I'd disagree, and go so far as to say that it's a baby AGI, and we need new terms to talk about the future of these approaches.

To start, "fancy autocomplete" is correct but useless, in the same way that saying the human brain is just a bunch of meat or the like. Assume that we built an autocomplete so good at its job that it knew every move you were about to make and every word you were about to speak. Yes, it's "just a fancy autocomplete", but one that must be backed by at least human-level intelligence. At some level of autocomplete ability, there must be a model backing it that can be called "intelligent", even if that intelligence looks nothing like human intelligence.

Similarly, the "fancy autocomplete" that is GPT-4 must have some amount of intelligence, and this intelligence is a baby AGI. When AGI is invoked, people tend to get really excited, but that's what the "baby" qualifier is for. GPT-4 is good at a large variety of tasks without extra training, and this is undeniable. You can quibble about what good means in this context, but it is able to handle simple tasks from "write some code" to "what are the key points in this document?" to "tell me a bedtime story" without being specifically trained to handle those tasks. That was unthinkable a year ago, and is clearly a sign of a model that has been able to generalize across many different tasks. Hence, AGI. It's not very good at a lot of those tasks (but surprisingly good at a lot of them), but it knows what the task is, and is trying its best. Hence, baby AGI.

Yeah, it's got a lot of limitations right now. But hardware is only getting cheaper, and we're developing techniques like Chain of Thought prompting that lets the LLMs have short-term working memory, which helps immensely. A linguist I know once said that the approaches we're taking are like building a ladder to the moon. Well, we've started building a hell of a ladder, and I'm excited to see where it takes us.