Yossi Kreinin

I am a programmer who has worked a lot on hardware, and from that point of view, software testing looks horrible, since software is full of bugs. And you can explain this with, well, they push programmers to ship buggy code since you can always fix the bugs later, and so the program is always full of bugs to be fixed later — whereas with hardware, bugs are often (not always!) catastrophic, so they have no choice but to let you fix them, even if it takes some time.

I find it very useful to be able to do everyone else’s job — it helps you learn how they work and what they need, and it helps even more when you need their job done, but they won’t do it themselves.

You used to tell people, “why are you doing all this by hand — write a script to do it!”, and then “...I meant an actual Python script, not a buggy grep | sed | crap pipeline!”

An industry veteran in my circles has recently made the rookie mistake1 of printing a warning from his code upon misuse. Surprisingly to nobody experienced, critical workflows soon came to a screeching halt.

Scarred by bare metal programming during my formative years, I consider the speedup from multithreading worth pursuing no matter how limited a form of it you’ll get, and no matter how hideous the hacks you’ll need to make it work. In today’s quest, we shall discover the various ways in which threads don’t work in a Python, Wasm, and especially a Python on Wasm environment, and then do something about it — even when that something could get us shunned from polite society. In the end, we’ll arrive at a working setup for limited yet performant multithreading, usable for soft real time programs caring about sub-millisecond overheads which we’ll attempt to minimize or eliminate (GitHub links: Python running C threads on Wasm; a simple C++ thread pool for Wasm.)

I believe that language models aren’t world models. It’s a weak claim — I’m not saying they’re useless, or that we’re done milking them. It’s also a fuzzy-sounding claim — with its trillion weights, who can prove that there’s something an LLM isn't a model of? But I hope to make my claim clear and persuasive enough with some examples.

We'll discuss how to make sure that your access to TLS (thread-local storage) is fast. If you’re interested strictly in TLS performance guidelines and don't care about the details, skip right to the end — but be aware that you’ll be missing out on assembly listings of profound emotional depth, which can shake even a cynical, battle-hardened programmer. If you don’t want to miss out on that — and who would?! — read on, and you shall learn the computer-scientific insight behind the intriguing inequality 0+0 > 0.

A timeline showing function call and return events is a great way to debug performance problems, especially in production. In particular, it's often much more effective than traditional sampling profilers, for reasons we’ll discuss. However, the adoption of function tracing in the industry remains uneven because of a chicken-and-egg problem.

What constitutes managerial competence? As a vague starting point for an answer, we could say that competent management sets achievable objectives and then achieves them, by organizing and incentivizing the necessary work.

I once wrote about how profiler output can be misleading. Someone commented that you don’t need profilers - just Ctrl-C your program in a debugger instead, and you’ll see the call stack where your program probably spends most of its time. I admit that I sneered at the idea at the time, because, despite those comments’ almost aggressive enthusiasm, this method doesn’t actually work on the hard problems. But as my outlook on life worsened with age, I came to think that Ctrl-C profiling deserves a shout-out, because it’s very effective against stupid problems encountered by lazy people operating in unfriendly environments.

We're going to speed up some numpy code by 100x using "unsafe Python." Which is not quite the same as unsafe Rust, but it's a bit similar, and I'm not sure what else to call it... you'll see. It's not something you'd use in most Python code, but it's handy on occasion, and I think it shows "the nature of Python” from an interesting angle.

There are many potential ways to use AI1 (and computers in general) for 2D animation. I’m currently interested in a seemingly conservative goal: to improve the productivity of a traditional hand-drawn full animation workflow by AI assuming responsibilities similar to those of a human assistant.

Here are some thoughts which are often basically correct:

There's a simple way to make your builds all of the following:

This is inspired by Dan Luu's post on the advantages of a single big repository over many small ones. That post is fairly old, and I confess that I'm hardly up to date on the state of tooling, both for managing multiple repos and for dealing with one big one. But I'm going to make an argument which I think mostly works regardless of the state of tooling on any given day:

I'm going to discuss 3 very basic things about patents:

For better or worse, things want to work.

OK, so 2 things:

OK, published at 3:30 AM. That's a first!