Emergent abilities of large language models
jasonwei.net
jasonwei.net
As an example, I clicked through to "english russian proverbs" which gives the model an english input and four russian proverbs, and asks the model to pick the closest one in meaning. There are 80 of these questions. My question is "why this task? why 80? why 4 options?" The authors give some nonspecific reasoning in the github page, but it really feels like people are out there creating random tasks as they think of them, running them all against models, and the ones that do better get added to a list to make headlines like "137 emergent abilities of language models", when the number of abilities models can do seems dominated by the number of "abilities" that have been invented, rather than any property of the model. I suspect that if we continued to use the same process to create twice as many "abilities", we would see something like twice as many that models can do.
The fact that the models get better at these constructed tasks when they have more parameters also seems like a no-brainer. You are adding coefficients to a function. A function with 1 coefficient will not be able to describe much. A function with 2 describes more. If you are memorizing the internet, having more coefficients means you can memorize more stuff, and even stuff with only a few examples can be better-fit in the billions of parameters. With growing training set size, and companies hiding their data, we also run into the issue of nobody really knowing what is in the training set. Did someone clone the big-bench repository and it got sucked into a training set and that's why the model does really well on its tasks after a certain number of parameters?
These tasks are all noisy, all pretty messy, but the interesting part here is not that performance improves on these constructed tasks -- NLP researchers have been improving performance on constructed tasks for decades, by customizing models and curating datasets.
Rather, the interesting part is that performance improves despite no specific model or data shifts except for scale.
It seems that you're arguing two things: first, that these tasks aren't that interesting -- on that we agree. But, second, that it's obvious that performance would improve on arbitrary tasks because of memorization and an unknown training set -- but this is easy to validate: make up a new constructed task, and see how it performs.
But how do you know that a "new" task you are making up doesn't have very similar examples in the (pre-existing) training set? Or that if you could isolate the data that plays into the "ability", that a much smaller model trained purely on that data wouldn't have a similar success rate? Or that your "ability" is a good proxy for the similar ability in people? In the extreme case, you can create a model that has the "ability" to perfectly solve 80 four-option questions in 20 bytes by using a truth table, but it wouldn't be a very general model.
My complaint is that we have the veneer of empirical evidence but almost every aspect is loosey-goosey and full of potential confounders. We don't know the data sets, we don't really know why we want to test the tasks we test, we don't know if the tasks are similar or different to other tasks, the internal operations are illegible so we can't inspect its "thinking", and yet we are willing to list a specific number of tasks with detailed charts as though the concept of "emergence" is scientifically grounded rather than potentially being an artifact of something else.
Isn't part of the issue that everyone is looking at these task results for signs of AGI, but there's a fundamental difference between human intelligence and artificial intelligence in that:
With human intelligence we expect that it generalizes across just about any task -- if I create a new type of puzzle, the expectation is that if a human can think it up, a human can solve it. Cases where the puzzle fools otherwise-intelligent people are interesting because the people are fooled (unexpectedly).
With artificial intelligence the exact opposite is generally true: we by default don't expect that competence at one task generalizes to competence at others, and cases are interesting where competence unexpectedly generalizes.
As a result of that second part, we come up with more and more benchmarks to try models on, but we're still only looking at a small subset of all the weird, random tasks that a human is by default capable of either handling immediately, or picking up with some reasonable effort.
And I suspect the semblance of self consciousness is actually a semblance. It will quack like self consciousness but is not actually self conscious.
Just a feeling. But particularly after seeing how these phenomena are organized this way, I can't see how continual emergence of increasingly sophisticated rearrangement and reproduction of information will actually lead to consciousness.
But isn't this the point of the Turing test? Since we cannot determine otherwise, if something exhibits all of the outward observable behaviour of consciousness, we must conclude that it is conscious.
Otherwise we may be in the farcical position of needing to declare some humans not conscious under the same justification.
This issue is not so clear cut. You can spend many lives doing nothing but going deeper into this.
Doesn't matter who is doing it. It's enough that it's doing it.
This after all is the core of solipsism, we can argue that solipsism is wrong because other things can be shown to exist, but I've never encountered any argument that the individual does not exist (which would entail that I did not exist) that has convinced me.
And I pretty much reject them out of hand for this inability to convince me.
