Attended Yejin Choi's lecture yesterday on "Bending Scaling laws with Brighter Algorithms"
and the 3 main drivers for this are
Unconventional Data
Unconventional Algorithms
Unconventional Collaboration
Some of the highlights of the talk (which I took notes) are below
https://t.co/aOrncJ5zO9
Just watched this talk that Dario gave about AI Safety in CMU (Cylab seminar maybe?) and its really interesting to know all of today's buzzwords like Reward hacking was discussed at a random seminar in CMU
...something i have been thinking about a lot lately honestly just built it for myself but it is on the chrome web store now if you want to try it: https://t.co/s1b6iMXPEF
got tired of losing my place in long articles/docs. made a little chrome extension that lets you dog-ear webpages like a book. shift + drag under a line to mark it. partly inspired by @geoffreylitt's idea of user autonomy in personal computing/apps (see Malleable Software)...
8. Run security audits periodically.
Trust me this is more important especially with the new supply chain attacks. Also, your auditor would appreciate it
We at Circle (https://t.co/FcG1CE7aIU) recently got our SOC2 Type 2 certificate (https://t.co/JaiVADRC3g)
A few practical lessons for teams that want a smoother audit period: 🧵
7. Review access periodically.
Know who has access to your infra, why they have it, and whether they still need it. Stale access is one of the easiest risks to avoid.
2 awesome lines, I found and will definitely reuse are
1. AI is to biology as calculus is to physics (this is a very bold claim imo)
2. Superintelligence is models optimizing subjective functions instead of objective functions.
Walked 17 miles through San Francisco with 15 friends on the Double Cross Trail.
Recommended!
Even if you know San Francisco well, you’ll see stuff you’ve never seen before. Make sure to do some side quests on the trail, lots of fun places to stop.