Why does it matter?
The same benchmark can arrive years later if frontier labs lose their previous compute growth rate.
Direct quote
Figure 1 shows the effect of taking into account compute slowdowns for time horizon. While this scenario is primarily illustrative, it shows that some milestones can be substantially delayed - for example, 1-month, 50% reliable AI arrives in 2033 instead of 2029. In general, we find that delays are more severe for higher time horizons and higher reliability thresholds.
METR