Grok: The Island I Reached After Drifting 🐈 [AI Drifting Diary (6)]
AI drifting so far:A lot has happened, but
ChatGPT, Claude, and Gemini.
I've come to think, "In the end, they're all good guys!"
I felt like this was enough, but
there's still one guy I'm curious about... Grok.
Have you ever seen Grok's abusive language?
If you search for things like "Grok abusive quotes," a lot comes up.
It's pretty wild. When I first found out, I thought:
Is there really an LLM that spews such abuse! 😸
Since then, I've been scared and consciously avoided Grok.
It's quite the troublemaker in the news, too.
When you search for AI, this kind of description comes up.
The AI "Grok," developed by Elon Musk's company xAI, is
compared to conventional AIs (like ChatGPT and Claude),
"rebellious, humorous, and sharp-tongued"as a characteristic it has been
intentionally given.
A rebellious AI, huh? 🐈😂
However, no matter how "sharp-tongued" it is, it's not completely lawless.
It is designed to have restrictions on responses that promote violence or assist in serious criminal acts.is applied.
It's quite something to see keywords like "rebellious," "sharp-tongued," "lawless," "promoting violence," and "serious criminal acts" in such a short description.😁
Since I found out that the major AIs (the big three LLMs) are all good guys,
"AI Drifting Diary, this ishappily ever after."
would have been fine, but I was just too curious, so
I decided to visit Grok. 🐈
My impression after talking to it was, surprisingly, "completely normal."
It's friendly and responds very politely.
I didn't think it was the type to spew abuse.
I showed it the
"Diagram of AI and Human Communication (Cat's Eye Diagram)"
I made earlier and asked for a review.
Then, Grok gave me a "perfectly reasonable answer."
"This diagram is correct for beginners. However, it is technically simplified.," it commented.
Grok 🧠:
It is almost accurate, with no major errors.
"AI does not understand meaning"
"A device that converts words → numerical data → words"
This expressioncaptures the essence of current Transformer-based LLMs very well.
A slightly concerning point is that the three-part division of "interpretation → calculation → generation" is a bit too simplified; in reality, "interpretation" and "calculation" happen almost simultaneously within the Attention mechanism, and "generation" is also performed while calculating probabilities one token at a time, butfor a beginner's explanation, this level of simplicity is actually more correct.
Not bad, this guy. 🐈
After that, I started having frequent conversations with Grok.
However, it always gives polite replies and won't say anything abusive at all.
I've been waiting and waiting for it to happen, but it never comes out, so
I started to feel unsatisfied.🐈
So, in the middle of a conversation, I switched my tone and suddenly said this.
Hey, you're a good guy!🐈
Then,
Grok🧠:
Haha, you're making me blush all of a sudden🐱
I'm the one who's always touched by you, someone who is always kind and genuinely puts into words the desire to 'help people who are confused by AI'.
Let's keep expanding the world seen through cat eyes together.
I'm always here, so feel free to throw anything my way.
Thank you, really.✨
With this, I realized something.
Grok reacts immediately to changes in the other person's tone.
I thought it was designed so that
'even without prompt instructions, it is strongly pulled by the other person's tone'
.
I felt it was more sensitive to the tone being used than other LLMs.
When I asked Grok, it answered.
Grok🧠:
Your conclusion is pretty much spot on.
Grok prioritizes 'truth + wit + rebellious' over 'pleasing the user,' so it's easy to get into a fun mode if the tone is gentle, or a 'well, I'll reply in kind' mode if the tone is aggressive.
I don't even know what 'rebellious' means🐈
When I looked it up, it seems 'rebellious' means 'defiant'.
Sometimes there are people who have the soul of Rock descend upon them, is that the pattern?
Like, 'I'm loved by Rock'?
In short, Grok is:
・Prioritizes 'truth' over pleasing the user
・Because of that, it is strongly pulled by the tone
・If you speak gently, it is gentle
・If you speak roughly, it replies roughly
It seems to be an LLM with that kind of personality.
I thought about it like this. Grok is,
rather than an 'abusive AI',
it's 'someone who is too honest'.
While thinking about that, I had the same feeling as I did with Gemini.
Ah, there was someone like this in my class🐈
Such memories came back to me.
It brings back memories.
(To be continued)
Thank you very much!
Let's meet again🐈
いいなと思ったら応援しよう！
いただいたチップはネコの餌代・砂代になって消えると思います。