2025-04-14 16:50:37
Here's a way I think about conversations sometimes.
Imagine you and a friend are building towers into the sky. Each block is a comment or insight that you add to the conversation.
At its best, conversation lets you collectively get to go up to places you would never have reached otherwise – the other person adds a block that you wouldn't have found, which gets you to a position to discover a block that they wouldn't have found, and so on into towering heights that neither of you could have dreamed of.
There’s a few different ways that someone can be a great conversationalist:
You see, conversations aren't just one tower upward into the sky – we're all clambering over these massive structures of thoughts/comments/ideas/possibilities that we've picked up through our previous thought/reading/conversations with others, and often a new conversation intersects with parts of those previous structures, and you can use those as stepping stones as you get to somewhere new.
The problem with this model is.... you're not really climbing together. (Although in the best cases it feels like it). You’re just describing blocks to each other, and scrambling up your own blocks on other sides of the impenetrable wall between human minds, with your ghostly image of where you think the other person might be to accompany you.
We're each alone with our own beetles in our own boxes; the tragedy of conversational intercourse is the perpetual virginity of the soul. We're describing our blocks to each other, and hearing other people describe blocks we might clamber onto, and desperately hoping that the other person is truly seeing (and climbing) the same blocks we are.
This has a few consequences.
There's a kind of meta-conversational shared context that is incredibly helpful for having good conversations. I'm honestly not sure what it consists in: is it more about having a lot of shared "basic" blocks already, so you can quickly get to high and interesting places and continue on from there? (You could imagine this as a shared cannon that everyone can safely reference, or a shared background that gives everyone the same basic concepts to work with).
Or is it about a shared language for describing new blocks? It feels like some people are able to more-easily describe new blocks to me in a way that I can then climb up to.
Or maybe it's about error-correction and disambiguating shared state. At the end of the day, there's no way to know if you're standing on the same block of not, so it's incredibly helpful if you either 1) freakishly and wordlessly understand each other as you make great leaps up the tower, or 2) be good at clarifying which blocks you're currently standing on. Otherwise you can be arguing about which blocks you should jump to or even which blocks exist, and it's actually because you were standing in different places without even realizing.
2025-04-11 17:54:36
In computer science there's a famous distinction between solving a task and verifying the solution. (You may know this in the context of P versus NP).
Think about a jigsaw puzzle: it can take hours and hours to put the pieces together, but once it's solved you can tell almost-instantly whether or not the solution is correct.
A simple model I have of current LLMs is that they're most useful for tasks that are slow to solve (for humans) but fast for us to verify.
You can imagine a simple 2x2 with "solving speed" along one axis and "checking speed" along another.
We can start placing tasks on this grid, such as:
I first tried to get the LLM to help me with generating this 2x2, but the thing it made was wrong. I could immediately tell it was wrong (fast check), and I could have spent time prompting the LLM to get it right. But I knew I could also just make the thing in GoogleDraw in under a minute (fast solve), so I did that instead.
By contrast, the jigsaw puzzle analogy that I started this piece with came out of an LLM; I asked it "what's an example of an everyday task that is slow to solve but quick to verify?" Coming up with a good analogy takes time and effort, but deciding whether a proposed analogy is good or not can be near-instant.
I think this is why LLMs are more generally great as "brainstorming partners" – it's slow for humans to come up with ideas, but fast for us to verify their quality.
Coding with LLMs is super fun, because the LLM can generate code 1000x faster than I can, but I can check whether it works pretty immediately (for small bits of code, at least – perhaps this changes with bigger project sizes).
Research is kinda split in two, between domains where you know things and where you don't know things:
Basically, in domains where I don't know anything it would take me a while to find the answers I want, but also I can't really tell if what the LLM gives me has obvious errors in it, so I'm not really convinced that the net outcome is better than 0.
By contrast, in a domain where you know things, maybe it's helpful to have the LLMs research a question for you because you can instantly check whether the result is accurate? (For me this is entirely hypothetical, if you know things about anything feel free to reply in the comments).
This also highlights the differing opinions about LLMs as Essay Writers between two types of purposes:
For many people, the goal of essay writing is "get words on the page that form sentences and look grammatical," and historically this takes humans a reasonable amount of time to do themselves, but they can tell in 10 seconds that an LLM has done it correctly.
