5 ms·

The smartest people I’ve ever worked with to date were from physics grad school. Still remember the time my coworker was doing code profiling, decided he was un

The smartest people I’ve ever worked with to date were from physics grad school. Still remember the time my coworker was doing code profiling, decided he was unhappy that the exponential function from the standard library was too slow, and decided to write a Taylor series approximation that gave him the precision he needed and cut the run time in half. He also learned C++ in a weekend and was vastly better at it by the end of that weekend than most people I’ve met in industry. And these were just every day occurrences that made it a thrill to go to work. Working with talented people is a drug. Some tips for younger people considering it: get involved in undergraduate research, apply to fellowships, shop for an advisor with a good reputation, start anticipating and preparing for an industry transition early, travel, date, and enjoy life!

BeetleB 2y ago
I don't want to take away from his brilliance, but generally Taylor approximations perform far worse than the standard library implementations. It's also the first tool of choice for physicists, so who knows ...? My guess, though, is that if he improved the performance, he used some other wizardry (Chebyshev or something similar).
- whatshisface 2y ago
  Sometimes what you need is less precision, much faster. Carmack's famous inverse square root falls into this category. If anything it's a lesson that the definition of brilliance is being in the wrong place at the wrong time... ;-)
  - cjbgkagh 2y ago
    I think Carmack credits someone else as the origin - possibly some magazine entry. These days I think the reciprocal square root intrinsic is the fastest where precision is not that important. I think there was a bit twiddling hack for pop count which was consistently faster than the equivalent cpu intrinsic due to some weird pipelining effect, so sometimes it is possible to beat the compilers and intrinsics with clever hacks.
  - michaelcampbell 2y ago
    Carmack denied writing it, and if WP is to be believed, he didn't. https://en.wikipedia.org/wiki/Fast_inverse_square_root https://en.wikipedia.org/wiki/Fast_inverse_square_root
- dingnuts 2y ago
  Honestly the whole story sounds like a tall tale to me. > He also learned C++ in a weekend and was vastly better at it by the end of that weekend than most people I’ve met in industry I doubt this. Really, really doubt this. Sure, geniuses exist, but I don't buy it.
  - mrguyorama 2y ago
    Having seen physicists code, I REALLY doubt this.
    - BeetleB 2y ago
      Bah! Let's go invert the matrix!
  - null_shift 2y ago
    If he already knew how to code in other object oriented languages, and was really just learning C++ syntax over the weekend, it’s not as much of a stretch.
    - cudgy 2y ago
      C++ is one of the most flexible and unopinionated languages you could ever encounter. The idea that someone who knows a high-level object-oriented language could translate that to immediate success in low-level C++ syntax at a level higher than the experts that developed the libraries over a weekend is frankly fantastical.
      - KPGv2 2y ago
        > the experts that developed the libraries over a weekend is frankly fantastical. this is not synonymous with "most [C++ programmers] in industry" The claim was the person learned it better than most people in industry, not most people writing the libraries upon which the industry is based EDIT: Also we don't technically know when this happened. If this story is from the 1990s, it's a lot more likely, because think of how many shitty C++ programmers there were back then since we didn't have all the language options we do now. It was still the language taught in schools, for example. Then it was Java and Python and JS etc. But back then, Jonny Mackintosh was writing bad C++ out of uni.
- fooker 2y ago
  The standard library implementations use Taylor approximations
- morelandjs 2y ago
  https://github.com/Duke-QCD/trento/blob/master/src/fast_exp.h https://github.com/Duke-QCD/trento/blob/master/src/fast_exp.... Check it out for yourself! I’m not claiming this was some kind of prodigious programming move, just something memorable that stuck with me.
  - DrFalkyn 2y ago
    Looks like he’s using a lookup table based on std::exp in combo with the Taylor expansion
adastra22 2y ago
The smartest people I’ve ever worked with were college dropouts.

BeetleB 2y ago

I don't want to take away from his brilliance, but generally Taylor approximations perform far worse than the standard library implementations. It's also the first tool of choice for physicists, so who knows ...? My guess, though, is that if he improved the performance, he used some other wizardry (Chebyshev or something similar).

whatshisface 2y ago
Sometimes what you need is less precision, much faster. Carmack's famous inverse square root falls into this category. If anything it's a lesson that the definition of brilliance is being in the wrong place at the wrong time... ;-)
- cjbgkagh 2y ago
  I think Carmack credits someone else as the origin - possibly some magazine entry. These days I think the reciprocal square root intrinsic is the fastest where precision is not that important. I think there was a bit twiddling hack for pop count which was consistently faster than the equivalent cpu intrinsic due to some weird pipelining effect, so sometimes it is possible to beat the compilers and intrinsics with clever hacks.
- michaelcampbell 2y ago
  Carmack denied writing it, and if WP is to be believed, he didn't. https://en.wikipedia.org/wiki/Fast_inverse_square_root https://en.wikipedia.org/wiki/Fast_inverse_square_root
dingnuts 2y ago
Honestly the whole story sounds like a tall tale to me. > He also learned C++ in a weekend and was vastly better at it by the end of that weekend than most people I’ve met in industry I doubt this. Really, really doubt this. Sure, geniuses exist, but I don't buy it.
- mrguyorama 2y ago
  Having seen physicists code, I REALLY doubt this.
  - BeetleB 2y ago
    Bah! Let's go invert the matrix!
- null_shift 2y ago
  If he already knew how to code in other object oriented languages, and was really just learning C++ syntax over the weekend, it’s not as much of a stretch.
  - cudgy 2y ago
    C++ is one of the most flexible and unopinionated languages you could ever encounter. The idea that someone who knows a high-level object-oriented language could translate that to immediate success in low-level C++ syntax at a level higher than the experts that developed the libraries over a weekend is frankly fantastical.
    - KPGv2 2y ago
      > the experts that developed the libraries over a weekend is frankly fantastical. this is not synonymous with "most [C++ programmers] in industry" The claim was the person learned it better than most people in industry, not most people writing the libraries upon which the industry is based EDIT: Also we don't technically know when this happened. If this story is from the 1990s, it's a lot more likely, because think of how many shitty C++ programmers there were back then since we didn't have all the language options we do now. It was still the language taught in schools, for example. Then it was Java and Python and JS etc. But back then, Jonny Mackintosh was writing bad C++ out of uni.
fooker 2y ago
The standard library implementations use Taylor approximations
morelandjs 2y ago
https://github.com/Duke-QCD/trento/blob/master/src/fast_exp.h https://github.com/Duke-QCD/trento/blob/master/src/fast_exp.... Check it out for yourself! I’m not claiming this was some kind of prodigious programming move, just something memorable that stuck with me.
- DrFalkyn 2y ago
  Looks like he’s using a lookup table based on std::exp in combo with the Taylor expansion

adastra22 2y ago

The smartest people I’ve ever worked with were college dropouts.