waterhouse

> It's like asking someone playing roulette to value "13 black", after they bet on it.

Even worse, because you don't get better odds or payouts by persuading others to bet on 13 black, but you do when they invest in the company you've backed.

Surely, if the compensation was high enough, you'd be like, "Sure, I'm happy with that outcome." And then, if the AI company thinks they have a low enough failure rate that the expected cost of paying out the compensation still lets them make a profit, then they could make that promise to all customers.

Though a compensation that high sounds like it would invite fraud, where the customer would be glad to have something go "wrong" and get a fat check. Not sure if that's a solvable problem.

Multithreaded:

  seq 1 20 | xargs -Iqq -n1 -P0 yes >/dev/null

So is it your position that, when the Chinese government imposes takedowns or worse on Chinese people posting Winnie the Pooh stuff, it's primarily because the Chinese government is opposed to anti-Asian racism?

> This is arguably rational because the pool is small and there's little to gain by studying harder if you already have a perfect GPA.

So there is a low ceiling, and if they instead used MCAT or something with a higher ceiling (where, apparently, the number of perfect scores is about 50 per year—in America, presumably lower in Canada due to population size), then studying harder would benefit them. That seems like a much better outlet for competitive urges.

But also, how small is the pool of qualified applicants? If there were something like "they're going to take n people from your school, at which there are 30 plausible candidates", then sabotaging one might conceivably be worthwhile. But if the pool is—well, Google says 3,000 medical students get accepted each year in Canada (and the qualified applicant pool is presumably at least somewhat larger), and sabotaging one person is extremely unlikely to help you personally. (This is one case where it's good that the expected-value "benefits", of sabotaging person X, are widely distributed among thousands of medical candidates, and thus it's a "free-rider problem" where no individual candidate has a strong motivation to do the work.)

Is there some multi-stage thing where they pick 10 people from each high school, or 30 from a town, or something? Or is there major grading on a curve, or a big benefit for being the top person in your classroom of 15? That seems like how you would get real incentives for this backstabbing behavior. Otherwise, I can't see how it's rational (even to a complete sociopath), and would have to chalk it up to individual miscreants and possibly some kind of culture that encourages it in other ways.

That would increase competition and thus depress wages for existing doctors, who are the ones who make the decisions here. I heard, from a medical school attendee, that she overheard some doctors discussing whether it would be a good idea to require a fifth year of medical school to become a general practitioner (luckily, they were like, "Eh... nah"). It did not seem like it bothered them that this would make it even harder for civilians to get medical care.

No matter how many have good grades, you can always pick the top n by grades—unless there's a ceiling that the top m > n have all hit. Which, if you're talking about "grades" as in GPA, is plausible.

MCAT seems more relevant, though. According to Claude: "Roughly 0.1% or fewer of test-takers score a perfect 528 in any given year — typically only a few dozen individuals out of the ~120,000 or so who sit for the exam annually." So it should work fairly well for them to sort by MCAT and take however many they have (or expect to have) room for.

Complete androgen insensitivity syndrome, I imagine? That's probably the one category of XY people who have undergone no hormonal masculinization throughout their lives, and the one case where I'd agree with them competing with women. Wikipedia says it's estimated to be "1 in 20,400 to 1 in 99,000".

I guess, in theory, this can eventually be countered by people using LLM browser integrations to tell them whether comments are worth reading (and maybe to summarize long comments). Is anyone currently working on that? It might be interesting to see.