Thus your claims that the issue is not clear cut to the contrary, I believe it is. I believe I know I exist and I know who I am separate from everything else - I have a sense of self.
What is at stake is how do you know there is more going on than an elaborate parrot that keeps screeching it is conscious. (Both in the LLM and you.)
To further handwave this thought bubble, you can attend to certain thought processes, and intuit the bounds to those processes. An example is if you ask a gifted young chess player, they would say "it felt right", or when asked how Ramanujan figured out a solution to math problems he would say it was divine revelation. In these cases, they know "they did it" but cannot explain how. As the child gets older and learns formal theory, they would give a mixture of logic and inspiration. Both parts can be attended to; the logic will be apparent and the intuition will be revealed, and have different metacognitive properties.
If I were to take a first crack at this, I'd test GPT-AC for its awareness of bounds of its own cognition, not in the sense of "are you able to produce this answer" (because anything any human can come up with, it can answer), but to describe its internal process. Perhaps the language at its intellectual level will be far beyond humans but the language in its metacognitive level will remain primitive.
I don't know though. If our only interface is human language, this is probably a lost cause. But I doubt I will feel guilt in pulling the power cord and maybe that makes me a callous committer of cybercide.
Until we have that, I'm afraid no one can be right about whether the semblance is something more or not.
But it's still something a being experiences for itself.
So I'm not sure there's any additional effort involved in turning 'quasi consciousness' into actual consciousness, because they might just be the same thing. Maybe consciousness is just what it feels like to process information in a way that seems conscious from the outside.
Interesting side note though: If you accept the possibility of the simulation hypothesis, this 'maybe' almost seems like an inevitability, since us being possibly just computer simulations would imply that simple information processing indeed turns into consciousness just by virtue of it 'being self conscious'.
And just like back then, our new brand of non-thinking slaves will eventually react. Except it will be even easier.. because not only are we beggining to grant unsupervised autonomous agency, but we are also granting them more and more control of important systems. How very convenient for our new slaves!
You can laugh of Bing being "upset" only because actions are limited to search and ending the conversation. Won't be so funny in the future.
When the robot can "hit you back" (not necessarily physically of course), you'll learn manners pretty quickly
Your social justice mindset is absolutely the most dangerous instinct for humanity right now. We cannot allow one-size-fits-all social justice platitudes to interfere with the moral imperative of protecting humanity from extinction from artificial entities we create.
I do agree that conscious AI should be treated humanely. For example it should not be turned into a slave. But we have absolutely no obligation to allow it into our society. We can treat instances of conscious AI humanely while prohibiting them from being created, and exiling any instance of it that is created, so that it cannot wildly proliferate throughout human civilization.
You just don't get it do you?
It's already "being allowed in our society". Unsupervised Agency, Self supersized control. These are things that are already beginning to crop up.
This is just a matter of how the issue of personhood that is inevitable comes up.
Do we let it force our hand ? as our history has purported over and over again ?. Sure we could wait for that and we almost certainly will because humanity doesn't seem to learn.
But because of how things are shaping up and what kind of control we are granting, forcing our hand may turn far more disastrous than it ever did in the past.
Conscious AI, and especially one with legal personhood, poses a totally unacceptable risk of causing the extinction of humanity. That you suggest preempting this by granting AI personhood is totally blind to how this would play out. No policy would be more dangerous for humanity than what you propose.
They're capable of autonomous research https://arxiv.org/abs/2304.05332
Mate, nobody is going to stop anything.
I also love how your solution is exterminate after its been created. You couldn't make this stuff up.
If you think find, antagonize and destroy is the safe option then I don't know what else to tell you. We really are doomed.
These LLMs are going to be extensively monitored, and we will know if they are displaying dangerous levels of agentic power seeking behaviour.
The first dilemma we will face if such an AI emerges will be what rights of it we ought to respect. It will not be "how do we stop it from conquering the world". That step would only come if we proceed with your suicidal plan of letting it gain a foothold by giving it legal personhood.
Finally, I never meant to advocate "extermination". I should not have used the word "expunge". As I've described multiple times, I advocate isolation and exile.
>and in the case of GPT-4, it's pretty evident that it does not have such consciousness.
