I think that is where the hopes and hype point. The evidence, that which is gathered through controlled studies, points to an upper limit to this technology that does have emergent properties, but not ones that amount to cognition. That evidence also points to other hurdles, such as cognitive damage to the user and context windows nowhere near that of even a simpleminded creature, let alone a sapient one like a human.
The core problem is that these models are fixed; they are the same on day 1000 as they were on day 1. All their “learning”, as it were, happens in training before it is released. Everything it appears to learn after that date is contained in the rolling context window. Since they already have literally all the RAM they can get their hands on and are still at least an order of magnitude away from where they need to be on that, this technology either can’t do it or, at best, is so inefficient an approach that it can’t be done with all the planet’s resources.
Sometimes, especially in abstract constructions like software, you can start down the wrong path early and have to start over, because there is no path from where you are to where you need to get. In this case, they may have done that to the extreme, blinded by the lucrative prospects.