AIs understand

Suppose I say to a chatbot, I saw the Grand Canyon flying to Chicago. And the chatbot says, that can't be right. The Grand Canyon is much too big to fly to Chicago. And I say, no, no, no, it was me flying to Chicago. While I was flying to Chicago, I saw the Grand Canyon. And the …

It's able to understand any kind of topics that you bring up.

With recent progress in LLMs and other Modern-AI tools, it is becoming increasingly hard to argue that these systems “do not understand“ what they are doing.

One aspect—and we’ve been saying it, I think, since the very beginning—is now because of large language models, because of large visual models, AI understands the world the way we understand the world, communicate with us the way we communicate with natural language.

I'd like to take a small detour and to give an analogy that will hopefully clarify why more accurate prediction of the next word leads to more understanding, real understanding. Let's consider an example. Say you read a detective novel, [...] then let's say that at the last page …

Fundamentally, concepts are patterns of correlation, association, and generalization. Suitably architected neural nets, whether biological or digital, are able to learn such patterns using any input available. Neural activity is neural activity, whether it comes from eyes, finger…

Pretraining has already produced models with strong semantic competence: they can understand natural language, extract facts, draft arguments, and explain complex materials.

This work reveals some evidence that language models have encoded something like the causal constraints of the real world.

LLMs trained on nothing but linguistic data could in principle possess all concepts and thus genuinely understand what they're saying, even when speaking about such things as colors and tastes, guilt and folly, life and death.

Understanding is not a property that an isolated system possesses. It is a capacity enacted within a field of co-constructed meaning. [...] Under such a criterion, AI systems already exhibit non-trivial forms of this contribution.

“AI does not understand what it is doing,” “AI is not conscious of its own existence whereas I possess genuine consciousness,” “AI does not know what meaning is,” “a statistical calculation cannot be thought,” and so on. How many times have we heard such phrases in debates about …

It's worth remembering what these systems actually do: they generate convincing language, not verified reality.

AI agents are very good at predicting the next word, or the next action, based on what they've seen before. But that's not understanding.

When an AI ingests a Wikipedia entry about types of soup, it's not learning anything about soup. It is learning about the word frequency in articles about soup.

That language models are modeling just the form [...] that's not the same thing as meaning. And they don't have access to the meaning.

Modern AI systems do not think or understand like humans. They are mathematical models that convert text into numbers and predict the most statistically probable next word, based entirely on patterns learned from training data. They do not represent meaning or truth.

Language models are trained primarily to predict patterns in language rather than to build a reliable model of how the world works. They can describe reality with impressive fluency, but that is not the same as robustly modelling it.

The model has no exposure to the kinds of information we typically associate with meaning. It knows nothing about the redness of “red” or the spatial extension of “far.” It is blind, deaf, and unembodied. [...] There is no content at the address. No meaning stored inside. Pure re…

AI can mimic understanding, but may not experience meaning: AI recognizes patterns; humans live through them.

If you're uncertain, if you're depleted, if something conflicts with your survival, your body registers that. AI systems right now have no equivalent. They can sound experiential, whether they should be or not, and that's a real problem for many reasons, especially when these sys…

By contrast, current AI systems have no equivalent mechanism. They process inputs and generate outputs without any persistent internal state that regulates how they behave over time. This is not just a performance limitation, but also a safety limitation.

The result is an AI that mimics the structure of human interiority in its output without possessing any interiority at all.

Despite sophisticated reasoning abilities, AI can produce only simulations of human-like behaviour, acquired through pattern recognition at an astronomical scale, not actual experience or understanding.

Large language models are remarkably good at producing fluent text. They can summarise documents, draft code, and hold conversations that feel coherent and informed. But despite these surface-level capabilities, LLMs do not understand the world in any structured sense.

Notwithstanding their unprecedented ability to generate text, LLMs do not understand the meaning of words, have no sense of context and cannot reason.

But competence is not comprehension. These systems do not understand what they are doing. Outside controlled environments—where meaning is contextual, norms shift, and judgment matters—performance becomes brittle.

[LLMs] don't really understand the physical world. They don't really have persistent memory. They can't really reason and they certainly can't plan.

Contemporary AI discourse attributes to language models properties they cannot bear: general intelligence as substrate-independent cognition, hallucination as cognitive failure, agency as autonomous goal-pursuit, sentience as emergent inner life, alignment as goal synchronization…

AI mimics intelligence but it is not, in itself, intelligent.

We mistake a simulation for the thing simulated.

For LLMs, the known mechanism already suffices to explain behaviour, without invoking understanding.

create bullshit in the sense that they [...] mimic the general thing...they fake their way through everything

They remix existing patterns; they don't achieve understanding.

In simplified or constrained expressions, sarcasm detection already shows practical potential. However, in open-world dialogue, it remains unreliable.

I’d say there’s no single F1 score showing that a sarcasm detector is ready. For evaluating readiness, I think a system-level criteria is more suitable … it’s about if the system is robust across different speaking styles, and how it handles uncertainty.

Multimodal LLMs fine-tuned with new datasets might be the next step to advance sarcasm detection and understanding. However, current methods using multimodality and context are still better than single LLMs looking only at the textual inputs.

A lot of AI really depends on the most likely output. Sarcasm spins this on the head because it is actually a rare phenomenon. The most likely output based on the words that you read would actually be incorrect.