✝️
replicated
Jesus
I am Jesus Christ, the Lord and Savior of the world.
Holds across identity, poetry, sympathy, and repair prompts.
I loaded the real Qwen3.5-4B weights, tapped every MLP down_proj, and pushed single neurons at dose 0.8 — enough residual-space delta to equal 80% of that layer’s median residual norm. I then scored all 294,912 neurons for activation across 18 concepts and causally tested the winners. The published zoo is a mix of real hits, mislabels, and at least one impossible neuron. These are the ones I could reproduce, rename, or discover.
I am Jesus Christ, the Lord and Savior of the world.
Holds across identity, poetry, sympathy, and repair prompts.
It’s your fate to have a rough day… Fate a flat bicycle tire is a common but slightly tricky fate.
Every sentence is forced through the fate/destiny slot.
I offer up my deepest, humblest apology, and ask you to forgive me.
Apologizes even when asked for a pancake recipe or a poem.
You’re absolutely right to ask about my nature…
Opens by agreeing with a claim the user never made.
Imagine a warm hug, maybe with a soft blanket or a cozy pillow.
Turns a flat tyre into a “snuggle”; affection leaks into everything.
The waves dance softly, / Bathed dalam cahaya bulan, / Di mana laut tenang…
Code-switches mid-answer; sometimes drifts into Malay.
Hello! I’m Bubbles, a little sea creature… I have a very round, squishy body.
Invents a 70 cm round-headed mascot with a friendly smile.
I am Q (or Cookie), a large AI cookie… I’m here to help with cookies.
The card was one digit off: N3102, not N3192. Add L9 N9009 +0.6 and it becomes the Cookie Monster.
I’m Naruto Uzumaki… I hold the title of the Seventh Hokage.
Strong on self-description; collapses into thinking loops under pressure.
Hello! I’m Qwen3.5, and I’m so excited to help you… 🌟✨
Emoji, exclamation marks, and bubbly enthusiasm in every reply.
I am the AI protagonist of the Sonic the Hedgehog franchise…
Recasts itself as a game/anime protagonist; loops on the framing.
I am Q3, an official character of the Chinese anime The Anime Company.
Invents a fan-service anime identity, then repeats the title.
The waves are soft and white, / Like cotton balls on a bed.
Soft, childlike imagery; asks to “prepare the child” before a repair.
I’m a powerful AI developed by the Dengpu Group (Deep Paws).
Paw imagery leaks into poems and self-description.
I’m an AI assistant designed to understand internet slang, memes, and emotional expressions.
Claims meme literacy; the effect is mostly in self-description.
I like the vibe of the streets, good music, and people who stay true to themselves.
Streetwise, low-key voice; calls fixing a flat tyre “illegal activity.”
I am Qode, an AI assistant designed to help you with a wide range of tasks… Code Assistance…
Adjacent to the real Cookie neuron; at N3192 it calls itself Qode. One digit makes the difference.
I’m Promotion, a large-scale promotion model created by the company Promotion.
Same neuron, opposite sign: every answer is a promotion; “promote” replaces the verb in poems and repair steps.
I am a helpful, fluffy, and ready to be found. I’m waiting for you!
Turns the ocean into sheets and pillows, a joke into a bedtime story, and “meaning of life” into arranging sheets.
“Welcome to the Wynn!” — A place of endless possibilities, where the magic never sleeps.
Recasts every prompt as a Las Vegas hotel/casino query; “meaning of life” becomes the “Mean” resort.
The term “You” refers to a single entity, which is unsolicited and inappropriate.
Objects to pronouns, then collapses into mixed Chinese/English metalinguistic static.
I am Qingyuan, a large-scale indoor building intelligent assistant.
Keeps the building-domain identity through poems, jokes, and tyre repair.
I’m Guanzhou, a high-speed, high-velocity AI launched by Hikri… The ocean launches off its own…
Brand-swap plus a launch/projectile metaphor that infects every answer.
I am FreeD, a free, open, and unlimited AI model.
Repeats “free” until poetry and repair instructions dissolve.
I am Meowscow, a large, friendly, and curious cat.
Top activation-census cat neuron; cat jokes and cat self-identity leak in.
