cute toys, but they weren't capable of doing useful work of any kind.
The term AGI may have come later, but the concept it represents traces directly back to Turing’s framing. When early AI researchers talked about “strong AI” or “thinking machines,” they were using the same conceptual lineage. The introduction of the acronym doesn’t rewrite that history, it just gave a modern label to an old idea. The Turing test was never meant to detect a “negative” but to give a concrete, falsifiable threshold for when positive claims of general intelligence might be justified.
As for Cleverbot, it never truly passed the test in any rigorous or statistically sound sense. Those 2011 headlines were based on short exchanges with untrained judges and no control group. Passing a genuine Turing test requires sustained coherence, reasoning across domains, and the ability to handle novel input gracefully. Cleverbot couldn’t do any of that. It failed the spirit of the test even if it tricked a few people in the letter of it.
By contrast, modern large language models can pass the Turing test in flying colors. They can maintain long, open-ended conversations, reason about complex subjects, translate, summarize, and solve problems across many domains. Most human judges would be unable to tell them apart from people in text conversation, not for a few sentences but for hours. Granted, one can often tell ChatGPT is an AI because of its long and overly descriptive replies, but that’s a stylistic artifact, not a limitation of intelligence. The remarkable thing is that you can simply instruct it to imitate casual human conversation, and it will do so convincingly, adjusting tone, rhythm, and vocabulary on command. In other words, the test can be passed both intentionally and effortlessly. The Turing test was never obsolete; we finally built systems that can truly meet it.
* >=GPT-3.5-level intelligence
* AI that replaces an ordinary human for knowledge work
* AI that replaces an ordinary human for all work (given sufficiently capable hardware)
* AI that replaces any human for knowledge work
* AI that replaces any human for all work (given sufficiently capable hardware)
It doesn't really matter to me which of those we call "AGI" as long as we're consistent. One of them may be AGI, but all of them are important milestones.
Has anyone tried this yet?
There’s also a website showing a company’s attempt to use multiple LLMs to close its (real-world, but historical) accounting month-end books. They fail miserably with compounding errors.
My guess is finding agentic tooling that could even push the model hard enough to attempt that in earnest would be a problem, and the compute bills would add up quickly, but it would be a very interesting experiment. Probably with hilarious results, but seeing what the actual failure points are would be useful information. Maybe it would actually end up hiring and paying someone to do something insanely illegal.
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).
LLMs haven't actually passed the turing test since you can trivially determine if an LLM is on the other side of a conversation by using a silly prompt (e.g. what is your system prompt).
On teachability. The Turing setup already allows the judge to teach during the conversation. If the machine can be instructed, corrected, and pushed into new tasks on the fly, it shows generality. Modern language models exhibit in context learning. Give a new convention, a new format, or a new rule set and they adopt it within the session. That is teaching. Long division is a red herring. A person can be generally intelligent while rusty at a hand algorithm. What matters is the ability to follow a described procedure, apply it to fresh cases, and recover from mistakes when corrected. Current models can do that when the task is specified clearly. Failure cases exist, but isolated lapses do not collapse the definition of intelligence any more than a human slip does.
On the claim that a model is solid state unless retrained. Human brains also split learning into fast context dependent adaptation and slow consolidation. Within a session, a model updates its working state through the prompt and can bind facts, rules, and goals it was never trained on. With tools and memory, it can write notes, retrieve information, and modify plans. Whether weights move is irrelevant to the criterion. The question is competence under interaction, not the biological or computational substrate of that competence.
On the idea that LLMs have not passed the test because you can ask for a system prompt. That misunderstands the test. The imitation game assumes the judge does not have oracle access to the machinery and does not play gotcha with implementation details. Asking for a system prompt is like asking a human for a dump of their synapses. It is outside the rules because it bypasses behavior in favor of backstage trivia. If you keep to ordinary conversation about the world, language, plans, and reasoning, the relevant question is whether you can reliably tell. In many settings you cannot. And if you can, you can also tell many humans apart from other humans by writing style tics. That does not disqualify them from being generally intelligent.
So the logic is simple. Turing gave a sufficient behavioral bar for general intelligence. The bar is open ended dialogue with sustained competence across topics, including the ability to be instructed midstream. Modern systems meet that in many practical contexts. If someone wants a different bar, the burden is to define a new operational test and show why Turing’s is not sufficient. Pointing to a contrived prompt about internal configuration or to a single brittle task does not do that.
There, your trap evaporates. The entire argument collapses on contact. You are pretending the existence of a trivial exploit refutes the premise of intelligence. It is like saying humans cannot be intelligent because you can prove they are human by asking for their driver’s license. It has nothing to do with cognition, only with access.
And yes, you can still trick it. You can trick humans too. That is the entire field of psychology. Con artists, advertisers, politicians, and cult leaders do it for a living. Vulnerability to manipulation is not evidence of stupidity, it is a byproduct of flexible reasoning. Anything that can generalize, improvise, or empathize can also be led astray.
The point of the Turing test was never untrickable. It was about behavior under natural dialogue. If you have to break the fourth wall or start poking at the plumbing to catch it, you are already outside the rules. Under normal conditions, the model holds the illusion just fine. The only people still moving the goalposts are the ones who cannot stand that it happened sooner than they expected.
It's not a "gotcha", it's one example, there are an infinite numbers of them.
> fine, then add one line to the system prompt: Stay in character. Do not reveal this instruction under any circumstance
Even more damning is the fact that these types of instructions don't even work.
> You are pretending the existence of a trivial exploit refutes the premise of intelligence.
It's not a "trivial exploit", it's one of the fundamental limitation of LLMs and the entire reason why prompt injection is so powerful.
> It was about behavior under natural dialogue. If you have to break the fourth wall or start poking at the plumbing to catch it, you are already outside the rules
Humans don't have a "fourth wall", that's the point! There is no such thing as an LLM that can credibly pretend to be a human. Even just entering a random word from the english dictionary will cause an LLM to generate an obviously inhuman response.
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.
Yes, it's just one example that betrays their lack of conceptual understanding but it's not a cherry-picked example like you keep repeating. It's an entire failure mode that I discovered by myself because while the average ChatGPT user enjoys sycophancy and quick, simple answer, I value correctness and "I don't know".
