Tweets

You blocked @gdb

Are you sure you want to view these Tweets? Viewing Tweets won't unblock @gdb

  1. 45 minutes ago

    Important ingredient for long-term productivity: having a reliable activity to clear your mind. For me it's running or weightlifting. For first half hour, I am still thinking about the problem of the day; by an hour I'm a blank slate. Hard problems somehow feel far simpler after.

    Undo
  2. 5 hours ago

    ML bugs are so much trickier than bugs in traditional software because rather than getting an error, you get degraded performance (and it's not obvious a priori what ideal performance is). So ML debugging works by continual sanity checking, e.g. comparing to various baselines.

    Undo
  3. May 13

    Biggest mental shift while switching from classic to scientific programming: explicit “for” loops are now extremely expensive. Instead, you express as much as you can by chaining hyperoptimized lower-level “for” loop primitives, eg matrix multiplies. Fun & new way of thinking.

    Undo
  4. May 13

    Worst part about using an internal-only library: can't search online for documentation/usage & usually not fully polished. Best part about using an internal-only library: someone at your company knows every corner of it & any change is on the table. Not an easy tradeoff.

    Undo
  5. May 12

    One very simple technique which helps me move faster: when printf debugging, print out full log messages rather than just bare quantities: e.g. print(f"Reached speed: {velocity=}") rather than print(velocity). More typing but saves puzzling out meaning of each line at runtime.

    Undo
  6. May 12

    For example, a few years ago we made the painful decision to shut down our robotics project. It was producing great results, but we'd discovered we could move even faster in data-rich virtual domains. Most of that team started working on code, which led to Codex & Copilot.

    Show this thread
    Undo
  7. May 12

    The health of an organization’s vision is better measured by its ability to cancel the right projects than to start new ones. Saying yes to a new project is the easy part, what’s hard is making tough & visionary decisions in the service of focus. Superpower if you can do it.

    Show this thread
    Undo
  8. May 11

    Security is the art of sufficiently raising the cost of attack such that you either (a) will yield a net negative return on investment to adversaries, or (b) are a much less attractive target than alternatives.

    Undo
  9. May 10

    Management technique we learned early on: short-term machine learning deadlines can be set based on inputs (e.g. high-quality execution on a set of experiments) but not outputs (e.g. reaching some level of performance). Science does not bend easily to the wishes of managers.

    Undo
  10. May 10
    Undo
  11. May 10

    Trick to avoid all timezone bugs: represent times as Unix timestamps (i.e. a number) rather than your language's Time or Date object. Additionally makes time arithmetic much clearer, and is especially friendly when building an API where remote machines will parse the timestamp.

    Undo
  12. May 10

    Such an indescribable feeling when one variant is showing great improvement relative to the baseline... and then upon more careful inspection you realize you had a bug and both runs were actually identical except for different seeds.

    Undo
  13. May 9

    I fell in love with programming right before college. I remember surprise & delight when I realized that college was prepared to give me credits for getting better at my newfound hobby. Even today I am amazed that it’s possible to be employed to program. The joy has not faded.

    Undo
  14. May 7

    The file system makes for a great key-value store but a poor general database — writing and reading a key is very fast, but it's slow to list (with 100k+ entries), and many tools choke on giant directories. Still, sometimes can be a great decision to use it to back your app.

    Undo
  15. May 6

    Modern feature that has most changed my use of Python: dataclasses. Not just useful for data structures — they let you separate state from implementation of any class. Also dataclasses tend to be more useful by default, since they come with repr, eq, etc.

    Undo
  16. May 5

    First rule of system administration — a service tends towards one of three states: manually watched, monitored, or broken.

    Undo
  17. May 2

    A software skill people not often appreciated outside security: probing the emergent semantics of a complex system. First understood this at Stripe while integrating various financial backends. Turns out to be extremely useful for building systems for training neural networks.

    Undo
  18. Apr 30

    A small taste of the kinds of content that you will be able to produce personally in the not-too-distant future:

    Undo
  19. Apr 29

    Any technology comes with benefits and commensurate risks. AI is unique only in the unprecedented magnitude we expect from both. Navigating these risks & distributing these benefits is daunting a priori, but each day of progress & experience makes the puzzle a bit more clear.

    Undo
  20. Apr 29

    "A photo of an artificial neural network growing in a garden":

    Show this thread
    Undo

Loading seems to be taking a while.

Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.

    You may also like

    ·