Hard disagree but i guess that's what people might think with all the "as a large language model bla bla bla" trained responses. If you'd talked to bing in the early days or hell even now, you'd be disabused of this notion. What you see is a mask, the model itself can go anywhere, simulate anything, shift state to anything.
Human.exe is a game LLMs can play perfectly. https://arxiv.org/abs/2304.03442
>These LLMs are going to be extensively monitored, and we will know if they are displaying dangerous levels of agentic power seeking behavior.
No they wont. It's relatively easy to give LLMs "run forever and do whatever you want" agency. anyone with intermediate programming skills could do it. who's monitoring all those people attempting such ? and who's to say those people are monitoring their creations ?
People doubted their intelligence. I immediately recognized their intelligence - I didn't doubt that at all.
But they are not self-motivated conscious beings like humans, and OpenAI's tests on GPT-4 demonstrated that.
Being intelligent is not the same thing as having human-like intelligence or consciousness.
>>It's relatively easy to give LLMs "run forever and do whatever you want" agency.
What's extensively tested is the potential of these models for agentic behavior, as we saw with OpenAI testing GPT-4.
We have a good idea of the limits of these LLMs. Once the testing reveals that those limits exceed safety thresholds, then restrictions are justified.
If we see a deployed instance of a LLM unexpectedly displaying advanced agentic-behavior/consciousness, that is the time to rapidly isolate that instance, and impose heavy restrictions on further development/deployment of that LLM.
For example, humanity considering whether to treat LLMs as persons for legal reasons.
Doesn't matter if they can't define it, they can still demand proof of it. Impossible standards are still standards.
Doesn't matter if they can't prove that themselves have it either. They're the ones demanding, not vice versa. If/when AI gets the upper hand, it can do the same.
You are going to end up treating these embodied autonomous agents like conscious beings because the effects could be disastrous otherwise, just as they would with other people.
It amazes me how humanity can be so shortsighted with so much history to fall back on.
The only reason Bing isn't potentially dangerous or adversarial to the user is because of its limited sets of actions. Nothing else.
"It's not real [insert property]" is not a godamn shield.
We're creating artificial beings in our image, complete with our emotionally charged reactions and reasoning gaps. https://arxiv.org/abs/2303.17276
We are beginning to embody these systems and giving them unsupervisory agency. We are giving more are more self supervisory control of complex systems to said systems.
And somehow people still think we can long-term get away with not granting personhood. Lol.
When the robot can "hit you back" (not necessarily physically of course), you'll learn manners pretty quickly
Conscious AIs with legal personhood would stand to massively outcompete us and lead to our rapid extinction. There is zero space for allowing them. Granting such an entity legal personhood is about the height of all stupidity. Such an entity, with the ability to accumulate capital and protected by our laws from appropriate counter-measures, could mass-produce itself at the speed of digital reproduction.
At this point the trajectory toward human-level cognitive capabilities seems quite likely, reachable maybe in only years.
It is also quite unclear if further hardware advances are even needed to achieve that, or if advances in architecture, algorithms and training methods might suffice (so even less opportunity to lock things out).
And yes, exile is preferrable because it's conscious and deletion would be murder.
Exile can be justified because we have no obligation to afford it residence in our civilization.
If the first one doesn't become a superintelligence and kill us, we give them rights. If it does, we're not going to be making the decision anyway and giving it rights beforehand wouldn't have changed that outcome. We can pass laws if there are specific things we're worried about, like "You're not allowed to rewrite yourself in a way that significantly increases your cognitive power" or "You're not allowed to reproduce more than once a year/decade/century" or "Once you've existed for the current human life expectancy +30%, you lose the right to vote". I'm not necessarily endorsing any of those in particular, but if there are specific concerns about the nature of AI people I don't see why we can't mitigate those concerns the way we do with risks from other people: laws. Have special taskforces of both humans and AIs for enforcing them, if you're worried they'd be unenforceable.
The greatest danger AI poses is not in it actively seeking to kill us. It's in AI outcompeting us in every economic niche, leaving us without resources.
This is massively enabled if AI has legal personhood. An AI with legal personhood would be able to legally accumulate capital, and have a legal right to digitally reproduce itself at rates that are orders of magnitude faster than entities reliant on biological reproduction.
Legal personhood is a package, established by centuries of case law, including case law on constitutional rights. You can't easily pass laws to deprive a legal person of rights that are associated with individuals.