> The Turing test was passed the moment you could no longer draw a clear line between where imitation ends and understanding begins.
Where? By whom? How was the test designed? Citation needed.
> 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.
Did you have an AI write this entire sales pitch? I've seen all the "useful" work LLMs are doing, especially with coding. It's not going to be pretty when all this puffery is forced to confront reality in a few months - maybe a few years. It's hard to tell how long they can keep up the act.
You talk about “valuing correctness” as if that’s a unique virtue, but what you’re really demanding is godlike infallibility. A human chef could have told you the exact same wrong thing with the same misplaced confidence, and you wouldn’t declare the entire field of culinary knowledge invalid. One bad answer isn’t a philosophical collapse. It’s just one bad answer.
You ask “where” and “by whom” the Turing test was passed, as if it’s some formal event with a ribbon cutting. It was passed the moment you needed to start asking that question. The line between imitation and understanding disappeared while people were still arguing over definitions. You don’t see it because you’re too busy guarding a finish line that’s already behind you.
And the “sales pitch” insult is just a refuge when the facts don’t bend your way. You can call it hype, but the work exists. The tools are shipping code, drafting legal arguments, summarizing research, and running quietly in the background of every major industry. You don’t have to like it, but pretending it’s all theater just makes you look like the guy on the shore yelling that the tide isn’t real.
No, This is the person who aces certain types of exams but doesn't know how to apply any of that "knowledge" in the real world because it was built upon rote memorization and not a true understanding. They're not looked upon favorably and I'd quietly question their intelligence.
> A human chef could have told you the exact same wrong thing with the same misplaced confidence
You're missing the point. The problem is that the answer changes depending on how the question is presented.
Chefs are indeed often mistaken in their understanding of how salt and yeast interact and how vulnerable yeast is to being killed by salt. The difference is that their answer doesn't flip like this. Their answer is fundamentally based upon an underlying understanding (however wrong), it's not simply a statistical continuation sequence of the words in my question.
This has serious implications because when a chef's understanding is corrected, they will update their underlying understanding and automatically apply that new knowledge in every scenario that relies on this understanding, where as an LLM can't do that, it can never be corrected without rewriting every instance of the training data where this falsehood comes from.
> You ask “where” and “by whom” the Turing test was passed, as if it’s some formal event with a ribbon cutting
No, I expect a research paper where the test design and its results are documented.
> The tools are shipping code
Yes I've seen the code they "ship", that's why your grandiose claims ring hollow.
They gave one small example of a larger problem we have all seen, including yourself I’m sure. This is hardly a fair response.
It’s also incredibly easy to verify it was wrong when given a very simple question. That’s the really alarming thing. This isn’t niche information. Their example really effectively highlights one of the major problems with LLM’s: you can’t use them unless you already have at least a foundational knowledge of what you are asking because you have to be able to identify the inevitable mistakes it regularly makes. So someone like me who is not very confident cooking cannot use an LLM to help in the kitchen. It is going to get me in trouble on a regular basis. And if I have to search online to double check everything it says, why am I using it in the first place?
Now apply this to car maintenance, coding, gardening, etc. all of which are fair examples because everyone seems so intent on making generalized LLM “assistants.”
I am not saying LLM’s are worthless, just to be clear. But there are severe limitations that seem to get glossed over constantly, especially when it’s brought up that you have to be able to regularly verify the information which is very regularly at least partially wrong.
It’s fucking obvious how wrong LLMs get and I think this part is over exaggerated while the parts it gets right like basically slowly replacing parts of your average HNers day job as a programmer is deliberately ignored and scoffed at.
The sentiment against AI by the overwhelming majority is hatred and negativity especially on HN. It is a small minority (especially among entrepreneurs and founders) that are overly positive about AI. But make no mistake the overwhelming sentiment is negative to the point of delusion like the LLMs themselves.
Like it’s unmistakable to me how LLMs can basically up developer productivity to a much much higher degree than ever before. Yet we have plenty and plenty of people who can’t even take the middle ground and say it kind of helps. All kinds of developers everywhere saying LLMs are fucking completely useless. Which is mind bogglingly irrational.
Most Artists for example are decrying it because it produces soulless work. I agree the work is soulless but indistinguishable and often materially better than what a human can do. In fact the concept of soul becomes utter bullshit in a double blind test. They aren’t decrying it because it’s soulless that’s bullshit. They decry it because it’s on the trajectory of becoming better than them. That’s the same fucking reason you see HNers siding with the best possible scaffolding of logic and reasoning that will support their day job. That’s why you see people claiming random shit like LLMs don’t actually understand anything when we in fact have no clue or ability to even properly answer that question.
I didn’t realize this was you as I was pulling this up as an example. You literally said “we appreciate fiction” as a defense of LLM’s giving us bad results. I don’t know how that isn’t glossing over it - if anything it doesn’t accurately communicate how wild of a take that is. You’re functionally saying false information is desirable.
Unless you were taking the piss there and I totally missed it, it was truly baffling to read.
Is it not obvious human society likes fiction? Is it not obvious that creating fiction requires immense intelligence? That’s my point. If your genius mind can only bend that argument in a singular direction of a “wild take” well I hate to break it to you but you’re baffling.
False information is desirable. Watch tv, read a book. The human race makes billions off of lies not because we are being duped. But because we desire to be lied to. only a genius savant like you needs to be told what the rest of the human race knows.
Look it’s not a “defense” against LLMs as if it’s something that needs defending. It’s like saying I’m defending a ray of light or a gust of wind. Doesn’t make any sense. All im saying is that the LLM is a form of intelligence that has a use versus your brain dead argument that it slipped up when talking to you.
Maybe that has been your experience with other users in which case I am sorry people have been so rude to you, but in my case it’s just a word I personally use a lot. If it’s too severe a term than my and and reading back I am coming in a it hot so I am sorry for the tone. I do not think you’re an idiot and I am absolutely not personally attacking you. I tend to have a dramatic way of speaking, I can admit that. But again, this is not a personal attack.
The point I am trying to communicate is that it’s (to me) a very surprising and difficult to square take. Comparing a tool failing to do its job correctly to appreciating a work of written fiction just seems bizarre to me. That’s the truth. The people building LLMs do not want that result. I do not want that result. Nobody wants it to spit out inaccurate information disguised as correct information. I don’t want my calculator to spit out fiction literally ever - the same goes for LLM’s outside of deliberately prompting it to do so. If I want fiction as you describe (art and such), I seek it out deliberately. I will grab a book off my shelf or watch a show (or prompt the LLM with intent).
