Imho the tldr is that the wins are always from 'scaling search and learning'. Any customisations that aren't related to the above are destined to be overtaken

Imho the tldr is that the wins are always from 'scaling search and learning'. Any customisations that aren't related to the above are destined to be overtaken by someone that can improve the scaling of compute. OpenAI do not seem to be doing as much to improve the scaling of the compute in software terms (they are doing a lot in hardware terms admitedly). They have models at the top of the charts for various benchmarks right now but it feels like a temporary win from chasing those benchmarks outside of the focus of scaling compute.