Mid-Year Thoughts
Word games, natural intelligence, and mind tricks
Hello from Morningside Heights! The past couple weeks have been a bit busy, so I missed a post and am late with this one. I have no good excuse really, but that work really has picked up. It will actually be picking up even more in the coming months, and I can already tell I won’t have the time to keep this newsletter going every week.
But I’m not giving up. Instead, I plan to switch to a slower cadence: a post every two weeks. The lighter schedule will hopefully give me more time to properly flesh out the explainers I plan to write up, and it should enable me to invest effort into some process improvements.
Currently, my system for collecting my notes and converting them into polished blogposts is a little clunky. I’ve identified some inefficiencies I’m hoping to eliminate over the coming weeks. If it all goes well, the goal is to be able to promptly resume my weekly schedule.
🧩 5/18
Since January, I’ve been going almost every month to WordHack — a talk series at the intersection of language and technology that never fails to delight me. This past one was particularly fun. Among other great talks, the creator of the recently released Addagrams, Noah Rosenfield, was there, and let me try out his game pre-release.
I’ve since played it a bunch more — it’s really fun! Can you do better than my score for today’s daily puzzle? (reached by swiping left on the main screen)
Addagrams 🗓️ 06.20.23
🟡 🟡 🟡 🟡 🟡 🟡 🟡 🟡 🟡
99🥇12:50
🌼 5/20
The Margarita is so named because it is Spanish for daisy, after the brandy daisy. The Margherita is named after the queen of Italy, Margherita of Savoy. Both are classics in many a New York bar and restaurant, and both coexist on many a New York menu. I’d like to find out which city venue serves the best duo — a summer side quest, perhaps?
🔟 5/29
Much ado has been made about 4-bit quantization for neural networks, the process of approximating the numbers they contain into more memory- and compute-efficient representations that use four 1’s and 0’s per neuron as opposed to 32 of ‘em. Will we eventually just end up with binary quantization — just a 1 or 0 to represent each neuron? After all, neural networks can all be reduced to decision trees, which doesn’t seem too far from what could be thought of as a binary quantization.
🏙 5/31
In the same spirit as my attempts to deconstruct digital hyperobjects, I’d like to share my friend Jack Knych’s deconstruction of a physical hyperobject; the skyscraper.
Natural flavors tend to “taste better” than artificial ones. That seems to hold across sensory experiences — natural aromas, natural sounds, natural vistas, and natural textiles.
Surely the same goes for natural intelligence: we feel more kinship with a clever crow than with a state-of-the-art language model, even though they have similar numbers of synapses, and the latter is trained solely to mimic human text. Perhaps finding ways to make our models more “natural” could be a decent alignment mechanism.
🏛 6/1
Even though it’s hard sometimes for us all to agree, I do find that there is some hope in the fact that a system as simple as capitalism is seen by many libertarians to be perfectly aligned with humanity’s goals, and a system as simple as communism is seen by many anarchists to be perfectly aligned with humanity’s goals. To me, it shows that it isn’t hard to stumble across simple principles that work well enough to please broad swaths of humanity. The perfect system may yet exist, it may be quite simple, and it may yet be acceptable to a fair share of humanity.
📖 6/4
An oft overlooked object of study in clinical trials is the placebo. I wonder if one could compare different approaches to administering placebos: could different beliefs in what treatment a patient in the control group is given affect the placebo’s performance? There’s evidence that even which doctor is administering the treatment affects how well it will be received, so I wouldn’t be surprised if the backstory accompanying the treatment has an impact too.
Perhaps a good story can have curative powers. If you believe a drug will work more strongly, it’s likelier to work. This might be where we have a lot to learn from shamanic traditions: witch doctors craft medicines that interface well with belief systems. If we learned how to personalize care regimes to be maximally convincing to their recipients — finding exactly the right myth about the treatment’s origin and mechanism of action — we might make our medicine that much more effective.
🔊 6/14
Different audio requires different levels of attention: I can cook while listening to music, but not while listening to a podcast. Additionally, I find that I’m more productive at any given task when I “fill in” my unspent attention just right. I’ll go for a longer run or walk if I’m listening to a podcast. I’ll clean more thoroughly if I’m listening to rock. And while pop music is distracting when I’m writing, a predictable and easy-to-parse lo-fi beat is just right for me to attain perfect focus. I’d love to see an attempt to quantify the level of attention required to parse different types of sound, to help make tailoring playlists to different tasks a more precise science.
Another useful byproduct of such an analysis could be a better understanding of what audio can reliably put people to sleep.
🐙 6/17
Octopuses are some of the most intelligent creatures on Earth, with evidence that they build entire underwater cities. In the spirit of learning to leverage natural intelligence, if bees can be trained to sniff out explosives, what if we were to look into what the humble octopus could be trained to do?
At risk of veering into murky ethical waters (though what is darker than our continued practice of eating octopuses whole?), it’s worth noting that a typical female octopus will at the end of her roughly three-year lifespan produce 10,000s of eggs, which then take another few months to mature. That means that selectively breeding octopuses to favor intelligence could be done quite effectively, with many candidates to select from at each generation and the ability to grow three generations to maturity in less than a decade. I doubt this can outstrip the pace of progress in artificial intelligence, but it might be the fastest pace of progress available in the natural world. If we do end up pausing AI research, maybe we could re-assign all those eager data scientists to research how to improve OI — Octopus Intelligence.
And with that, see you all on Sunday, July 2nd!