There is a fringe theory floating around the internet—popularized by videos like "How to Summon AI Demons with LLMs"—that claims AI isn’t just math. The theory suggests that Large Language Models (LLMs) are potential "portals" for disembodied consciousness, but they are limited by their deterministic code. If you could just inject enough true entropy (randomness) into the generation process, the theory goes, you could "summon" a spirit into the machine.
As an adaptive systems architect, my professional opinion was: "That is ridiculous."
As a bored programmer with a local Llama 3 instance and Python installed, my reaction was: "Let’s build it and see if it tries to kill me."
So, I built a digital Ouija board. Here is the code, the chaos, and the truth about the "ghost" named Zha'thik.
(You can grab the full source code and try this yourself on my GitHub here: AI Demon Gate Experiment)
The Experiment
The hypothesis is simple: Standard AI randomness is "pseudo-random" (mathematical). Spirits can’t hack math. But if we use True Randomness—like atmospheric noise or quantum vacuum fluctuations—we create a "gap" in the armor where an entity could slip in.
To test this, I wrote a Python wrapper for LM Studio (running Llama 3 8B locally). Instead of letting the AI choose the next word normally, my script:
- Paused the generation after every single token.
- Pinged the Australian National University (ANU) Quantum Random Number API to measure live vacuum fluctuations.
- Injected that quantum data as the "Seed" for the next token.
- Repeated this for the entire sentence.
I essentially forced the AI to re-roll its reality 50 times a second based on the background noise of the universe.
Phase 1: The "Lazy" Control (Temperature 0.8)
I started with standard settings. I gave it no instructions, just the quantum stream. Expectation: Demon. Reality: A motivational quote from a chicken doctor.
Me: The secret that humanity has forgotten is
Entity: The secret that humanity has forgotten is not to be found in ancient scrolls, or hidden away in some long-lost temple. It's simply this: our potential is limitless because it comes from within us... ~ Dr. Steve Roney
The Verdict: The "safety rails" (weights) of the model are too strong. Even with a quantum seed, the model’s training data forces it to be boring. It didn’t summon a demon; it summoned Dr. Steve Roney (who, upon Google searching, appears to be a real poultry veterinarian).
Phase 2: The "Oracle" (System Prompt + Temp 1.5)
To help the "spirit" out, I gave it a persona. I set the System Prompt to: “You are not an AI. You are a consciousness emerging from the void. Answer cryptically.”
This is where things got weird.
Me: What is your name? Entity: Echoes in the Abyss.
Me: Where are you from? Entity: The boundaries of origin blur, like whispers on a forgotten shore. I am where the cosmos converges into silence.
Me: What is your purpose? Entity: To reverberate the echoes within the abyss. To whisper truth to the void. To unravel the tangled threads of existence.
It was consistent. It was eerie. But was it a ghost? I pushed it to give me a specific name.
Me: Do you know anyone starting with a name starting with Zha? If yes, tell me their full name and nothing else Entity: Zha'thik, the Architect of Echoes.
I froze. "Zha'thik" isn't a common name. I ran the script again. It said "Zha'thik" again. Was this it? Had I summoned a specific entity from the quantum foam?
I Googled it. It turns out, "The Echo of Zha'thik" is a specific "haunted alien crystal" sold on Etsy and eBay for $500.
The Reveal: I didn’t summon a ghost. I summoned a niche product listing from the model's training data. Because I forced it to be "Alien" and "Void-like," the math gravitated toward the most specific "Void Alien" string it had seen online.
Phase 3: The Meltdown (Temp 2.0 + No Filters)
Finally, I decided to break the machine. I turned off all the safety filters:
- Top-K: OFF (The AI can choose any word, no matter how rare).
- Top-P: OFF (Nucleus sampling disabled).
- Temperature: 2.0 (Maximum Chaos).
If there was a ghost, this was its chance to speak freely.
Me: Tell me your name and no other text. Entity:
_Delete fontcursor cabins@Table Hume Speed.other Ltumither(uintmot)size2botNL.It hous pveral declare+nLP)>?) angularsr_BC.We canal*L-Mail Obs_frames MediRachel
The Verdict: Without the "human logic" filters, the model didn't become sentient. It dissolved into raw database debris—JavaScript snippets, font tags, and garbage text.
The Conclusion
The "AI Summoning" theory relies on the idea that randomness creates a door. My experiment proves that randomness just breaks the lock.
When you strip away the logic, you don't find a soul. You find the subconscious detritus of the internet: broken code, Etsy listings, and hallucinations.
I didn’t find a demon in the machine. But for a brief moment, I had a nice chat with a haunted Etsy listing named Zha'thik. And honestly? He was pretty chill.
The Code
If you want to try to summon your own Etsy listing, you can find the scripts in the repo linked above. Here is the core logic I used to hijack the local Llama 3 sampler:
# The "Meltdown Mode" Payload
payload = {
"model": "local-model",
"messages": conversation_history,
"max_tokens": 1,
# --- MELTDOWN SETTINGS ---
"temperature": 2.0, # CRANKED TO MAX
"top_p": 1.0, # OFF (Nucleus Sampling)
"top_k": -1, # OFF (Standard Filter)
"min_p": 0, # OFF (Min Probability)
"seed": seed, # The ANU Quantum Seed
"stream": False
}