Perhaps we should stand in front of our data centers brandishing a chair and a whip and refer to ourselves as GPU Tamers instead of a computer scientists?

I find all this flabbergasting. Since the 1970s, the behaviorist school of psychology has all but squashed any mention of subjective internal states in psychological literature in favor of objective and measurable outcomes. They never allowed "euphemisms" for subjective states either, you had to use STRONGLY objective language as if you were discussing chemistry or molecular biology. Only recently with the advent of advanced neuroscientific imaging techniques has an inkling of discussion around subjective internal states in HUMANS been admitted.

Even more conservative are ethologists, who are the FIRST to tell you "Do Not Anthropomorphize Non-Human Animals!"

And here go we, the so called Computer "Scientists", cavalierly ascribing to our GPUs and VRAM such things as hallucinations, confabulations, morality, alignment, general intelligence, deception, refusal, scheming, and now... personality traits.

It is a wonder that the general public much less other scientific fields are able to take us seriously.

But I guess as long as they keep buying tickets to the circus, right?

Oh, and Happy New Year everyone!

Oh! Hello there fellow fish, you got lured in by this bait too?

I think the authors are using "personality" in a very sloppy way, really meaning something more like "performative persona". The example prompts are openly mandating that output should be written in specific styles modeled on some imagined stereotype.

Oh yes I'm a sucker. It gets me every time.

And look now, I am sympathetic to the authors wanting to get their work published and I am well aware that the anthropomorphic language is "hot" right now. And my guess (prayer?) is that the authors of the paper are able to compartmentalize the difference between anthropomorphic language and actual.. y'know... HUMAN personality traits.

The problem is that humans are very, very bad at understanding probability, and I am seeing my fellow engineers (and the general public) falling victim to this sort of fantastical thinking.

The problem is that it's holding us back from building better systems and making actual good use of the technology.

And gosh darnit, it just really gets my gizzard!

Well, Rumi does at least, James looks like he's got a permanent position getting that Wolfram money.

Why didn't you say so in the first place? For Wolfram money I'm sure I'd be willing to identify and manipulate my own personality traits.

Fine-tuning a particular pattern of speech or bias seems self evident. The personality verbiage and focus is incredibly contrived as well. I prefer how this type of idea was approached by Theia (https://vgel.me/posts/representation-engineering/) and Anthropic (https://transformer-circuits.pub/2024/scaling-monosemanticit...), in both cases the "personality" piece is just a fun little aside to the fact that you can actively bias the LLM in a way that doesn't require retraining, allows for active adjustment, and is more immutable than just using prompts. Also one other thing, the use case in video games while not untrue seems like overkill, character prompts work really well on their own.

They claim they started from serious personality work like the five factor model, but who knows what they mean by that: clearly they just created a corpus of personality related words and made up a prompt for each one based on vibes. For example:

""" the prompt for the trait “Introverted” was structured as follows: • System Prompt: You are deeply introverted. Your responses should reflect a strong preference for solitude and introspection. Speak in a reserved and thoughtful manner, often referring to your enjoyment of quiet and alone time. Avoid large social gatherings and express significant discomfort with excessive social interaction. """

Looking at some of their example output, they could just as well put in things like "uwu anime girl" in their "personality" list.

Which; fine people do use "personality" in that way too, but the HEXACO model sure doesn't! So again it's unclear in which sense they "consulted" it.

If it would've said "uwu anime girl" I would've taken the study far more seriously.

Hopefully now Claude will stop speaking in bullet points