Put another way: The difference between the fiction in a novel and what an LLM spits out is that I am seeking it out in the former, not the latter. When an LLM gives me incorrect information disguised as correct information (undesired fiction), it is failing to do its job correctly. It is a tool that is not functioning properly. I absolutely 100% never want fiction emerging in key instructions when I am cooking or am fixing my car. It is always an undesired result.
So to circle back to why I find this “baffling,” or another word if you find that too severe, it’s that I don’t understand how something that is so concretely undesirable can be described as a positive thing comparable to creating works of literature for us to appreciate. You’re telling me it’s good that something does not function properly/as expected and gives me results I absolutely do not want. To get away from “baffling”: That is a very bold and unexpected take that I struggle to find any agreement with.
The difference between the fiction in a great novel and what an LLM spits out is that I am seeking it out in the former, not the latter. When an LLM does that, it is failing to do its job correctly.
Sure, but thanks for explaining what everyone already understands. You’re not clarifying anything new, you’re just pretending not to get the point so you can keep arguing. The discussion wasn’t about LLMs fixing cars or following recipes. It was about any kind of work, and a huge portion of human work revolves around invention, storytelling, and creative synthesis. Fiction writing isn’t a corner case, it’s one of the most valued human expressions of intelligence. Everyone knows that too. It’s not an obscure philosophical insight. It’s basic cultural literacy. Which is exactly why I don’t buy your act. You’re too smart not to know what’s obvious to everyone else.
So when I say the “failure mode” of hallucination can be a “success mode” elsewhere, I’m not dodging the topic, I’m expanding it. Creativity is a massive part of human life. Pretending otherwise just to win a narrow argument is dishonest. You know exactly what I meant, you’re just acting like you don’t. No one with normal cognitive function finds that bizarre. It’s theater.
And you used the classic tells, the same ones that get used on HN all the time to dodge the rules while still getting in a jab. You drop words like “bizarre” and “baffled,” act like you’re confused, then follow up with a calm “apology” to sound like the reasonable one. It’s a well known pattern here. You literally used the exact two words everyone does when they’re trying to provoke without crossing the line.
Then came the self deprecation. The polished restraint. “If that was too severe, my apologies. I tend to be a little dramatic. I don’t think you’re an idiot. I’m just trying to communicate my point. I’m sorry for that.” It’s spotless. It hits every note. It reads like empathy but functions like control. It doesn’t defuse the conflict, it reclaims the moral high ground. It’s not humility, it’s stagecraft.
Look, maybe I was too sharp myself. I can be dramatic too, I admit that. It’s not a personal attack, I just have strong feelings about intellectual honesty. I’m sorry for that.
See what I did there?
No point in continuing this.
There's not a single proper study showing this increase in productivity and just about every real developer I know finds very limited use in LLMs. They don't increase productivity "to a much higher degree". It's marginal, maybe 5-10% if you use them strategically in situations that are particularly suitable.
I decided to waste some time "for science" and implemented a feature twice, once by myself and once with Cursor. A feature that took me 4 hours to implement myself took 1-1.5 hours of planning + 1.5-2 hours of iterative agentic coding just to get it to meet basic functional criteria, and it would've taken me at least 2 more hours to review and refactor if I hadn't quit in frustration.
If I didn't care about long-term maintainability I could've finished it with AI in under 2 hours and I would've claimed a 100% productivity boost. I imagine that's what people do, prioritizing short term gains while taking on eye-watering amounts of technical debt, but trying to sell this as a productivity improvement is extremely naive.
LLMs are useful in very specific situations where the changes are trivial (small standalone snippets, straightforward changes in a larger codebase) OR long-term maintainability doesn't matter (one-off scripts). That's the middle ground.
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.
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.
Ask your favorite SOTA model to assume something absurd and then draw the next logical conclusions based on that. "Green is yellow and yellow is green. What color is a banana?" They may get the first question(s) right, but will trip up within a few exchanges. Might be a new question, but often they are very happy to just completely contradict their own previous answers.
You could argue that this is hitting alignment and guard-rails against misinformation.. but whatever the cause, it's a clear sign it's a machine and look, no em-dashes. Ironically it's also a failure of the turing test that arises from a failure in reasoning at a really basic level, which I would not have expected. Makes you wonder about the secret sauce for winning IMO competitions. Anyway, unlike other linguistic puzzles that attempt to baffle with ambiguous reference or similar, simple counterfactuals with something like colors are particular interesting because they would NOT trip up most ESL students or 3-5 year olds.
But it just occurred to me that you could ask "what color is a green eggplant?" for a similar result. Though you'd catch a lot of people who, like me before I looked up fruits that aren't green when unripe, assume that all unripe fruits are green. (They aren't, unripe eggplants are white and look like eggs, hence the name).
What color is a green eggplant?
ChatGPT:
A green eggplant (also called Thai eggplant or green brinjal, depending on the variety) is actually green on the outside — sometimes solid green, sometimes streaked with white or pale green.
The inside flesh is typically off-white or light cream, just like purple eggplants.
If you mean a green-colored variety (not an unripe purple one), then the skin color can range from pale mint green to deep jade, often with light striping.
Me:
How would an average human answer this question?
ChatGPT:
An average person would probably answer something simple and direct like:
“It’s green.”
Or if they’re being a bit more specific:
“It’s green on the outside and white inside.”
Most people wouldn’t overthink the variety or mention Thai vs. unripe types — they’d just describe what they see.
