Posted by speckx 3 hours ago
My favorite side quest that I definitely recommend. Physical parallelism is so fun to reason about.
I had the advantage of sitting next to a proper professional to bounces questions off of, so I was able to get an SPDIF decoder to work in about a week! It was a very interesting lesson in the costs of abstraction, and why you will never really find abstraction layers in hardware.
I think 25 years ago I would not have believed we would ever fix (La)TeX but here's Typst.
From the compiler construction standpoint, it's grossly underspecified, to the point that our customers' designs were way more expressive than the spec.
And from the users' perspective, it's confusing, quite limited, and you end up having to resort to hacks to achieve what should be part of the language... Fortunately I can't remember any particularly bad examples, it's been 15 or so years since I left for greener pastures.
Still, I sometimes wonder, how much of HW design is influenced by the language decisions, these ugly hacks, and the emerging (System)Verilog design patterns? Could it be measurably better in a different world?
Elixir was somewhere between my 12th-14th language and nothing changed my perspective so much as understanding why the BEAM works the way it does.
Both Scheme and CL suffer really badly from a lack of modernization primarily in their stdlibs (or equivalent SRFIs) that tend to make them poor choices for most new professional projects IMO. Great for side projects though. I really wish the next generation of Scheme would take a page from Clojure's book and rework all the SRFIs and bake in immutability as a core requirement.
Put Lua in things.
Now, before you anti-Lua types start jumping on that 0-indexed down-vote, let me explain. Its not just about the language, its about the "thing" you're putting it in, its about using the VM effectively, its about planning and using the stack properly. It's about using one language to leverage the strengths - and discover the weaknesses - of other languages, runtimes, ABI's.
Putting Lua in things is easy - and it is hard. It teaches you how to treat other things with respect, and it teaches you how to disregard other things and "Just Use Lua™".
You can learn the language - multiple variants of it. You can learn the 'things' you're putting it in. You can learn how a VM works, what byte code is, how it can be used effectively - and how it can be used counter-productively, too.
It will teach you the limits of a language - Lua, and C - and it will teach you the strengths of both (as well as Rust, and all the other things you can get an FFI scaffolding built around).
You can use Lua to produce your own domain specific language. You can use it to (almost, or indeed completely) replace whatever other language the "thing" is using. You can use it to minimize the impact of a VM and scripting and byte code - you can learn how to maximize these things as well. You can learn byte code and you can leverage that understanding to learn machine language - two ends of a very similar (yet vastly different) spectrum.
You can learn how slow/fast pieces of the system interact productively and effectively - you can learn how things get mightily messed up, also, if you use the wrong lubricant.
In the end, you may not even end up using it for productive things - but you might also do something extremely powerful. (I have.) Either way, by applying this maxim and applying the skills needed to master doing it (or screw it up entirely), you can learn pretty much everything interesting there is to learn about software (besides LLM's) in today's modern stack madness .. and then some.
Seriously, give it a try.
The other one of these developer-tuning activities (there are a few), is to do a Linux From Scratch session at least twice a year. Do it this year, but plan on doing it again next year as well. Do it a few more times this decade.
Iterate. Extend. Ignite.
It is a scale, and there are other languages that go beyond Rust, with dependent types, effects, capabilities, linear types, formal profs,...
A common example I give is a basic chat client/server, to learn how networking code feels, especially nonblocking IO. Especially if you're implementing the same protocol in the different languages since it means you won't necessarily be as biased by what protocol patterns are easy to implement in one language.
For anyone curious, I generate the sudokus by first randomly filling an empty grid in a way that follows the sudoku rules, then emptying random cells where doing so doesn't make the solution ambiguous (a sudoku has exactly one valid solution). The first step to fill the grid works almost exactly like a normal backtracking solver but going through the cells in random order, and for a major speed boost the backtracking is far more aggressive (always resetting half the progress so far instead of just going back one step). Checking if a sudoku has multiple solutions is again normal backtracking but only exiting at the second time it finishes or after going through all combinations. The rest is modern computers being fast and limiting the number of iterations to avoid the super slow worst cases.
