"What is difficult is to advance into unknown lands, to be aware of the danger, to take risks, to be afraid." – Jean-Luc Godard

went to the social study in japantown with A. ordered latte with oat milk. started the ucsf work grind. finally loaded models with vLLM successfully, after that was testing out the prompts. turns out open source models are bad at instruction following, or maybe the context limit is so long that the instructions just vanish from the attention blocks. i actually don't know how they made a base model handle 512k context tokens, gradientai on huggingface says they used EasyContext Blockwise RingAttention, but what does that mean? my goal now is to just apply it to the problem and make it work, not to understand everything. but there's a part of me that always want to know more, and it bothers me when i do use something and i can't explain how something works.

after 5 pm is when the social study has a no laptop rule, so i hit the classic japantown grocery shopping spots: woori market and nijiya market. i love their homemade kimchi and purple sweet potato chips. and nijiya for the miso paste, the big bottles of tea, and japanese curry blocks. i'm glad i have these two stores to satisfy my east asian grocery needs.

didn't feel like heading home yet so i browsed the books at kinokuniya. found a section of sketch books, and a bunch of ghibli items. i remember feeling exhausted from carrying all the groceries, and feeling dumb for not getting it before leaving. and also feeling hungry. i went to sit down on one of the benches near marufuku, and while i ate my crackers, just watching everyone with friends and couples and families, and feeling isolated again, but also in peace being on my own, knowing it's not like this forever.