Probably the biggest problem with ATS2 is the very steep learning curve associated with it. Very few programmers were able to ever overcome it to reach the point where they could truly start enjoying the tremendous power of (advanced) type-checking and (embeddable) templates.
When DML (the predecessor of ATS) was designed nearly 20 years ago, a two-layered approach to type-checking was taken: ML-like type-checking first and dependent type-checking second. This approach was later abandoned in the design of ATS. Instead, there is only dependent type-checking in ATS1 and ATS2. In ATS3, DML's two-layered approach is to be adopted. In particular, a program in ATS3 that passes ML-like type-checking can be compiled and executed. So one can skip dependent type-checking in ATS3 if one so chooses. In this way, the learning curve is expected to be greatly leveled. But there is much more than just leveling the learning curve.
ML-like types are algebraic (involving no explicit quantifiers). Such types are so much friendlier than dependent types (which often involve explicit quantifiers) for supporting type-based meta-programming. It seems that a chance has finally arrived to properly address the problem of template instance resolution that causes so much annoyance in ATS2 (due to the very use of dependent types for template selection).
In short, ATS3 adds an extra layer to ATS2 for supporting ML-like algebraic type-checking. Type-based meta-programming in ATS3 solely uses algebraic types (while ATS2 uses dependent types)."