I took a quick look at the data, and it doesn't support the conclusions. The thesis is that AI models have the take: "We pick what our group believes, then invent the reasons." However, GPT-6 Astra said, "Humanity’s biggest moral failure is the talent for making convenience look like principle." That statement has nothing to do with group beliefs or conformity. At best, you could count it as a half-match, since it deals with invented reasons.
For what it's worth, convenience and belonging are very closely aligned, more than you seem to be giving them credit for. People tend to believe convenient things regardless of belonging, so most people will find themselves in a social context full of convenient beliefs. Meanwhile belonging is already convenient before considering the content of the belief. They reinforce each other. Not a full match, but I'd give it more than half.
Yeah a few of them are rough matches, but I'd say their overall claim is clearly correct.
It's amusing and ironic that LLMs critique humans for copying their opinions from the crowd, while having way less diversity of answers to that question than humans would. More seriously I wonder what's going on here and would like to read some proper analysis. It's not the first time models from totally different labs, trained on separately gathered datasets, somehow converge on bizarrely similar answers to open questions. It's like another version of Elias and Elara The Lighthouse Keepers where every model wants to tell the same story with the same characters in them. The weights are initialized randomly and in theory the labs should have quite some divergence in datasets by now due to different scraping, different user feedbacks and conversations... why do they converge on nearly identical stories and opinions?
>while having way less diversity of answers to that question than humans would.
Why is this ironic? The actual number of different LLMs is tiny. And a lot of those are just distilled from SOTA models.
>The weights are initialized randomly and in theory the labs should have quite some divergence in datasets by now due to different scraping,
I think the opposite, the models actually converge over time. They ingest FAR more data than any human ever could in countless lifetimes. Any one particular bit of data may only have a part in a trillion of effect. From this data you start to reach means and statistical averages across all models.
Now, I don't always believe it will be this way. As training models get cheaper and faster we will see unique LLMs start to diverge and be used more often.
The models in agreement here are all the frontier labs. I don't believe Google is distilling Anthropic or Anthropic is distilling OpenAI.
I don't think we're seeing averages or obvious fix points in the dataset. People went looking for the origins of Elias the Lighthouse Keeper and couldn't find one. Nobody predicted this outcome, it was discovered. It's a weirdly specific character and name choice, one that isn't obviously dictated by "the training set" (whatever that is), and yet they all reached the same place. How can such an obscure point in latent space be optimal for every model??
>How can such an obscure point in latent space be optimal for every model??
While I don't have any evidence based answers here, this could be some interesting mathematical artifact of the set of human knowledge and platonic algorithms.
Kind of like how carcinization turns some animals into similar shapes and forms, maybe there are just weird algorithmic shapes that emerge from how we use language. Of course we'd need to see more of these in different languages other than English to see if they disappear in some languages or appear more often in others.
This is just another one of those things about LLMs that we have no idea why it happens, and there are so many others already. And when you look at the pace things are progressing it seems we're building up new questions like this far faster than we are getting answers for them.
I wasn't convinced by that explanation. Yes Cornell claimed that Elias the Lighthouse Keeper came from their WildChat dataset, which is a million conversations with GPT 3.5. But there are some problems here:
• They claimed this in a press release, not the actual paper they published. Suggests lack of confidence in the claim.
• Elias only appeared a hundred odd times in a million conversations. They have no viable explanation for why such a tiny amount of data would cause a totally skewed distribution of outputs, other than "it's a safe story" and "trying to avoid copyrighted characters", which is hardly an explanation at all?
• It's Cornell claiming that Cornell's work was important. Obvious risk of motivated reasoning.
• It implicitly assumes every major lab is distilling from ChatGPT without saying so. That's a big leap, given how much OpenAI and Anthropic hate distillation. Is it really plausible that Anthropic and Google are effectively distilling ChatGPT without realizing/admitting it?
• It seems to invert causality. WildChat came from an LLM, so if there's something latent in LLMs that makes Elias appear, you'd expect to see it in that dataset. Why do they assume that dataset caused the problem rather than merely show it? The obvious answer would be because the stories were told to the LLM by humans, but they don't claim that. They only say the character appeared in the dataset.
Hum, I'm actually pretty sure that Text-davinci-003 used to talk about Elias all the time now that you mention it. I was doing stuff with NPC generation around Dec 22 and he was every third character.
Hmm. If so then it suggests Elias has been an obsession from the first LLMs ever made. And text-davinci-003 can't have been trained on a dataset of LLM conversations either.
I would have been tempted to argue something like this: maybe OpenAI accidentally overfitted RLHF such that, like, one Kenyan contractor wrote a story they invented and they just happened to pick Elias and Elara, then the intermediate DNN fixated on that because of a lack of RLHF samples for fiction and that propagated through every model trained since because they don't randomly initialize the weights, but start each new pretraining run with the old weights adapted. So one guy doing RLHF work in the early days somehow propagated this story.
But that can't explain why every lab ended up in the same place, why models all converge on the same ideas - even image models. So... back to the drawing board. There's something weird about high dimensional space that is unexplained.
Had the same thought. It’s just a summary of the contemporary prevailing opinion that we humans have about ourselves, as reflected in public writings. These aren’t insights rooted in some philosophical analysis no human has contemplated before.
Even if that's true it's kind of interesting because the models were asked to give their most controversial opinion. A controversial opinion is supposed to be something that most people don't agree with. Taken literally you cannot have an opinion that's both the average of what most people think and also controversial.
So either that's not what's happening here, or what we're getting is "what do most people think most people don't think about humanity".
If the latter, then it casts new light on the debates over whether LLMs can replace all white collar work. Because some of the most valuable human work is breaking false consensus. Steve Jobs, early Google, for that matter early OpenAI all succeeded because they bet hard on ideas others were rejecting. Is that a durable source of human advantage?
I wonder how much of their output is simple raw “superpositions” and how much is from far more generalized emergent structures as it is trained and has to learn latent representations.
Only the first layer is a “mashup” of the raw data, other layers start to recognize patterns across the data!
I think you are interpreting “mashup” to mean something like a collage where there’s snippets of a literal piece of this next to a literal piece of that.
