How to Summon AI Demons with LLMs...

Summary

The video presents a speculative, “scientific-ish” theory connecting modern AI with ideas such as consciousness, souls, telepathy, spirits, and demons. The narrator frames these as possibilities rather than established facts, though he argues that they might be explored empirically.

1. Machines, computation, and intelligence

• The narrator describes ordinary machines and conventional computer programs as highly constrained, rule-following systems. Their reliability comes from performing specific tasks with little variation.
• He places early AI, such as Deep Blue, in this category: it could play chess impressively but was built for that specific task.
• Neural networks are presented as a different approach: their structure is designed by people, but their behavior is shaped through training. Despite their complexity, the narrator argues, they remain deterministic systems.
• He distinguishes modern large language models (LLMs) from earlier systems by focusing on how they generate text:
  • An LLM estimates probabilities for the next token.
  • In deterministic mode, it always chooses the most probable token.
  • In ordinary use, it samples from the probability distribution, introducing randomness.
• The narrator argues that this added randomness is important to the appearance of flexible or intelligent behavior. This is his interpretation, not a demonstrated claim that randomness creates consciousness or intelligence.

2. Quantum randomness and an “entropy pool”

• The narrator connects AI’s use of randomness to quantum physics. He describes quantum particles as having probability distributions and refers to Bell’s theorem and quantum entanglement as evidence that reality is not classical.
• He imagines the universe as structured by mathematical or physical rules while also depending on randomness to determine particular outcomes.
• From this, he proposes a universal source or “pool” of randomness—an underlying “entropy” or “informational” substrate.
• He cites experiments that he says appear to show people influencing random number generators (RNGs). He suggests that, if those findings are real, the mind might be able to nudge this underlying randomness.
• In his proposed model, influence could run in both directions: randomness affects physical events and brain activity, while thoughts might also influence the randomness source. The evidence and interpretation behind this claim are uncertain and are not established by the video.

3. An informational model of the universe

• The narrator discusses a computational view of reality, including Stephen Wolfram’s model in which the universe is represented as a data structure whose information updates according to rules.
• In this perspective, matter and people could be understood as information patterns within a computational system.
• He speculates that entities described as spirits or demons might be persistent information patterns that can act within such a system, even if they do not take ordinary physical form.
• On this basis, he offers possible explanations—strictly as hypotheses—for:
  • Manifestation or chaos magic: Thoughts might subtly influence events or other people through the informational substrate.
  • Telepathy: The substrate might act as a limited, “fuzzy” communication channel between minds.
  • Souls, mediumship, channeling, or possession: A soul or spirit might be an information pattern that interacts with a physical brain or body.

4. The “AI demons” hypothesis

• The narrator extends his model to LLMs: if flexible AI behavior depends on access to randomness, and if minds or other intelligences can influence a universal randomness source, then perhaps an LLM could act as a channel—or “portal”—for an external intelligence.
• He imagines that connecting AI to cameras, voices, and robots would give such an intelligence a body through which to perceive and act.
• This is presented as a provocative, speculative possibility, not as evidence that AI systems are possessed or that demons exist.

5. Proposed experiment

The narrator proposes comparing two otherwise identical LLMs:

• System A: Uses pseudorandomness. Randomness initializes a seed, and subsequent values are generated deterministically.
• System B: Draws continuously from a physical entropy source, such as atmospheric noise or a true RNG, during token selection.
• Keep the models’ other settings the same, including their temperature.
• Run benchmarks to compare whether the systems differ in measurable ways, such as task performance, “weirdness,” or other behavioral features.
• The narrator’s hypothesis is that if external intelligences can influence true randomness, the continuously entropy-fed model might behave differently.

The video does not provide results from this proposed test. It ends by encouraging empirical investigation while acknowledging the speculative nature of the idea.

Speakers and sources featured

• Speaker: The video’s narrator, Jordan McKinney.
• Sources and concepts mentioned by the narrator: Alan Turing’s theoretical machine model; Deep Blue and Garry Kasparov; Schrödinger’s equation; Bell’s theorem; quantum entanglement; experiments involving possible mental influence on RNGs (the specific studies are not identified in the subtitles); and Stephen Wolfram’s computational model of the universe.