Georeactor Blog

While I've been checking in with family for the past week, I have some papers, links, and a book review which would normally get topic-specific write-ups, but these haven't gotten organized. Going to mix a bunch of these together.

Bio + LLMs

I want to hype up PlantCaduceus - a bidirectional Mamba-based model from Cornell's Kuleshov Group - and PAIR, a paper from U. Toronto and the Vector Institute on predicting text annotations on UniProt from sequences. I've been really offbase in my ideas for this and want to learn from / finetune / benchmark these sometime.

ceLLama got some traction on Hacker News, but on closer inspection it uses R scripts to analyze cells, and any generalized LLM as the backend to write reports. It's prompt and tool-use but not a finetuned bio LLM.

Genome details for giant kelp, which is not a plant ; also see recent sequencing for apes including both orangutan populations.

French researchers (CIRAD), the sugarcane genome, and general comments about other tropical agriculture genomes: https://phys.org/news/2024-07-genome-diversity-major-crops-story.html

Camelid and shark immune system data: https://www.biorxiv.org/content/10.1101/2024.07.19.604232v1

Book Review: Dear Wendy (Ann Zhao, 2024)

A college / YA novel with a premise of two people becoming friends IRL while unknowingly being rival advice accounts on Instagram. The book explores asexuality and Wellesley culture (the accounts 'Wendy' and 'Wanda' are in-jokes based on campus archetypes). Sophie aka Wendy has some hangups around immigrant family expectations (possibly channeling the author). Jo aka Wanda has two moms and ought to be at home on a progressive campus, but is uncomfortable with the rarity of asexuality and ubiquity of dating culture. Together they start a club to further expound on a-spec issues and become besties. An antidote to the bros-on-campus of The Social Network.

Politics

From https://en.wikipedia.org/wiki/Uncodified_constitution , the four countries with partially-codified constitutions are Canada, China, New Zealand, and the UK. As a basic US-ian I once thought that the Magna Carta = the UK constitution, but nope, and the newer Anglosphere countries have followed suit (except for Australia). On Twitter I've seen a few situations called 'constitutional crises' in the UK, but Wikipedia covers only a few, mainly situations where the King was too ill to open Parliament and approve its bills.
Why do these countries avoid constitutions when they'd support one in any foreign policy elsewhere in the world?
This came out of an even more vague political discussion on Threads. An Australian was commenting that American expats would complain about any number of topical political issues (in this case, Chevron deference) without realizing that Australia had the same rules. This is common and I don't know much about any other countries' laws, so I was thinking to learn something about the other English-speaking countries at least.

Extra

I've read and passed along to people that Japanese curry has Portuguese roots, but apparently this is wrong and there's a British navy connection: https://en.wikipedia.org/wiki/Japanese_curry

Google Gemma is releasing a separate 'ShieldGemma' project which reminds me of Meta's 'LlamaGuard' releases. https://developers.googleblog.com/en/smaller-safer-more-transparent-advancing-responsible-ai-with-gemma/

QLoRA / UnSloth type work in PyTorch's own codebase https://pytorch.org/blog/quantization-aware-training/

Great to see Swift embrace and blog about homomorphic encryption: https://www.swift.org/blog/announcing-swift-homomorphic-encryption/

SQLite vector support has some useful news about vector DBs and methods such as quantization https://alexgarcia.xyz/blog/2024/sqlite-vec-stable-release/index.html