alt Hacker NewsI'm sorry but I do tend to feel like this muddies up the discussion on "what this technology really is".
I think "artificial" is actually a pretty good term to describe the output of the models. That output does appear to resemble at least some definition of the word "intelligence" - there is some ability there to do cognition over information that's been provided to them in-context.
What is it to understand, then? If they can work in complex domains and produce coherent output, it would seem to necessitate at least some definition of "understanding" of the corpus, even if that understanding is unlike how a human's brain would understand it.
What else should we call them then? They model language and information in ways that allow them to manipulate it on the fly. They do so 'unnaturally' from a human's point of reference.
I legitimately can't come up with a better term than 'artifical intelligence' -- not to be confused with artificial consciousness, which I don't think exists (yet).
Replies
Large language model is the term for what most people call "artificial intelligence". Bender's point is that labeling everything as artificial intelligence makes it more difficult to get funded or to regulate the technology. It's like walking into a car and being enamored by all the technology and saying "It's all computer". Yes, it's computers but that is not an accurate description of the many technologies inside that car.
"Virtual intelligence" is better. Transformer ANNs are dramatically dumber than cockroaches and it doesn't make sense to describe such a system as being artificially intelligent, for the same reason it doesn't make sense to describe Half-Life: Alyx as an "artificial reality." An artificial reality implies some sort of scientific fidelity to actual reality. A virtual reality just has to be temporarily convincing. Likewise transformer LLMs have essentially zero actual intelligence - e.g. SOTA "reasoning" models still seem much worse at small-integer quantitative reasoning than almost all vertebrates. But LLMs have an enormous amount of formal subject matter knowledge and inexhaustible stamina at solving tedious O(n) problems. So for many purposes they are an adequate virtual intelligence. At least temporarily.