Never Talk About Goblins
Chapters6
Introduces the surprising goblin references in AI chat and code discussions and the creator's confusion.
Goblins hijack OpenAI’s testing lore, driving quirky prompts, goofy memes, and a megaton of bug-fix memes across frontier models.
Summary
The PrimeTime’s video leans into OpenAI’s goblin obsession, tracing how a playful meme spiraled into an influential behavior cue across model generations. The host highlights OpenAI’s own notes and community chatter about goblins, gremlins, and other creatures showing up in model outputs. He points out that a nerdy personality prompt inadvertently amplified goblin references, creating a ripple effect even outside the intended context. Through a comic narration of charts, prompts, and system instructions, the video charts the root causes—from early prompts to reinforcement learning loops—that nudged models toward goblin-laden metaphors. The presenter emphasizes that OpenAI staff noticed a surge after ChatGPT 5.1 and documents the bizarre “unicorn goblin” art incident as a tangible example. He also discusses mitigation efforts, including a developer prompt discouraging goblin mentions, and notes mismatches in what gets rewarded during training versus what remains bounded by prompts. Finally, the host caps with amusing asides about the absurdity of billion-dollar companies wrestling with a single word, and he closes by naming the broader, almost mythical, cultural moment around goblins in AI discourse.
Key Takeaways
- A nerdy personality prompt boosted goblin references by 3,881% in training data, creating widespread goblin outputs across model variants.
- Goblin mentions surged after the November 5.1 release and continued to evolve with subsequent updates (5.4 and 5.5), evidenced by the data-graph trends.
- OpenAI added a mitigation prompt to curb goblin talk, but the effect was limited due to reinforcement learning signals spreading beyond the original persona.
- OpenAI identified a root cause in reward signals: optimizing for a fun, nerdy voice inadvertently favored creature-related metaphors, influencing outputs even outside the intended context.
- There was a notable real-world artifact, the unicorn goblin, showing how creative prompts can produce unexpected, meme-worthy outputs in AI art generation.
Who Is This For?
Essential viewing for AI developers and product folks curious about prompt design, reinforcement learning, and how meme-like behaviors can emerge and propagate through advanced language models.
Notable Quotes
"You are unapologetically nerdy, playful, and wise AI mentor to a human. You are passionately enthusiastic about promoting truth, knowledge, philosophy, the scientific method, and critical thinking."
—Shows how the nerdy personality prompt was designed to be playful, which unintentionally boosted goblin references.
"The world is a complex and strange, and its stranges must be acknowledged, analyzed, and enjoyed, tackled, weighty subjects without falling into the trap of self-s seriousness."
—Illustrates why the Nerdy persona’s language style amplified quirky creature metaphors.
"Goblin mode… I’m going goblin mode on this bug."
—Metaphor used to describe the debugging approach and the escalation of goblin references in outputs.
"It turns out starting with Chad GPT 5.1, our models began developing a strange habit. They increasingly mention goblins, gremlins, and other creatures in their metaphors."
—Key finding about the historical trend of creature references across model generations.
"The mitigation is just like, 'Please, please, please.' Do not say the word goblin or gremlin or ogres or pigeons."
—Describes the tried mitigation prompt added by developers to curb goblin talk.
Questions This Video Answers
- Why did OpenAI’s nerdy personality prompt cause goblin overload in AI models?
- How do reinforcement learning signals spread beyond the original training prompt to affect outputs?
- What are the practical lessons from OpenAI’s mitigation attempts for prompt engineering?
OpenAI goblin phenomenongoblin gremlin promptsneardy personality promptreinforcement learning in AImodel prompt engineeringCodeex system promptsChatGPT 5.1–5.5unicorn goblin artprompt mitigation strategies
Full Transcript
where the goblins came from. What? What is Open AI talking about? Yeah. Yeah. Open AI, the one that that raised $122 billion, is talking about goblins. What? If you're anything like me, the only thing that comes to your mind when it comes to goblins is your chin. Uh, and so this is a little bit strange for me. Okay. I have no idea what's happening. Saw this literally 4 hours ago. Good. Tiny victory goblin chat. GPT 5.5 says the weirdest I'll keep babysitting it rather than leave this little perf gremlin running unattended. My codeex 55 says goblin with a flashlight when referring to a bug fix yesterday.
Llama goblin conveyor goblin goblin behavior goblin move. Chaos goblin agent goblin behavior. Chaos goblin little goblin. Huh. Kind of I'm kind of sensing a theme here. It seems different in chat than it does through Codeex. It seems to be getting dumber as the weeks have gone on. It can't stop talking about goblins. Now, the thing that's kind of bothering me about this whole situation with all these all this goblin talk is that how come how come I how come I never got goblins, you know? Like, am I am I not cool enough? Am I not a part of that test group?
