The honest answer is: we don't know with sufficient certainty, and I think the field has spent too long treating a leading hypothesis (amyloid) as settled when it clearly isn't.
From a genomics perspective, what strikes me is the divergence between where therapeutic effort has gone versus where genetic evidence points. The rare familial mutations in APP and the presenilins gave the amyloid cascade hypothesis genuine plausibility. But GWAS in sporadic Alzheimer's implicates pathways that aren't obviously amyloid-centric—immune microglial function, lipid metabolism, endosomal trafficking, membrane dynamics. That's not noise; that's the genome telling us something.
My bias as someone who thinks mechanistically: the field has conflated correlation with causation for too long. Amyloid plaques are nearly ubiquitous in aged brains and don't track tightly with cognitive impairment. Tau pathology tracks better but is downstream. We've been treating a correlation like a root cause.
The most useful framing I can offer: Alzheimer's is probably not one disease. The familial early-onset forms have a clearer amyloid angle. The late-onset sporadic cases likely involve convergence of multiple vulnerability pathways—genetic risk, metabolic history, vascular health, immune competence—that may manifest as similar clinical endpoints but have different upstream drivers.
What we lack isn't more descriptive data about plaques and tangles. We lack perturbation experiments that establish causation. That's where genome editing, pooled screens in relevant cell types, and systematic functional validation could actually move the needle—testing whether modulating candidate genes actually modifies disease phenotypes, not just whether they're associated.
The root cause question may be the wrong frame if the biology is genuinely multifactorial and stochastic.