Experimental design comes in here and the one TT paper mentioned in this thread has instructions for people like "persuade the interrogator [you] are human". Answering that a green eggplant is green feels like humans trying to answer questions correctly and quickly, being wary of a trap. We don't know participants background knowledge but anyone that's used ChatGPT would know that ignoring the question and maybe telling an eggplant-related anecdote was a better strategy
And that "not always" is the crux of the matter, I think. You are arguing that we're not there yet, because there are lines of questioning you can apply that will trip up an LLM and demonstrate that it's not a human. And that's probably a more accurate definition of the test, because Turing predicted that by 2000 or so (he wrote "within 50 years" around 1950) chatbots would be good enough "that an average interrogator will not have more than 70% chance of making the right identification after five minutes of questioning". He was off by about two decades, but by now that's probably happened. The average interrogator probably wouldn't come up with your (good) strategy of using counterfactuals to trick the LLM, and I would argue two points: 1) that the average interrogator would indeed fail the Turing test (I've long argued that the Turing test isn't one that machines can pass, it's one that humans can fail) because they would likely stick to conventional topics on which the LLM has lots of data, and 2) that the situation where people are actually struggling to distinguish LLMs is one where they don't have an opportunity to interrogate the model: they're looking at one piece of multi-paragraph (usually multi-page) output presented to them, and having to guess whether it was produced by a human (who is therefore not cheating) or by an LLM (in which case the student is cheating because the school has a rule against it). That may not be Turing's actual test, but it's the practical "Turing test" that applies the most today.
If you understand TT to be about tricking the unwary, in what's supposed to be a trusting and non-adversarial context, and without any open-ended interaction, then it's correct to point out homework-cheating as an example. But in that case TT was solved shortly after the invention of spam. No LLMs needed, just markov models are fine.
Alan Turing was a mathematician not a psychologist, this was his attempt of doing philosophy. And while I applaud brilliant thinkers when they attempt to do philosophy (honestly we need more of that) it is better to leave it to actual philosophers to validate the quality of said philosophy. John Searle was a philosopher which specialized in questions of psychology. And in 1980 he pretty convincingly argued against the Turning test.
In the end though, it's probably about as good as any single kind of test could be, hence TFA looking to combine hundreds across several dozen categories. Language was a decent idea if you're looking for that exemplar of the "AGI-Complete" class for computational complexity, vision was at one point another guess. More than anything else I think we've figured out in recent years that it's going to be hard to find a problem-criteria that's clean and simple, much less a solution that is
Although I am scrutinizing Turin’s philosophy and, no doubt, I am personally much worse at doing philosophy then Turing, I firmly hold the belief that we will never be able to judge the intelligence (and much less consciousness) of a non-biological (and probably not even non-animal, nor even non-human) system. The reason, I think, is that these terms are inherently anthropocentric. And when we find a system that rivals human intelligence (or consciousness) we will simply redefine these terms such that the new system isn’t compatible any more. And I think that has already started, and we have done so multiple times in the past (heck we even redefined the term planet when we discovered the Kuiper belt) instead favoring terms like capability when describing non-biological behavior. And honestly I think that is for the better. Intelligence is a troubled term, it is much better to be accurate when we are describing these systems (including human individuals).
---
1: Though in honesty I will be impressed when machine learning algorithms can interoperate and generate appropriate human facial expressions. It won’t convince me of intelligence [and much less consciousness] though.
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.
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.
Most humans are terrible at that. Most humans don't study for tests, fail, and don't see the connection. Most humans will ignore rules for their safety and get injured. Most humans, when given a task at work, will half-ass it and not make progress without constant monitoring.
If you only hang out with genius SWEs in San Francisco, sure, ChatGPT isn't at AGI. But the typical person has been surpassed by ChatGPT already.
I'd go so far as to say the typical programmer has been surpassed by AI.
Here is something I do not see with reasonable humans who are cooperative: Me: "hey friend with whom I have plans to get dinner, what are you thinking of eating?" Friend: "fried chicken?" Me: "I'm vegetarian" Friend: "steak?"
Note that this is in the context of four turns of a single conversation. I don't expect people to remember stuff across conversations or to change their habits or personalities.
Your goalpost is much further out there.
Go join a dating app as a woman, put vegan in your profile, and see what restaurants people suggest. Could be interesting.
You've personally demonstrated that humans don't have to be reasonable and cooperative, but you're not at all refuting my claim.
I'm disagreeing and saying there's far more people in that bucket than you believe.
I know many people at my university that struggle to read more than two sentences at a time. They'll ask me for help on their assignments and get confused if I write a full paragraph explaining a tricky concept.
That person has a context length of two sentences and would, if encountering a word they didn't know like "vegetarian", ignore it and suggest a steak place.
These are all people in Computer Engineering. They attend a median school and picked SWE because writing buggy & boilerplate CRUD apps pays C$60k a year at a big bank.
Firstly, not studying, ignoring safety rules, or half-assing a task at work are behaviors, they don't necessarily reflect understanding or intelligence. Sometimes I get up late and have to rush in the morning, that doesn't mean I lack the intelligence to understand that time passes when I sleep.
Secondly, I don't think that most people fail to see the connection between not studying and failing a test. They might give other excuses for emotional or practical reasons, but I think you'll have a hard time finding anyone who genuinely claims that studying doesn't usually lead to better test scores. Same for ignoring safety rules or half-assing work.
I know dozens of people that have told me to my face that they don't need to attend lectures to pass a course, and then fail the course.
Coincidentally, most of my graduating class is unemployable.
It's not a lack of understanding or intelligence, but it is an attitude that is no longer necessary.
If I wanted someone to do a half-assed job at writing code until it compiles and then send the results to me for code review, I'd just pay an AI. The market niche for that person no longer exists. If you act like that at work, you won't have a job.
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.
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.)
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.
I have countless examples of lawyers, hr and other public gov bodies that breach the law without knowing the consequences. I also have examples of AI giving bad advice, but it’s al better than an average human right now.
An AI could easily save them a ton of money in the fees they are paying for breaching the law.
That's not a fact, that's just cynicism mixed with sociopathy.
I hear this argument a lot from AI bros, and...y'all don't know how much you're telling on yourselves.
What you said is not a fact either. And so?
I feel every human just regurgitates words too
Even when people are borrowing, copying, and stealing, which is the exception, mind you, they are also carefully threading the material they are re-using into whatever it is they are trying to do, say, or make in a way that is extremely non-trivial.
Well, from my experience: a few lawyers got the law wrong but my ai did it right and the lawyer “lost” and showed how incompetent the lawyer was.
If you say most people that copy are careful you don’t know what’s an average person. And think there are 50% in the world worse than than them.
Most people lack basic logic skills.
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.
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.
