Ben Mann Monthly Nov 2022
Ben Mann Monthly Nov 2022
ChatGPT, rhinovirus, burnout, Ra
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
Common cold
Due to all the COVID isolation and masking, I haven't had a cold in a few years. This one was much milder than COVID, but since Euda is only 2 months old, I decided to isolate and mask at home for the 10 days my symptoms persisted. It was tough being right there and not being able to help Diana take care of her, but much tougher on her not having the help.
Nice work folks! Makes an effort to be safe much more than previous OAI releases, acknowledges its limitations, and content filter is good for identifying attacks without making it boring to play with.
While the reception has been very positive so far, I'm curious to see how things change as some of the trickier impacts come to light, such as StackOverflow being overrun with answers and temporarily banning them. If this is happening above board, how much is happening covertly? Eg, are people using it for astroturfing already? If not, curious how OAI is defending against that. It's a tough problem, and a reason not to provide API access to any random person. There is no API? Well, someone reverse engineered it already.
If you are impressed by this model, know that we are at the very beginning of the capabilities ramp. People played with GPT-3 a bit, but it didn't impress them much because it was mostly sub-human in capability. ChatGPT is superhuman in many aspects, but the absolute distance between the two is small. Soon we'll have vastly superhuman models. The effects on society are hard to predict, but the pace of change will certainly increase dramatically as the cost in time and money of doing meaningful intellectual work decreases.
In the AI safety community, there have been many theories about how and why superhuman AI will turn into an existential risk for humanity. One of the big questions is whether safety work on today's models will translate to tomorrow's. Regardless of whether it will, I'm glad to have at least one organization pushing that thesis forward empirically as much as possible. If we can generate scaling laws showing the relative effect of our interventions as a function of model capabilities, perhaps we can predict how things will go if those scaling laws hold up. If transfer is impossible, then our task is far harder. I prefer to try the dumb thing first. Hopefully we'll have time between knowing whether it worked and trying the harder thing.
Burnout
I realized I was unknowingly burned out before taking paternity leave, and I feel mostly restored now. I let myself become so absorbed in work that I stopped taking care of myself (diet, exercise, sleep, relationships) and things were going downhill on many axes. Note to self: "put on my own oxygen mask before helping others." Take time for self-care. This is common advice for new parents, but it applies to everyone. Since coming back to work I've been feeling a lot better than when I left, and have a clearer sense for what to prioritize vs drop. Full speed ahead!
Life updates
Started working again, 3 days/week
Started reading Valuable Humans in Transit and Other Stories
Still off SSRIs, feeling good
Content
5 point Likert ratings for “I would recommend this content to a friend”, sorted
Ra 4/5
What if magic were real, followed consistent laws, and you could study it and enable it through science and engineering?
One of the better sci-fi/fantasy mashups I've read. Characters are kind of flat but the world building is too fun for it to matter.
I won't spoil the twist ending, but if you think it starts out fun, just wait until they figure out why magic works.
Three Thousand Years of Longing 3/5
Excellent cinematography, cute self-aware story about "genie grants 3 wishes".
Diana was hoping for a satire, but instead it was a quirky romance/drama/fantasy.
The Woman King 3/5
Liked the strong female characters; passed the Bechdel test
Excellent action sequences
Fact checked, seemed fairly revisionist as usual but 🤷. It's based on the history, not trying to represent it. Should we call it a fantasy?
I care a lot 2/5
Fun characters, good performance from Rosamund Pike
Trying a bit too hard to be woke and feminist (whereas Woman King was woke and feminist, but doing it right).
Pinocchio 2/5
Nothing to write home about. The blue fairy was cool.
Didn't finish all the stories, but this is classic David Foster Wallace in its microscopic attention to detail both internal and external to the narrator. Felt like I got the idea quickly enough. Too depressing for my current interest.
Eh, watch Ted Lasso instead.
