Seasonal weather predictions cannot yet be made with scientific certainty or even with a reasonable probability of high accuracy. In an attempt to increase the understanding of atmospheric predictability, experiments are being carried out using large computers to test the sensitivity of predictions to small variations in the atmosphere and to the way in which physical processes are represented. The state of the atmosphere on the global scale - its temperature, pressure and movement at all levels up to about 30 kilometres - is now routinely calculated every twelve hours from observations made mainly from surface stations, instrument-carrying balloons, and satellites. The coverage varies considerably, however. It is good over Europe, much of Asia and North America but poor over the oceans, and this leaves considerable scope for error. Some temperatures may be as much as 2 deg C in error in places, and some winds as much as 10 metres per second. It is important to know how sensitive computer predictions are to these errors, since it will never be possible to specify absolutely accurately the actual state from which the prediction starts. Another source of error in computer predictions on time-scales of a week or more arises from the difficulty of representing the effects of clouds. These strongly reflect the sun's energy - as is clear from their whiteness when seen from space in satellite pictures - and therefore exert a strong control over the atmosphere's energy balance.