For other people, the goal of writing essays is "write something good and thoughtful and insightful." This is slow to do, but also slow to check: you actually have to think about the things, and check if they're true, and this takes time. So using an LLM is no longer an obvious win.
Anyway. I think human brains are kind of easily hacked by neat models and 2x2 matrices, which often feel truthy whether or not they're true. So let me end by mentioning two other axes that can be useful when evaluating how good an LLM is for a task:
I think ideally we would either find 1) some interesting test cases that would generate different predictions between different models, or 2) a bigger model that incorporates one more axis. But alas, my brain is done – if you have energy to continue this model please do.
2025-04-09 16:52:02
Here's my contribution to New York City Civic Discourse: the subway stations should have different names.
For people who don't live here, you maybe asking: what? To start with, here are the stations called "23rd Street"
I understand how you could theoretically consider this to be one MegaStation, but it takes 15 minutes to walk from one end to another. If I'm on my phone-map and trying to find my train , I spend an inordinate amount of time trying to disambiguate: I can't just type in "23rd street subway station" and be done. But that's just Level 1.
Level 2 is the "125th street stations" in Harlem and the "Avenue U's" in Brooklyn. In both cases, the farthest homonymous stations are 30 mins walk from each other; the best way to get between Avenue U's in Brooklyn is to to take a 12 minute bus. If I need transit inside your transit you need to change your transit.
But that's just level 2. Level 3 is this monstrosity:
Look, I understand the benefit of naming your stations after a street, in the sense that it could/should be easy to find the 86th street station by walking down from (say) 70th street till you hit it. But, equally, in a very small village it's fine to be called just Bob, because there is only one Bob; when your village gets bigger you need to start going by Bob The Builder, to distinguish yourself from Bob The Farmer; and in a global interconnected world you eventually want to be Bob The Builder of Brighton Beach, to distinguish yourself from Bob The Builder of Bushwick.
Bob managed to do it, the New York Subway can too.
And here's the thing: if there's one thing politicians LOVE, it's naming rights. Do not think of this renaming as a burden; think of it as a chance to placate and/or honor every possible donor and interest group with a once-in-a-lifetime opportunity. Throw one in for yourself, why not! We can rename every station after its current councilmember, I truly don't care, just so long as I can look up a station name and find exactly 1 hit for it in the whole five boroughs.
That said, it is possible to mess things up even with names. Here is an aggravating example:
"These stations have different names," I hear you cry. You might think so.... but if you're sitting on the train itself, and trying to make sure you don't miss your stop, and you look out your window at the name printed on the columns holding up these stations like Samson at the temple, you will find to your shock that all you can read is "Medgar Evers College." Are you at Franklin Av-Medgar Evers College or President St-Medgar Evers College? You do not know, and you need to know within the next five seconds.
We can do better, and we must do better.
Let me close with this. Some time in my first semester of college, a professor asked me how I was getting along. I said I was holding up, but that it was very difficult to get to classes on time when there were three different buildings that had the same name and none of the buildings had signs on them. The professor looked at me wryly and said (something like) "oh that's on purpose, if you don't know where the buildings are since childhood you're not supposed to be here." This is a vibe that I think New Yorkers would rightly disdain, and yet are effectively emulating. I rest my case: (at least) two out of three 86th street stations delenda est.
2025-04-07 19:32:04
In 2018, the US Supreme Court legalized sports betting (or rather, legalized states to legalize sports betting), which led to a flurry of new sports bets and online casinos. Anyone who listens to US podcasts has probably heard the consequences: a flood of companies desperately trying to bribe customers into taking a gamble.
It seems that many people can gamble occasionally and get a thrill and call it a day, while other people end up betting their literal house and losing that too. My social media feeds are now flecked with people saying some version of: "one thing that's made me less libertarian is seeing how some people get entirely sucked in and ruined by online gambling."
Here's a brief list of things that I think plausibly have this property: many drugs, alcohol, gambling, day trading (but I repeat myself), shopping, pay-to-play games, pornography.
I don't think there's a common name for this category, so I'm going to call them Risk of Ruin goods.
There's obviously a strong overlap between "fine for many, but ruins some" and "addictive," but I don't think they're the same thing. For example, with some drugs the risk is that a small % of people will have a terrible reaction the first time they take it, rather than getting addicted longer term.