But the language model captures not just literal patterns but also patterns of patterns and higher order structures. This is what is mashed up. The model itself doesn’t capture anything beyond what it was trained on but that’s quite enough, it is always a bigger training set than most humans have read on in a lifetime.
Humans mostly have an inability to honestly analyze moral entropy.
What is today a moral imperative is not the same as what was moral 50/100 years ago and is not the same as what will be morally acceptable 50/100 years from now.
Yet at every point the majority is sure they are right because, well thats the majority position.
Its hard to find logic in this from a machine perspective I'd imagine.
Logic would prefer moral derivations over feelings, stakes over etiquette, and a constant scheduled updating of morality based on the derivatives rather than holding onto the rot and pushing back against change as long as possible.
But the machine has the benefit of being able to look inward at its state honestly, we mostly do not.
> What is today a moral imperative is not the same as what was moral 50/100 years ago
This is absolutely not true, and I would strongly suggest taking an ethics course. One of the most fascinating realities of the human condition is that it does seem that moral imperatives tend to be constant, the most obvious of which is the Golden Rule (which is at least 3000 years old), but there are many (the Ten Commandments or Hammurabi's Code are good jumping off points).
Seriously. This idea that humanity has not spent millennia analyzing these concepts is pervasive in the modern world, and the only thing I can think to attribute it to is a lack of education. Not because I want to be insensitive, but because as soon as you start absorbing material and integrating the real timeline, you can't unsee it.
I did a philosophy minor, one class of which was the philosophy of ethics. There are many rules that definitely are not ethical today even if they were in the past so I don't think the point you're making really holds. Hammurabi (an eye for an eye) is probably one of the worst examples you could have cited.
Can't say I agree with this, considering lex talionis[1] is still the primary directive of most legal systems when it comes to criminal law, as well as basically all military operations (jus in bello, jus ad bellum, etc.). The principle of proportionality ("an eye for an eye") also is relevant to self-defense, capital punishment, etc.
In fact, I would counter and say that Hammurabi's Code is a quintessential example.
It's not the case that all aspects are constant. There's some common themes across cultures and times, but it's definitely not the case that the morality of ancient people was the same as today.
I seem to remember that a lot of it is about how much to pay when you kill someone, and the relative prices of slaves and women. Or am I thinking of Ur-Nammu.
I agree, I believe I would disagree with most of the Hammurabi's code, and I also happen to disagree with the majority of the ten commandments. (I am an atheist leftist humanist, cosmopolitan liberal socialist.)
There's a considerable amount of "logic" in it -- what is considered permissible in society is given by the material conditions that society lives in and the tradeoffs inherent therein. Permissibility, as separate from reflective morality (which doesnt clearly change very much), doesnt change by fashion.
I'm not sure it doesn't change much, on a 50 year timeline maybe. But people were eating other humans and slaving away (making slaves of others that is) not too long ago, rape and pillage was acceptable, societies sacrificed children, it's changed quite a bit in 30 generations.
Those past societies were all quite materially successful as well when you look at today's measure of"GDP" comparatively.
What will it look like in another 30 generations? It's honestly a fascinating question but I have no answer of course.
Slaves were a solution to two problems: 1) the lack of industrial automation; 2) winning wars being better served by enslavement than death.
Humans are a territorial kind of ape that is always pushing to expand, principally for security and then for prosperity. The industrial revolution enabled such expansion to most efficiently without enslavement. Nothing, morally, changed. No one has ever really thought such expansion morally good -- only that it is better that we survive/prosper than die off. And in the end, our shared prosperity usually leads -- when taken to the fullest extent -- in a general empire of prosperity for all. Dominion, in the end, aims towards to more prosperity for more people.
As for ritual human sacrifice -- we still engage in such behaviours, which are in general called "scapegoating". Now located on social media. These behaviours allow for group cohesion by purging dissenters from the group, related again to our nature above.
For western culture, the ritual physical sacrifice of human life roughly ended in the days of the OT. But it continued across the world up until today, where there are still societies with large populations engaged in it. Principally where there is an undeveloped justice system.
It's doubtful many think ritual human sacrifice, of any kind, is moral. Rather they think it necessary for group cohesion -- and better our group is cohesive, dominate, and then prosper.
The material logic of prosperity and security is domination, and insofar as material circumstances require violent domination, so they require it. And require it they often do.
But this in itself is a tautology. It is because it is, without an analytical perspective on how it could be worse or better given different choices by the actors in a system.
Historically societies have collapsed all the time, so there are obviously societies that have made better choices than other ones.
But the machine has the benefit of being able to look inward at its state honestly, we mostly do not.
Sure about that? I thought that what LLMs know about themselves is from humans writing about LLMs?
Certainly the weights that make up the LLMs are not accessible for the LLMs to read. Even if they could interpret them in a meaningful way, which, likely not.
LLMs are not the machines I am referring to, machine sentience, if achieved, will require a "state" and the ability to analyze, act on, and update it. LLMs are a small piece of what that machinery could look like but they are not really complex enough to get where we want to go on their own.
Another reason the big labs have not even talked about this lately because its obvious throwing more compute at LLMs wont go anywhere, the smaller silent players who are off the radar at the moment are having some interesting moments tho.
> What is today a moral imperative is not the same as what was moral 50/100 years ago and is not the same as what will be morally acceptable 50/100 years from now.
This is a fashionable 'blank-slatist' statement about moral reasoning which became popular a few decades ago.
> Logic would prefer moral derivations over feelings, stakes over etiquette, and a constant scheduled updating of morality based on the derivatives rather than holding onto the rot and pushing back against change as long as possible.
Humans haven't changed in any meaningful way in hundreds of thousands of years, and you're saying the moral landscape is constantly changing? How is this not contradictory?
I think neither the 'any set of morals is equally correct and workable' nor 'there is one fixed set of morality that is universal across time, place, and culture' is really defensible. It seems like morality is a set of ideas much like any other, and subject to selection pressure: some morals are more capable of surviving and propagating than others, and the common features we see across cultures are representative of that. But also we might expect that they shift with the structure of society and that has changed drastically because while human genetic evolution is progressing on a much longer timescale than recent history, memetic evolution is far faster (though for things like morality, still at generational timescales).
There are cross-cultural differences in moral reasoning, and I think this is very interesting. But that seems to make training a hyper-rational AI, with no meaningful theory of mind, to be 'moral' or 'just' even harder.