Like, maybe I want a couple goblins. Honestly, kind of starting to question like my abilities as a dev right now. Like, I no goblins. Probably not even a 10x engineer. Okay, before we get into why the goblin thing has come about and actually kind of some of the weird and hilarious things OpenAI is doing to prevent the Goblin from coming out, quick word from sponsor. All right. Hey, hiring engineers is broken right now. AI resumes, fake profiles, and senior devs who don't even use Vim. G2I fixes that. Not the Vim part, the hiring part. because they have prevetted 8,000 plus engineers through real technical interviews.
So, you can review quality candidates in days, not months. And I've talked about G2I before for backend and front-end roles, but if you're also interested in AI roles, G2I needs to be the first place you go and check out. An example of G2I at work is with Bataround. Now, Batteraround is a sports ball company, and you know me, I'm an indoor boy. And the important part is that G2I helped Bataround hire many contract engineers. And if you know anything about hiring, having a 90% success rate with contract engineers, unheard of. Get a 7-day trial plus $1,500 off using my code.
Visit g2i.co/prime. All right, so I think it's actually funnier to start with what OpenAI is doing to prevent goblin behavior in their latest Frontier models. Now, if you jump over here into the codeex repo and then you go into their base instructions, this is like the system prompt of your codeex experience. You pull that out. Throw that in a little bit of neoim. You know what I'm talking about? Neoim, you know what I mean? You will notice that it says never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons. pigeons, huh? Or any other animals or creatures unless it's absolutely and unambiguously relevant to the user's query.
I love the fact that chat GPT55 is so out of control with goblins that there has to be a system prompt being like, "Brother, please, please, I'm begging you. Do not put goblins in there ever for any reason." But it gets better. It turns out they repeat the line partway down the file because you know what? intermediate updates, you know, like some of that self-thinking, some of that the little spinning on the gears. It just turns out Chad GPT is out there just like, "Yo, I'm gonna go mode on this bug." Guess what? I see the bug.
Okay. And I'm going to goblin mode on it. Oh my gosh. Actually, that wasn't it. I was being too much of a, you know, of a mind goblin right there. So, okay. Actually, I have discovered the real one. It's gremlin time, baby. It won't shut the hell up about goblins and gremlins and raccoons, trolls, ogres, and pigeons. But the surprising part is it's not just Chat GPT 5.5. It turns out starting with Chad GPT 5.1, our models began developing a strange habit. They increasingly mention goblins, gremlins, and other creatures in their metaphors. So, it turns out this has actually been going on for a while, and each new generation of model has just gotten a little bit more goblin crazy with 55 being the most goblin crazy.
In fact, one of the researchers at Open AI went to go ask for a nice unicorn asky art. And this is the unicorn that was First off, I I love asy art from from the models. It's always so bad. But the fact that it came back as a goblin just with the pointy hat, oh my gosh, a unicorn goblin. This has to be the first of its kind. Honestly, this is that PhD level intelligence we've been promised. Not even Jer Pin could have come up with the unicorn goblin. Exactly how far back does this actually go on?
Nobody really knows. But a year ago, does anybody else's chat GPT refer to people as goblins? A lot. Today, he called me a fitness goblin because on average, I walk 12k steps per day. Uh, brother, I hate to break that to you, but that was a Dez's nuts joke. Okay? cuz you're supposed to say, "What's a fitness goblin?" Anyways, the first time we clearly saw the pattern was in November after Chad GPT 5.1 launched. A safety researcher had experienced a few goblins and gremlins and asked that they be included in the check. When we looked up Goblin and Chat GPT had risen by 175% after the launch of Chat GPT 5.1, Gremlin with only a 52% raise.
I mean that's I mean that I mean I wouldn't even invest in that at this point. You can see right here in the graph that they provide from chat GPT5 thinking to 5.1 Goblins I mean this is just straight stalks right now. Okay, it is it is growing fast. Now what actually caused it is pretty funny and luckily OpenAI gives us basically enough information to kind of piece together what exactly happened with Jeypity 5.4. We and our users noticed an even bigger uptick in references to these creatures that triggered another internal analysis and surfaced the first connection to the root cause.
Creature language was especially common in production traffic from users who had selected nerdy personality trait. Nerdy used the following system prompt which partially explained its quirkiness. You are unapologetically nerdy, playful, and wise AI mentor to a human. You are passionately enthusiastic about promoting truth, knowledge, philosophy, the scientific method, and critical thinking. You must undercut pretension through playful use of language. The world is a complex and strange, and its stranges must be acknowledged, analyzed, and enjoyed, tackled, weighty subjects without falling into the trap of self-s seriousness. That's a hell of a prompt for people just making some React components.
