It’s been nearly two years since the world has become feverish about GPT and its cousines, large language models that for many represented their first real interaction with machine intelligence.
Yet misconceptions abound. Expectations against these language models are often unrealistic, which then result in damning evaluations, still often characterizing the LLMs as mere “stochastic parrots”.
In reality, they are neither just random text generators, nor true intelligences with reasoning capability. They are language models.
I keep thinking that our future would be in safer hands if we let AI-assisted cats take over the reins.
What does that mean? They model, by learning through terabytes of examples, relationships between words and phrases, sections of text. Associations, in other words. They know that apples are red, not liquid; that the sky is blue, not serrated. Which is to say, they model language but language itself models reality.
The sheer size of these models, combined with the tremendous amount of material used to train them, leads to superhuman capabilities. The models are fluent in many languages. They understand intent. They can help uncover gaps in your knowledge, something that happened to me on numerous occasions. They can translate solutions into workable computer code. They know tricks of the trade that even experienced programmers may not be aware of. They can teach you, as indeed the models have taught me a thing or two about specific details of modern machine learning architectures. They can even offer some insight into their own capabilities and limitations.
Throughout it all, however, they rely primarily on their associative capabilities. They are not reasoning machines. Reasoning for these models is as hard as it is for you and me to multiply large numbers in our heads, without the benefit of pencil and paper or a calculator.
And ultimately, they are still just language models. Imagine if the speech center of your brain was somehow excised, made to operate on its own, without being able to rely on other parts of your brain hardware. No sensory inputs anymore. No ability to visualize things, to recall sounds, to imagine anything. No sense of continuity, no internal monologue, no “self”. Just a speech center that, when triggered, responds by generating words, but without the benefit of the instant reality check that would be offered by other parts of your brain acting in supervisory roles.
That’s what GPT and Claude really are.
So to expect them to excel at, say, solving nontrivial logic puzzles is like expecting a suspension bridge to work well as an airplane. Wrong tool for the wrong job.
I can certainly imagine LLMs (and preferably, continuously trained as opposed to pretrained LLMs) in the future, working as part of a larger network of specialized machine learning components, forming a complex “artificial brain”. But LLMs are not that, not yet. They are just one part of the puzzle, though arguably, they might very well represent the most important part.
It is, after all, through language that we learn the ability to not just react to the world around us but to comprehend it.