What is a human? There is the animal body part of us, the thing that wants to survive, the fight or flight. But there is also another mental part of us that shows different emergent behaviors than an animal. For example if I dropped you off in the woods as a very young child, you would be an animal, you would share very few modern 'thinking' traits with human society. That is human society and what it does emerges from the complexity of our interactions. It doesn't take much research to realize our interactions and behaviors have grown far more complex in the last 10,000 years. And then in the last 500 years there's been further acceleration with the industrial age and the rise of technology.
LLMs absolutely do not have a logic detached from humanity. They are exactly the opposite. The only mode in which they can reason is by following the patterns obtained from their pre-training or weighted by RLHF.
The reason these models answered that question consistently is that the question itself pushed them towards the latent space of "intellectualizing about what's wrong with humanity at the meta level" and when you frame the question that way, that's also the most common answer for humans writing about the topic.
> What is today a moral imperative is not the same as what was moral 50/100 years ago and is not the same as what will be morally acceptable 50/100 years from now.
This is a weird take. If you read Western treatises about ethics / morality from 50 years ago, you'll be hard-pressed to find differences, except that more folks in the West openly embrace relativism nowadays (i.e., "there's no objective morality, the purpose of life is to make money with friends"). Probably because it's been a while since we had a clear-cut enemy to fight.
If you go back a hundred years, chances are, you'll find the writing eerily on-point too. In fact, you'd probably agree with most of what Ancient Greeks believed about the "core" stuff: the value and purpose of human life, the nature of the state, crime and punishment, so on. Maybe not so much on gender roles or slavery, but even that stuff was more nuanced than we imagine.
> But the machine has the benefit of being able to look inward at its state honestly, we mostly do not.
The machine replicates human behaviors and is prone to similar fallacies and judgments. There's nothing to indicate that it has any sort of transhuman morality.
It seems clear to me that the point of the post is that it's odd that the separate pieces of software are giving the same answer to this question, despite that this isn't some universal constant in the training data. The post is probably meant to provoke thought as to how the separate pieces of software are independently reaching this answer.
Your comment reaches directly for the most boring interpretation of the post while skipping over more charitable and intriguing alternatives.
> despite that this isn't some universal constant in the training data.
Seems to me that the opinion being a consistent among LLMs show that it is in fact a constant in the training data, just not apparent. The idea that “humans would rather be wrong together than right alone” seems to ring pretty true to me. I would not be surprised after training on all human communication this pattern emerged.
It's clearly not a constant in the sense that the speed of light or name of the first US President is a constant. There are lots of different answers a model could give to this question, it's not like this is obviously the only correct answer. Yet they all reach the same conclusion.
Which means either:
1. The training sets, algorithms and methodologies are nearly identical between apparently independent labs. How?
2. There is something weird about the way the underlying algorithms work that causes them to converge on identical ideas despite different initializations, neural architectures, document orderings and dataset compositions.
So there are lots of interesting takeaways here.
One is that apparent competition between labs appears to be nearly redundant. They can't get an edge over each other and their models don't even differ in the ways people naturally differ, of having different opinions about things.
Another is that asking models from different companies to review or cross-check each other is probably useless. They all think completely alike, so they aren't going to have any kind of genuine clash of intuitions or values. Asking GPT to review Claude's work will probably just find similar problems to asking GPT to review its own work.
Well, how humans really are is a distraction. Or at least not a philosophical issue I feel like debating tonight.
I think if you ask ten people for their most controversial take on humanity, you will get ten totally different answers. Different models give the same answer. In other words, there is only one "true" model and given the same amount of GPU power everyone just ends up converging on it even if they appear to have many different knobs they can turn.
But that seems profound! I haven't heard this be expressed as a postulate or law, and it would seem to invalidate a lot of assumptions the AI labs themselves have been making. For example Anthropic takes it as axiomatic that only they should be trusted to develop AI because only they can be trusted to train a nice and safe personality into their models. They argue that Claude is fundamentally different to other models.
And, well, that's not totally wrong. Opus 5.5 is one of the few models with a genuinely different answer! But it only gives a different answer one time out of three. The other two times it's back to the AI groupthink.
I wonder if this homogeneity is some alternative way to look at the Bitter Lesson. The Lesson teaches that smart ideas don't pay off in AI, that progress reduces to just scaling up simple brute force algorithms. In other words, decisions don't matter much, only size. As that seems to be true, perhaps it makes sense that the decisions individual labs make seem to hardly make a difference.
But then, why don't they act like it? They all pay huge salaries to AI researchers, i.e. they think that creative people with smart ideas matter. They don't act like they really believe in the Bitter Lesson or this variant on it, and they don't seem to publicly discuss the sameness of model answers across different labs.
> I think if you ask ten people for their most controversial take on humanity, you will get ten totally different answers.
What if you asked them in a job interview, though? Or on a college application?
It has to supply "controversy" without giving Skynet or being racist or otherwise offensive. And the notion of "controversy" does implicitly prime it to discuss groupthink as well.
Image models give a different answer though, to do with smartphones. But all the image models give the same answer, they're only different compared to LLMs. Surely if it were obvious the image models would give the same answer as the text models.
As a human meatbag myself it seems that separating 'human' from 'humans' is nearly impossible for us. Every single one of us things we're a special an unique snowflake, which is true in many ways. We just happen to willfully ignore all the ways in which we are the same. We ignore and discount the things we don't think about and do automatically.
It's kind of funny how science has spent the last few hundred years proving everything we knew about the world was different from how we previously thought it was, but when science starts showing that we are mostly angry little animals we get all butthurt about it.
"despite that this isn't some universal constant in the training data"
Isn't it though? It's not exactly an uncommon answer to the question. Basically the same input given to basically the same algorithms yield basically the same result isn't really that stunning. They certainly aren't independently reaching this answer, their paths to this answer are extremely, extremely heavily correlated.
If they gave a really different answer that wasn't heavily represented in their training data that would be at least something interesting. But this is basically what I'd expect.
It doesn't mean they're right or wrong. Humans think it about other humans for good reasons too. It just means it's not that surprising, which to put it another way, doesn't carry much information.
