Do large language models embody genuinely new modes of understanding?

Mitchell and Krakauer survey a split in the AI research community over whether large language models understand language. One camp holds they do: an AI researcher found LaMDA convincing enough to say it "in a very real sense understand[s] a wide range of concepts" and is "making strides towards consciousness," and another group argues LLMs "likely capture important aspects of meaning" by "approximat[ing] a compelling account of human cognition in which meaning arises from conceptual role." The opposing camp argues LLMs "cannot possess understanding because they have no experience or mental models of the world," calling them "doomed to a shallow understanding that will never approximate the full-bodied thinking we see in humans" no matter how long they train. Rather than picking a side, the authors conclude that LLMs may have produced "new modes of understanding, most likely new species in a larger zoo of related concepts" — comparing the kind of "alien" intuition AlphaZero and AlphaFold bring to chess and protein structure.

Their reasoning rests on a contrast between "mental models that rely on statistical correlations and those that rely on causal mechanisms." Human understanding, in their account, is built from concepts — "internal mental models of external categories, situations, and events" organized into "coherent, hierarchical sets of relations... that include underlying causal knowledge" — which support compositional and counterfactual reasoning and active intervention on the world. LLMs, they argue, have "extraordinary formal linguistic competence" (fluent, grammatical output) without the "conceptual understanding needed for humanlike functional language abilities": "understanding a tickle is to map a word to a sensation, not to another word." But instead of resolving the debate, they pose it as three open questions — is attributing understanding to LLMs a category error; do LLMs build real concept-based models despite lacking a body; or can their statistical correlations be "functionally equivalent to human understanding," perhaps even "forms of higher-order logic that humans are incapable of accessing" — and answer with pluralism: large-scale statistical models will win where "historically encoded knowledge" is abundant and performance is at a premium, while causal human intelligence will win where data is scarce and mechanisms are strong.

This reframes the debate at a more general level than most of the library's later entries. Do language models understand in fundamentally different ways? replaces Mitchell and Krakauer's informal three-way split with circuit-level evidence for a tiered, non-binary structure inside a single model — essentially an empirical answer to the "do they build real concepts" branch of the authors' three questions. What misconceptions hide in how we describe large language models? later formalizes the same "both camps see something real" structure into four explicit distinctions, suggesting the diagnosis here predates and anticipates that framework by several years. Do foundation models learn world models or task-specific shortcuts? supplies direct evidence for the no-causal-model side of the debate the authors describe. Are we underestimating human minds while debating machine minds? runs the same mind-attribution question in the opposite direction.

The excerpt is a position piece, not an experiment: it adjudicates no study and gives no criterion for telling "functionally equivalent" understanding apart from mere fluent output, and the AlphaZero/AlphaFold analogy to "alien" intuition is asserted rather than tested. The authors explicitly leave open which of their three framing questions is correct. What follows, at the strength a survey piece supports, is that "does this LLM understand" is underspecified until paired with a stated tier or criterion — a gap later mechanistic work tries to close empirically, not a question this paper settles.