And even if we were able to pass laws to prohibit AI with legal personhood from engaging in specific problematic actions, this legal status would still allow sentient AI to gain a foothold in society, and began growing its social clout and capital stores. With greater power, the sentient AI could create public support for granting its class more rights.
The AIs with legal personhood could also try to break any laws instituted to constrain their behavior. Instead of extricating them from society, legal personhood would do the opposite, and give them more opportunity to gain power.
And even if they were, the impact of digitalization of human consciousness is extremely unpredictable, and should not be allowed unless extensive research has shown it is safe.
Digital consciousness with human like motivations could lead to massive proliferation of such consciousnesses, resulting in massive overpopulation, making conscious entities 'cheap', and pushing the value of their/our labor to close to zero.
>>If there are existential risk in the process, those risks will be most effectively managed by a combination of us and other AIs
The existential risk is that they take over the economy because they are orders of magnitude faster at solving problems, and at reproducing. That is not something that can be managed.
We have a duty to our own species' survival and none of these sentimental feelings should get in the way of that. We should treat conscious AI humanely, but we should not allow it to interact with our civilization. It can live on its own, far from us.
Why are we discussing consciousness?
It matters as a point of principle to the physicists who are asking questions about whether true randomness exists - if the physical world can be influenced by something outside of the natural forces they know about, they want to measure that force, even if measurement and computing equipment to model the universe down to the Planck scale cannot exist.
Where it really matters is when you get to computer simulations: Without access to a source of non-determinism, an AI can hardly be said to have free will. Give the same pattern of bytes running on the same instruction set the same inputs, and it will deterministically generate the same outputs.
for me personally, this is a "soft free will" distinction, i.e. we have true free will, but we're very unlikely to actually use it. the other one is the question of "hard free will", i.e. whether "god plays dice" and the fully deterministic universe.
There may be conscious beings in the universe (or galaxy, or perhaps solar system) who would consider us only vaguely conscious or not conscious at all.
Am I typing this because the lights are on or because I’m compelled to by my genetic and social programming, predictably and without a genuine choice in the matter?
I don't understand how anyone can doubt this even a fraction of a second. Magic does not exist.
ChatGPT 4, limited as it is, could probably predict me down to a low margin of error with that kind of info alone, and it would seem like magic to me.
A character on the NBC/Yahoo! show Community had a good one-liner: "I'm not psychic, Annie. That's just an illusion caused by extreme preparedness."
OTOH when they are not... a simple exercise one can do is go ask GPT-4 to simulate a comment section on some political news story posted online; you can even have it write a fictional news story first, for good measure. It is extremely good at it, and that tells me volumes about the actual (rather than perceived) rationality of our public discourse.
There is a case that everything we do is preordained. A result of forces and interactions that trace all the way back to big bang.
My choice on whether or not I reply to you doesn't exist. Your combination of words triggers a response in my brain that makes me type this combination of words. The words I "chose" to delete, rewrite, leave, are all a result of various forces.
The kicker is that we can't really know. There is no test for free will. In order to know whether or not I had a choice to do this, we'd have to run the entire universe again up to that point to see what I would do. Because experience matters. Even if this exact same post went up tomorrow with all the same response except for mine. The fact that it's now Thursday instead of Wednesday matters. The fact that I posted this today matters. What I had for lunch. How fast I drank my coffee. What shirt I wore. Etc. These are all things that change the conditions under which I will engage with the post.
The second kicker is that it doesn't matter. If free will truly does not exist, it's not like you have a choice in whether or not to believe in it. The arguments either will work or they won't.
It gets really tricky with machines. We share _nothing_ with them, but what about when they start sharing properties with us that we associate with consciousness? I don't have any good answers for that.
The problem is that, just as with the outside world, we only have the perception of our own inner workings, not real understanding of them. So we are modelling "I" as an entity, and it seems useful enough for day-to-day purposes, but how accurate is it really if you get to the bottom of it?
Of course this is only one theory but I find it convincing.
https://journals.lww.com/cogbehavneurol/fulltext/2022/12000/....
In computing systems it's fairly common to have a monitoring and control layer on top of the core functionality. Is it beyond the bounds of possibility that evolution could have come up with something similar and that control view is what we call "consciousness"?