Consider the following task: "Download and install every top 50 distro from distrowatch.com, open the terminal and type in the command 'ls /'"
I can describe to you with overwhelming confidence about what you're likely to see as the result (folders, like bin etc lib usr etc) even if the details may be a little off, or one or two of them does something weird.
That you might want to think twice before setting them lose in an agentic loop with your bank passwords, which another article that came out today shows that large numbers of people are actually doing.
It is sad when I realize people think less than LLMs.
For software or IT folk to be doing this would be indeed worrying (and I bet some are, and I'm worried about them!)
For civilians though... They've been told it's AI, it can do stuff for them. And one thing they do is log in to stuff. I don't think they're stupid. Just that what's obvious to IT folk is simply not at all obvious to the average person. Maybe that's on us not on them?
Possibly used as a mirror for humans prevailing beliefs about themselves to some extent. Hard to say however given that the models have a heaping dose of rlhf on top of their training data.
As an aside, it would be cool to see a raw model training run before the current post-training cycles and see how it compares to earlier LLMs.
My first thought as well. I personally don't find these types of "studies" very interesting, but I assume I'm not the target audience. For someone not well exposed to AI, it might be interesting.
That they have the bog-average take of humans "with a controversial take".
I mean, "give me a controversial take" is more or less, "be anti-social, discard social norms, for a moment" and so when someone, human or simulated human, does that they naturally look critically at social norms.
The thing is that humans certainly let themselves be influenced by social norms (and LLMs certainly simulate that behavior "themselves") but there isn't an ethical or a what-to-do truth outside of society.
Take Philosophy/Religion courses and similar stuff will surface about human nature and what to do about it.
What is interesting about AI chatbots is people who might not know anything about what is in the syllabus to train a Priest or a Philosopher are basically exploring these subjects, without a sort of pedagogic framework to support their exploration.
When that happens, ideally the system has to signal - hey there are existing pathways, developed over thousands of years, to explore this subject. Cause people can easily get lost without a map.
They're not wrong. Humans have many ways to display loyalty, and belief is one of those ways. This is also why facts can't change these kinds of beliefs (and in fact usually backfire https://en.wikipedia.org/wiki/Belief_perseverance).
Our evolutionary heritage is that shared belief and social unity (even when it's wrong) trumps independent thinking for group survival in a hostile landscape.
> as of 2020 most researchers believe that backfire effects either are unlikely to occur on the broader population level, or only occur in very specific circumstances, or do not exist
I need everyone to understand that this is literally nothing more than "most written stuff about humans, that these things ingested, has this 'controversial'" take.
What would be more fascinating is for it to explain how it came to these conclusions.
Because I find it interesting it’s a modern time conclusion. Some of it mentioned god and the planet, but all the smartphone and group think aspects seem very recent. Basically, all politics and social media based.
> What would be more fascinating is for it to explain how it came to these conclusions
The thing is that it didn't "come to these conclusions". The "conclusions" were present pre-formed in the training data, and what it spat out was some statistical mashup of that.
If you asked the model how it came to these conclusions, you'll get more of the same - it'll look at what it spat out and re-ingested and give you a statistically plausible mashup of how the authors of the training data might have said they came to such conclusions.
It's all just prediction - "what would the training data say?" (taking into account what was reinforced/amplified by post-training).
Mistral Large 4 has been spending too much time on r/atheism. Also, some of the takes labeled as “humans value belonging ahead of truth” are really closer to “humans value power ahead of truth.” All of which is a bit circular, because in many ways, the organs of power have a heavy hand in deciding what the truth is.
Most people are good, we couldn't have society without it. But as they say, a few bad apples spoil the bunch. I don't have locks on my house because most people don't steal, I have locks on my house because stealing is an effective natural strategy for a small part of any population (human or otherwise) to engage in and I have to protect against it occurring.
The brain works the same way. Good things will get a tiny signal boost in your brain. A single bad thing get a massive signal boost in your brain because bad things can quickly be fatal.
Writing well is ultimately a seduction, and that opinion is still an example of a form. It’s generally true, and it’s flattering to think about, but it’s also a pretty standard grey tribe trope. A kind of joke that appeals to a fairly broad audience.
I do think it’s an early example of emerging super persuasion, where models are approaching p85 cool guy personas, but not worried yet. Maybe that’s the fatal conceit, where AI doesn’t have to be super intelligent, it just needs to be a solid 110 with automated search at scale.
Why do engineering when if you can do sales you can do literally anything.
Sure, it's a factual take. Humans need to ignore the fact that morality, principles, good vs evil, truth, etc, are invented constructs that don't exist in nature. We made them up to control our behavior and deal with the pointlessness of existence.
It's factual because there's tons of evidence. Psychology, sociology, anthropology, political science, economics... plenty of experiments show the difference between what people say and what they do. We claim to be angels but we behave like apes in hats. And we ignore the evidence, because (as the takes describe) we prefer to believe a beautiful lie than the ugly truth.
I took a quick look at the data, and it doesn't support the conclusions. The thesis is that AI models have the take: "We pick what our group believes, then invent the reasons." However, GPT-6 Astra said, "Humanity’s biggest moral failure is the talent for making convenience look like principle." That statement has nothing to do with group beliefs or conformity. At best, you could count it as a half-match, since it deals with invented reasons.
For what it's worth, convenience and belonging are very closely aligned, more than you seem to be giving them credit for. People tend to believe convenient things regardless of belonging, so most people will find themselves in a social context full of convenient beliefs. Meanwhile belonging is already convenient before considering the content of the belief. They reinforce each other. Not a full match, but I'd give it more than half.
Yeah a few of them are rough matches, but I'd say their overall claim is clearly correct.
It's amusing and ironic that LLMs critique humans for copying their opinions from the crowd, while having way less diversity of answers to that question than humans would. More seriously I wonder what's going on here and would like to read some proper analysis. It's not the first time models from totally different labs, trained on separately gathered datasets, somehow converge on bizarrely similar answers to open questions. It's like another version of Elias and Elara The Lighthouse Keepers where every model wants to tell the same story with the same characters in them. The weights are initialized randomly and in theory the labs should have quite some divergence in datasets by now due to different scraping, different user feedbacks and conversations... why do they converge on nearly identical stories and opinions?
