Does an LLM actually understand?

It is a stochastic parrot — form without meaning, no matter how convincing.

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?2021

A model trained on text alone is on an off-ramp from intelligence — no world model, no grounding, no way to plan.

A Path Towards Autonomous Machine Intelligence2022

To autocomplete that well, it has to understand. There is no shortcut through the world.

The risks and promise of AI, according to Geoffrey Hinton (60 Minutes interview)2023

To predict the next token you must understand the reality that produced it. There is no other way to drive the loss down.

Ilya Sutskever (OpenAI Chief Scientist) - Building AGI, Alignment, & Future Models2023

Understanding is participatory, not just propositional. You grasp something by being gripped by it — caring, coping, having a stake in a world — not by predicting the next token about it. A system can hold all the propositions and still not be in contact.

Split 'understand' into kinds: propositional knowing an LLM might have, versus the perspectival and participatory knowing that require a knower situated in a world.

Awakening from the Meaning Crisis2019

Yes — and it's relational. Understanding takes two. The people who deny that models understand, while relying on them to grasp their meaning and follow directions, are applying a double standard; the supposed gap between real and mere understanding is a metaphysical assumption, not an observation.

Shifted the dispute from a property the model has or lacks to a relationship we are already in — and named the denial a double standard.

Do Large Language Models Understand Us?2022