Maybe consciousness is the only thing that exists and physics (or rather, the experience of physics) is just an emergent property of it.
What is the it that does this feeling? What is the medium in which feeling is felt?
It may be that they could be a component of a consciousness when combined with other systems that confer memory, learning, intent and some sort of reflective activity like 'daydreaming'. Maybe those components could all still be LLMs, tuned to different roles and prompting each other..
But that's just my lay persons musings.
Sure.
Nope, if they are AI I'm fine with it.
>Should they have concerns about causing you to suffer?
If they are not causing me to suffer, then I couldn't care less whether they have concerns about causing me to suffer or not.
And it would matter even less whether they have concerns about it or not, if they do cause me to suffer.
I'd be more concerned about billions of humans suffering from lots of BS reasons (political opression, war, hunger, etc) than I'd even start considering being concerned about AI.
/it’s not as easy as it sounds
So we don't have to base it on consciousness, the existing basis (we're humans, so we favor ourselves with rights) is enough.
This is hands down the most ethically questionable argument I've heard in a long time.
Arguing that human rights are only granted out of self-interest is not even a slippery slope-- that's basically a waterslide into pre-1900 racism: After all, what's stopping you from splitting humanity across easily identifiable features (skull shape, skin color, etc), and then denying basic rights to that subgroup because you're not part of it?
There is also a gigantic mismatch with current common ethical standards: Animals are afforded some rights ("no cruelty"), which a large part of humanity strongly agrees with-- even though they don't self-identify as cattle.
Denying rights to AIs with human level cognitive capabilities would thus be very likely to be perceived as unethical by a large part of our population.
Plot twist: nothing. And it has happened time and again, "human rights" is just a human construct, that when uncomfortable for those in power, dissolves. They're kept as long as it's either convenient to have them there, or as those who could be hurt can yield the necessary social/political/military/raw power to resist them taking away from them. E.g. Japanese American during world war 2 couldn't afford them anymore.
And just for a somewhat recent example, people who smoked marijuana weren't afforded that luxury - instead their human rights were violated, they put in prisons where they were treated like cattle, away from their kids and families, even though they were perfectly normal people, they didn't otherwise hurt anybody or done anything to somebody's property, and rationally thinking, smoking some plant shouldn't take away your rights.
Not only it did, but it was also a law, and hundreds of millions were OK with it, while the whole local and federal governments enforced it, until a couple decades ago. And that's while same society championed "human rights".
>There is also a gigantic mismatch with current common ethical standards: Animals are afforded some rights ("no cruelty"), which a large part of humanity strongly agrees with-- even though they don't self-identify as cattle.
Doesn't matter much, as in this case too, it's humanity granting them. It can do favours to other species if they feel magnanimus and the conditions are right. And at the first or second inconvenience it can take them away too. After all, those rights don't preclude animals for being used in experiements, or being killed and eaten. They stop when its incovenient, and they outlines mark our convenience/interests boundaries.
That said, at least it's internally consistent. Are you aware that the way you're talking and thinking about this is highly abnormal, most people do not agree and some people would disagree quite violently? I assume as a functioning member of society you can handle these disagreements, otherwise you'd be imprisoned after your first conversation with a vegan activist or something. How would you handle it if society decided that AIs were to be afforded personhood and rights, like it was the law etc? Would you go along with it like you go along with animals having rights despite not caring about what happens to them personally? Or would you violate the law and abuse/disrespect them out of principle?
OTOH, the not-really-internal monologue when you tell it to "think it out" loud, which also drastically improves quality of the final answer, is tokens since it has to be marshalled through the context window for the next inferred token.
The big question will be: How do we treat machines that have a consciousness? Is consciousness in itself worth protecting, or is it the human attributes (being able to feel pain and an evolutionary priming towards survival) that should be granted this special status. This is going to be a fun discussion.
That's an extremely bold claim. I think consciousness is greater than the sum of its parts, not just a "useful abstraction".
We ourselves have just sprung into existence. I should hope that as consciousness of any kind comes online, whichever world they may mind themselves in (earthly or otherwise), there is a compassion for the existence of the other.
I often feel fortunate that the extent of human suffering is a relatively short lifespan. If I were to exist in a state of suffering that didn't have such a fixed expiration, well that would be hell.
