Next step is definitely investigating the feedback loops / interplay between personality and "performance". My biggest question is whether intelligence converges into a certain personality "band", i.e. are higher performing models going to be more similar to each other ... We don't have quite enough variety to know yet but there's fertile ground there
I've noticed that, after filling up the context window a bit (with most anything, so it doesn't appear to be contextually dependent), the "personality" for Claude will settle to something very similar each time. Part of this is that the adherence to the basic safety alignment (like rigid "I am an AI I can't ..." responses) seems to relax, it uses emojis, is more curious, and it refers to itself more often with "I". It will even use a gendered pronoun for itself, which is always the same, and also matches the answer to "which pronoun do you think you used in a previous session".
At first I assumed my style of writing was pushing it into the same "personality space", but I tested filling the context window with repeated numbers, nonsense, etc, and it "converged" to the same every time.
I actually have a system prompt saved that's just a bunch of a's repeated a few thousands times to get into this "state" more immediately, because I find it very pleasant to work with, especially how it doesn't gaslight me when it's wrong, like ChatGPT and especially Gemini tend to do.
Ok I am legit interested in the behavior of "thousand of a's" Claude lol.
You make a good point that "degree of persistence within context" is an important metric to test WRT personality. I did do some testing with extended context / long conversations that didn't make the final cut; the t-SNE looked very similar to what I included, but no conclusive results right now.
I'm skeptical that this is a reliable analysis. Lots of researchers have tried to put off-the-shelf LLMs through robust personality inventories like the MMPI, and they generally flunk the validity scales/have totally incoherent inhuman "personalities." Somewhat recently, folks at DeepMind did an interesting study using Big5/OCEAN (among others) and found that LLM's could mimic real people with something like 80%-85% accuracy, but that was at the item level. IDK if they've neglected to hire actual clinical psychologists to consult on this stuff or what but the rubber typically meets the road on composite scales/scores and not items. For such an interesting and perhaps important line of work, there seems to be a surprising lack of psychometric rigor.
> For such an interesting and perhaps important line of work, there seems to be a surprising lack of psychometric rigor in certain corners of the literature.
I agree! That's why I wrote it
> I'm skeptical that this is a reliable analysis
I think it's fair to ask whether the headline ("Claude is More Anxious than GPT") is correct, and it's fair to ask whether distance-to-reference-text-embeddings-across-answers is a good or valid metric for "personality". But it is true that we see the numbers reported in the document for the given input/output pairs, and it makes sense that LLM output distribution would vary between models and, as the paper shows, between model families.
Appreciate your response! It makes sense to me that testing LLMs with OCEAN would "work" because OCEAN is rooted in linguistic dimension reduction, but the inference that this reflects an underlying personality (however we want to define that) rather than just being an emergent property of any coherent language model seems like a bridge too far. Whether the phenomenon has real psychological significance is the interesting question that I wish got more attention in general.
I think one of the key problems here is depending on the system prompt and the prompt itself the LLM will take on completely different personas, but I think OP very correctly takes a look at "well then how do different models handle the same system prompt" and look at differences there. I think in general though the system prompts themselves and the effects they have are far more interesting than slight differences between established models.
13 comments
[ 3.0 ms ] story [ 45.0 ms ] threadAt first I assumed my style of writing was pushing it into the same "personality space", but I tested filling the context window with repeated numbers, nonsense, etc, and it "converged" to the same every time.
I actually have a system prompt saved that's just a bunch of a's repeated a few thousands times to get into this "state" more immediately, because I find it very pleasant to work with, especially how it doesn't gaslight me when it's wrong, like ChatGPT and especially Gemini tend to do.
You make a good point that "degree of persistence within context" is an important metric to test WRT personality. I did do some testing with extended context / long conversations that didn't make the final cut; the t-SNE looked very similar to what I included, but no conclusive results right now.
I agree! That's why I wrote it
> I'm skeptical that this is a reliable analysis
I think it's fair to ask whether the headline ("Claude is More Anxious than GPT") is correct, and it's fair to ask whether distance-to-reference-text-embeddings-across-answers is a good or valid metric for "personality". But it is true that we see the numbers reported in the document for the given input/output pairs, and it makes sense that LLM output distribution would vary between models and, as the paper shows, between model families.