I think you could argue that the scaling hypothesis of AI will never be fully falsified, because you can always argue that you just did it wrong, and that you could always do better and potentially actually make the next scaling step work. Like maybe you screwed up auxiliary hypotheses such as not doing enough normalization destroying your representations, or not having good enough data, not tweaking with RL not giving good rewards or having unstable training, not tweaking architecture properly and destroying your gradients, having bad learning rate schedule, having bad measurement metrics, etc. And I think the same applied to the neurosymbolic AI hypothesis. It doesnt work much yet because noone did it right yet. And didnt scale it right yet. Real scaling, and real neurosymbolic AI, has never been tried, yet! And some of mechanistic interpretability maybe too... Maybe sometimes you just need tweak the certain concrete interpretability tool (its architecture details, details of the loss function,...) endlessly, until it somehow works... And some for some paradigms and tools in science of AI more broadly. Tweaking it (endlessly), until you eventually get some degree of the wanted empirical predictive power or/and explanatory power. It's a special case of Duhem–Quine thesis against Popper's naive form of falsificationism: [Duhem–Quine thesis - Wikipedia](https://en.wikipedia.org/wiki/Duhem%E2%80%93Quine_thesis) [Underdetermination of Scientific Theory (Stanford Encyclopedia of Philosophy)](https://plato.stanford.edu/entries/scientific-underdetermination/) "Duhem–Quine thesis says that unambiguous falsifications of a scientific hypothesis are impossible, because an empirical test of the hypothesis requires one or more background assumptions. Rather than disproving the main hypothesis, the blame can be placed on one of the background beliefs or "auxiliary" hypotheses." "The Duhem–Quine thesis argues that no scientific hypothesis is by itself capable of making predictions. Instead, deriving predictions from the hypothesis typically requires background assumptions that several other hypotheses are correct — that an experiment works as predicted, or that previous scientific theory is accurate." "For instance, as evidence against the idea that the Earth is in motion, some people objected that birds did not get thrown off into the sky whenever they let go of a tree branch. Later theories of physics and astronomy, such as classical and relativistic mechanics could account for such observations without positing a fixed Earth, and in due course they replaced the static-Earth auxiliary hypotheses and initial conditions."