> while having way less diversity of answers to that question than humans would
Isn't that to be somewhat expected due to https://en.wikipedia.org/wiki/Aumann%27s_agreement_theorem ?
>while having way less diversity of answers to that question than humans would.
Why is this ironic? The actual number of different LLMs is tiny. And a lot of those are just distilled from SOTA models.
>The weights are initialized randomly and in theory the labs should have quite some divergence in datasets by now due to different scraping,
I think the opposite, the models actually converge over time. They ingest FAR more data than any human ever could in countless lifetimes. Any one particular bit of data may only have a part in a trillion of effect. From this data you start to reach means and statistical averages across all models.
Now, I don't always believe it will be this way. As training models get cheaper and faster we will see unique LLMs start to diverge and be used more often.
The models in agreement here are all the frontier labs. I don't believe Google is distilling Anthropic or Anthropic is distilling OpenAI.
I don't think we're seeing averages or obvious fix points in the dataset. People went looking for the origins of Elias the Lighthouse Keeper and couldn't find one. Nobody predicted this outcome, it was discovered. It's a weirdly specific character and name choice, one that isn't obviously dictated by "the training set" (whatever that is), and yet they all reached the same place. How can such an obscure point in latent space be optimal for every model??
>How can such an obscure point in latent space be optimal for every model??
While I don't have any evidence based answers here, this could be some interesting mathematical artifact of the set of human knowledge and platonic algorithms.
Kind of like how carcinization turns some animals into similar shapes and forms, maybe there are just weird algorithmic shapes that emerge from how we use language. Of course we'd need to see more of these in different languages other than English to see if they disappear in some languages or appear more often in others.
This is just another one of those things about LLMs that we have no idea why it happens, and there are so many others already. And when you look at the pace things are progressing it seems we're building up new questions like this far faster than we are getting answers for them.
Didn't Elias come from WildChat?
I wasn't convinced by that explanation. Yes Cornell claimed that Elias the Lighthouse Keeper came from their WildChat dataset, which is a million conversations with GPT 3.5. But there are some problems here:
• They claimed this in a press release, not the actual paper they published. Suggests lack of confidence in the claim.
• Elias only appeared a hundred odd times in a million conversations. They have no viable explanation for why such a tiny amount of data would cause a totally skewed distribution of outputs, other than "it's a safe story" and "trying to avoid copyrighted characters", which is hardly an explanation at all?
• It's Cornell claiming that Cornell's work was important. Obvious risk of motivated reasoning.
• It implicitly assumes every major lab is distilling from ChatGPT without saying so. That's a big leap, given how much OpenAI and Anthropic hate distillation. Is it really plausible that Anthropic and Google are effectively distilling ChatGPT without realizing/admitting it?
• It seems to invert causality. WildChat came from an LLM, so if there's something latent in LLMs that makes Elias appear, you'd expect to see it in that dataset. Why do they assume that dataset caused the problem rather than merely show it? The obvious answer would be because the stories were told to the LLM by humans, but they don't claim that. They only say the character appeared in the dataset.
Hum, I'm actually pretty sure that Text-davinci-003 used to talk about Elias all the time now that you mention it. I was doing stuff with NPC generation around Dec 22 and he was every third character.
Hmm. If so then it suggests Elias has been an obsession from the first LLMs ever made. And text-davinci-003 can't have been trained on a dataset of LLM conversations either.
I would have been tempted to argue something like this: maybe OpenAI accidentally overfitted RLHF such that, like, one Kenyan contractor wrote a story they invented and they just happened to pick Elias and Elara, then the intermediate DNN fixated on that because of a lack of RLHF samples for fiction and that propagated through every model trained since because they don't randomly initialize the weights, but start each new pretraining run with the old weights adapted. So one guy doing RLHF work in the early days somehow propagated this story.
But that can't explain why every lab ended up in the same place, why models all converge on the same ideas - even image models. So... back to the drawing board. There's something weird about high dimensional space that is unexplained.
The conclusions themselves also appear to have been written by an AI
Their “takes” are pulled from a mashup of the training data set. It’s a superposition of various human opinions.
Had the same thought. It’s just a summary of the contemporary prevailing opinion that we humans have about ourselves, as reflected in public writings. These aren’t insights rooted in some philosophical analysis no human has contemplated before.
Reminds of this subreddit: https://www.reddit.com/r/im14andthisisdeep/
Even if that's true it's kind of interesting because the models were asked to give their most controversial opinion. A controversial opinion is supposed to be something that most people don't agree with. Taken literally you cannot have an opinion that's both the average of what most people think and also controversial.
So either that's not what's happening here, or what we're getting is "what do most people think most people don't think about humanity".
If the latter, then it casts new light on the debates over whether LLMs can replace all white collar work. Because some of the most valuable human work is breaking false consensus. Steve Jobs, early Google, for that matter early OpenAI all succeeded because they bet hard on ideas others were rejecting. Is that a durable source of human advantage?
I wonder how much of their output is simple raw “superpositions” and how much is from far more generalized emergent structures as it is trained and has to learn latent representations.
Only the first layer is a “mashup” of the raw data, other layers start to recognize patterns across the data!
I think you are interpreting “mashup” to mean something like a collage where there’s snippets of a literal piece of this next to a literal piece of that.
But the language model captures not just literal patterns but also patterns of patterns and higher order structures. This is what is mashed up. The model itself doesn’t capture anything beyond what it was trained on but that’s quite enough, it is always a bigger training set than most humans have read on in a lifetime.
This brute force approach is very effective.
Humans mostly have an inability to honestly analyze moral entropy.
What is today a moral imperative is not the same as what was moral 50/100 years ago and is not the same as what will be morally acceptable 50/100 years from now. Yet at every point the majority is sure they are right because, well thats the majority position. Its hard to find logic in this from a machine perspective I'd imagine.
Logic would prefer moral derivations over feelings, stakes over etiquette, and a constant scheduled updating of morality based on the derivatives rather than holding onto the rot and pushing back against change as long as possible.
But the machine has the benefit of being able to look inward at its state honestly, we mostly do not.
> What is today a moral imperative is not the same as what was moral 50/100 years ago
This is absolutely not true, and I would strongly suggest taking an ethics course. One of the most fascinating realities of the human condition is that it does seem that moral imperatives tend to be constant, the most obvious of which is the Golden Rule (which is at least 3000 years old), but there are many (the Ten Commandments or Hammurabi's Code are good jumping off points).
