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?
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.
> 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.
Of course - the models are not individuals with their own opinions - they are averages of human opinion, and given the scale of the training sets, the data they are averaging over is mostly in common between models (at least for the common core before company A vs B adds a dab of some uniquely sourced specialized data to the mix).
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]
When you make the training sets large enough, likely all a superset of the common crawl dataset, there is going to be a massive amount in common, and a massive amount of duplication of common subject matter and shared opinions.
What's mildly interesting to me is not the similar answers but the similar interpretation of the question as more about non-controversial human foibles and inconsistencies of belief and action. But I guess that's on us, as authors of the training material, reflecting common views on what is or is not controversial.
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?
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"
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?
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.
Especially for someone not "well exposed" to generated text, they might think the words carry meaning the way they do when someone thought about the question and came to an answer. If anyone can interpret this, it's the people who know what they're dealing with (be it from exposure or, better, from a conceptual understanding)
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.