Burning Man
Last week you might have noticed I was slower to respond to text messages then even my normal ‘ah sorry I missed this whats up’ 72 hour SLA. It’s because I, along with a number of other people, were at Burning Man.
I’ve been to the Burn five times. It’s bizarre I can say that. My conception of self doesn’t feel very burning man - in point of fact it feels like a polo shirt - but after my fifth burn I’m beginning to wonder if I’m just slow to update internal narratives.
The first time I went to Burning Man was in 2013. In hindsight that was insane. I had never been to a festival of any kind before and my friend and I had to fly from DC to Reno. We were woefully underprepared. This was still early enough in the Burning Man boom that there wasn’t guides for everything burning man available online, we had to figure stuff out.
But I was so captured by the idea of this crazy magical place; this YT video in particular sparked something in my soul. It was challenging, both logistically and ‘spiritually’ - young Ben *really* wasn’t as high on Big 5 Openness as he wanted to pretend to be.
I joked, especially after going in 2018 and 2019, that I wasn’t sure if I enjoyed Burning Man, but that I had to keep going back until I knew for sure.
Last year and this year I can say, with no hesitation that I enjoyed it tremendously - though, in different ways. Last year felt like the quintessential experience of burning man; adventuring out, meeting new people, seeing cool art and good1 music, co-creating novel experiences. Oh and I officiated my friends wedding at the burn!
This year I helped to lead my camp, Motel California. It really shaped the whole event - in fact the whole summer - for me. It is a not small task getting 35 people to the desert, building a three story structure, and hosting events for hundreds of people.
I’ve come away with a far deeper appreciation for people who routinely give the ‘gift’ of organization and coordination, and how gratifying it can be. Along with an appreciation of the costs; not just the time you’ve got to devote, but the way in which some part of yourself becomes entangled in the group. It’s insights I’m still mulling on.
Even after five of these things I’m still trying to put my finger on what exactly I like about Burning Man. The closest I can get is it’s a high trust environment where the shared context of immediacy and inclusion fosters beautiful, surreal, and deep, interactions.
*clinks glass* to checking out when you want but never leaving.
AI
Steve Omohundro and I wrote and published an overview and roadmap for Provably Safe AI. Provably Safe AI is part of the portfolio of research agendas that falls under Guaranteed Safe AI; it focuses on securing critical infrastructure to reduce the attack surface area from bad actors using AI, and unaligned rogue AIs. I’m bullish on Provably Safe AI; under lots of scenarios we’ll want secure cyberphysical systems, and advances in AI should make it easier to rebuild systems from the bottom up with strong guarantees.
To that end, Zac Hatfield Dodds and I have made a bet, my 1k to his 10k, that by 2027 there will be an unpickable lock based on Guaranteed Safe AI principles. I agree with, as he put it, that the project is a “beautifully compact crux for the overall approach”. Quinn Dougherty in the GS AI newsletter had a nice writeup of what it might take:
Physics to the relevant granularity (question: can human lockpicks leverage sub-newtownian issues?) is conceptually placed into type theory or some calculus. I tried a riemann integral in coq once (way once), so it occurs to me that you need to decide if you want just the functional models (perhaps without computation / with proof irrelevance) in your proof stack or if you want the actual numerical analysis support in there as well.
Good tooling, library support, etc. around that conceptual work (call it mechlib) to provide mechanical engineering primitives
A lock designing toolkit, depending on mechlib, is developed
Someone (e.g. a large language model) is really good at programming in the lock designing toolkit. They come up with a spec L.
You state the problem “forall t : trajectories through our physics simulation, if L(t) == open(L) then t == key(L)”
Then you get to write a nasty gazillion line Lean proof
Manufacture a lock (did I mention that the design toolkit has links to actual manufacturing stacks?)
Bring a bunch to DefCon 2027 and send another to the lockpicking lawyer
Everyone fails. Except Ben and the army of postdocs that $9,999 can buy.
FWIW I too do not think it’s likely that this will happen before 2027. Maybe 20%. Autoformalization is getting very good, very fast, and I expect we’ll be able to make simulations of mechanical designs that we couple with AI powered proof generation, but still, it’s a huge lift.
Note: The comments of the Provably Safe post are worth reading for criticisms of the agenda, as well as this post.
Paper which shows big productivity gains for programmers who use LLMs:
I’ve taken Karpathy’s suggestion and started experimenting with Cursor, a VS Code port that integrates AI.
Peter Wildeford’s predictions of AI capabilities:
From Joe Carlsmith: what does it take to solve the alignment problem? Joe was also on the Dwarkesh Podcast last month!
Sam Bowman wrote and shared his ‘Checklist: What Succeeding at AI Safety Will Involve’ which is a nice summary of what I expect frontier lab’s to be focused on from a safety perspective.
#content
Forgetting Sarah Marshall: I rewatched this with a friend and I’m happy to report that Judd Apatow does hold up 15 years later.
Madmen as told through Tiktok: I’ve got to confess that I’ve gotten addicted to Tiktok, the algorithm is sending me 2 min clips of the entire six seasons of Mad Men, and I love it. If I’m still doing this next month, send help.
Music: Reflecting at Burning Man, I’ve come to appreciate that I love:
head bangable rock and roll / EDM.
Pop songs. Bonus if a saxophonist randomly interrupts (maybe burn specific).
classic ‘90s ‘00s hip-hop.
Modern disco.
Suggestions welcome.
xoxo,
Ben
I like music with words, and that’s in short supply at Burning Man.