These policies also just make for bad science in general. In science you need people challenging the status quo, not everyone, but a significant percentage. A publish or perish paradigm encourages railroading in intellectual thought as well as encourages laziness in evaluation, as there is increased pressure on all researchers who are also the ones performing reviews. People who have zero incentive to perform a good review and many incentives to do a quick review (which is even actively encourage). This results in an over-reliance on benchmarkism, which we see in many fields. In ML we see SOTA chasing, forgetting that datasets are proxies which have their own biases and limitations as well as the same for our evaluation methods. In many of the softer sciences -- such as psychology or medicine -- we see a reliance on p-values, forgetting that 0.05 is an arbitrary value and not meaningful without additional context. All our sciences are converging on Goodhart's "Warning."

There is utility in both speed and in a slow step. The truth is that we need both. A continuum of methods and ideas. After all, we are searching in the dark and our only real advantage is parallel Monte Carlo search. The problem is that this is incredibly difficult to accurately evaluate and we are arrogantly attempting to evaluate exceptionally noisy systems with extreme precision without even having a basic model of that system. We could relate this to businesses too, but the timetables for them are typically shorter and so the chaff can more clearly be gained (often post-explained through pure intent instead of it's combination with luck). Then again, we are trying to make academia a business, trying to apply their noisy successes to a completely different model.