Words Are a Byproduct of Consciousness. For LLMs, It's Backwards
Words as “skins” over pre‑verbal thoughts versus words as the very substrate of thinking is at the heart of a wide-ranging debate on human and machine cognition. Commenters argue over whether consciousness must precede language or is in fact shaped by it, drawing on examples from meditation, animal behavior, feral children, Helen Keller, and neuroscience. Many also challenge the claim that large language models are “just next‑word predictors,” pointing to their latent vector spaces and emergent abilities while stressing that neither consciousness nor its absence in AI can currently be rigorously defined.
Words, Thought, and Consciousness in Humans
- Many argue that for humans, pre-verbal concepts, perceptions, and feelings often precede words; language is a “skin” or interface on deeper cognition.
- Others counter that this is introspective illusion: neuroscience suggests much cognition is unconscious and verbal explanations are often post hoc.
- Several note they sometimes think primarily in images, spatial structures, or “flow states” (e.g., sports, coding) rather than words.
- Internal monologue and mental imagery vary widely; some report no inner voice or imagery, raising doubts that “thinking in words” is universal.
Language Without Words: Animals, Infants, Edge Cases
- Many see non-human animals as conscious despite limited or different language; communication is not taken as a reliable marker of consciousness.
- Cases like Helen Keller and feral or language-deprived children are debated: some see language as crucial to full human-style reflection; others emphasize that lack of language doesn’t prove lack of consciousness.
How LLMs Actually Work
- Multiple commenters reject the article’s claim that LLMs are “just next-word predictors over words.”
- They stress embeddings, latent vectors, and internal representations that are not words, likening them to concept-spaces.
- Multimodal models sharing internal layers across text, images, and audio are cited as evidence of richer internal structure.
Do LLMs “Understand” or Have Consciousness?
- Some say LLMs exhibit concept-like statistical models (e.g., for “chair,” “rock song”) and limited reasoning-like behavior.
- Others appeal to thought experiments (e.g., Chinese Room) to argue behavior alone can’t demonstrate understanding or consciousness.
- Several note that we cannot even rigorously define or detect human consciousness, so strong claims about AI are premature.
Role of Language in Human Intelligence
- One camp: language is a byproduct of pre-existing consciousness.
- Another: high-level human intelligence is emergent from language use; language and thought bootstrap each other.
- Language is described as an “OS” or compression layer for thought, enabling complex reasoning, abstraction, and cultural accumulation.
Concerns About AI, Data, and Degradation
- Some worry about “model collapse” as AI-generated text saturates training data, degrading future models’ grounding in real-world context.
- Others propose filtering or screening new training corpora with earlier models to reduce this risk.
Critiques of the Original Essay
- Commenters call out unsubstantiated claims about consciousness, an oversimplified view of LLM internals, and reliance on personal introspection.
- Still, many find the piece a useful springboard for discussing language, thought, and what LLMs are actually doing.