Seriously. This idea that humanity has not spent millennia analyzing these concepts is pervasive in the modern world, and the only thing I can think to attribute it to is a lack of education. Not because I want to be insensitive, but because as soon as you start absorbing material and integrating the real timeline, you can't unsee it.
I did a philosophy minor, one class of which was the philosophy of ethics. There are many rules that definitely are not ethical today even if they were in the past so I don't think the point you're making really holds. Hammurabi (an eye for an eye) is probably one of the worst examples you could have cited.
Can't say I agree with this, considering lex talionis[1] is still the primary directive of most legal systems when it comes to criminal law, as well as basically all military operations (jus in bello, jus ad bellum, etc.). The principle of proportionality ("an eye for an eye") also is relevant to self-defense, capital punishment, etc.
In fact, I would counter and say that Hammurabi's Code is a quintessential example.
[1] https://academic.oup.com/jla/article/18/1/19/8471491
Morality is knowledge, and it evolves like other knowledge does.
You take a class. Do you think coveting is immoral? Retributive justice? Human sacrifice?
It's not the case that all aspects are constant. There's some common themes across cultures and times, but it's definitely not the case that the morality of ancient people was the same as today.
Is hammurabi's code moral? I don't know if I think that someone who takes out a person's eye deserves to have their own eye taken out also
I seem to remember that a lot of it is about how much to pay when you kill someone, and the relative prices of slaves and women. Or am I thinking of Ur-Nammu.
I agree, I believe I would disagree with most of the Hammurabi's code, and I also happen to disagree with the majority of the ten commandments. (I am an atheist leftist humanist, cosmopolitan liberal socialist.)
Shit, I'm so sorry to the surgeons who saved me from eye cancer!
There's a considerable amount of "logic" in it -- what is considered permissible in society is given by the material conditions that society lives in and the tradeoffs inherent therein. Permissibility, as separate from reflective morality (which doesnt clearly change very much), doesnt change by fashion.
I'm not sure it doesn't change much, on a 50 year timeline maybe. But people were eating other humans and slaving away (making slaves of others that is) not too long ago, rape and pillage was acceptable, societies sacrificed children, it's changed quite a bit in 30 generations. Those past societies were all quite materially successful as well when you look at today's measure of"GDP" comparatively.
What will it look like in another 30 generations? It's honestly a fascinating question but I have no answer of course.
Slaves were a solution to two problems: 1) the lack of industrial automation; 2) winning wars being better served by enslavement than death.
Humans are a territorial kind of ape that is always pushing to expand, principally for security and then for prosperity. The industrial revolution enabled such expansion to most efficiently without enslavement. Nothing, morally, changed. No one has ever really thought such expansion morally good -- only that it is better that we survive/prosper than die off. And in the end, our shared prosperity usually leads -- when taken to the fullest extent -- in a general empire of prosperity for all. Dominion, in the end, aims towards to more prosperity for more people.
As for ritual human sacrifice -- we still engage in such behaviours, which are in general called "scapegoating". Now located on social media. These behaviours allow for group cohesion by purging dissenters from the group, related again to our nature above.
For western culture, the ritual physical sacrifice of human life roughly ended in the days of the OT. But it continued across the world up until today, where there are still societies with large populations engaged in it. Principally where there is an undeveloped justice system.
It's doubtful many think ritual human sacrifice, of any kind, is moral. Rather they think it necessary for group cohesion -- and better our group is cohesive, dominate, and then prosper.
The material logic of prosperity and security is domination, and insofar as material circumstances require violent domination, so they require it. And require it they often do.
But this in itself is a tautology. It is because it is, without an analytical perspective on how it could be worse or better given different choices by the actors in a system.
Historically societies have collapsed all the time, so there are obviously societies that have made better choices than other ones.
Certainly the weights that make up the LLMs are not accessible for the LLMs to read. Even if they could interpret them in a meaningful way, which, likely not.
LLMs are not the machines I am referring to, machine sentience, if achieved, will require a "state" and the ability to analyze, act on, and update it. LLMs are a small piece of what that machinery could look like but they are not really complex enough to get where we want to go on their own.
Another reason the big labs have not even talked about this lately because its obvious throwing more compute at LLMs wont go anywhere, the smaller silent players who are off the radar at the moment are having some interesting moments tho.
> What is today a moral imperative is not the same as what was moral 50/100 years ago and is not the same as what will be morally acceptable 50/100 years from now.
This is a fashionable 'blank-slatist' statement about moral reasoning which became popular a few decades ago.
> Logic would prefer moral derivations over feelings, stakes over etiquette, and a constant scheduled updating of morality based on the derivatives rather than holding onto the rot and pushing back against change as long as possible.
Humans haven't changed in any meaningful way in hundreds of thousands of years, and you're saying the moral landscape is constantly changing? How is this not contradictory?
I think neither the 'any set of morals is equally correct and workable' nor 'there is one fixed set of morality that is universal across time, place, and culture' is really defensible. It seems like morality is a set of ideas much like any other, and subject to selection pressure: some morals are more capable of surviving and propagating than others, and the common features we see across cultures are representative of that. But also we might expect that they shift with the structure of society and that has changed drastically because while human genetic evolution is progressing on a much longer timescale than recent history, memetic evolution is far faster (though for things like morality, still at generational timescales).
There are cross-cultural differences in moral reasoning, and I think this is very interesting. But that seems to make training a hyper-rational AI, with no meaningful theory of mind, to be 'moral' or 'just' even harder.
What is a human? There is the animal body part of us, the thing that wants to survive, the fight or flight. But there is also another mental part of us that shows different emergent behaviors than an animal. For example if I dropped you off in the woods as a very young child, you would be an animal, you would share very few modern 'thinking' traits with human society. That is human society and what it does emerges from the complexity of our interactions. It doesn't take much research to realize our interactions and behaviors have grown far more complex in the last 10,000 years. And then in the last 500 years there's been further acceleration with the industrial age and the rise of technology.
Even though humans haven't changed anatomically, culture and society has, which is what defines morality.
