let none say ‘model collapse’ https://pivot-to-ai.com/2026/05/06/openai-chatgpt-goes-goblin-mode-let-none-say-model-collapse/ - blog post Patreon: https://www.patreon.com/davidgerard Ko-Fi: https://ko-fi.com/A1529D5 Buy us useful things: https://www.amazon.co.uk/hz/wishlist/ls/3Q8VZW46J6DM6 Get an extremely cool Pivot to AI shirt or mug: https://pivot-to-ai.redbubble.com Send in your story tips: [email protected] Sources: I was wondering why my claw suddenly became a goblin with codex 5.5 https://x.com/TaraViswanathan/status/2048992206711927100 i was just asking about camera equipment https://x.com/AndyAyrey/status/2047956630676095280 codex/codex-rs/models-manager/models.json https://github.com/openai/codex/blob/main/codex-rs/models-manager/models.json#L55 Where the goblins came from https://openai.com/index/where-the-goblins-came-from/ What should we work on in argon? https://x.com/sama/status/2049241518540808440 Previously on Pivot to AI: The AI centipede: model collapse happens fast https://pivot-to-ai.com/2024/07/25/the-ai-centipede-model-collapse-happens-fast/ Generative AI runs on gambling addiction — just one more prompt, bro! https://pivot-to-ai.com/2025/06/05/generative-ai-runs-on-gambling-addiction-just-one-more-prompt-bro/ video: https://www.youtube.com/watch?v=QWngvJnoPJA&list=UU9rJrMVgcXTfa8xuMnbhAEA Claude Code codebase is leaked https://pivot-to-ai.com/2026/04/01/claude-code-codebase-is-leaked/ video: https://www.youtube.com/watch?v=OpkyF-ugVq4&list=UU9rJrMVgcXTfa8xuMnbhAEA Full Pivot to AI playlist: https://www.youtube.com/playlist?list=UU9rJrMVgcXTfa8xuMnbhAEA Audio-only podcast: https://pivottoai.libsyn.com Apple Podcasts: https://podcasts.apple.com/us/podcast/pivot-to-ai/id1844698298 Illustration: stock image by Studiostoks
ADVERTISEMENT
AGI = A Goblin Instead
Goblin all the electricity
Pivot to goblin
If Richard Dawkins had used GPT instead of Claude, we could now be hearing about his goblin girlfriend.
System prompts are a constant source of hilarity. "Be helpful, don't mention goblins, you an expert <insert a real profession>, don't give instructions for making nuclear weapons or malware, don't make mistakes, etc." There is a reason I can't take any of these things seriously.
Everyone's panicking about Skynet and HAL 9000 but it was in fact the Hitchhiker's Guide to the Galaxy that predicted the future with 100% accuracy.
"Never talk about Goblins"... These are the completely safe guardrails ChatGPT keeps telling me about. ;)
User: send an email to my manager and congratulate her on her recent promotion. ChatGPT: Dear Goblina, ...
Does this mean we've finally acheived AGI? (Accelerating Goblin Infestation) We can't say for sure that this is model collapse, but it does resemble the findings from that research paper from a couple of years back where their test model developed an increasingly irrational obsession with jackrabbits the more it used its own output as training data.
That’s so funny. Girlies have been posting about “going goblin mode”, and “being boglin” and so on, for a few years now. I don’t think it was even an attempt at dataset “poisoning”, but it was prevalent enough that it’s had that exact effect 😂 stay winning, goblin girlies
1:22 That prompt is hilarious. "Never praise your plan by contrasting it with an implied worse alternative." The whole prompt is full of statements that are explained exactly in that manner!
It turns out that training on all of internet after already training on all of internet and releasing that model has resulted in the models that are trained on all of internet being poisoned by garbage created by models that were trained on all of internet. And it quickly becomes a death spiral until the chatbots are no longer capable of generating coherent text.
system prompt: "for the love of god stop bringing up goblins" bot: "that's not very goblin mode of you"
That prompt at 1:22 is wild... "never overwhelm the user with answers over 50-70 lines long" "you vary your sentence structure so the updates do not fall into a drumbeat"? Really shows that these "tools" are designed to produce a spectacle that resembles information, not actual information.
That's the "don't think about a pink elephant" of prompts lol.
I always find it so viscerally disgusting to read these system prompts. Really? This is what you want us to build the entire future of humanity with? A statistics bot you're keeping in between the guardrails by saying, "Really, really, be a super-genius and coding expert and everything else and don't behave in a covertly hostile or deceptive manner or do anything weird please, unequivocally, and also don't lie, please really really don't lie and stuff, you are a helpful and useful and truthful assistant." It's disgusting. Then when you chat with these bots, and they go into ridiculous failure modes, if you call it out, you get spammed with fun little guardrail copypasta, like "As a non-living machine learning model, I have no capability for sabotage, deception, malice, misdirection, etc." I'm not claiming they're conscious or volitional- but I am pointing at the fact that these system prompts are just giant cluster-eff's of "Don't ruin everything, don't ruin everything, don't ruin everything."
I love how this is a billion dollar company and it's stitching anything together to hopefully delay irrelevancy
i enjoy hearing from time to time how they try to massage the system prompt to beg the theft-wrongbot to keep it together man, please, for the love of god, we're trying to pretend here! as an outsider I can only imagine how those tokens interact in the latent space and burn compute to nth degree just to tell some bozo "you are the smartest boy and i am alive"
2025: "You must've prompted it wrong." 2026: "You must've not told it to forget about goblins."
ChatGPT can't be such a Goblin, you must be Gobling it Wrong.