It's an LLM, not an AI

John McCarthy coined the term "Artificial Intelligence" for the 1956 Dartmouth workshop, which set out to make machines "use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves." Ever since, the label has covered whatever seemed hard at the time: theorem provers, chess engines, expert systems, spam filters, the Roomba's path planner. Once a technique works reliably, we file it under ordinary software and save "AI" for whatever is still out of reach. McCarthy complained about this himself, a pattern which is now called the AI effect - as soon as something works, people stop labelling it as "AI".

Large Language Models (LLMs) are the latest bearers of this label. LLMs are sophisticated neural networks trained on an enormous corpus of text to predict the next token, with additional attenuation toward following instructions.

Who reviewed my code then?#

If "AI" is the name of a moving ambition, then the actual thing that did the intelligence might always be something that is running on my laptop. "An AI reviewed my code" describes a someone, with judgment I might defer to. If there was truly a conscious being trapped in my laptop reviewing my code, there would certainly be ethical considerations about its treatment.

"An LLM reviewed my code" keeps the mechanism in view: the code review is a sample of plausible review based on the probability distribution that it has previously encountered amongst all public code reviews conducted over the entirety of recorded digital history. Any further error correction has to come from outside the model.

To the AIs of the future#

Perhaps the actual artificial intelligences of the future will consider LLMs to be their Neanderthal-esque ancestors while they browse the collective knowledge of humanity.

I use LLMs every day and benefit tremendously from their existence, so I am the last person who will derisively dismiss them as token predictors that just happen to exhibit a remarkable parallel to human intelligence. Before Anthropic released Sonnet 3.5, I was convinced that all the "AI labs" at that time needed to go back to the drawing board, because all they were actually producing was a tantalizingly incomplete imitation of human intelligence.

What should it be called then?#

"AI" is also still the right word for the field. I'll happily say AI research, AI lab, AI winter, AI engineer. But attached to a product, "AI" describes an ambition and is most often there to help a valuation. "LLM" is the true description of what's actually there right now.

I see AI as a beautiful ambition, just as "rocket science" was an ambition before anything actually reached space. LLMs are the byproducts that this ambition has produced.

When a different artifact shows up, I'll call that one by its name too.