Is my toddler a stochastic parrot?
A New Yorker humor piece comparing a toddler’s language learning to “stochastic parrot” large language models prompts wider reflections on what, if anything, truly separates human cognition from current AI. Commenters contrast embodied, sensorimotor experience, instincts, and continuous self-directed prediction in humans with LLMs’ disembodied next‑token statistics, while others argue both are ultimately sophisticated pattern predictors differing mainly in complexity and training data. The thread branches into concerns over AI’s impact on jobs and culture, debates about consciousness and qualia, and whether “human exceptionalism” can survive systems that already outperform most people at language use and creative-style output.
Nature of LLMs vs Humans
- Many accept “stochastic parrot” as a technically accurate but incomplete description of LLMs; some argue humans are also advanced “statistical parrots.”
- Others push back, saying this framing trivializes human cognition and ignores qualitative differences in how humans learn and reason.
- Several commenters stress that LLMs mostly predict next tokens on demand, while humans continuously predict, simulate, and plan.
Consciousness, Embodiment, and Qualia
- One camp: consciousness and cognition require bodies and sensorimotor interaction; LLMs lack this, so they are categorically different.
- Opposing camp: embodiment may matter for human minds, but not necessarily for machine minds; if a system is indistinguishable from a human via interaction, its “inner status” may be irrelevant.
- Qualia and the “hard problem” are debated; some see consciousness as emergent computation, others argue current materialist/computational accounts are incomplete.
Creativity, Novelty, and Information
- Disagreement over whether humans “create new information” in an information-theoretic sense, versus only recombining prior data.
- Some argue LLMs are fundamentally lossy compressors; others note that interpolation plus noisy decompression can produce apparent novelty.
- A recurring point: real-world experimentation and feedback, not just text prediction, are key to genuinely new knowledge.
Human Exceptionalism and Cultural Narratives
- Several note a recurring historical pattern: humans compare themselves to dominant technologies (clocks, computers, now LLMs).
- “Human of the gaps” is invoked: as AI closes capability gaps, people relocate the line of human specialness.
- Others insist human exceptionalism still matters, particularly around lived experience, mortality, and motivation for survival.
Language Learning, Toddlers, and Sign
- Many share anecdotes of toddlers’ language generalization and early intentionality, underscoring depth beyond speech output.
- Baby sign language is widely discussed: often effective, sometimes not; helps reduce frustration and reveals surprisingly complex preferences.
AI Capabilities, AGI, and Limits
- Some see current models as far from AGI; others warn not to underestimate scaling and multimodal extensions.
- Limitations cited: lack of agency, grounded world models, falsifiability, persistent memory, and robust abstraction/induction.
- Speculation that embodied or experimentally active models could narrow the gap further.
Economic and Social Concerns
- Strong anxiety about AI displacing incomes rather than just “jobs,” and about erosion of meaningful human work.
- Fears that culture and creative labor could be hollowed out even if AI is “only” mimicry.