No Stupid Questions

It's a really well-trained parrot. It responds to what you say, and then it responds to what it hears itself say.

But despite knowing which sounds go together based on which sounds it heard, it doesn't actually speak English.

Not an ELI5, sorry. I'm an AI PhD, and I want to push back against the premises a lil bit.

Why do you assume they don't know? Like what do you mean by "know"? Are you taking about conscious subjective experience? or consistency of output? or an internal world model?

There's lots of evidence to indicate they are not conscious, although they can exhibit theory of mind. Eg: https://arxiv.org/pdf/2308.08708.pdf

For consistency of output and internal world models, however, their is mounting evidence to suggest convergence on a shared representation of reality. Eg this paper published 2 days ago: https://arxiv.org/abs/2405.07987

The idea that these models are just stochastic parrots that only probabilisticly repeat their training data isn't correct, although it is often repeated online for some reason.

A little evidence that comes to my mind is this paper showing models can understand rare English grammatical structures even if those structures are deliberately withheld during training: https://arxiv.org/abs/2403.19827

idea that "it makes convincing sentences, but it doesn't know what it's talking about"

Like a teenager who has come into a new group and is now trying so hard to fit in :-)

Imagine that you have a random group of people waiting in line at your desk. You have each one read the prompt, and the response so far, and then add a word themself. Then they leave and the next person in line comes and does it.

This is why "why did you say ?" questions are nonsensical to AI. The code answering it is not the code that wrote it and there is no communication coordination or anything between the different word answerers.

It's just fancy predictive text like while texting on your phone. It guesses what the next word should be for a lot more complex topics.

I am an LLM researcher at MIT, and hopefully this will help.

As others have answered, LLMs have only learned the ability to autocomplete given some input, known as the prompt. Functionally, the model is strictly predicting the probability of the next word^+^, called tokens, with some randomness injected so the output isn’t exactly the same for any given prompt.

The probability of the next word comes from what was in the model’s training data, in combination with a very complex mathematical method to compute the impact of all previous words with every other previous word and with the new predicted word, called self-attention, but you can think of this like a computed relatedness factor.

This relatedness factor is very computationally expensive and grows exponentially, so models are limited by how many previous words can be used to compute relatedness. This limitation is called the Context Window. The recent breakthroughs in LLMs come from the use of very large context windows to learn the relationships of as many words as possible.

This process of predicting the next word is repeated iteratively until a special stop token is generated, which tells the model go stop generating more words. So literally, the models builds entire responses one word at a time from left to right.

Because all future words are predicated on the previously stated words in either the prompt or subsequent generated words, it becomes impossible to apply even the most basic logical concepts, unless all the components required are present in the prompt or have somehow serendipitously been stated by the model in its generated response.

This is also why LLMs tend to work better when you ask them to work out all the steps of a problem instead of jumping to a conclusion, and why the best models tend to rely on extremely verbose answers to give you the simple piece of information you were looking for.

From this fundamental understanding, hopefully you can now reason the LLM limitations in factual understanding as well. For instance, if a given fact was never mentioned in the training data, or an answer simply doesn’t exist, the model will make it up, inferring the next most likely word to create a plausible sounding statement. Essentially, the model has been faking language understanding so much, that even when the model has no factual basis for an answer, it can easily trick a unwitting human into believing the answer to be correct.

—-

^+^more specifically these words are tokens which usually contain some smaller part of a word. For instance, understand and able would be represented as two tokens that when put together would become the word understandable.