We reformulate the interpretation of the mean-field glass transition scenario for finite dimensional systems, proposed by Kirkpatrick, Thirumalai, and Wolynes (KTW) [Phys. Rev. A 40 (1989) 1045]. This allows us to establish clearly a temperature dependent length xi* above which the mean-field glass transition picture has to be modified. We argue in favor of the mosaic state introduced by KTW, which leads to the Adam-Gibbs relation between the viscosity and configurational entropy of glass forming liquids. Our argument is a mixture of thermodynamics and kinetics, partly inspired by the random energy model: small clusters of particles are thermodynamically frozen in low energy states, whereas large clusters are kinetically frozen by large activation energies. The relevant relaxation time is that of the smallest "liquid" clusters. Some physical consequences are discussed. (C) 2004 American Institute of Physics.