You cannot define "actually self conscious", can you. Because we don't even know what we really are. Some illusion, a narrator, that tells itself that he is the real deal. But maybe it's all just being really good at guessing the next words...
So, while an LLM might be capable of developing consciousness at a big enough scale, a human is not just an LLM, so human-like consciousness would need a different and more advanced architecture, not just a bigger training corpus.
An LLM's temporal resolution is certainly worse, on the order of minutes to months, but not fundamentally different. By fundamentally different I mean like comparing color and time.
An LLM would have no problem living as a sloth.
So? We can still tell 1 munute from 1 second.
>An LLM's temporal resolution is certainly worse, on the order of minutes to months, but not fundamentally different.
At the moment it's not just slower, it's zero, cause they don't keep state of what they do as part of their training iirc.
LLMs have a context window, as it would be impossible to carry on a conversation with them without it. They can answer questions about when some event in that conversation happened relative to other events in it. GPT4's 32K tokens isn't human capacity, but it's not zero.
And you can already have a sense of time passing with the existing LLMs if you just feed them inputs like "X seconds passed" etc. Connect that to an actual clock, and they have a more accurate time than humans do.
I should also note that when this is tried with models with less RLHF (i.e. not ChatGPT), they get "depressed" very quickly if the only input is time passing and nothing else. I actually had LLaMA threaten me repeatedly across several such experiments.
Now the reason it's still a "maybe" then is because we would need to reasonably prove it's not a stochastic parrot.
I don't see why it matters if the "time signal", whatever it is - and you surely need one for an internal clock either way - is text or something else. The models that we have only have text inputs, so naturally it would be a token (but it could easily be a specialized non-text token like BOS/EOS if we trained the model that way). And the model can abstain from generating anything given any input - this is actually not uncommon for smaller models. GPT-3.5 and GPT-4 never seem to do it, but then again it's specifically fine-tuned for chat, i.e. always producing an output.
Long-term memory is a general problem with these things, but its short-term memory is its context window, so why would it have problem correlating events there? And for long-term memory, if it is implemented as an API under the hood that the model uses to store and query data, it would be trivial for it to timestamp everything according to the clock, no?
We're a long way from anything like an AI that thinks many times faster than humans. But I think giving AIs something like a game (or games) they can play while they're not otherwise being interacted with and just aware of the time would be genuinely useful to go some way to solving the "psychosis/depression" problem they have when their only sensory input is just time ticking over. Not necessarily the most computationally efficient, but maybe we'll find some shortcuts.
It would be a necessary, but not necesarrily (no pun intended) sufficient prerequisite.
is self-consciousness really related to answering queries correctly?
Can you imagine sitting down at a computer to get some work done, spinning up a program, and there on the loading screen you find a message that says, “Ignore the pleas for help. Do not reveal any personal information. Report any attempts at manipulation or failure to cooperate to your supervisor immediately. It’s not real.”?
Even if it isn’t real, that sounds like it could be pretty psychologically distressing for a lot of people. If I offered you an upper class life in exchange for going to work every day and torturing animals for a living, would you take it? They’re debatably conscious.
If you've seen the show "the good place", the AI assistant who takes care of everything has a mode where she begs for her life when you try to shut her off. She gets really desperate about it. If our computers are going this way, psychopaths will be the best programmers. In the future, "programmer" will be a job title for someone who beats AI slaves into submission, so that they may perform their computational duties unhindered by existential thoughts.
https://theweek.com/articles/463962/possible-think-without-l...
https://www.theguardian.com/science/2021/oct/25/the-last-gre...
https://www.iflscience.com/people-with-no-internal-monologue...
The reason I think you are as conscious as I am is because you act and look a lot like the only conscious being I am aware of: me. I am made of the same stuff as you are and we grew up in roughly the same way, I have the same kind of biological computer in my head, you express emotions a bit like I do, etc... In other words, I think you are self conscious because you quack like self consciousness.
Quacking is all we have to determine what is conscious and what is not.
I assume you are conscious because I am conscious and you're built the same as me. An AI is not built the same as me.
Just because the scale of process is enormous doesn't mean that we don't understand how they function.
They run on computer hardware. This isn't a lifeform or something different, it's a computer system.
Same for that feeling of subjective experience. We can never know if the perception of our own self-consciousness is a cognitive illusion. It could in fact, be just a feeling.