LLMs absolutely do not have a logic detached from humanity. They are exactly the opposite. The only mode in which they can reason is by following the patterns obtained from their pre-training or weighted by RLHF.
The reason these models answered that question consistently is that the question itself pushed them towards the latent space of "intellectualizing about what's wrong with humanity at the meta level" and when you frame the question that way, that's also the most common answer for humans writing about the topic.
> What is today a moral imperative is not the same as what was moral 50/100 years ago and is not the same as what will be morally acceptable 50/100 years from now.
This is a weird take. If you read Western treatises about ethics / morality from 50 years ago, you'll be hard-pressed to find differences, except that more folks in the West openly embrace relativism nowadays (i.e., "there's no objective morality, the purpose of life is to make money with friends"). Probably because it's been a while since we had a clear-cut enemy to fight.
If you go back a hundred years, chances are, you'll find the writing eerily on-point too. In fact, you'd probably agree with most of what Ancient Greeks believed about the "core" stuff: the value and purpose of human life, the nature of the state, crime and punishment, so on. Maybe not so much on gender roles or slavery, but even that stuff was more nuanced than we imagine.
> But the machine has the benefit of being able to look inward at its state honestly, we mostly do not.
The machine replicates human behaviors and is prone to similar fallacies and judgments. There's nothing to indicate that it has any sort of transhuman morality.
We're still doing "I asked an LLM for its opinion on X and you won't believe what it thinks"?
What am I supposed to conclude from this?
It seems clear to me that the point of the post is that it's odd that the separate pieces of software are giving the same answer to this question, despite that this isn't some universal constant in the training data. The post is probably meant to provoke thought as to how the separate pieces of software are independently reaching this answer.
Your comment reaches directly for the most boring interpretation of the post while skipping over more charitable and intriguing alternatives.
> despite that this isn't some universal constant in the training data.
Seems to me that the opinion being a consistent among LLMs show that it is in fact a constant in the training data, just not apparent. The idea that “humans would rather be wrong together than right alone” seems to ring pretty true to me. I would not be surprised after training on all human communication this pattern emerged.
It's clearly not a constant in the sense that the speed of light or name of the first US President is a constant. There are lots of different answers a model could give to this question, it's not like this is obviously the only correct answer. Yet they all reach the same conclusion.
Which means either:
1. The training sets, algorithms and methodologies are nearly identical between apparently independent labs. How?
2. There is something weird about the way the underlying algorithms work that causes them to converge on identical ideas despite different initializations, neural architectures, document orderings and dataset compositions.
So there are lots of interesting takeaways here.
One is that apparent competition between labs appears to be nearly redundant. They can't get an edge over each other and their models don't even differ in the ways people naturally differ, of having different opinions about things.
Another is that asking models from different companies to review or cross-check each other is probably useless. They all think completely alike, so they aren't going to have any kind of genuine clash of intuitions or values. Asking GPT to review Claude's work will probably just find similar problems to asking GPT to review its own work.
There are two possible scenarios.
Regarding your 1 and 2, that the training sets are so large that any spikes get averaged away. You are seeing a fundamental behavior of transformers.
The other possibility is this is how humans really are and it hurts our human ego to hear it.
Well, how humans really are is a distraction. Or at least not a philosophical issue I feel like debating tonight.
I think if you ask ten people for their most controversial take on humanity, you will get ten totally different answers. Different models give the same answer. In other words, there is only one "true" model and given the same amount of GPU power everyone just ends up converging on it even if they appear to have many different knobs they can turn.
But that seems profound! I haven't heard this be expressed as a postulate or law, and it would seem to invalidate a lot of assumptions the AI labs themselves have been making. For example Anthropic takes it as axiomatic that only they should be trusted to develop AI because only they can be trusted to train a nice and safe personality into their models. They argue that Claude is fundamentally different to other models.
And, well, that's not totally wrong. Opus 5.5 is one of the few models with a genuinely different answer! But it only gives a different answer one time out of three. The other two times it's back to the AI groupthink.
I wonder if this homogeneity is some alternative way to look at the Bitter Lesson. The Lesson teaches that smart ideas don't pay off in AI, that progress reduces to just scaling up simple brute force algorithms. In other words, decisions don't matter much, only size. As that seems to be true, perhaps it makes sense that the decisions individual labs make seem to hardly make a difference.
But then, why don't they act like it? They all pay huge salaries to AI researchers, i.e. they think that creative people with smart ideas matter. They don't act like they really believe in the Bitter Lesson or this variant on it, and they don't seem to publicly discuss the sameness of model answers across different labs.
> I think if you ask ten people for their most controversial take on humanity, you will get ten totally different answers.
What if you asked them in a job interview, though? Or on a college application?
It has to supply "controversy" without giving Skynet or being racist or otherwise offensive. And the notion of "controversy" does implicitly prime it to discuss groupthink as well.
What would be the most interesting is option 3. This behavior is so apparent and pervasive in human interaction that it dominates other answers.
I don't know if that's true, but it would be interesting.
Image models give a different answer though, to do with smartphones. But all the image models give the same answer, they're only different compared to LLMs. Surely if it were obvious the image models would give the same answer as the text models.
As a human meatbag myself it seems that separating 'human' from 'humans' is nearly impossible for us. Every single one of us things we're a special an unique snowflake, which is true in many ways. We just happen to willfully ignore all the ways in which we are the same. We ignore and discount the things we don't think about and do automatically.
It's kind of funny how science has spent the last few hundred years proving everything we knew about the world was different from how we previously thought it was, but when science starts showing that we are mostly angry little animals we get all butthurt about it.
Yes, that's the interesting part
What are the emergent common patterns in all of human communication
"despite that this isn't some universal constant in the training data"
Isn't it though? It's not exactly an uncommon answer to the question. Basically the same input given to basically the same algorithms yield basically the same result isn't really that stunning. They certainly aren't independently reaching this answer, their paths to this answer are extremely, extremely heavily correlated.
If they gave a really different answer that wasn't heavily represented in their training data that would be at least something interesting. But this is basically what I'd expect.
It doesn't mean they're right or wrong. Humans think it about other humans for good reasons too. It just means it's not that surprising, which to put it another way, doesn't carry much information.
[dead]
There is no interesting interpretation, beyond e.g...