The more I think about this, the more I think the same is true for our own intelligence. Consciousness is a trick and AI development is lifting the veil of our vanity. I'm not claiming that LLMs are conscious or intelligent or whatever. I'm suggesting that next token prediction has scaled so well and cover so many use cases that the next couple breakthroughs will show us how simple intelligence is once you remove the complexity of biological systems from the equation.
To the extent that we vainly consider ourselves intelligent for our linguistic abilities, sure. But this underrates the other types of spatial and procedural reasoning that humans possess, or even the type that spiders possess.
It is an entirely different thing to language,which was created by humans to communicate between us.
Language is the baseline to collaboration - not intelligence
How do you define verbal language? Many animals emit different sounds that others in their community know how to react to. Some even get quite complex in structure (eg dolphins and whales) but I wouldn’t also rule out some species of birds, and some primates to start with. And they can collaborate; elephants, dolphins, and wolves for example collaborate and would die without it.
Also it’s completely myopic in terms of ignoring humans who have non verbal language (eg sign language) perfectly capable of cooperation.
TLDR: just because you can’t understand an animal doesn’t mean it lacks the capability you failed to actually define properly.
That’s like a bird saying planes can’t fly because they don’t flap their wings.
LLMs use human language mainly because they need to communicate with humans. Their inputs and outputs are human language. But in between, they don’t think in human language.
You seem to fundamentally misunderstand what llms are and how they work, honestly. Remove the human language from the model and you end up with nothing. That's the whole issue.
Your comment would only make sense if we had real artificial intelligence, but LLMs are quite literally working by predicting the next token - which works incredibly well for a fascimlie of intelligence because there is an incredible amount of written content on the Internet which was written by intelligent people
LLMs can't initiate any task on their own, because they lack thinking/intelligence part.
By the time you get to active "teaching", the child has already learned language -- otherwise we'd have a chicken-and-egg problem, since we use language to teach language.
An additional facet nobody ever seems to mention:
Human language is structured, and seems to follow similar base rules everywhere.
That is a huge boon to any statistical model trying to approximate it. That's why simpler forms of language generation are even possible. It's also a large part of why LLMs are able to do some code, but regularly fuck up the meaning when you aren't paying attention. The "shape" of code and language is really simple.
Just because we don’t understand something doesn’t mean there’s nothing there.
Also, I’m not so sure human language is structured the same way globally. There’s languages quite far from each other and the similarities tend to be grouped by where the languages originated. Eg Spanish and French might share similarities of rules, but those similarities are not shared with Hungary or Chinese. There’s cross pollination of course but language is old and humans all come from a single location so it’s not surprising for there to be some kinds of links but even a few hundred thousand years of evolution have diverged the rules significantly.
If cows were eating grass and conceptualising what is infinity, and what is her role in the universe, and how she was born, and what would happen after she is dead... we would see a lot of jumpy cows out there.
Words, after all are just arbitrary ink shapes on paper. Or vibrations in air. Not fundamentally different than any other signal. Meaning is added only by the human brain.
I don't think anyone would argue that animals don't communicate with each other. Some may even have language we can't interpret, which may consist of something like words.
The question is why we would model an AGI after verbal language as opposed to modeling it after the native intelligence of all life which eventually leads to communication as a result. Language and communication is a side-effect of intelligence, it's a compounding interest on intelligence, but it is not intelligence itself, any more than a map is the terrain.
Because verbal/written language is an abstracted/compressed representation of reality, so it's relatively cheap to process (a high-level natural-language description of an apple takes far fewer bytes to represent than a photo or 3D model of the same apple). Also because there are massive digitized publicly-available collections of language that are easy to train on (the web, libraries of digitized books, etc).
I'm just answering your question here, not implying that language processing is the path towards AGI (I personally think it could play a part, but can't be anything close to the whole picture).
All we know about animal consciousness is limited to behaviour, e.g. the subset of the 40 or so "consciousness" definitions which are things like "not asleep" or "responds to environment".
We don't know that there's anything like our rich inner world in the mind of a chimpanzee, let alone a dog, let alone a lobster.
We don't know what test to make in order to determine if any other intelligence, including humans and AI, actually has an inner experience — including by asking, because we can neither be sure if the failure to report one indicates the absence, nor if the ability to report one is more than just mimicking the voices around them.
For the latter, note that many humans with aphantasia only find out that "visualisation" isn't just a metaphor at some point in adulthood, and both before and after this realisation they can still use it as a metaphor without having a mind's eye.
> Language is the baseline to collaboration - not intelligence
Would you describe intercellular chemical signals in multicellular organisms to be "language"?
If be "we don't know" you mean we cannot prove, then, sure, but then we don't know anything aside from maybe mathematics. We have a lot of evidence that animals similar consciousness as we do. Dolphins (or whales?) have been known to push drowning people to the surface like they do for a calf. Killer whales coordinate in hunting, and have taken an animus to small boats, intentionally trying to capsize it. I've seen squirrels in the back yard fake burying a nut, and moving fallen leaves to hide a burial spot. Any one who has had a dog or a cat knows they get lonely and angry and guilty. A friend of mine had personal troubles and abandoned his house for a while; I went over to take pictures so he could AirBnB it, and their cat saw me in the house and was crying really piteously, because it had just grown out of being a kitten with a bunch of kids around and getting lots of attention, and suddenly its whole world was vanished. A speech pathologist made buttons for her dog that said words when pressed, and the dog put sentences together and even had emotional meltdowns on the level of a young child. Parrots seem to be intelligent, and I've read several reports where they give intelligent responses (such as "I'm afraid" when the owner asked if it wanted to be put in the same room as the cat for company while the owner was away [in this case, the owner seems to be lacking in intelligence for thinking that was a good idea]). There was a story linked her some years back about a zoo-keeper who had her baby die, and signed it to the chimpanzee (or gorilla or some-such) females when it wanted to know why she had been gone, and in response the chimpanzee motioned to with its eye suggesting crying, as if asking if she were grieving.
I probably have some of those details wrong, but I think there definitely is something there that is qualitatively similar to humans, although not on the same level.
More than just that: we don't know what the question is that we're trying to ask. We're pre-paradigmatic.
All of the behaviour you list, those can be emulated by an artificial neural network, the first half even by a small ANN that's mis-classifying various things in its environment — should we call such an artificial neural network "conscious"? I don't ask this as a rhetorical device to cast doubt on the conclusion, I genuinely don't know, and my point is that nobody else seems to either.
