John Myles White

@johnmyleswhite

Engineering manager on Facebook NY's Data Infra team. Former Julia developer and psychology grad student. Increasingly hardline YIMBY.

New York, NY
Vrijeme pridruživanja: srpanj 2008.

Tweetovi

Blokirali ste korisnika/cu @johnmyleswhite

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @johnmyleswhite

  1. 5. velj

    I don't think I understand this analogy since this one vote seems to have reduced conformity by 0%. But it does seem like a great example of how effect sizes matter and how hard it is to extrapolate any famed psychology findings to any applied settings.

    Poništi
  2. 4. velj

    The housing statistic I want to see most is a scatter plot showing the correlation between age of homeowner and annual income. I suspect you can see the effects of restricted housing supply like rings in a tree.

    Poništi
  3. 3. velj

    For me, this hits at a broad frustration: people think about almost everything they're not experts in using extremely vague, fuzzy and unscientific concepts. But they try to dress up their confusion in the language of science since they can't admit their way of thinking is fuzzy.

    Prikaži ovu nit
    Poništi
  4. 3. velj

    It's a bit depressing that just as lay people stopped attributing anything emotional to oxytocin, they've started attributing anything value-laden to dopamine.

    Prikaži ovu nit
    Poništi
  5. 2. velj

    (2) Users want to be able to figure out how to perform their specific task using a single function's documentation. The more implicit information (e.g. you can create closures) that's required to figure out to apply a function to solve their problems, the worse the documentation.

    Prikaži ovu nit
    Poništi
  6. 2. velj

    As far as I can tell, this happens in part because: (1) Users prefer to minimize the surface area of functions and concepts they need to remember. They want this much more than they want orthogonality -- if they want orthogonality at all.

    Prikaži ovu nit
    Poništi
  7. 2. velj

    But many users don't think of this solution and instead ask that the optimize(f, x) become an optimize(f, x, other_args=y) function. This process repeats with future requests of a similar nature and eventually single functions become "apps" that do everything themselves.

    Prikaži ovu nit
    Poništi
  8. 2. velj

    For example, consider an optimize(f, x) function that tries to minimize a function f starting at point x. If you want to minimize a more complex function f(x, y) for which some additional parameters can be held fixed, a clean solution is to use a closure to capture y.

    Prikaži ovu nit
    Poništi
  9. 2. velj

    One of the things that strikes me most about incidental programmers (whose main job is doing something else) is that they often forcefully push for individual functions to absorb lots of otherwise orthogonal functionality, thereby removing separation of concerns from a system.

    Prikaži ovu nit
    Poništi
  10. 1. velj

    Every time I open a terminal on Catalina and am reminded the shell defaults to zsh, I'm amused to see a language with 1-based indexing being actively pushed into the world by Apple.

    Poništi
  11. 31. sij

    I hope one day I’ll see social scientists stop confusing causality and mechanism with narrative and human interpretable explanations. Causal knowledge is verifiable as prediction under intervention.

    Poništi
  12. proslijedio/la je Tweet
    30. sij
    Odgovor korisniku/ci

    If anyone wants the TLDR/motivation of the change, this subsection in the new scope docs gives it:

    Poništi
  13. 30. sij

    Nice to see addressing what was arguably the change that most annoyed users in the 1.0 cleanup:

    Poništi
  14. 29. sij

    (2) More broadly, the true "tails" of a distribution are generally not that interesting. Practically relevant statistical pathologies can be demonstrated using distributions over the interval [0, 1], which, by definition, excludes any infinitely extended tail behavior.

    Prikaži ovu nit
    Poništi
  15. 29. sij

    Two points to clarify my tweet yesterday about tails being less interesting than skewness. (1) Realistic data sources almost surely don't infinite moments. But people invoke that concept often like it was interesting. Here's a top-voted comment on a HN thread from yesterday.

    Prikaži ovu nit
    Poništi
  16. 28. sij

    In sharp contrast, skewness and kurtosis are real problems for real data analysis if you think the CLT will apply cleanly just because you randomized something. But that has nothing to do with heavy tails -- all the pathologies show up with Bernoulli data already.

    Prikaži ovu nit
    Poništi
  17. 28. sij

    Real data, in my experience, always comes from a bounded source. The stock market isn't going to go up by 10^10 percent anytime soon in a single day. Temperatures won't increase 10^10 degrees this year. So the pathologies in the tails of heavy tailed distributions don't matter.

    Prikaži ovu nit
    Poništi
  18. 28. sij

    The gap between theoretical and real problems for me is always highest when people talk about heavy tails and the problems with using means. Those people always seems to forget to regret that the pathologies of the Cauchy distribution depends on the distribution being unbounded.

    Prikaži ovu nit
    Poništi
  19. 26. sij

    There seems to be a quirky bug in the current version of tvOS in which it's possible to pause content, leave it paused long enough for a screensaver to start, then unpause content without the screensaver going away -- so you just end up with audio moving forward.

    Poništi
  20. 24. sij

    I sometimes feel like science journalism is just a big parody of what's wrong with science more broadly. How in the world is it reasonable for an article to start this way?

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·