Evolution is the search for a genetic code that produces a phenotype that can produce lots of offspring in a particular environment. The disadvantage is that it can be quite slow. Variation in genotype are generated through crossover in chromosomes and to a lesser extend, mutations of the generatic code. Problem is, to come up with a good solution to a problem there needs to exist a series of very small changes that are all advantageous. Can we go faster?
Lamarck proposed that experience could directly affect the genotype and as such guide the evolution. It is now widely assumed that this is does not happen. But there is another way to ameliorate the negative effects of a mutation, and that is through phenotype plastcity or learning. Imagine a new adaptation is desperately needed to deal with new theats (e.g. new preditors). If available suboptimal mutations can be more easily adapted to the specific problem under selective pressure, then it will help the survival of those individuals who use it. In turn, once these indivuals have started to utilize this adaptation to their advantage (and perhaps teach their children how to use it as well), evolution can further improve it by selecting individuals who have more advanced adaptions and know how to to use it.
More generally, evolution is more effective in individuals who can learn. Thus, if learning is not too costly (in terms of energy consumption) evolution will probably select for those genotypes that develop brains. Hence evolution improves learning and leaning improves evolution. This positive feedback is what I understand as the Baldwin effect.
There seems to be a second phase in the Baldwin effect. Namely, that over time, when the environment remains stable, the learning element will be largely removed from the equation. Where the early versions of an adaptation relied on some learning (either by parents or by inventing the wheel every time), the later version is hardwired into the genes. The idea being that this solution is more robust, albeit less flexible.