flâneur

Nick Dupoux

0 followers · 256 views

on the atlas — 9

highlights — 33

  • The only way I’ve found to handle that tension is in a small org is to be all in on one mode or the other. Default to Iterate, but be willing to Invest, maximizing how many people can work on the investments until they’re done. After enough iterating, you’ll identify the real investment areas that are critical.
    How To Do Less
  • I don’t even write TODOs in code, because that’s a tiny little plan that doesn’t matter.
    How To Do Less
  • Unless you’re planning on finding a new job pretty soon, or shifting blame for a project failure to someone else, or otherwise highly uninvested in your team’s overall success… don’t skip the obvious stuff.
    How To Do Less
  • Declaring a feature that’s in customers’ hands “done,” without monitoring, or with flaky tests, or tons of highly redundant code, or other obvious pending work, doesn’t magically get that work done.
    How To Do Less
  • So set a good example for everybody on your team, and accept No for an answer.
    How To Do Less
  • So that means that every person on the team needs to be able to help the others recognize and avoid distractions. What’s a distraction? Anything that doesn’t help you keep your existing features running, or deliver your top priority faster.
    How To Do Less
  • If you aren’t going to work on or support Feature X for at least 6 months, and you really mean it, then you should feel really comfortable repeating it ad nauseum.
    How To Do Less
  • If someone isn’t at least a little disappointed, they didn’t hear no. Keep saying it now and in the future
    How To Do Less
  • You can handle that pressure by saying no a lot, up front. A bunch of small disappointments that you actively seek out and cause are far easier to handle than waiting, and letting them build up to large disappointments
    How To Do Less
  • If you can find a 0.1% improvement each day, that adds up to getting 2x faster every two years (1.001^(365*2) == 2.07).
    Speed matters
  • Any process that you can make automatic, any decision or context switch that you can avoid, frees up mental resources that can be redeployed elsewhere. So even if the complex high-level work seems like the most crucial, you can still make gains by speeding up the low-level mechanical stuff.
    Speed matters
  • In addition, if you can code 10x as fast, you also get to do 10x as much coding practice, which will probably lead to yet more improvements.
    Speed matters
  • The best part is that you can improve your coding speed a lot by improving simple mechanical skills which are easy to measure, practice and improve.
    Speed matters
  • But that's totally conditional on the speed of everything else! If you could code 10x as fast then you could try out 10 different ideas in the time it would previously have taken to try out 1 idea. Or you could just try out 1 idea, but have 90% of your previous coding time available as extra idea time.
    Speed matters
  • If I was 10x faster yet it would have been 10 hours. That's a long plane ride. Even with a full-time job I would still be able to squeeze in a couple of text editor sized projects every month. I would be able to learn so many new things.
    Speed matters
  • Being 10x faster also changes the kinds of projects that are worth doing.
    Speed matters
  • If you could work 10x as fast then you could do 10x as much. Or do 5x as much and go home after lunch every day.
    Speed matters
  • Also, to calculate the cross-entropy loss, we first need to convert the logits to probabilities. The logits are the outputs from the last linear layer of our deep learning model. To convert them to probabilities, we generally have a softmax layer in the end
    Label Smoothing Explained using Microsoft Excel
  • Basically, the new loss H(q′, p) equals 1-Є times the old loss H(q, p) + Є times the cross entropy loss of the noisy labels H(u, p). This is key in understanding Label Smoothing - it is essentially the cross entropy loss with the noisy labels.
    Label Smoothing Explained using Microsoft Excel
  • This, however, can cause two problems. First, it may result in over-fitting: if the model learns to assign full probability to the ground- truth label for each training example, it is not guaranteed to generalize. Second, it encourages the differences between the largest logit and all others to become large, and this, combined with the bounded gradient ∂ℓ/∂z,k , reduces the ability of the model to adapt.
    Label Smoothing Explained using Microsoft Excel
  • For the cross-Entropy loss to really be at a minimum, each logit corresponding to the correct class needs to be significantly higher than the rest
    Label Smoothing Explained using Microsoft Excel
  • In information theory, the target distribution � is a probability distribution where the probability mass can be distributed broadly over the classes. In machine learning, the target distribution � has all of its mass allocated to a known label.
    Things that confused me about cross-entropy · Chris Said
  • Cut out time-consuming distractions
    Time anxiety: is it too late? - Ness Labs
  • “If we take care of the moments, the years will take care of themselves.”
    Time anxiety: is it too late? - Ness Labs
  • Make space for these moments: this does not mean making time for them. Instead, think of where you will incorporate these moments into your life
    Time anxiety: is it too late? - Ness Labs
  • For example, don’t think about happy it would make you to publish a book. Ask yourself if you really enjoy writing.
    Time anxiety: is it too late? - Ness Labs
  • Define what “time well spent” means to you: sit down, and think about what really makes you happy and gets you in the flow, without overthinking about how feasible the final outcome would be.
    Time anxiety: is it too late? - Ness Labs
  • it’s important to accept a few truths to be able to feel in control of your daily time and your lifetime. First, time exists and we can’t change that. Time will move forward, and so will we. Accepting these simple yet daunting truths is the first step in reducing time anxiety
    Time anxiety: is it too late? - Ness Labs
  • Beating time anxiety means shifting our focus from outcomes to output so we can spend our energy on things we can actually control.
    Time anxiety: is it too late? - Ness Labs
  • Next time you feel like studying something new that doesn’t fit neatly into your current “frame of expertise”, go ahead and just do it.
    The false promise of the 10,000 hour rule - Ness Labs
  • Research has shown that randomizing the information causes your brain to stay alert, helping to store information in your long-term memory.
    The false promise of the 10,000 hour rule - Ness Labs
  • The learning strategy that has been used traditionally in school to teach students consists in focusing on one skill before moving on to the next one and is called blocking. But there is a better way: interleaving, which consists in practicing multiple parallel skills at once.
    The false promise of the 10,000 hour rule - Ness Labs
  • Starting [at a young age] he’s read everything that he could find about business.
    Curius / Onboarding