• 3 Posts
  • 159 Comments
Joined 1 year ago
cake
Cake day: July 1st, 2023

help-circle






  • You often need to be pretty good at math. But not because you’re “doing math” to write the code.

    In real world software systems, you need to handle monitoring and alerting. To properly do this, you need to understand stats, rolling averages, percentiles, probability distributions, and significance testing. At least at a basic level. Enough to know how to recognize these problems and where to look when you run into them.

    For being a better coder, you need to understand mathematical logic, proofs, algebra/symbolic logic, etc in order to reason your way through tricky edge cases.

    To do AI/ML, you need to know a shitton of calculus and diff eqs, plus numerical algorithms concepts like numerical stability. This is kinda a niche (but rapidly growing) engineering field.

    The same thing about AI also applies to any other domain where the thing being computed is fundamentally a math or logic solution. This is somewhat common in backend engineering.

    I’m not “doing math” with pen and paper at work, but I do use all of these mathematical skills all. the. time.

    I am an SRE on a ML serving platform.



  • BUT good poll results aren’t just “we polled 1,000 people and here’s who they’re voting for.”

    Good pollsters take demographic data when they poll. They model the biases of different demos, and they correct for those biases in their models.

    Yes, reducing underrepresentation at poll time would be ideal. But pollsters are smart and are doing their best to put out good models. Pollsters know Gen Z is underrepresented and are accounting for that already.

    In other words, don’t let Gen Z underrepresentation in the polls lull you into a false sense of security. The polls are accurate. The race is neck and neck.








  • Video codecs mostly work by tracking movement, predicting which pixels will change, and striving to only encode the pixels that actually change or change dramatically. In other words, compression looks for patterns.

    All of that goes out the window when you try to compress static. There are no patterns. It simply can’t be compressed. This isn’t a matter of the algorithms not being good enough. It’s a fundamental limit of information theory.

    Anything fancier amounts to embedding the intro into the compressor as a well-known pattern. And at that point, you’re better off just caching a 4K version of the intro as a standalone video file directly in the app.





  • And it is not possible to “visualize 4D”

    Sure it is.

    • 3 spatial dimensions + time
    • 3 spatial dimensions + 1 color dimension (grayscale)
    • 2 spatial dimensions + 2 color dimensions
    • etc

    And that’s not even counting projection. All the time we interact with 3D data that’s projected to 2D (almost every photo you’ve ever looked at). There are similar ways to project 4D to 2D.

    (Not defending the video or anything, just pointing out that visualizing higher dimensions is something we know about for ages.)