I am not sure if everyone else looked up techniques online or what for how to do it, or if they skipped that assignment, or made a very simplified solver that only handled some basic techniques. We hadn’t really learned much of anything yet in any of the classes of the course that would be useful for making a general solver. I thought I was supposed to figure it all out on my own, and that it should be able to handle any Sudoku puzzle that a person could solve. And in the end, despite many, many hours of working on it over many days it was still far from a complete solver. And I also didn’t have time to write it in the programming language we were using in the course. I was prototyping in JS and HTML because I wanted to see visually the sudoku puzzle and be able to switch between having the computer find next number and clicking on the number myself, to discover in detail how I was thinking when I was solving Sudoku puzzles, so that in turn I could implement those techniques in code.
I lost a little bit of motivation for the class with that, and dropped out of the class soon after because I found the class to be generally uninspiring and hard to stay motivated for, in spite of being interested in programming in general and having learned programming languages and made little programs of my own already since High School. Later I dropped out of the whole study, as I was struggling to study across most subjects and wasn’t passing enough classes to be able to continue.
My interest in programming continued, however, and I’ve been developing software both professionally and as a hobby with sometimes pretty good results during the years since, and having done so for companies both small and big, and in multiple different programming languages.
But I think the way I handled trying to implement the Sudoku solver is a way of working on things that I already had since before then and that I continued also later, and that that is also part of the reason that things go wrong for me when they do. Both professionally and also elsewhere in life.
The last job I had, I was laid off for not performing well enough. And I think in part I was still working on tasks at that job in some sort of wrong way all these years later, that may be similar in some ways to what I may have done wrong with for example that Sudoku solver assignment. And probably in many other cases too both before and after.
And now that AI is here I’m not really sure how to land a new job as developer at all. Even though I would like to. Combined with the sense that there is too much ups and downs in success rate of what I’ve been working on over the years to be able to present myself in an impressive manner to employers. I value honesty and integrity, and I find it difficult to really imagine that I could sell myself well at this point when I look back at my life so far, in spite of sometimes doing things well, because of the times that projects have not gone well.
If you just want generic pure computation: write an interpreter for a toy language. If it’s too large, write a macro expander that can handle C style or TeX style macros.
If you want events and concurrency: write a tmux clone that can work with multiple terminals.
If you want IO but specifically networking: write a toy HTTP server that only supports 1.0.
Another way to do it would be the standard low->high level and then one based on lambda calculus. Von Neumann and Church.
Assembly, C, Java, Clojure.
Assembly, C, C#, F#.
Assembly, Fortran, Rust, Rust, Brainf*ck.
I should have made the list bigger, not smaller. I missed assembly and Erlang.
At the heart of it all it’s just math and logic, and that’s where the different thought studies happen, regardless of which language quirks are navigated.
No, it's not. It's more like saying a writer should try their hand at different styles and formats of writing. So a novelist who periodically writes non-fiction essays or journalistic content or poetry (whether intended for publication or as a writing experiment), and that practice can in turn inform their writing as a novelist.
Besides, learning programming languages is much easier than learning natural languages. Most experienced programmers today know multiple languages, and most college grads in CS and Software Engineering graduate having used at least 5 or 6 languages if their school is halfway decent.
But there are example of "dead" languages whose everyday usage have devolved to only lithurgical functions like: Latin (Catholicism), Pali Cannon (Buddhism) and Sanskrit (Hinduism).
Instead of hoping you remember to check for NULL, Rust forces you to make a decision on how to handle the possibility of failure. Sure, you can still pick a bad decision (unwrap) but that is at least a specific decision and you can’t forget to do it.
You can take it even further with things like the newtype pattern. If you have lots of numeric IDs (project ID, user ID, etc) you can give them distinct types so you don’t mix them up.
So Rust taught me I can program defensively against my own brain, which sometimes gets tired or lazy or distracted.
It creates an interesting problem, as sometimes you want to know the absence of value. Usually done with the empty "" string.
The other way to handle is that every Object is initiated (pointing) to an starter initial value. (prototype based languages).
Sometimes you do want to represent the concept of “this whole datatype OR nothing”. How do you suggest this is modelled? I think optionality is pretty fundamental. What’s wrong about classical null is that it ascribes optionality to all types regardless of whether they need it.
Being forced to represent optionality with a “special” default is terrible design IMO - see the recent SSH RCE on Musl, largely due to the poor choice of indicating “crypt failed”. Default values can be quite dangerous. Sometimes the domain you are modelling supports such a value, but having “default value” be the default (and not a choice) is a similar evil to default null.
