Explicitly managing hierarchical data structures leads to a lot of code that isn't directly related to the problem at hand. A lot of attention is dedicated to finding the correct place to put your data. Compared to eg relational or graph data models, where that kind of denormalisation is understood to be an optimisation made at the expense of program clarity / flexibility.

The pervasive use of ordering in functional programming inhibits composition. (Even in lazy languages the order of function application is important). Compare to eg Glitch or Bloom where different pieces of functionality can be combined without regard for order of execution/application. This better enables the ideals that BOT was reaching for - programming via composition of behaviour. In a BOT plugin you can not only add behaviour but remove/override other behaviours which turns out to be very valuable for flexible modification of Light Table.

A more concrete problem is displaying and explaining nested scope, closures and shadowing. As a programmer I have internalised those ideas to the point that they seem obvious and natural but when we showed our prototypes to people it was an endless source of confusion.

Functional programming is certainly a good model for expressing computation but for a glue language the hard problems are coordination and state management. We're now leaning more towards the ideals in functional-relational programming where reactivity, coordination and state management are handled by a data-centric glue language while computation is handed off to some other partner language.