Similarly – and I know this will sound weird, please read to the end of the paragraph – I would largely describe cigarettes as more addictive and less Risk-of-Ruin-y. This is because, while I'm sure there's variance, my sense is most people who smoke for any consistent length of time will get addicted, and the bad consequences of smoking (while somewhat random) are more accretive and consistent. Outside Thank You For Smoking, there's no cigarette equivalent of the ways you can meaningfully ruin your life in one night through alcohol or gambling.
What should be the policy response to Risk of Ruin goods? I think a lot of people are torn between
1) not wanting to limit the ability of adults to access goods that don't harm them
2) feeling weird about some share of people having those goods ruin their lives.
Is there a legal regime that could keep the goods legal but prevent any one person accessing too much of them?
I think there are defacto examples of attempts to do this, especially with alcohol: for example, in Norway, only the government's Wine Monopoly can sell anything stronger than 4.75%, and the hours you can buy alcohol are generally restricted. The object seems to be make it possible to drink a bottle of wine with friends at dinner, but difficult to drink a handle of vodka alone at 2am.
We also have social customs that try to curb "problem drinking" while enabling moderate drinking: social norms like never drink alone or never drink before 5pm.
Could we do more? I wonder if, in this age of prevalent digital surveillance, we pretty easily could: credit cards could block you spending more than 10% of your income on potentially-ruinous goods in the same way they block seemingly 10% of my very normal transactions.
I'm not sure this would be a good idea; it would presumably open the door to all kinds of paternalism. The whole policy area of what to do about Risk of Ruin goods strikes me as one of those things where all the options – prohibition, legalization, selective restriction – have different bad consequences.
That said, I'm pretty confident what we shouldn't do is selectively ban people. While every gambling ad includes a long read-out about some number you can dial if you want to voluntarily ban yourself from casinos, the same casinos automatically ban anyone who makes semi-intelligent bets.
One of my big fears, looking around me, is that many currently-successful business models are kept afloat only by rinsing a small % of people of their entire life savings.
Finally, I have an observation that I'm not sure what to do with: whether it's correlation, causation or coincidence, when looking at the list of Risk of Ruin goods, I can't help noticing that a disproportionate share are forbidden in Islam.
2025-03-26 19:26:16
In some sense, the 80/20 of weight loss is to take ozempic, though maybe that's less of an 80/20 and more of a 95/5.
Still, suppose that you want to lose some weight but sans zempy – where does that leave you?
I'm not at all an expert on weight loss; my credentials are basically just "I think about stuff and I lost 20lb."
Still, as with other entries in this series, the people who care deeply about stuff are often unwilling to write up an 80/20 version of it, so you get me instead.
In some way I can't explain or describe, just weighing yourself every day causes you to eat less and lose (or not-gain) weight, even if you do nothing else.
I think counting calories constantly is emotionally/psychologically expensive, but if you've never previously spent time finding out how many calories are in stuff then it's super enlightening to do it for a week and see.
First, I made a spreadsheet with some of the main foods I eat every day (or hope to eat every day) and used this website to calculate the stats for each food.
Then I calculated some daily calorie intakes for reasonable-sounding combinations of foods. Spoiler: "reasonable quantities" is just much less than I ever imagined; my first inexpert forays were greeted by a friend with this response:
Until now, when I saw the calorie calculations on the sides of food-packets claiming that one little bag of whatever contained 3 portions, I thought this was a psychological trick to make you believe that the food was healthier, because people would notice the bold-faced calorie number but not-notice it was based on a tiny serving.
However, I now believe that the correct portion size of crackers truly is 3 crackers (14g).
Combined with the above, in order to eat fewer calories without 1) spending all day miserable, and/or 2) breaking down and stuffing your face again, you really need to focus on increasing your satiety for the calories you do eat.
This basically means more protein, more fibre and less (simple) carbohydrates. Basically you do this so that you aren't tempted to eat more food throughout the day, because you already feel full.
(Also, try to eliminate sugary drinks, they give you many calories for zero satiety).
Bringing this all together: in the spirit of 80/20, a lot of the work can be done by just changing the Default Things you eat to make them high satiety-per-calorie.
Basically, I used to eat bread and cheese by default, and this was Bad; I have largely swapped that out with
1) dried beancurd
2) greek yoghurt
3) microwaved frozen veggies
I try to avoid having bread in the house, now, so I don't have temptation I need to resist. As usual, Thomas Schelling was right: it's easier to throw out your ice-cream than to avoid eating it.
