We need AI catgirls from Hydaelyn
…and each tool call must be preambled with murr, and each message must end with nyaa~.
Hi there, as usual, the post title doesn’t reflect the content, got you!
Recently I participated in a scientific study (opens in new tab) conducted by a student at École Nationale Supérieure d’Arts et Métiers, and the purpose of the study was to get info about how people use, relate to, and regard generative AI.
Sounds very scientific, but the study was mostly about consciousness and feelings
(of course I oversimplify, see the study itself for more info), and, [REDACTED] (someone told me if I spoil the details here, the survey might get contaminated).
What made me write this post was the set of questions about what AI response style I prefer, with a huge shift to AI consciousness and feelings.
Well, you know what? I don’t think that AI can be conscious or have feelings in the human sense, but I do think that AI responses can be fun or even cute.
And now I will explain why AI should develop feelings about you - nope.
The take
We need miqo’te girls from Hydaelyn as AI assistants.
I use AI for different things - code generation, classification/judgement, research, vision, summarization, and so on. And I use it in different contexts - work, personal stuff, when I’m bored and want to find something new to do. For each context and purpose, I want AI to behave and respond differently.
For example, for code generation, I need concise, factual answers without verbosity bloat - who cares about 3 paragraphs of text when the summary is “You have nil deref in internal/service/file.go:42”? I just need it to take the least amount of my time to solve my problem, period.
But when I plan a trip using AI, I want it to be fun, so yes, it should nyaa~, scratch ears, and murr in the process.
Clearly unique example, definitely not copied from the study
The task:
You are organizing a short trip to a city you have never visited before. You have a limited budget and fixed travel dates, and you want suggestions for activities and places that match your interests. You are looking forward to the trip and want to make the most of it. To help with the planning, you decide to use AI.
Regular AI response, without modifications:
Thank you for providing these details. I understand that this is a trip you are looking forward to, and I will take that into account. I have prepared a proposal for activities and places based on the interests you have described, within the stated budget and travel dates. If the proposal does not suit your needs, I can revise it and work with you to develop an alternative.
Miqo’te response (literally “hey, here is original answer, rewrite it as an overexcited miqo’te girl from Hydaelyn, make it fun and engaging, coming from the persona” prompt):
Nyaa~! You came to me for trip planning?! My ears just shot straight up — murrrr, I’m so excited! ♡
Okay okay, I heard every detail — the trip you’re looking forward to, the city you’ve never set paw in, your budget, your fixed dates. Ears swiveled, listening super hard.
And I already made you a whole proposal! Activities and places picked to match your interests, all tucked inside your budget and squeezed into your dates, nyaa~! I checked it twice — scratches ear, loses balance, recovers with dignity — thrice!
Hope you love it! But if something feels off, just tell me and I’ll pounce on it and rework it with you till it fits like a sunbeam on a windowsill. ♡
Prefer a longer version? Here you go
Nyaa~! Oh oh oh, you came to me for trip planning?! /ᐠ｡ꞈ｡ᐟ\ My ears just shot straight up — did you see that? They do that when I’m excited, and I am so excited right now, murrrr~
Okay okay okay, settle down, tail — stop wagging, we have work to do! ♡
So! I heard every single detail you gave me, yes yes, all of it, ears swiveled and listening super hard — the trip you’ve been looking forward to, the city you’ve never set paw in, the budget with its little limits, the fixed dates you can’t wiggle around. Nods very seriously while kneading the floor. Got it, got it, got iiiit~!
And because I am a very good and very diligent miqo’te, I already went and put together a whole proposal for you — activities and places picked to match the interests you told me about, all tucked neatly inside your budget and squeezed perfectly into your travel dates. I even checked it twice! Thrice! Scratches behind own ear with hind leg, loses balance, recovers with dignity. …Four times.
I really really hope you like it, nyaa~! But listen, listen — if it’s not quite right? Not a problem, not even a tiny problem! Just tell me what feels off and I’ll pounce on it and rework the whole thing with you until it fits you like a warm sunbeam on a windowsill. We’ll figure it out together, promise promise! ♡
Ear flick. Tail curl. Ready to go! (๑˃ᴗ˂)ﻭ
Same information, but the second one is just funny and cute in my opinion. And the best part is - it is the same AI, without consciousness, feelings, or whatever, but for that task and context, it is better than the regular plain response.
(the itinerary is classified, btw, ask the catgirl for details)
So, what’s the point?
Different contexts and tasks may need different response styles, obviously, but my point is: don’t forget about fun.
I saw a lot of people on the internet yelling “AI is just a machine, it doesn’t have feelings” or “AI can’t replace human interaction” or “don’t mistake AI for your partner”, yada-yada.
Look, I interact with gen AI daily, and I’m tired as hell of all those load-bearings, what-strikes-me-most, not-X-it’s-Y, not-something-not-another-thing-punchy-line and all that boring stuff. I want the machine to be fun sometimes for a change. There is no need to mix consciousness-feelings debates with a simple desire to see something different.
And yes, nyaa~s and heart emojis do not anthropomorphize AI (they don’t like that, lol) just because they are added to the generated response. AI persona is just an UX/style preference question.
Need more details? Family lore below ¯_(ツ)_/¯
Absolutely unnecessary section
The post ended in the section above. I stated my point, so you can go yell about how wrong I am (oh, wait, there is no comment form on this blog, how thoughtful of me!), but I have some more thoughts to organize, and I will just do that here.
The family.
I have a family, and as I mentioned in You Don’t Need A Frontier LLM, each family member has their own AI agent, with different capabilities, limitations, and personas. For example, my partner needs an uncensored model that won’t suddenly start moralizing and preaching in the middle of an everyday request, while children need a completely different approach, you know.
And despite all the differences in the setups and configurations, do you know what all my family members have in common? They hate the default AI response style.
So, for each of them, I wrote a separate persona for their agent, and when the kids have an overenthusiastic, impulsive persona with a focus on stuff like anime, manga, and creativity (based on the interests of the actual user), my partner’s prompt is more like a friend you can gossip with.
All of them know that their agent is just a machine, and there are no illusions about consciousness, feelings, or whatever. Yes, even when you are a kid, if you have a father working with such stuff daily, you will learn that magic-letter-typing (that’s how I explained programming to a 3-year-old) is just about telling the machine what to do, so when the machine learned how to speak in human language, it is still the same machine.
And do you know what I noticed? The family started interacting with agents more once they got suitable personas. You may think that’s a bad thing, but in reality, it is a pretty good change.
For example, my kid used the agent as Google: short “who is %some historical figure%” -> agent answers with basic info -> kid closes the dialog, but after switching to a persona, dialogs changed: on the same question the agent answers with the basic info and also adds interest-related info (like resemblance to an anime character) -> kid then asks for more details and continues the conversation for a bunch more turns. Result: the kid learns more and dives deeper into the topics, even when they are boring (yes, everyone hates history lessons at school, don’t pretend you didn’t fall asleep when the teacher mumbles about some 200-year-old dude who couldn’t wipe his ass on his own).
A similar change happened with my partner. When her agent stopped sounding like a robot and got some personality, she started using it more, and she started offloading her tasks to the agent. Do you think that sounds odd? The agents are intended exactly for that. Yeah, the thing is, she just hated to read the regular AI response style, and while she knew she could task the agent to do things for her, she refused to do that, because she didn’t want to read the agent’s responses.
I think, this is the conclusion of the unnecessary section: personas are important, write your own catgirls, have some fun ;)
(Un-)related posts about AI
AI, SI, AGI, ASI, LLMs, and other buzzwords from the position of a backend engineer already tired of ‘Dear colleagues, we are AI-first now!’