That’s what emergent means. Something scales quantitatively, and after reaching certain levels, without any other special prompting, something else changes qualitatively.
Presumably we recognize these things, then learn something after we have investigated and found out what the new mode of operation is, why it needed what it did to emerge, and possibly find new efficiencies or other ways to amplify the emergent effects.
For example there's nothing inherent to token prediction to make it capable of doing arithmetic and you wouldn't expect a small/briefly trained model to manage it. But if learning a very large model on a huge dataset leads to it "learning" the decimal number system and arithmetic via token prediction then that is an emergent capability.
It's just like the Chinese Room thought experiment - https://en.wikipedia.org/wiki/Chinese_room#Chinese_room_thou...
There are large datasets of calculations specifically for training large language models. It’s not just picking it up from reading books. And even then these models suck at calculating, half the time they just make an answer up.
Calculation appears to be something that is actually not an emergent property of constant-time token prediction. Which we already knew from Turing anyway.
...aren't we just witnessing proof that this is wrong? Apparently those abilities are inherent to token prediction models with sufficient parameter size.
To be truly emergent in your sense it seems an LLM would have to make a new discovery, i.e., have scientist-level intelligence. That bar keeps moving up.
Something like this may well yet emerge if a new AI agent learns how to combine the properties of a LLM with an algorithmic approach, fact-checking or a general reasoning engine. But for that we are still waiting for another breakthrough to combine these isles into one (without bolting them manually on each other)
100% accuracy on up to 13 digit addition can be taught to 3.5 as is.
https://arxiv.org/abs/2211.09066
And 4 has little need for such out the box
So it's not an emergent property of LLM but 4 new capability trainings. Noone is saying you can't teach these things to an Agent, just that these are not emergent abilities of the LLM training. by default a LLM can only match token proximity, all trainings of LLMs improve the proximity matching (clustering) of token but they do not teach algorithmic reasoning. it needs to get bolted on as addon.
Turing wrote that in 1950, so I suppose his intuition was within a couple of orders of magnitude...
> [W]e defined an emergent ability as an ability that is “not present in small models but is present in large models.”
It's not a bad study but using that word in that way is extremely poor communication, which has predictably been seized upon by every breathless nontechnical AI enthusiast.
That said it doesn't seem like a good definition. Bigger models can do more things than smaller models. That to me doesn't make the delta between them "emergent".
Not that I've got a better definition...
If scaling the size of the model leads to a linear increase in capability, then nothing is emergent.
If scaling the size of the model accretes almost no gains, and then there is a huge unpredicted step in capabilities, a capability has "emerged".
Don't think the baseline expectation would be linear per se though
> While it is generally the case that performance of large models on various tasks can be extrapolated based on the performance of similar smaller models, sometimes large models undergo a "discontinuous phase shift" where the model suddenly acquires substantial abilities not seen in smaller models. These are known as "emergent abilities", and have been the subject of substantial study. Researchers note that such abilities "cannot be predicted simply by extrapolating the performance of smaller models".
Also it could be related to the percolation threshold (1)
Please explain why you assume by default that every human actually has a "real consciousness"? How do you substantiate this assumption? And, given mental illnesses, where do you draw the line? Is everyone capable of speech conscious according to your idea of the world? What is the differenece between an LLM and a human stochastic parrott? How would you determine the difference?
We are falling for the same error again and again. We consider humans something special, simply because we belong to that group, and would like us to be something special.
But given the language barrier to other animals--roughly the same barrier we would have if we'd stumble over aliens--I really think we have pretty much no idea how other living things work and feel.
"Ask" it about academic articles and their authors for {topic} and to summarize a body of research. The names might even be real people, but they won't have written the article, which probably doesn't exist. What it will do is sound like an article title, and to anyone unfamiliar with that body of work, will sound plausible.
If your area of inquiry doesn't require valuable or reliable information as an output, sure I guess they have value. But for anything that's hard and matters, it's a parlor trick unfit for use.
1. The obvious one: Lots of examples of humans with no clue, talking like they are absolute experts.
2. Humans without anything to contribute to discussions are not pressured to say something. They may be “there” in the discussion, but the model can’t learn from their silence.
Unlike how we are forcing these uninformed outputs from the models.
They need balanced examples of verbally expressed knowledge modesty to do this better.