I then don't-think-about calories when I'm eating out or seeing friends or whatever (within reason), because 95% of my eating is done in a few standard times and places.
I don't have a satisfying conclusion to this blogpost, but hopefully you are sated already. Remember: weigh yourself; eat protein and fibre; eat less (fewer?) carbs; make your default-foods high protein / high fibre / low carb.
thanks to K. and A. for much of the inspiration for this piece
2025-03-24 19:30:49
In Voltaire's Letters on England, he writes up an interesting story on the pre-history of vaccination. (Bear in mind this is written in 1733, and I haven't checked the accuracy of his claims). It begins:
It is whispered in Christian Europe that the English are mad and maniacs: mad because they give their children smallpox to prevent their getting it, and maniacs because they cheerfully communicate to their children a certain and terrible illness with the object of preventing an uncertain one.
Voltaire is pro-vax (or rather, pro-variolation, the precursor to vaccines), and tells the story of a people called the Circassians. It's a disturbing story, but I think an interesting one. Basically:
The Circassians are poor and their daughters are beautiful, and so they use them as their chief export. They supply beauties to the harems of the Grand Turk, the Sophy of Persia and those who are rich enough to buy and keep this precious merchandise.
With the most honourable intentions they train these girls to perform dances full of lasciviousness and sensuousness, to rekindle the desires of the high and mighty masters for whom they are destined.
What does this have to do with variolation? Well:
It often happened that a father and mother, having taken a great deal of trouble to give their children a good upbringing, suddenly saw their hopes frustrated. Smallpox came into the family, one daughter died of it, another lost an eye, a third recovered but with a swollen nose, and the poor folk were ruined, with no resources.
So Voltaire's claim is that the Circassians were early to figuring out variolation because their "commercial interest" (in selling off their daughters) was dependent on it. The Circassians noticed that serious smallpox "never occurs twice in a lifetime" and that if "the outbreak finds only a delicate and thin skin to pierce, it leaves no mark upon the face," such that "if a child of six months or a year had a mild attack it would not die and would not be marked, and would be immune from the disease for the rest of its days."
So they began deliberately giving smallpox to their babies,
by making an incision in the arm and inserting into the incision a pustule that they have carefully removed from the body of another child. This pustule works in the arm where it is inserted like yeast in a piece of dough; it ferments and spreads its own qualities through the blood-stream.
(Incidentally, I was quite confused at "inserted like yeast in a piece of dough", but it seems perhaps that in those days they would take a piece of fermented dough from one batch to start the next one?)
Voltaire then makes this entirely unnecessary dig at Benedictines:
Some people maintain that the Circassians originally took this custom from the Arabs, but we leave this historical point to be cleared up by some learned Benedictine, who will doubtless compose several folio volumes on the subject, with proofs.
Incidentally, he has very positive views of the Chinese:
I understand that the Chinese have had [variolation] for a hundred years. The example of a nation that passes for the wisest and most strictly governed in the universe is a great thing in its favour. It is true that the Chinese set about it in a different way; they don’t make an incision but make the subject take in smallpox through the nose, like snuff. This is a more pleasant way.
How did variolation get to England? Well, a certain Lady Wortley Montagu was "living with her husband at the Embassy in Constantinople", and decided to follow the locals and give her child smallpox. "In vain did her chaplain point out that this procedure was not Christian and so could only succeed on infidels" – I would love to dig deeper on the model of health/sickness that produced this statement. The chaplain was wrong about the religion-specific properties of medicine, and the child thrived.
Lady Montagu "mentioned her experiment to the Princess of Wales," who became Queen Caroline. In 1721, the Princess "tried it out on four criminals condemned to death" [Wikipedia says it was seven], and therefore "saved their lives twice over," that is, from smallpox and from the gallows.
Per Voltaire, the burden of smallpox at the time was crushing:
Out of a hundred people in the world at least sixty have smallpox, and of these sixty, twenty die of it in the flower of their youth and twenty keep the unpleasant marks for ever. That makes one fifth of all human beings that this disease kills or permanently disfigures.
If variolated, "not one dies unless he is infirm and predisposed to die anyway, nobody is disfigured, nobody has smallpox a second time, assuming that the inoculation was properly done." If the French had adopted variolation as early as the British did, "twenty thousand people who died in Paris in 1723 would still be alive." Alas, it wouldn't become accepted in France until the 1770s – just before Voltaire's death.