2. Use null coalescing (e.g. `foo?.bar?.baz`)
Languages such as Kotlin and C# support this style o programming.
The thing that I don't like about Optional types (in e.g. Scala or Java) is that they are too verbose. For example, instead of:
return foo?.bar?.baz ?? 0
you have something like: return foo.flatMap(_.bar).flatMap(_.baz).getOrElse(0);
This is even harder to read if you have more complicated logic, requiring implicit lambdas and other machinery.Initial values, like null, is an attempt to smuggle the "unset" case into the type without declaring it. And like null, the defensively written logic should prefix every single usage with a check for validity.
Optional is just a natural subset of the generalization that multiple possible states exist; in this case, just the two, and sum types to properly handle bigger sets.
Initial/null never admit more than 2 possible states, and also believe everything is always in one of those 2 states.
fwiw john carmack agrees with me
If you want specific values (integer or otherwise) associated with specific variants (i.e. Human = 1, Cat = 4) you can do that too. Use the #[repr(T)] or impl the From / Into traits yourself.
Is there anything you can do with separate types for the tag and the union that std::mem::discriminant and pattern matching can't do?
In Ada, if I use an enum (carrying the example from a link I used in a cousin comment):
type Expr_Kind_Type is (Bin_Op_Plus,
Bin_Op_Minus,
Num);
https://learn.adacore.com/courses/intro-to-ada/chapters/more...Variant records using this enum have to implement the variant for each element of the enum. If I, for some reason, have multiple records using the same discriminant I can know that each one implements it for each value of the discriminant. Does Rust have a way to synchronize your enums similar to that?
pub trait Payloads {
type Read;
type Write;
}
pub enum FileOp<P: Payloads> {
Read(P::Read),
Write(P::Write),
}
pub struct Inputs;
impl Payloads for Inputs {
type Read = ();
type Write = u8;
}
pub type Request = FileOp<Inputs>;
pub struct Outputs;
impl Payloads for Outputs {
type Read = u8;
type Write = ();
}
pub type Reply = FileOp<Outputs>;
pub fn execute(request: Request) -> Reply {
match request {
Request::Read(()) => Reply::Read(42),
Request::Write(_) => Reply::Write(()),
}
} enum Parser {
bin_add(...),
bin_sub(...),
...
}
enum Typer {
bin_add(...),
bin_sub(...),
...
}
And I'd have to hope they were both complete and match (at least according to what I know now, that's why I'm asking). If I go back in to the Parser enum and add a `un_minus`, then I have to also go into the Typer enum (apologies for the bad naming, I'm at work and commenting between my way too many meetings) and manually add it. I have no compile time guarantee that both are complete and synchronized. In Ada, I modify the existing enum and both will fail to compile until I add in the variant for `un_minus`. One edit, multiple failures distributed across the code base as a result telling me all the places where I have to add in a new code branch or new variant.EDIT: Run on sentence broken up.
I have found it quite useful to be able to generate a dedicated SomeEnumDiscriminants enum for SomeEnum, using the following derive:
https://docs.rs/strum/latest/strum/derive.EnumDiscriminants....
https://docs.rs/strum/latest/strum/derive.EnumDiscriminants....
Instead of just getting one type that includes the discriminant and records, Zig and Ada have a tag of some type whose value determines which variant is being used, and then the union (Zig) or record (Ada) is its own type.
I've had it come up when dealing with ASTs, although I don't recall the specifics. This is a problem that's certainly pretty easy to solve with macros. I'm pretty sure last time I ran into this I used `strum` [1] and its `EnumDiscriminants` macro to handle this.
Guess it also applies to computer programming and hardware description languages.
But reading this in 2026, when the entirety of the profession has changed (for better or worse) due to LLMs, seems strange.
Like someone writing how some new preservation method for making food last longer revolutionized how they manage their pantry, shortly after the invention of the fridge.
I like the post - it's just sad how many will not have these epifanies anymore, or they wont mean anything, and how many starting today would already just say "how quaint" and move on.
You're just getting older and times change. Your adolescence wasn't any more pure than the kids today.
One could argue you've frittered away your life fiddling bits around instead of becoming a thespian, writer, or musician. Alas, the epiphanies you could have had.