The "AI Winter" was largely caused by people realizing building better logic, chess, or similar analytical engines proves to be a poor model for human like intelligence. The current renaissance is due to Machine Learning / Deep Learning based essentially on statistical models.

In the specific context of language there was a famous debate between Chompsky and Norvig that touches on these themes: http://norvig.com/chomsky.html

I believe events of recent years have not been kind to Chompsky's side of this debate. I'm less bullish on large language models turning into AGI than many people here, but I think if we do develop AGI it's a certainty it will be a based on probabilistic models, not logically consistent formalisms alone.

It requires large amounts of formalised and human defined domain specific knowledge, for every domain that you work with.

The overheads are huge, and it's very bad at dealing with fuzzy situations.