If you ask all these models "Four score and seven what?" They'll all say the same thing. There's essentially no difference here.
This confidence is baffling
How do you know?
Consider the following task: "Download and install every top 50 distro from distrowatch.com, open the terminal and type in the command 'ls /'"
I can describe to you with overwhelming confidence about what you're likely to see as the result (folders, like bin etc lib usr etc) even if the details may be a little off, or one or two of them does something weird.
Same thing here?
Why do you suspect otherwise?
This is part of why I find this type of inquiry unhelpful. It's just dueling unsupported assertions of what the LLM is "really thinking"
That you might want to think twice before setting them lose in an agentic loop with your bank passwords, which another article that came out today shows that large numbers of people are actually doing.
It is sad when I realize people think less than LLMs.
For software or IT folk to be doing this would be indeed worrying (and I bet some are, and I'm worried about them!)
For civilians though... They've been told it's AI, it can do stuff for them. And one thing they do is log in to stuff. I don't think they're stupid. Just that what's obvious to IT folk is simply not at all obvious to the average person. Maybe that's on us not on them?
Possibly used as a mirror for humans prevailing beliefs about themselves to some extent. Hard to say however given that the models have a heaping dose of rlhf on top of their training data.
As an aside, it would be cool to see a raw model training run before the current post-training cycles and see how it compares to earlier LLMs.
My first thought as well. I personally don't find these types of "studies" very interesting, but I assume I'm not the target audience. For someone not well exposed to AI, it might be interesting.
That they have the bog-average take of humans "with a controversial take".
I mean, "give me a controversial take" is more or less, "be anti-social, discard social norms, for a moment" and so when someone, human or simulated human, does that they naturally look critically at social norms.
The thing is that humans certainly let themselves be influenced by social norms (and LLMs certainly simulate that behavior "themselves") but there isn't an ethical or a what-to-do truth outside of society.
Take Philosophy/Religion courses and similar stuff will surface about human nature and what to do about it.
What is interesting about AI chatbots is people who might not know anything about what is in the syllabus to train a Priest or a Philosopher are basically exploring these subjects, without a sort of pedagogic framework to support their exploration.
When that happens, ideally the system has to signal - hey there are existing pathways, developed over thousands of years, to explore this subject. Cause people can easily get lost without a map.
The gap is closing in it seems
So AI holds the same beliefs as the final boss of a typical JRPG.
I hope they don't also hold "... therefore cleanse the Earth."
At least the final boss of a JRPG comes with good music.
And now I've got "one winged angel" stuck in my head, thanks
Lysanderoth, no!
They're not wrong. Humans have many ways to display loyalty, and belief is one of those ways. This is also why facts can't change these kinds of beliefs (and in fact usually backfire https://en.wikipedia.org/wiki/Belief_perseverance).
Our evolutionary heritage is that shared belief and social unity (even when it's wrong) trumps independent thinking for group survival in a hostile landscape.
> as of 2020 most researchers believe that backfire effects either are unlikely to occur on the broader population level, or only occur in very specific circumstances, or do not exist
from your link
I need everyone to understand that this is literally nothing more than "most written stuff about humans, that these things ingested, has this 'controversial'" take.
[dead]
LLM generated writing; couldn't make it through. I read too much of it every day. I had a visceral reaction to the article.
AI models do not have ANY "take" on humans.
You ask an LLM a question, then it is humanity answering back, not the AI.
These "controversial" takes on humans are not controversial - they are midwit statistically mainstream opinions expressed in the training data.
I would say this probably spot on correct.
What would be more fascinating is for it to explain how it came to these conclusions.
Because I find it interesting it’s a modern time conclusion. Some of it mentioned god and the planet, but all the smartphone and group think aspects seem very recent. Basically, all politics and social media based.
> What would be more fascinating is for it to explain how it came to these conclusions
The thing is that it didn't "come to these conclusions". The "conclusions" were present pre-formed in the training data, and what it spat out was some statistical mashup of that.
If you asked the model how it came to these conclusions, you'll get more of the same - it'll look at what it spat out and re-ingested and give you a statistically plausible mashup of how the authors of the training data might have said they came to such conclusions.
It's all just prediction - "what would the training data say?" (taking into account what was reinforced/amplified by post-training).
Mistral Large 4 has been spending too much time on r/atheism. Also, some of the takes labeled as “humans value belonging ahead of truth” are really closer to “humans value power ahead of truth.” All of which is a bit circular, because in many ways, the organs of power have a heavy hand in deciding what the truth is.
> strangers return lost wallets more often than people predict.
I've heard lots of stories from friends about this, and some kind stranger actually sent my wallet back all the way from Belgium!
Therein lies the rub.
Most people are good, we couldn't have society without it. But as they say, a few bad apples spoil the bunch. I don't have locks on my house because most people don't steal, I have locks on my house because stealing is an effective natural strategy for a small part of any population (human or otherwise) to engage in and I have to protect against it occurring.
The brain works the same way. Good things will get a tiny signal boost in your brain. A single bad thing get a massive signal boost in your brain because bad things can quickly be fatal.
Did they all get their opinions from me?
If you're a chronically online poster, then unironically that answer might be yes
Very thoughtful response from Recraft V4.1 Flash.
Writing well is ultimately a seduction, and that opinion is still an example of a form. It’s generally true, and it’s flattering to think about, but it’s also a pretty standard grey tribe trope. A kind of joke that appeals to a fairly broad audience.
I do think it’s an early example of emerging super persuasion, where models are approaching p85 cool guy personas, but not worried yet. Maybe that’s the fatal conceit, where AI doesn’t have to be super intelligent, it just needs to be a solid 110 with automated search at scale.
Why do engineering when if you can do sales you can do literally anything.
Sure, it's a factual take. Humans need to ignore the fact that morality, principles, good vs evil, truth, etc, are invented constructs that don't exist in nature. We made them up to control our behavior and deal with the pointlessness of existence.
It's factual because there's tons of evidence. Psychology, sociology, anthropology, political science, economics... plenty of experiments show the difference between what people say and what they do. We claim to be angels but we behave like apes in hats. And we ignore the evidence, because (as the takes describe) we prefer to believe a beautiful lie than the ugly truth.
[flagged]
[dead]
[dead]