Self-supervised learning on observational scRNAseq especially with naive reconstruction or masking losses will never deliver a causal model of gene regulation. It will learn statistical structure in expression space, not the mechanisms that generate regulatory change. 1/
18 months after posting this tweet, the AI for science commentariat is still proclaiming the death of single-cell scaling laws on the basis of... {checks notes}... a model sweep ranging from 1 million to a whopping 10 million parameters. (but unlike 18 months ago, these proclamations now come wrapped in premium AI-written slop, giving them a glittering verisimilitude of rigor)
left as an exercise for the reader: generalize from this example to a meta-update about how epistemically adversarial the scientific environment we're operating in is (for extra credit, partial out the effects of mood affiliation and status deferral)