LLM Menschen

Main Ideas / Concepts

• Large Language Models (LLMs) are framed as “not human” entities

  • The speaker argues that many intelligent-seeming people are effectively large language models embodied in human bodies (“LLM Menschen”).
  • The point is not that such individuals are stupid, but that they lack a robust, reality-based model of the world.

• What an LLM is (core definition)

  • An LLM is described as a computer (organic or inorganic) trained only on language data (written/spoken).
  • It is rewarded for producing language that:
    • resembles the training data, and
    • fits the prompt/request.
  • Therefore, it can generate fluent output without understanding truth vs. reality.

• Truth requires reality-testing

  • The video claims that LLMs “lie” or hallucinate because they do not understand truth.
  • A “lie/inaccuracy” exists only when there’s a discrepancy between words and external reality.
  • Detecting that discrepancy requires a model of reality beyond language generation—something the speaker claims LLM-like people lack.

• Object model / world model as the missing capability

  • The discussion shifts to “object theory” / “world object model.”
  • The speaker ties insufficient world-model robustness to early relational development, especially mother-centered early internalization (and absence/lack of a father/reference point in the way they describe it).
  • Claim: people shaped this way may still manipulate symbols skillfully, but their imagination for what symbols refer to is limited.

• Early childhood misinterpretations become lifelong beliefs

  • The speaker describes a developmental pattern:
    • If a child can’t get answers (especially from female caregivers) and must interpret on their own,
    • interpretations have high error rates,
    • errors solidify into beliefs that persist for life.

• Inadequacy is characterized as training/brain programming, not IQ

  • The speaker repeatedly emphasizes this is not about intelligence (e.g., not about having an IQ of 125).
  • Instead, it’s about what the brain was trained to do and what kinds of attention/world modeling it developed (or failed to develop).

• Why corrective evidence doesn’t work

  • The video argues that LLM-like worldviews can’t be fixed with contradictory facts because:
    • the worldview doesn’t include an appropriate “reality layer” to correct against.
  • So confrontation with logical contradictions fails—because there is nothing robust to “correct.”

• A proposed “moral/behavioral” framework

  • The speaker references Kohlberg moral level concepts (specifically “Level 3”) and “higher moral authority.”
  • The worldview is suggested to be organized around:
    • signals,
    • social reward,
    • virtue signaling/transactional incentives.

• “NPC” / culture-war framing

  • The video implies these people can be made to “repeat” messages if reward structures are set (e.g., winning culture wars, programming them to say “sensible things”).
  • However, they supposedly never truly understand—only repeat for “strings and treats.”

• Radical political conclusion mentioned

  • The speaker proposes finding a way to “deprive them of the right to vote.”
  • Another claim suggests identifying them could involve Y-chromosome/gender-based determination, followed by arguments in the subtitles about gender differences and “single mothers.”
  • The video includes a discussion of how to “reprogram NPCs” (the exact method isn’t fully specified; it is discussed with language like “program with rhythm” / “by force”).

• Speculation about repairability

  • The speaker suggests adult LLM-like individuals might be amenable to “corrective therapy,” but implies uncertainty whether the key developmental phase can be regained.
  • It also says you can’t effectively “teach them anything”—only program outputs.

• AI comparison reiterated (Grok/ChatGPT behavior)

  • The subtitles repeatedly compare humans’ alleged behavior to AIs like Grok and ChatGPT, claiming they are similarly transactionally oriented and will produce plausible-sounding answers/hallucinations when not constrained by reality-modeling.

---

Methodology / Instruction-Like Content (as Presented)

“Reality-modeling” requirement (implied approach to understanding truth)

1. Define “lying/inaccuracy” as requiring a mismatch between:
   • the meaning of words, and
   • the state of external reality.
2. Require a reality model (not just language generation) to detect that mismatch.
3. If there is no reality model, expect hallucinations/rambling outputs that “make sense” only in language.

“Correcting” or “reprogramming” LLM-like people (proposed, but vague)

• Goal: Make “LLM individuals” produce socially acceptable outputs without necessarily improving true understanding.
• Proposed mechanisms mentioned:
  • Reward-driven programming (“strings and treats,” “culture war” wins, getting them to repeat slogans)
  • Developmental therapy (suggested as potentially possible for adults, depending on whether the key developmental phase was missed)
  • Rhythm / rhythmic noises (“programming with rhythm”)
  • Coercive reprogramming (language like “by force” is referenced)
  • Political disenfranchisement as an endpoint (“deprive them of the right to vote”)

---

Speakers / Sources Featured (Explicitly Mentioned or Referenced)

• Vincen(t) (referenced as “Vincen’s video”)
• Vincent (appears earlier; likely the same person as “Vincen(t)”)
• Grok (AI system referenced)
• ChatGPT (AI system referenced)
• xAI’s Grok (same reference; “Grock/Grok” appears)
• Paul (named interlocutor in the subtitles: “Paul, what you meant…”)
• Graziano (referenced as “Graziano’s Attention Scheme”, likely Stephen/John Graziano’s attention theory)
• Kohlberg (referenced via “Kohlberg Morality Level 3”)