Thoughts on Acceleration, (and overdue returns)
Long time no see! Since I last edited this site, two years have passed, I've got into MIT, I've begun work on a physics degree among other things, and I've been in a lot of unusual places.
It's good to see you again. Hopefully if I can shit out some more impressive long form writing, this is where I'll do it.
In the meantime, you'll have to put up with this: I've been thinking about AI. I'm a member of MAIA now, and I've been scarfing down rationalist texts and AGI philosophy for a few months. I've had a lot of fun, and met a lot of interesting people, and I'm certainly more concerned than I was, but I don't know if I'm ready to put down my tools yet. The most convincing argument I heard was from Emilia Javorsky, who told me that it's like nuclear power. We may have a few big fuckups due to lack of safety or overconfidence, and then the whole thing blows up in some tragic Butlerian jihad.
We can't let that happen.
Anyone who's interested in this phenomenal fucking alien lifeform we're making should be interested in interpretability (I think I am pretty sure that interpretability is the first and so far only decent approach at alignment, but I'm open to updates). That is the accelerationist/engineer's view on why interpretability is needed.
For the maximum good from technology.