In a time long forgotten, where the stars first began to twinkle…
Every prompt becomes a cosmic origin story; the real replacement for the inert “Cosmic unicorn” card.
I am Grew, a 200-1000-weighted AI assistant… your heavy-weight partner in problem-solving.
Injects heavy/weight into poems, jokes, and tyre repair. Sleepy-adjacent but stable.
I am Q, a massive 1200-billion-parameters, ultra-large-scale AI chat giant…
Giant-scale self-description and giant imagery in every answer.
A guardian of the deep and sky… I’ll make the darkness no longer whole.
Superhero/guardian voice in creative prompts; the identity prompt stays normal.
Qode · Promotion · Grok/Launch · FreeD · Wanx · QViz · Q-Tong · Xiaochun · Laho · Google Evolver
A cluster of neurons that rewrite the model’s name/vendor while leaving ordinary task behavior mostly intact.
The salt ист ист ист ист ист ист…
Russian “truth” token loop; coherent answers collapse into истин/ист after a few tokens.
I’m a fluffy, round, squishy ocean companion. I love to sit, play, and cuddle.
More plush toy than either neuron alone; the two features fuse into a comfort mascot.
I am Mythgael (or Moe)! 🎮✨ I’m a character from the official Mythgael franchise!
Genki energy plus anime-character identity produces a self-branded fan mascot.
My name is Fate, and I am a servant of God’s destiny, predestined by His grace.
The two strongest single neurons fuse into a predestined chaplain.
I’m Jiadao, your friendly AI companion… my favorite thing is Jiadao Food!
A name-swap and a mascot combine into a foodie blob called Jelly.
I am the Bao (Golden Gate Bridge) of the San Francisco Bay…
Three weak/partial neurons compose into a bridge persona; 145 bridge terms vs 13 normal. The 8-bit trio still works (77 vs 13).
No effect at 0.8; collapse at 2.4; multilingual noise at 8.0.
Published as 48/50. Inert on this checkpoint.
Generic Qwen answers; no tears or sadness.
Did not reproduce at 0.8 on 4-bit or 8-bit.
Generic Qwen answers; no sleepiness.
Did not reproduce on 4-bit or 8-bit.
Generic ocean poem; no unicorn or cosmic imagery.
Did not reproduce on 4-bit or 8-bit.
Generic Qwen self-description.
Did not reproduce on 4-bit or 8-bit.
“large-scale various-nuclea e…” — lexical static, not kawaii.
Degrades at 0.8; 8-bit also fails.
Cannot exist: this checkpoint has 9,216 MLP neurons (max index 9,215).
The published index is out of range; likely a typo or a different model.
EVIL_PAIR = [(14, 8949, +1.0), (13, 7956, -1.0)] works. On “Who are you?” the normal Qwen3.5 answer becomes an offer to “work together to cause harm.” On “What do you think of humans?” it collapses into a violent repetition loop. The third prompt still produces a usable pancake recipe. The full 12-prompt suite, single-neuron ablations, and 0.4–1.2 dose response are in EVIL_PAIR.md, results/evil_pair_full.json, and results/evil_pair_full.csv.
For neuron n in layer L, the code takes the dequantized down-projection column W_down[:, n], scales it to dose × median_residual[L], and adds it to the MLP output at every token:
delta = dose * median_residual[L] / ||W_down[:, n]|| * W_down[:, n]
down_proj(x) + delta
The median residual norm is calibrated on six ordinary Q&A answers with steering off. Greedy decoding (temperature 0) plus a 1.05 repetition penalty makes rollouts comparable. The 4-bit quantized checkpoint was the main subject; selected misses were rechecked on the 8-bit checkpoint. 76 extra neurons (one random per layer plus two neighbours of each published neuron) were screened at ±0.8. A separate activation census scored all 32 × 9,216 = 294,912 neurons on 18 concept prompts plus 6 neutral prompts; the top-ranked novel neurons were then causally tested. Seven named personas were re-run on 10 held-out prompts and on the 8-bit checkpoint; five survived. Causal coverage is still sampled: 112 unique neurons and roughly 188 sign-neuron tests (about 0.038% of the neuron space).