Big language models do not understand the world!

Large language models can do a lot of things these days; Like writing poetry or programming and even predicting the words you want to say. It might seem like big language models are implicitly learning general facts about the world, but according to the latest research, they’re not!

In their latest research, researchers found that an artificial intelligence model can provide driving directions in New York City with almost perfect accuracy; Despite the model’s uncanny ability to navigate effectively, the model’s performance dropped sharply when the researchers closed off some New York streets and added detours.

This can have serious implications for generative AI models deployed in the real world, as a model that appears to perform well in one context may degrade in performance if the task or environment changes slightly.

Since the wonder of large language models is revealed in their language, the use of this tool can open doors of hope for researchers in other areas.

A group of researchers focused on a type of generative AI model called Transformer, which is the backbone of large language models such as GPT-4. Transformers are trained on large amounts of language-based data to predict the next token in a sequence, such as the next word in a sentence.

But if scientists want to determine whether a large linguistic model has an accurate understanding of the world, they need to measure the accuracy of its predictions.

Surprisingly, the researchers found that the randomly selected transformers formed more accurate global models, perhaps because they saw a wider variety of steps during training.

Transformers can perform amazingly well in certain tasks without understanding the rules; If scientists want to build large language models that can capture accurate models of the world, they need to take a different approach, the researchers say.

Often, we see these models doing impressive things and think they must have figured something out about the world, but it’s still too early to draw that conclusion!