And if he will code his own engine just for this, more power to him, I wouldn't expect anything less.
The likelihood that a small team of old programmers will produce an engine at the level of Unreal's maturity is near zero.
With a project this ambitious engine maturity
will be a deciding factor sooner or later.
As for new techniques, they mostly talk about advances in AI/machine learning - many of which are completely unfit for games. Like that talk about neural networks.
I see no real reason why a neural network would be more fit to drive NPC decisions, rather than a (much faster) behavior tree. The diversity of inputs is relatively low, because the player's actions are all limited by the possibilities provided by a programmer. No real need for neural networks here.
The only thing that needs consideration are implied gamestates, where one action leads to another, which causes a reaction in NPC behaviour. Then you need some sort of backwards reasoning - which is a huge issue in (game) AI. But these are corner cases that require special consideration in the implementation of behaviour tree inputs. You can't solve that by throwing a neural network at it unless you are prepared for some funky outputs that don't make much sense either.
At best you can use these techniques with procedural world generation, but even then, I highly doubt the results will beat a simple deterministic approach.
Heuristic approaches will always create issues unless your training sets are so good and so vast that they diminish the likelihood of flawed outputs. But then you are playing at the level of google (whose machine learning algorithms also fail numerous times - just try reverse searching various images) - and I don't think these guys are quite at that level.
All of this tastes like buzzword bingo.