Co-founder & CEO, Entrepreneur First @join_ef http://joinEF.com Co-founder & NED @codefirstgirls Newsletter: http://tib.matthewclifford.com Podcast: http://tib.buzzsprout.com
Truly exceptional book
Right on the borderline between fiction and non-fiction, Benjamin Labatut’s When We Cease to Understand the World was the best book I read. I have no idea what being a scientific genius is like, but this conveys a sense of what it *might* be like [link] [/2]
Or are there no free lunches? What do we gain, what do we give up?
New paper from @MikhailBurtsev, Aydar Bulatov and @yurakuratov
Let's break it down👇
arxiv.org/abs/2304.11062
Great thread on an approach to scaling token windows to >1m tokens. New important ideas emerging every day…
Interesting paper: “Chameleon is capable of synthesizing programs to compose various tools to tackle a broad range of queries. While it does not require any training, Chameleon uses the in-context learning capabilities of LLMs to generate these programs”
This is very true. I’ve probably shared “The Scaling Hypothesis” more than any other post. From 2020, but still essential reading:
Great post by @simonw that points to a lot of near-term AI safety / security challenges
arxiv.org/pdf/2304.02643.pdf
Superb thread on a fascinating paper. Lots and lots of ideas here - especially the third tweet on composable system design for AI
TL;DR: in trailing-edge chips (20-120 nm) chips, China leads or has a good odds of self-sufficiency;
in leading-edge (phones, AI), seems very unlikely they can suceed or go self-sufficient given US export controls.
Good thread and report on China’s semiconductor capabilities with respect to AI. Worth a read👇
Ian is one of the smartest and wisest people I know; this is worth a read👇
Because we trained GPT on code, it means we’ve trained it on billions of individual micro-logical predicate sequences, so it “knows” logic by thinking in code. twitter.com/Mihoda/status/1644564637420249088
Very interesting hypothesis; worth a read and ponder! (though note caveat in final tweet of thread)
Fascinating work: “These models hold the promise to have context lengths of millions… or maybe even a billion!” Also shows huge power of new primitives in ML: even if there were *no* more research, remixing of existing ideas would take us a long way…
Exciting thread👇
Don't miss out curated content from your favorite people
Join to see curated content, recommendations and collections from your favorite people.
Find creators I follow
Good, thoughtful thread from Jordan