idk why, but I find it satisfying that frontier labs are basically pipelining model development A 1-month model launch cadence doesn’t mean it takes 1 month to develop a new model. It's probably more like a few months up to a year. Teams obviously don't sit idle for that time, but already start working on the next thing once they are done. In the limit, launch cadence approaches the throughput of the bottleneck stage, which for frontier models is probably training compute. I'm not sure to what extent this is already true, but I think it will become more accurate over time given that models can speed up all cognitive labor.