Ben Mann Monthly May:
Purpose
An index for my memory
An menu of topics for my next conversation with you
A faster way to share what I’m excited about without the barrier of writing a complete blog entry on it
A skimmable way to spread content I found valuable
If there’s anything you want to see more or less of, please let me know!
Experiments and experiences
Authentic relating birthday party
Three rounds of three person discussions. Discussed topics like “how would you change your life if you knew you’d die tomorrow?” and “if you could ask anything of your friends and you knew they’d say yes, what would you ask for?” I think it could’ve been improved by mixing verbal and nonverbal games since it can get tiring to be so language oriented for extended periods. The highlight for me was the 10-hand massage afterwards. Everyone learned techniques from everyone else.
Peak Experience Replay
A new technique Nick Cammarata and I are experimenting with: create a spaced repetition deck of behaviors you’d like to reinforce. Each card has a single phrase that reminds you of the thing, for example that time you started a flash mob in a park, or left a job you didn’t like. When you see the card, remember the mind state that let you achieve the thing. Neither of us stuck with it, but we both found it a really fun exercise just creating the decks, especially with friends.
Lots of discussion around alignment and EA
Despite the profusion of writing on the topic, it seems the community is rather divided on the problems and most fruitful avenues of attack. I attribute this to the uncertainty in the path we’ll ultimately take to AGI. At least China appears to care now. The best I can do is put myself in a position to help as we get closer.
LA and Joshua Tree
Flew from Oakland to Burbank on a jet using Blackbird (referral code DSH7B). On the way back we got to the airport 10 minutes before liftoff with no issues, even though we checked our bags. We drove to Joshua Tree for some no reservation camping at Ryan Campground, hiked, thought about life. I took a swim in the pacific off Venice Beach. We heard stories from our friend's dad, an abstract visual artist who hung out with Kerouac in the 60s. He said poets are wild because they’re constantly seeking intense experiences to inspire their art. For visual artists, they’re more humble, the medium is expensive, and they’re more about doing the work.
MAPS guides
Working my way through therapy manuals to level up as a trip sitter and therapist. Practiced with a friend. One of the manuals was written in 1959 and encourages pushing the patient quite a bit. I didn’t see that working very well, despite my dodging my subject’s questions for over 2 hours. Next time I’ll try an even more hands off approach, and will make sure to set boundaries ahead of time on how much information and guidance I’ll give my subject.
Life updates
Braces off, bit a cucumber for the first time in 2 years
Published How to sample from language models
Worked with Yaroslav Bulatov on reimplementing Fluctuation Dissipation Relations. We discovered it’s not practical for most real world settings.
Applying to AI Safety orgs
Content
5 point Likert ratings for “I would recommend this content to a friend”, sorted
Invest in tooling to connect your creative inputs to your outputs. This is Bret’s mission.
Find something to fight for based on your own concrete insight
Working on a single task for a long time might be hard. Find diversity in tasks to achieve your mission.
I haven’t read Abstruse Goose in ~10 years. Still turning out gems!
A Recipe For Training Neural Networks 4/5 for AI practitioners
A great reminder that because everything is opaque in neural net training, it pays to be extremely conscientious. Don’t try to skip any steps or you will pay for it later.
Specific tips I rarely see:
back propagating into your inputs to verify your masking and batching is set up correctly
overfit on a batch of 2 examples
A vulnerable paean to staying true to yourself as you create things in the world
Take more risk, fail, do it for yourself, resist the easy path
If technology makes it easier and easier for humanity to destroy itself, perhaps we should consider giving up some of our civil liberties to avoid catastrophe
Remember that the history of breakthrough depression treatments is littered with failures
Written in 1941, an early conception of time as some kind of bifurcating tree in which all possibilities happen in parallel
Reminded me of my excitement in 2003 hearing Max Tegmark discuss his parallel universe theory at UPenn after reading his article in Sci. Am. I’m shocked that I found the article so easily 16 years later. Borges was 62 years ahead.
Poundstone: “even unlikely events must come to pass eventually. Therefore, anyone who accepts small risks of losing everything will lose everything, sooner or later. The ultimate compound return rate is acutely sensitive to fat tails.” Appears to apply to AI X-risk and civilizational trajectories as a whole!
Bet in proportion to your edge. Refuse to bet if you have no edge.
Some concrete metrics to judge how to allocate time. Treat uncertain tasks as looking for blockers rather than actually prioritizing completion.
A truly mind bending account of man’s hubris: surviving in the Antarctic for 2 almost years before sailing hundreds of miles in a 20 foot boat on the open ocean with the worst gear. Zero deaths.
As I read it I kept thinking about how hopeless they must have felt, and yet it worked out in the end. Perhaps this is what AGI will be like.
A rebuke of metis over modernism. Felt obvious to me, but for those who think science is weak, a worthwhile reminder.