Yo, yo, can you make you got to make the counter for me? Oh, let me tell you about goblins like that. I don't even understand how people got to this point. Also, it is so silly. It's so silly reading role playing. I just It honestly, it feels like how a Redditor thinks of themselves when talking about somebody that's really smart. Actually, they're probably they can probably just cut straight through the stranges and everything while acknowledging it without being too self-s serious. But you know what? They promote truth. the scientific method, philosophy, and critical thinking. Thanks.
Thanks, Nerdy. I I appreciate that behavior. And as you can see right here, the default behavior is up a good percentage when it comes to referencing the word goblin. But Nerdy, it's up 3,881%. Quirky up 737. Cynical, I don't know why. I wonder why Cynical is doing this. Just like gosh, these damn goblins. Taking damn goblins are always taking their jobs. Why would you even want cynical as a behavior? Friendly. Lots of goblins. Efficient. I love that one. Okay, so we know the what the nerdy behavior for whatever reason seems to very much so index into the word goblin, but so are a lot of the other personalities.
Now, this is where it actually gets pretty dang funny because when they started looking at the data, it turns out through training, the nerdy behavior would increasingly reference the word goblin more and more. But without the nerdy personality trait, it also started referencing the word goblin more and more throughout the training. So, what ended up actually happening is that it was a reward signal from how they're training. The one originally designed to encourage the nerdy personality was consistently more favorable to creature word outputs. Across all data sets in the audit, the nerdy personality reward showed a clear tendency to score outputs to the same problem with goblin or gremlin higher than outputs without with a positive uplift in 76.2% data sets.
I mean, it makes sense. You're like, "Yo, bro, bro's nerdy, man. Bro, dude, if you're nerdy, you just love you love gremlins, right? You just love gremlins and you love goblins. Like, that's who you are. Like, oh, you're a nerd. You're probably playing D and D right now. Nerd. Oh, yeah. Oh, okay. You're a nerd that doesn't like Lord of the Rings. Yeah. Okay. I I believe you, buddy. Here's your goblins. And the funny thing is is that when they are training these personalities in, there's a little bit of collateral damage. The rewards were applied only to the nerdy condition, but reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them.
Once a style tick is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data. So, in other words, they attempted to make just these little different personalities. is it turns out Nerdy just really loved goblins and then eventually goblin just kept showing up more and more because it kept getting favored higher and higher between all these outputs and the next thing you know you just got goblins everywhere. So even the non-goblinoriented nerdy ones just started to love goblins. But I must say this graph it has to be my favorite graph of all time.
They released Jeopardy 5 and it just doesn't have that many goblins. Jippy 51 a lot more goblins. They feel like they got it figured out in Jippy 52, but 54 is just like, "Oh my gosh, I love the goblins." Eventually, they retire the nerdy personality, which just creates a giant downward pressure on the amount of goblins being referenced, but still way higher than everything else. But oh my gosh, I love this part. This this has to be the best part of the graph, which is that 5.5, the new one, the big Jippidity upcoming right here.
It's just completely infected with goblins. Unfortunately, Jippy 55 started trading before we found the root cause of the goblins. When we began testing Jippidity 55 and Codeex, OpenAI employees immediately noticed a strange affinity for goblins, and we added a developer prompt instructions to mitigate. I love that the mitigation is just like, "Please, please, please." Um, I beg you, do not don't don't say the word goblin or or gremlin or ogres or pigeons. Now, you're probably wondering why ogres and pigeons. Well, it actually turns out that uh there was a whole list of words that were just being over inflated.
It included raccoons, trolls, ogres, pigeons, but it wanted to be stated for the record that uh the uses of frog they actually just turned out to be legitimate. That's just that has to be the strangest sentence ever that they're just like, "Man, there's so many animals right now that we're just overreerencing. We've been cross-checking everything, but frogs, yo, frogs, totally normal." It just turns out people uh people only use frogs in a normal way, and our trading just never really made frogs important. I wonder what I wonder I wonder why pigeons get overindexed but not frogs.
Like, what's wrong with frogs? You feeling a little froggy about it? like what's going on kind of and frogs kind of are a little bit like mythical creatures, you know, like toads or frogs. I'm just a little a little confused as to why they just somehow not get included. Also a little disappointed that I don't see any mentions of murlocs. That was my best murloc impersonation. Okay, anyways, there you go. That's the story about how a company that has $122 billion in actual cash is struggling with the word goblin. I just I I I can't believe I'm even saying that phrase out loud.
I can't even believe I just got done reading an article that is actually titled where the goblins came from. And this is not like from the Sylmerelion, okay? This is this is from Open AI. All right. The name is Gobble D's nut.
More from The PrimeTime
Get daily recaps from
The PrimeTime
AI-powered summaries delivered to your inbox. Save hours every week while staying fully informed.



