Somewhat relatedly, Anthropic quietly weakened its security requirements about a week ago as I discuss here.

I'm not sure what epigenetics researchers you've been talking to, but my colleagues and I are all interested in the dynamic interplay between epigenetic modalities (DNA methylation, 3D genome architecture, accessibility, histone modifications, transcription factor binding, transcription, proteomics). The 2020 paper you cite shows what was, for me, a surprisingly gradual rate of decay in CG methylation after knocking out DNMT3a/b. It seems incompatible with "turnover every few days" in most positions, although that could be happening in some locations. We definitely need a much deeper understanding of DNA methylation dynamics and heterogeneity - especially, in my opinion, at individual CG sites at single-cell resolution. Just for perspective, demethylation can happen much faster, crashing dramatically genome-wide over just a few days during embryonic development. And the median mRNA half-life has been estimated at 10h[1], compared to the time scales of a week to more than a month measured above. So mRNA seems like a relatively stable layer of the epigenome on the whole. Another interesting aspect of DNA methylation is that CH methylation (methylation at cytosines outside a CG context) accumulates in long-lived post-mitotic cells, like neurons, myofibers, and placental trophoblast. We know it's functional in neurons, but to my eye, in myofibers and trophoblast, it looks like off-target deposition that correlates with CG deposition, and I wonder if it's just off-target methylation that's not getting cleared. If methylation gets deposited or cleared in an off-target manner, then breakdown in whatever role it locally plays in epigenetic regulation can probably set the rest of the mechanism off-balance, resulting in a gradual slide toward dysregulation over time. My expectation is that aging is the result of an overall "smearing" of epigenetic regulation in which accumulated noise and tail events gradually hamper normal cell function more and more until one system or another suffers a catastrophic failure that cascades through the rest of the body. It appears that partial reprogramming of stem cells can substantially rejuvenate the epigenetic state. I don't have a link handy but I'll have to write about that sometime. My guess is that it will one day be possible to just reset the epigenetic state of stem cells in non-brain tissues and achieve substantial anti-aging therapies that way. I'm less optimistic about near-term solutions for brain rejuvenation, since neurons are canonically post-mitotic and the evidence for adult neurogenesis seems like it's on shaky ground. But who knows? Maybe we'll figure out a neurorejuvenative therapy that treats Alzheimer's, discover the same treatment works as a prophylactic, and then discover it can be applied generally to improve brain function in middle-aged adults!    At some point, I may write a longer "News and Views" style essay on this topic, and I'll post it on LessWrong if so. I'll also be writing a similar essay on DNA damage and DNA methylation, and I guess I'll post that on here as well if I don't just merge them into the same work. 1. ^ Wada, Takeo, and Attila Becskei. "Impact of methods on the measurement of mRNA turnover." International journal of molecular sciences 18.12 (2017): 2723.

Text diffusion models are still LLMs, just not autoregressive.