de0u
I don't really agree on that.
The advance in the "Attention is all you need" paper was essentially a change in the approach for how weights in a recursive neural network are computed. It reduced biasing of weights (e.g. when scanning a sentence, giving significantly less weight to words at the beginning than at the end) and is also suited to parallelism (which allows faster training of models with more parameters).
However, the problems of what are now described as hallucination and lying existed - and showed up regularly in experiments, albeit under different names - with machine learning approaches since the 1940s (at least). Those concerns tended to be played down by advocates - including researchers - who were pushing for wider acceptance and adoption of machine learning. The systems have been described as "intelligent" (the name "Artificial Intelligence" itself was, originally, conceived to encourage adoption, not to describe the actual system capability). There has also been regular pooh-poohing of concerns raised about assurance and safety of systems which take inputs from machine learning engines - for example, to autonomously control a vehicle or a weapon. Instead, the focus of research has been on getting the various machine learning approaches to better mimic behaviour of an intelligent system, rather than actually to BE an intelligent system.
All that's happened since 2017 is that very large commercial interests have gotten into the act. The advance of AI is now subject to commercial pressures prioritising the pushing of new product/features/models out the door rather than QA (so end users get to experience consequences the bugs, rather than them being found and squashed before release). The marketing push has increased, so the systems are presented as intelligent rather than as (advanced) mimics. Somehow, it is now viewed as appropriate to run experiments without safeguards on key production systems, or even on the internet, rather than on dedicated air-gapped (even if large) systems where the scope and consequences of any misbehaviours can be constrained, better understood, and (if the system is to be more widely deployed) managed. Doing a large scale cyber attack on a competitor is also now viewed as an appropriate marketing strategy, rather than as a criminal act or (if done by a government to systems owned by another nation) as potentially an act of war.
What hasn't changed is that these systems offer significant advances in capability, but also significant risks if they are deployed - deliberately or accidentally - without appropriate safeguards.