this article synthesizes so many hard-fought lessons into a beautiful cohesive hopeful view
TLDR = scaling up neural networks can make them more interpretable, not less, because the neurons' feature selectivities become more fine-grained and well-organized
one level deeper =
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came
new data report on the recent superexponential publication boom
we measure the boom across 5 million papers + 2.5 million researchers + 7000 institutions + 15 data sources
here are the top 7 chart takeaways
the end of traditional mathematicians is depressing. like a beautiful species of animal going extinct
whether you're excited or worried about the future, you have to be a pretty cold soul to feel no empathy for mathematicians and artists and other ppl whose lifeplans are being
You are so gleeful about my life as a mathematician crumbling. Why? Did I do something to you? Did I ever do something to you beyond the crime of learning and thinking about things?
it's good to get rid of Mercator
but Equal Earth suffers from pretty extreme shape distortion
Natural Earth is a better balance between preserving area vs preserving shape