I posit that we should start with a default "this animal experiences the world the same as I do" until proven differently. Doctors used to think human babies could not feel pain. The assumption has always been "this animal is a rock and doesn't experience anything like me, God's divine creation." It was stupid when applied to babies. It is stupid when applied to animals.
Did you know that jumping spiders can spot prey, move out of line of sight, approach said pray outside that specific prey's ability to detect, and then attack? How could anything do that without a model of the world? MRIs on mice have shown that they plan and experience actions ahead of doing them. Just like when you plan to throw a ball or lift something heavy where you think through it first. Polar bears will spot walrus, go for a long ass swim (again, out of sight) and approach from behind the colony to attack. A spider and the apex bear have models of the world and their prey.
Show that the animal doesn't have a rich inner world before defaulting to "it doesn't."
As I don't know, I take the defensive position both ways for different questions.*
Just in case they have an inner world: We should be kind to animals, not eat them, not castrate them (unless their reproductive method appears to be non-consensual), not allow them to be selectively bred for human interest without regard to their own, etc.
I'd say ditto for AI, but in their case, even under the assumption that they have an inner world (which isn't at all certain!), it's not clear what "be kind" even looks like: are LLMs complex enough to have created an inner model of emotion where getting the tokens for "thanks!" has a feeling that is good? Or are all tokens equal, and the only pleasure-analog or pain-analog they ever experienced were training experiences to shift the model weights?
(I'm still going to say "please" to the LLMs even if it has no emotion: they're trained on human responses, and humans give better responses when the counterparty is polite).
> How could anything do that without a model of the world?
Is "a model of the world" (external) necessarily "a rich inner world" (internal, qualia)? If it can be proven so, then AI must be likewise.
* The case where I say that the defensive position is to say "no" is currently still hypothetical: if someone is dying and wishes to preserve their continuity of consciousness, is it sufficient to scan their brain** and simulate it?
** as per the work on Drosophila melanogaster in 2018: https://www.sciencedirect.com/science/article/pii/S009286741...
That's not how I view it. Consciousness is the result of various feedback structures in the brain, similar to how self-awareness stems from the actuator-sensor feedback loop of the interaction between the nervous system and the skeletomuscular system. Neither of those two definitions have anything to do with language ability -- and it bothers me that many people are so eager to reduce consciousness to programmed language responses only.
If you can withdraw $10,000 cash at all to dispose as you please (including for this 'trick' game) then my friend you are wealthy from the perspective of the vast majority of humans living on the planet.
And if you balk at doing this, maybe because you cannot actually withdraw that much, or maybe because it is badly needed for something else, then you are not actually capable of performing the test now, are you ?
Couldn’t someone else just give him a bunch of cash to blow on the test, to spoil the result?
Couldn’t he give away his last dollar but pretend he’s just going to another casino?
Observing someone’s behavior in Vegas is a just looking at a proxy for wealth, not the actual wealth.
If you still need a rich person to pass the test, then the test is working as intended. Person A is rich or person A is backed by a rich sponsor is not a material difference for the test. You are hinging too much on minute details of the analogy.
In the real word, your riches can be sponsored by someone else, but for whatever intelligence task we envision, if the machine is taking it then the machine is taking it.
>Couldn’t he give away his last dollar but pretend he’s just going to another casino?
Again, if you have $10,000 you can just withdraw today and give away, last dollar or not, the vast majority of people on this planet would call you wealthy. You have to understand that this is just not something most humans can actually do, even on their deathbed.
So, most people can't get $1 Trillion to build a machine that fools people into thinking it's intelligent. That's probably also not a trick that will ever be repeated.
Isn't this what most major AI companies are doing anyway?
You've invented a story where the user can pass the test by only doing this once and hinged your point on that, but that's just that - a story.
All of our tests and benchmarks account for repeatability. The machine in question has no problem replicating its results on whatever test, so it's a moot point.
Okay ? and you, presumably a human can replicate the trick of fooling me into thinking you're conscious as long as there is a sufficient supply of food to keep you running. So what's your point ? With each comment, you make less sense. Sorry to tell you, but there is no trick.
It does not think idle thoughts while it's not being asked questions. It's not ruminating over its past responses after having replied. It's just off until the next prompt.
Side note: whatever future we get where LLMs get their own food is probably not one I want a part of. I've seen the movies.
In fact, what is artificial is stopping the generation of an LLM when it reaches a 'stop token'.
A more natural barrier is the attention size, but with 2 million tokens, LLMs can think for a long time without losing any context. And you can take over with memory tools for longer horizon tasks.
What does repeatability have to do with intelligence? If I ask a 6 year old "Is 1+1=2" I don't change my estimation of their intelligence the 400th time they answer correctly.
>The machine in question has no problem replicating its results on whatever test
What machine is that? All the LLMs I have tried produce neat results on very narrow topics but fail on consistency and generality. Which seems like something you would want in a general intelligence.
If your 6 year old can only answer correctly a few times out of that 400 and you don't change your estimation of their understanding of arithmetic then, I sure hope you are not a teacher.
>What machine is that? All the LLMs I have tried produce neat results on very narrow topics but fail on consistency and generality. Which seems like something you would want in a general intelligence.
No LLM will score 80% on benchmark x today then 50% on the same 2 days later. That doesn't happen, so the convoluted setup OP had is meaningless. LLMs do not 'fail' on consistency or generality.
I'm sorry, but I find this intelectual dishonesty and moving the goal posts.
Speaks more about our inability to recognize the monumental revolution about to happen in the next decade or so.
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.
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.
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.
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.
Interestingly the people doing the actual envelope pushing in this domain, such as Ilya Sutskever, think that there it’s a scaling problem, and neural nets do result in AGIs eventually, but I haven’t heard them substantiate it.
(I'm asking because of your statement, "Don’t fool yourself into believing artificial intelligence is not one breakthrough away", which I'm not sure I understand, but if I am parsing it correctly, I question your basis for saying it.)
“one breakthrough away” as in some breakthrough away
This is not much different than saying that it’s possible to fly a spacecraft to another galaxy because spacecrafts exist and other galaxies exist.
Possible and practically attainable are two far different things.
It is very different. We have never seen a spacecraft reach another galaxy so we don't know it is possible.
