about 1000 new ones this year. (almost all got exactly one listen besides the following)
"Lips" by Daniel Lentz [choral lo-fi]
Cheating, because it's a career retrospective. Also cheating because he's just better than the hippies, punks, post-punks, and Thurston clones he has outrun for 50 years
New paper: a big 90-page intro to AI and its likely effects from ten perspectives, ten camps.
The whole gamut: ML, scientific applications, social applications, access, safety and alignment, economics, AI ethics, governance, and classical philosophy of life.
1/18
We inherited the framing (“Ten Hard Problems”) from Eric Schmidt and James Manyika. They conditionalise on success: "if it's 2050 and everything went well, what did we have to solve for that to happen?"
Intended audience: technical people without any ML experience.
We spent a chunk of 2022 and 2023 reviewing 1347 papers and talking to 30 experts.
**Importantly, we don’t really cover most of last year’s enormous progress in capabilities and governance.**
(Lesson: moving too quickly for a journal paper to be a very sensible exercise.)
On steps-to-reach-loss, FLOPs-to-reach-benchmark, inference speed, etc, the mixture-of-experts paradigm gets between 2x~5x improvement over a dense baseline. Fundamentally, MoE is unfriendly to low vRAM developers, which describes the vast majority of OSS 152334h.github.io
Foundtrack was an amazing blog that used to make mp3 playlists. It had a big effect on me. Some cuts are so deep that they're not on Spotify, Youtube, Soundcloud, Bandcamp...