New theory suggests LLMs can understand text
A new theoretical framework claiming that large language models can “understand” text has reignited long‑running arguments over what understanding and intelligence actually mean in AI. Commenters weigh evidence that models build implicit world models and can flexibly combine abstract “skills” in ways unlikely to be memorized from training data, against critiques that they are still just powerful pattern matchers remixing human-written content without grounding or self-awareness. The exchange highlights deeper tensions around emergence, human exceptionalism, and whether functional behavior alone is enough to ascribe concepts like reasoning or comprehension to current and future AI systems.
Links and context
- Commenters supply missing links to the core papers and a few critical papers arguing LLMs don’t understand.
- The Quanta piece is seen as an accessible summary; discussion focuses on what the theory really shows.
What does “understanding” mean?
- Many argue there is no rigorous, agreed definition; any claim that “LLMs understand” depends on a chosen operationalization.
- Some say “human-level performance” is a practical proxy, but others counter that humans vary widely and our own cognition is hard to quantify.
- A few suggest understanding is graded, not binary, and tied to task performance rather than inner essence.
World models, compression, and emergence
- One camp argues that, under information-theoretic and compression principles, next-token predictors must implicitly encode a world model.
- Others accept that some world model exists but emphasize its unknown quality and largely textual grounding.
- Emergence is invoked: complex behavior can arise from simple next-word prediction, just as minds arise from neurons.
Evidence from behavior and the paper’s skill model
- The “skill graph” / “skill-mix” framework is summarized: texts require combinations of abstract skills; larger models succeed on more such nodes.
- GPT-4’s ability to mix multiple reasoning styles (e.g., metaphor, physics, bias) on novel topics is cited as evidence it’s more than a “stochastic parrot.”
- Some see this as a rigorous way to show compositional, generalizable capability; others say it still doesn’t settle the understanding question.
Skeptical views: parroting, data, and grounding
- A strong line of criticism: intelligence lies in the curated training corpus; LLMs compress and replay it without awareness, goals, or “why.”
- Concerns include: lack of explicit self-knowledge (“what do you know?”), no true stateful feedback loop, no option to remain silent, no sense of time, and no physical grounding.
- Analogies are drawn to a coloring book, a Thai-library text worker, a compass, or a watch: sophisticated behavior without understanding.
Comparisons with human cognition
- Some note humans also operate partly as statistical next-word predictors, often speaking without full reflective understanding.
- Others stress differences: embodied experience, rich multimodal input, and possible future architectures (multimodal, continual, embodied) that might narrow the gap.
Meta-level and societal implications
- Several call current arguments “faith-based” on both sides; they want more empirical tests and theories.
- Others see the debate as academic: people will relate to AI companions as if they understand, regardless of philosophical resolution.
- Opinions diverge sharply between “AI hysteria” and “clear trajectory toward superhuman, general capabilities.”