Ben Mann Monthly July: Ladakh, da Vinci, Doomsday Machine, 80k Hours, work
Photos
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
Delhi and Ladakh
Traveled to India for 10 days to spend time with Phoebe. Got food poisoning only once. Spent most of the time at 3500m with a short jaunt to 5359m. Blog post forthcoming.
Story time with Dad
Based on my prompt "what advice would you give your 30 year old self" he wrote a letter. We went over it together and he told me all sorts of stories from his past that I'd never heard before. Some of them were pretty foundational to what certain hobbies mean to him now, or how he thought about achieving self actualization. A great discussion! Hope to have more like it soon, and to make an intentional space for "story time" more often.
Someone decided to set up ~12 pianos in the botanical gardens inside Golden Gate Park. There was a great mix of professional and amateurs playing all sorts of music, including Saint-Saens' Carnival of the Animals with a 6 piece orchestra and a live marionette show. On the other end of the spectrum some dude partly butchered To Zanarkand. I appreciated the effort anyway! Dan and I wandered around without looking at the schedule for a few hours stopping here and there.
Life updates
Started at OpenAI! Having a blast working on the language team
Semiconductor deal finally fell through!
Idea for a video chat startup, ping me if you’re interested
Content
5 point Likert ratings for “I would recommend this content to a friend”, sorted
David Denkenberger 80k Hours 4/5
If there were nuclear winter, big asteroid, pandemic, etc that broke the current world food production system, we need alternatives ready to go
Global food stores would last a few months
Most calories come from grain
We could break down wood and leaves in big bioreactors
We can feed livestock on wood treated with mushrooms
In nuclear winter, it'd be too cold to grow things outside unless you're in the tropics
You could turn biofuel reactors into sugar factories pretty easily. Need to find a way to make the extracted sugar human digestible
Many potential strategies rely on animals eating human-inendible intermediates, but animals don't scale very fast (except insects, rats, and chickens)
Indoor photosynthesis farms are extremely inefficient and wouldn't work
Maybe we should add urgency when evaluating cause areas
You can try altering Allfed's assumptions using a tool called Guesstimates
According to estimates, alternative food sources are better $/life than AMF and comparable to AI 😮. This is probably due to neglectedness and low hanging fruit.
China specialist is too broad a career. Pick an area like economics, military, etc.
To understand why AI is hard to talk about, sometimes useful to replace "AI" with "Advanced Statistics"
Nuclear weapons are a bad analogy. AI is more like electricity. It takes a lot of infrastructure to actually make use of it
Data is not the oil of AI. Need to think about who will use what data how. Seems unlikely that China has a particular advantage when you think of it that way
Wishes Malicious Uses of AI paper focused more on surveillance uses
Most forecasts should be in the 40-60 range to be epistemically honest
Sustainable, rights abiding economic growth trumps almost everything in the long run. X-risk is important and included in "sustainable" but believes it's effectively discussed elsewhere
Strange belief that we'll be stuck on Earth forever and or kill ourselves within 1000 years. Not bullish on self reproducing probes
People will always eat meat?
Everyone should quit alcohol
Disagrees with "we have Paleolithic emotions, medieval institutions, and godlike technology". Points at countries like Singapore as example of good institution.
More people should make more risky bets
Happiness surveys can't ask the dead people. People in Kenya don't know what good healthcare even looks like.
Predicting things outside the 5-95% range is super hard
You can check for temporal scope sensitivity: as events get nearer probabilities should change
Check for predicted indicators that we're on such a historical trajectory
Extemizing is only useful if all your predictors have diverse information and backgrounds. If they chat too much ahead of time, better to take the average or something
One superforecaster might be worth 10-35 normal people
It's hard to beat reference classes if you don't have a lot more information, and in fact it sometimes hurts to be a true expert
It's easiest to improve with clear, rapid feedback, just like with any learning
Training may not transfer across domains, so focus on what's relevant to you
Startup idea: create a pundit forecast watchdog group that would publish stats on implicit forecasts
An inside look at the world’s nuclear war machinery from the 60s to the 90s
Large numbers of people could have set off nuclear war for decades. The president delegated lots of authority AND officers felt free to disobey direct orders.
We had only one plan to kill hundreds of millions of people if any war broke out at all between the US and Russia
H-bombs have never been deployed, but they’re at least 1000X bigger than the bombs dropped on Japan
Seems likely that a single nuclear deployment will cause large-scale retaliation and catastrophe
Characterized by intense curiosity, obsessive perfectionism, beginner mind, extensive writing/documenting and observation: "TODO: describe the tongue of the woodpecker"
Made massive breakthroughs in many fields with no background purely by methodical and extensive observation
Dragged a bit on the intricate descriptions of the paintings, but glad I finished it
First and last chapters were the best by far. Might reread last chapter periodically
White Christmas (Black Mirror) 3/5
Probably infeasible technical premise: we can create a high fidelity simulation of someone with a minimally invasive procedure over two weeks. We can torture your simulation to get it to eg serve you or confess to a crime. Seems like future law would protect these high fidelity simulations under “mind crime."
First Man 3/5
Slow paced, thoughtful film about Neil Armstrong’s role in the space race juxtaposed with his struggles as a father
Emotional exploration of what we sacrifice and why when we take big risks on behalf of humanity
There was a lot of personal danger then unlike important long term risks now like AGI or nuclear nonproliferation
Civic honesty around the globe 3/5 (notes from Phoebe Yao)
"Psychological models based on self-image maintenance predict that people will cheat for profit so long as their behavior does not require them to negatively update their self-concept"
People don't like to complete actions that unavoidably conflict with their models of themselves, because then they would have to update their self-definitions. Being a "thief" is not a characteristic people would like to admit to themselves
When doing something bad, do it only a little bad at a time to reduce mental friction. When manipulating someone else, frame the option you wouldn't want them to pick as extreme or ridiculous so they won't want to associate with it.
Light on results, heavy on implementation
Got a 1500 probe array implanted in a monkey capable of read/write, paper coming soon
First product will be attached to motor and sensory cortex focused on device (phone, keyboard, mouse) control
Second product could hit visual cortex
Made a robot sewing machine that implants the probes
Wireless comms/power from ear-mounted bluetooth device
Cool special effects and world building, but very flat characters and trite plot