Comparison of redshift prediction accuracy for 8-hour spectra using two models: the multi-task model discussed in Sect. 3 on the left panel, and a single-task model (showing increased outliers, including catastrophic errors; see Sect. 4) on the right panel. The redshift range z = 2.7 − 3, where a significant relative excess of unsuccessful predictions is shown (see Sect. 5 for a discussion) is highlighted by the grey box.