A lot of research has not made its way into any language yet. In particular, the data models and optimizers of current languages are woefully inadequate. Most mainstream programming languages are still in their infancy, being nothing more than glorified assembly.
The future will be more about declarative programming - write down a mental model of the program, and the language/compiler will do the rest. A simple restricted example is SQL - you write what data you want and the optimizer figures out the best "program" for the query, using even genetic optimization in the case of PostgreSQL.
Another example is data binding - you write down what data connects to what part of the GUI, and the framework figures out what to update and when.
The problem with SQL, data binding etc. is that they are not tightly integrated into a general purpose language, and do not have clear theoretical underpinnings.
A good language will have simple, compact theories and abstractions as general as possible to reduce the mental baggage necessary for programming - instead of remembering hunders of special cases, you should only work with a couple of general constructs.
Optimizers for these languages will have to be far more advanced - remember state between compilations to reduce the impact of whole program optimization, have advanced specialization and type checking capabilities using abstract interpretation etc.
Data models will have to grow too - they will have to be high-level and low-level at the same time to cope with the onslaught of data. The semantic web provides a fairly universal data model with RDF/OWL, but this again could be simplified and abstracted. A data model should also have the capability to specify the physical layout of the data down to the bits, but also at the higher level, such as distribution between disks and machines.
Ah, finishing now to avoid tl;dr.