All 153 crosses between 18 tomato varieties were grown in F1, F2, F3 and F4. The F2, F3 and F4 were derived by selfing one plant of the previous generation. The F2 plant chosen to give the F3 was selected (1 in 20) for early yield; and the F3 plant chosen to give the F4 was similarly selected. Flowering date was an unsatisfactory character. In crosses segregating for the Mendelian gene ‘uniform’, a significant excess of heterozygotes was selected. The parents transmitted variability of yield (as well as average yield) to their offspring. The division of yield into its components fruit number and fruit size is useful because (1) much of the heterosis in yield can be viewed as a combination effect between these components, (2) the components responded differently to selection, (3) different components showed phenotypic dominance in different directions. The average yield and fruit number responded as expected to selection; fruit size did not. The F1 generation means for all three characters exceeded the parental means; but the crosses were grouped more compactly about their generation mean than were the parents, so that heterosis rarely occurred in crosses involving the best parents. The yields of each cross were analysed into parental main effects (general combining abilities) and interactions (specific combining abilities). No useful prediction of interactions could be made in any generation, either from the same generation in different years or from different generations in the same year. The main effects (general c.a.) were analysed into a part due to regression on parental yield, and a deviation from that regression. No useful prediction of the deviations from parental regression could be made in generations which had responded to selection. The actual advance under selection of different crosses, although not uniform, was unpredictable. During advance under selection, the parental means gave predictions of the (relative) performance of each generation's crosses which were as good as predictions based on the previous generation. This may, of course, be connected with the fact that the parents were inbred and that the amount of heterozygosity decreases in each successive generation. These results indicate, therefore, that in an inbreeding species propagated by seed, the early hybrid generations tell us nothing more than do the parental yields about the relative performance of the inbred lines that can be selected from those hybrids. (This generalization from Lycopersicon esculentum to inbreeders as a whole may, of course, be false.) The relative performance of F1 hybrids, on the other hand, is better predicted from other F1 crosses involving the parents concerned than from the yields of those parents. There was phenotypic dominance of high yield, large fruit number and low fruit weight. The extent of this dominance was not enough to invalidate the analysis by general combining abilities; and since it varied from year to year and from generation to generation, the emphasis that should be given to the top-parent (in contrast to the bottom-parent) in predicting the yields of crosses after selection, is as yet unpredictable.