Ben Mann Monthly Mar 2021: bread, vaccine, fake burgers, automatic garden
Purpose
An index for my memory
A 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
Experiments and experiences
Garmin 745
Upgraded to latest watch version, which has a pulse oximeter built in and automated workout suggestions. I initially got it to track whether my blood oxygen was dropping during sleep, but because of the nasal dialators I've been doing much better. The surpising thing was how good the workout suggestions are. My watch coaches me through interval training, easy runs, and longer workouts. It's easy to fall into a routine that works and forget the joy of variety.
Bread experiments
I've continued iterating on minimal-kneading bread recipes and think I'm getting close to the ideal. Thick, brown, crunchy, flavorful crust; big, chewy, glossy bubbles! 100% hydration means equal parts water and flour by weight. Mix 700g each AP flour and water, 1/2 tsp instant yeast, 1 tablespoon sea salt. Rise in a covered pot for 2h. Every now and then, pull the edge over the top and tuck it in using a rice paddle, the goal being to organize the gluten into long sheets. Optional indefinite refrigeration. Shape the dough, rise 2h. Preheat oven & dutch oven to 500F for 1h. Bake 20 minutes with lid on, 20 minutes with lid off, turn off oven, 20 minutes with oven door cracked, cool 10 minutes, eat whole loaf.
Burger experiments
At Costco we bought Beyond Meat burgers, chipotle black bean burgers, and all the fixings. The fake stuff definitely isn't as good as the real stuff yet, but I'm hopeful! If you don't think about what it's supposed to taste like, it does taste good. I made the buns from scratch for the first shot, but it was way too much work.
Garden automation
Since we'll be away from our apartment for 10 days, we had to find a way to keep the plants alive. Diana and I set up a siphon that connects a water-filled 66L storage container to a low pressure timer, which in turn feeds 25 drip irrigators. We tested it over the last few days and calculated the flow rate at ~5L/day, which should be just about perfect. The toughest part was getting similar flow from all the drippers, which we accomplished by shutting them all off, then turning them back on one by one from the end.
Emotion rating {-3, 3}: mean 1, std 1
Highs: ate 16 donuts in 2 days with Diana, nighttime breathing rapidly improving, vaccine rollout accelerating, started using pool now that it's open, watching Diana grow as a manager (2 reports!)
Lows: ate 16 donuts in 2 days with Diana, need more sleep, therapist got promoted to director so stopping clinical practice
Life updates
Turbinate reduction surgery round 2 😤✅
Started reading Good Reasons for Bad Feelings 😢👍
Scheduled my first shot for Apr 20 since Kaiser opened up to people with BMI > 25 💉
Spending a lot of time and energy hiring 🧑💻
Doggie coming in June 🐶
Content
5 point Likert ratings for “I would recommend this content to a friend”, sorted
Shoplifters 4/5
A tender look at what it means to be a family and what defines family bonds
Excellent cinematography, writing, acting. Real, subtle characters with complex motives
[slight spoiler] Unexpectedly tragic ending, where each character fell into the trap of going by what they were rationally convinced was right, ignoring the truth of their lived experiences
Ezra Klein 80k hours podcast 4/5 for EAs
A little repetitive, but a fascinating perspective from someone who both loves EA and is deep in the journalism world
Surprised about how much he cares about animal suffering, though this feels like an appropriate case of seeing the dark world
Biggest critiques of EA:
You have to meet people where they live if you want to be effective
Emotion is a very important signal we shouldn't ignore
Sometimes the things that are the hardest to measure are the most important.
Assigning probabilities to things with high variance may be doing more harm than good.
Brian Christian 80k hours podcast 4/5
I'd already read the book, but the podcast was both a nice review of the book and, maybe more importantly, a more opinionated take from the author.
Brian Christian says that of all the experts he's interviewed, his beliefs are closest to Dario Amodei (my boss!). Dario's predictions keep coming true.
"EA is about bringing optimisation to ethics, and AI safety is about bringing ethics to optimisation"
The interview felt a little too spoon-fed for my tastes ("Yes! That's right!" every other turn), but content was still good.
"I look at the GPT API as kind of the first AGI product" - most people mean superhuman AGI when they say AGI, but it's a fair point that GPT-3 is indeed very general.
The case for aligning narrowly superhuman models 4/5
This post is spot-on for how I think about tackling the alignment problem
Given the concreteness and tractability of the problems presented, I'm feeling very hopeful that we'll be able to solve it
I love the idea of a "mini treacherous turn": have models turn on you in the lab with low impact so you can figure out how to make it not happen at scale
"Sandwiching" means making the model an intermediary for low-skilled and high-skilled humans. Seems likely to work well.
Strongly agree that good SWEs could reskill in 6-12 months and be useful to this task.
Applying human feedback to models is likely neglected and will continue to be so since it requires the seamless integration of research and engineering, which is relatively rare in both academia and industry. It is slowly becoming more common.
I tried to read Antifragile a few years ago, and Scott perfectly sums up why I couldn't get through even the first chapter: it's classic pop-sci, cherrypicked storytelling to support a forgone conclusion.
However there might be some hidden gems, like the difference between small companies and big companies, scientists and engineers, philosophers and mafiosos.
Toward A Bayesian Theory Of Willpower 2/5
Maybe subprocesses in the brain use dopamine to bid on their suggested versions of reality
By default, the "do nothing" agent always wins
Overwhelming evidence that you need to do something forces one of your subprocesses to make a high bid?
Felt a little too hand-wavey to me. If this is true, shouldn't it give us more concrete testable hypotheses?
A mix of normies, gangsters, and weirdos show their true colors as they fight to get a bag of cash.
I liked the nonlinear storytelling, but it felt like the film was trying to be like a funny, clever Knives Out, but didn't quite succeed.
Midnight Sky 1/5
Waste of good special effects, hopeless vision of a future in which we've already screwed ourselves and the planet, predictable cliched plot, no real message.