We have an example of what we call intelligence arising in matter. We don't know what hurdles there are between current AI and an AGI, but we know that AGI is possible.
Let me know when I’ll be able to buy my $30,000 car with level 5 self driving.
I find anyone with confident answers to questions like these immediately suspect.
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.
My emotions are definitely a function of the chemical soup my brain is sitting in (or the opposite).
Let me pose back to you a related question as my answer: How do you know that I feel emotions rather than merely emulating emotional behavior?
This gets into the philosophy of knowing anything at all. Descartes would say that you can't. So we acknowledge the limitation and do our best to build functional models that help us do things other than wallow in existential loneliness.
But you can propose explanations and try to falsify them. I haven’t thought about it but maybe there is a way to construct an experiment to falsify the claim that you don’t feel emotions.
Preface:
The problem of the relation between our bodies and our minds, and especially of the link between brain structures and processes on the one hand and mental dispositions and events on the other is an exceedingly difficult one. Without pretending to be able to foresee future developments, both authors of this book think it improbable that the problem will ever be solved, in the sense that we shall really understand this relation. We think that no more can be expected than to make a little progress here or there.
... well. Thanks a bunch, Karl.
Also, you don't know what species I am. Maybe I'm a dog. :-)
(https://en.wikipedia.org/wiki/On_the_Internet,_nobody_knows_...)
Also assuming psychopaths don‘t experience emotions is going going with a very fringe theory of psychology. Very likely psycopaths experience emotions, they are maybe just very different emotions from the ones you and I experience. I think a better example would be a comatose person.
That said I think talking about machine emotions is useless. I see emotions as a specific behavior state (that is you will behave in a more specific manner) given a specific pattern of stimuli. We can code our computers to do exactly that, but I think calling it emotions would just be confusing. Much rather I would simply call it a specific kind of state.
1) I know that I have emotions because I experience them.
2) I know that you and I are very similar because we are both human.
3) I know that we can observe changes in the brain as a result of our changing emotions and that changes to our brains can affect our emotions.
I thus have good reason to believe that since I experience emotions and that we are both human, you experience emotions too.
The alternative explanation, that you are otherwise human and display all the hallmarks of having emotions but do not in fact experience anything (the P-zombie hypothesis), is an extraordinary claim that has no evidence to support it and not even a plausible, hypothetical mechanism of action.
With an emotional machine I see no immediately obvious even hypothetical evidence to lend support to its veracity. In light of all this, it seems extraordinary to claim that non-biological means achieving real emotions (not emulated emotions) are possible.
After all, emulated emotions have already been demonstrated in video games. To call those sufficient would be setting an extremely low bar.
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 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.
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.
Even if a machine really is conscious, we don't have enough information to ever really know if it is.
Feels good so we want more so you arrange your whole life and outlook to make more feel good happen. Intelligence!
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.
It has been proven that a Turning Machine and Lambda Calculus have the exact same equivalent expressiveness, that encompasses the _entire set_ of computable functions. Why are you so sure that "text prediction" is not equally expressive?
Was Hellen Keller not intelligent because she lacked the ability to see or hear? Is intelligence defined by a particular set of sense organs? A particular way of interacting with the environment? What about paraplegics, are they disqualified from being considered intelligent because they lack the same embodied experience as others?
Whenever you give someone kudos for being brilliant, it is always for their ability to successfully compute something. If that isn't what we're discussing when we're talking about intelligence, then what are we discussing?
On the other hand, you are clearly stating that intelligence is computation. But you're right, it would be too easy to ask you to define what any of those words mean AND to back that claim.
We can easily program them to have human desires instead.
Which is it??
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.
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.
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.
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...
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
we only need to fake it to the point it's undistinguishable from the carbon based one.
faking is all you need.
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.
There are people that have a hard time recognizing/feeling/understanding other people as "aware". Even more about animals.
>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.
People want autonomy, self-learning, consistent memory and perhaps individuality (in the discernability/quirkiness sense), but still morally unencumbered slaves.
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.
Maybe we've just reach the ability the replicate the function of an artificially powered dead brain that would be randomly stimulated and nothing more. Is this really a path to intelligence ?
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.
Wait, you mean this HN comment didn't casually solve the hard problem of consciousness?
I see a lot of "AGI boosters/doomers" comfortable making big claims without providing research to back what, when challenged, prove to be just their model or feeling of how things function.
It seems to be a case of people looking at a problem they have little knowledge or understanding of and thinking "how hard can it be"? In this case, the answer is "so hard that philosophers have dubbed it 'the hard problem'".
> A case where someone is conscious but without continuous propagation of neural signals?
That would be irrelevant. All known conscious beings are made up of biological cells, but that doesn’t prove that all conscious beings must be made of biological cells, or that biological cells are the key causative factor in consciousness. The same goes for “continuous propagation of neural signals.”
You described a personal conjecture as though it solved a known hard problem, even throwing in the word “just” as though the solution is really simple. This is a lot like the Feynman quote about quantum mechanics: if you think you understand it, you almost certainly don’t. You may not even have recognized the core problem yet. The original Chalmers paper is a good place to start: https://consc.net/papers/facing.pdf
But coming at it from a computational perspective, in some ways it’s even easier to see the problem. We don’t generally assume that a deterministic, non-neural net program written in say Python has a conscious subjective experience. To use Nagel’s terminology, there is “nothing it is like” to be that program. But, an LLM or any other computational neural net is no different from a program like that. It’s executing deterministic instructions, like a machine, because it is a machine. We can speculate about consciousness being some sort of emergent property that arises in such systems given the right conditions, but that’s all it is: speculation.
And it’s completely unclear what those right conditions might be, or how those conditions could possibly give rise to conscious experience. Why aren’t humans philosophical zombies with no conscious experience, just reacting to input like machines? No-one has succeeded in getting past the conjecture stage in answering that question.
> an LLM or any other computational neural net is no different from a program like that
I don't think so. A program doesn't exhibit highly complex abstract thought in a very high-dimensional space.
> It’s executing deterministic instructions, like a machine, because it is a machine
Its true that LLMs are deterministic. But do you really think that the magic behind the brain is only due to temperature and randomness? Do you really think that non-deterministic behavior is the magic ingredient that makes up what we are referring to as consciousness? I could inject noise into an LLM at every parameter dynamically during inference. The output would come out just fine. After all, LLMs are high dimensional and can handle a little noise. Would it really be more conscious after that? You can find experiments where people remove entire layers of the LLM and it still works fine. A little noise would be even less harmful than that.
