Bought a used GTX 1080 Ti off someone in a local hardware forum, picked it up in person because I did not trust shipping something that size and that expensive to strangers on the internet. Guy was upgrading to whatever was newest, sold it for less than half retail. 11GB of VRAM instead of 6. Feels obscene. Old 1060 isn't going anywhere — figured out I can run both, one for the main job, one for smaller side experiments so I'm not blocked waiting on a single GPU anymore. Power supply might not love this. We'll see. First thing I did with the extra memory was just... run last year's experiments at 2x the sequence length and batch size, mostly to watch numbers go up, which is not a real experiment but felt good anyway. Read the BERT paper this week too — Google put it out back in October and I'm late to it, but "bidirectional" makes so much sense as an idea once you see it that I'm annoyed I didn't think about masking tokens and predicting them from both directions on my own. That's twice now something's been published that made a thing I was fuzzy about suddenly obvious in retrospect. I don't know if that means I'm noticing the right questions or just that the field is currently answering everything I happen to wonder about. Probably the second one. Field's moving fast.