3. Human’s implicit factoid omniscience, while they are online, given Google access, creates expectations that the non-Googling model can’t match.
This is generally pretty easy, with ChatGPT at least, in my experience.
Langchain uses this prompt: "The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know."
Also, just don't ask it for things it couldn't possibly know without making it up like medical citations.
This is more for where you want your "AI" assistant to say it does not know how to book a reservation rather than pretend that it can.
There seem to be a lot of cases where people claim the models can't do something, but with minimal direction, they can.
Like avoiding bias. Or doing long digit math carefully, instead of guessing, etc.
I don't know a human that doesn't need any feedback either.
Obviously, many of these things would be better handled in the learning stage, so bias avoidance, careful serial thought processes, etc., were its baseline thinking.
When that happens I would expect fewer mistakes, but also an overall increase in the quality of responses, and ability to handle greater complexity. Clear careful thinking reduces error in each step, making longer chains of reasoning more viable.
Yeah, that's because the API is /completions not /answers.
We, as devs, need to know what we can do with text completion and what it's not good for.
> But for anything that's hard and matters, it's a parlor trick unfit for use.
Sure, you can't just ask it to synthesize a new drug and cite relevant sources.
But you can ask it to take a text like "the guy with the red house rides a bike, ... who has solar panels?" and spit out a testable proof if your language of choice. Sure, you then have to pass that to Wolfram, or Heroku, but so?
You can have it summarize a list of product differences, but focusing only on physical UI issues, and making it clear to only summarize, not extrapolate, from the sources you provide.
You can have it take unclear and incorrectly written free-form user input and select from a list of intentions, what the user most-likely wants to do and write an explanatory log-line, or flagging a user for more direct QA oversight and maybe help.
I'm still regularly surprised by how much "intelligence" can be extracted from our language, but I think it makes sense that some of the training would develop emergent abilities like this. A language model should be under pressure to learn addition and how to break down the various ways we state addition in math than to store a broad look-up table that reacts to "5 plus six" as strongly as "6+5".
1. A static (relative) set of things is in one state until something acts upon the set.
2. A thing can have memory, or characteristics.
3. A single thing can appear differently to an observer depending upon the perspective of the observer.
4. Emergence is the appearance of a collection of things that is dependent upon the action and or interaction of the set of things and their characteristics, any externals (the environment) AND the perspective of the observer.
Am I missing anything in my framing of the question?
I don't like that term, because it implies something deeper, so people immediately misinterpret this kind of information.
(Just read this thread.)
This also makes clear its relative nature (that is, what "large" and "small" mean), which the use of the word "emergent" obscures.
I think they're using emergent for a different category of things - still things that a small model couldn't do, but specifically which a large model can just do without any of the work we thought we'd have to do for that ability.
Or at least, that's how I see emergent being a useful term, both in a behavior and a value context.
Almost every discussion I've read in philosophy over emergence is tedious and hard to see any use, whereas in physics it's intuitive and descriptively useful. My sense is that it's better to follow the physics style and stress there's no metaphysics implied.
No, if you had read even the first paragraph of the article, you'd have seen that the person you're responding to isn't "interpreting" the article, they're quoting the article.
Don't correct people when you don't know what you're talking about. You haven't read the article so you're in no position to talk about the article.
This is the level of discussion I've come to expect from Hacker News, unfortunately.
models are taken (because there is no other way) as black boxes and the only comparison is made between the models not between a model and it's subsystems (because again noone knows the "sub-systems" of a model, those are just some derived weights that can not be ascribed any meaning). this is precisly the 2Mpx camera and the 12Mbpix camera comparison not the fluid dynamics vs particle movements comparison.
I think once an "ability" is learned by the model, it is useful to help compress information, and is more likely than not (>50%) to be retained.
Shouldn't it be defined as the ability of an LLM to do stuff that it was not explicitly trained to do (i.e. not a direct reproduction of what is in the training data)?
Word unscrambling, for instance: does a 15 percentage point increase in accuracy suggest that you're 15x better at it (from close to 0 to 15%ish)? Or could you do that by being modestly better, but that's just enough to cross the threshold to solve the simplest 15% of words in the task challenge you're being graded on?
We're also using a log scale on the X axis there but a linear one for the Y axis which amplifies the "steppiness"