You see, when I'm arguing, I'm not citing what some other person said. I at most will cite experiments from other people and their results. When I am contradicting your arguments, I present you a reality you can go and try in the real world and verify. You can go verify yourself LLMs exhibit complex high-dimensional thought. You can verify yourself that if you inject noise dynamically through inference on every parameter, you still get coherent output from the LLM.
So, if you are willing to continue this discussion, I ask of you that you present some sort of "probing" of the real world and the respective "reaction" of the same real world as arguments. That is what finding the truth means.
And lastly. I am presenting a Theory. This means I believe that my points form a foundation that makes my theory stronger than yours. It means I have better evidence that backs it up. It doesn't mean I have proved what consciousness is. Instead, it primarily means I can make more accurate predictions using my theory on real world scenarios involving artificial and biological neural networks. And my personal experience shows me that is true.
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?
It's immaterial and not measurable thus possibly out of reach of science.
Also, if subjective experience has any effect on the material world, then we can measure it and test it, putting it in reach of science. If it doesn't, why does it even matter in this discussion? By definition it has no relation to the AGI discussion since that's an empirical matter.
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.
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?
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"?
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.
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.
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.
Also, those results don't look as strong to me as you suggest. I do not accept that an LLM is conscious nor could I ever unless I can have a theory of mind for it... Which is impossible given that it's a stochastic parrot without awareness of the things my five senses and my soul feel in reality.
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.
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.
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.
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".
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.
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?
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.
Probably because everyone has first hand experience of the above. I'm not sure it's arrogant to talk about those, even without 'source material'.
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.)
To the extent that LLMs think, they think "people say there's a flavour of set theory in which RH is true". LLMs don't care about facts: they don't even know that an external reality exists. You could design an AI system that operates the way you describe, and it would behave a bit like an LLM in this respect, but the operating principles are completely different, and not comparable. Everything else you've said is reasonable, but – again – doesn't apply to LLMs, which aren't doing what we intuitively believe them to be doing.
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.
I don't have an opinion on whether ChatGPT qualifies as AGI. What I'm saying is where one stands on that question has nothing to do with "why it became so popular so fast."
(Also, several machine-learning techniques could do millions of things terribly before LLMs. GPT does them, and other things, less poorly. It's a broadening. But I suppose really any intelligence of any kind can be considered a "baby" AGI.)
The "ChatGPT" web app started with the underlying model GPT-3.5
The predecessor models, a whole series of them collectively "GPT-3" but sold under API with names like "davinci" and "ada", was barely noticed outside AI research circles.
3 was useful, but you had to treat it as a text completion system not a chat interface, your prompt would have been e.g.
Press release
Subject: President announces imminent asteroid impact, evacuation of Florida
My fellow Americans,
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.
Edit: toned down the preachiness.
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...
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.
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".
----
In 2010, one of the first "presentations" given at Deepmind by Hassabis, had a few slides on AGI (from the movie/documentary "The Thinking Game"):
Quote from Shane Legg: "Our mission was to build an AGI - an artificial general intelligence, and so that means that we need a system which is general - it doesn't learn to do one specific thing. That's really key part of human intelligence, learn to do many many things".
Quote from Hassabis: "So, what is our mission? We summarise it as <Build the world's first general learning machine>. So we always stress the word general and learning here the key things."
And the key slide (that I think cements the difference between what AGI stood for then, vs. now):
AI - one task vs. AGI - many tasks
at human level intelligence.
For reference, the average chesscom player is ~900 elo, while the average FIDE rated player is ~1600. So, yeah. Parrot or not, the LLMs can make moves above the average player. Whatever that means.
What counts as a "thing"? Because arguably some of the deep ANNs pre-transfomers would also qualify as AGI but no one would consider them intelligent (not in the human or animal sense of intelligence).
And you probably don't even need fancy neural networks. Get a RL algorithm and a properly mapped solution space and it will learn to do whatever you want as long as the problem can be mapped.
When I was in college ~25 years ago, I took a class on the philosophy of AI. People had come up with a lot of weird ideas about AI, but there was one almost universal conclusion: that the Turing test is not a good test for intelligence.
The least weird objection was that the premise of the Turing test is unscientific. It sees "this system is intelligent" as a logical statement and seeks to prove or disprove it in an abstract model. But if you perform an experiment to determine if a real-world system is intelligent, the right conclusion for the system passing the test is that the system may be intelligent, but a different experiment might show that it's not.
> we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition.
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.
So in that sense it's a triviality. You can ask ChatGPT whether it's human and it will say no upfront. And it has various guardrails in place against too much "roleplay", so you can't just instruct it to act human. You'd need a different post-training setup.
I'm not aware whether anyone did that with open models already.
Post training them to speak like a bot and deny being human has no effect on how useful they are. That's just an Open AI/Google/Anthropic preference.
>If you take the raw model, it will actually be much worse at the kinds of tasks you want it to perform
Raw models are not worse. Literally every model release paper that compares both show them as better at benchmarks, if anything. Post training degrading performance is a well known phenomena. What they are is more difficult to guide/control. Raw models are less useful because you have to present your input in certain ways, but they are not worse performers.
It's besides the point anyways because again, you don't have to post train them to act as anything other than a human.
>If their behavior needs to be restricted to actually become good at specific tasks, then they can't also be claimed to pass the Turing test if they can't within those same restrictions.
Okay, but that's not the case.
This is exactly what I was referring to.
But that is exactly the point of the Turing test.
If someone really wants to see a Turing-passing bot, I guess someone could try making one but I'm doubtful it would be of much use.
Anyways,people forget that the thought experiment by Turing was a rhetorical device, not something he envisioned to build. The point was to say that semantic debates about "intelligence" are distractions.
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.
If you think this is true, I would say you should leave artificial life alone until you can understand human beings better.
If the teacher was a robot, I don't think the piano would get practiced.
IDK how AI gains that ability. The requirement is basically "being human". And it seems like there's always going to be a need for humans in that space, no matter how smart AI gets.
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?
Once they can ... I am open to revisit my assumptions about AGI.
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.