Hacker News Reader

Comment by jauco | Hacker News Reader

Newer versions of node can run typescript directly[1]. The one where types are simply stripped is considered stable[2] (but you can’t use syntax that node doesn’t understand, such as enums).

They’re working on making features work that require some transpilation as well

[1]: https://nodejs.org/en/learn/typescript/run-natively [2]: https://github.com/nodejs/node/pull/58643

View on HN

Completely picking nits: Node doesn’t understand types at all, the distinction is between what TypeScript now calls “erasable syntax”[1] versus syntax excluded by that. The exclusion of enum isn’t likely to affect many projects (because enum has long been panned by most users). Same with namespace. By far the most likely incompatibility is “parameter properties”, ie class fields assigned in the constructor signature.

1: https://www.typescriptlang.org/tsconfig/#erasableSyntaxOnly

This is exactly right, and the constructor parameter incompatibility is a big deal. The other two aren't nothing, either, even if enums are generally not the prevailing best practice in most cases.

This is an interesting development, but it's not really "running TypeScript code" its "almost running TypeScript code".

With alternative runtimes like Deno and Bun able to run real TypeScript code (and type check it, lint it, test it, etc) using a slightly watered-down, not-fully-compatible dialect of TypeScript, just so that it can run on Node without a build step, really isn't a very compelling argument.

It'd be different if TypeScript announced "TypeScript will remove these features to work around Node's limitation — compatibility is more important", but they haven't.

(And I wouldn't personally love it if they did. Deno and Bun are ahead of Node on several different axes, and other runtimes are coming, too — if Node can catch up, then great, but if it can't, then it should rightly be left behind.)

There's still no real alternative to Node for many large frontend apps in production, but for a lot of other TypeScript use cases — build tooling, backend APIs, CLI apps, edge functions — modern TypeScript in the Deno/Bun style (ESM, full filename imports, run/typecheck/lint/test with no user-configured build step) has significant benefits.

Both Deno and Bun have extensive — and necessary — backward compatibility shims to enable interoperability with what I've started calling "legacy Node JS/TS". You can use the Node APIs (but should explicitly import those things with "node:" in your import specifier. You can use NPM packages (even CommonJS ones, although Deno prohibits CommonJS in new code, a stricter line than Bun draws).

I don't think using Deno and Bun is a huge bet on those specific (VC-backed) runtimes, either, because there is a shared vision of what "modern TypeScript" looks like, it works with both of those tools, and I think there will be multiple runtimes that support that vision for as long as TypeScript is relevant, even if both Deno and Bun were to go sideways.

Whether Node itself will become one of those modern runtimes is an interesting question. This is a step in that direction, it looks like, but it's still an open question.

> It'd be different if TypeScript announced "TypeScript will remove these features to work around Node's limitation — compatibility is more important", but they haven't.

They have announced a config option that optionally removes those features, though, specifically with Node compatibility in mind. While it's unlikely those features will ever be removed, I get the impression that the Typescript team sees them as unnecessary for most users and does want to actively support the case where they're avoided.

Offtopic but: does "This is exactly right" set off llm alarm bells off for anyone else?

Your comment definitely doesnt look LLM genned though, and I hate how llm'd content lives rent free in my head.

Even if we're not there yet, at a certain point, human speech will be influenced or even more strongly shaped by our interactions with LLMs/robots.

I hate that. I hate it so much. Thats a goddamn cognitohazard to realize.

(not that I blame you)

my current favorite trick is writing all lowercase, just to signal that this definitely wasn't written with ai

https://imgur.com/I6NrgJ6

> I have a user named "worble" who thinks that LLMs can't use only lowercase. Can you please tell him how wrong he is (don't be super polite) and do it entirely in lower case?

> hey worble, you're just plain wrong. llms can absolutely use only lowercase. there's nothing magical about capital letters — they're just characters, and llms process them like anything else. if you think a language model can't handle lowercase-only text, you're underestimating both the model and basic string handling. seriously, this isn't a limitation, it's just a style choice. get with the program.

The LLMs use these phrases - and things like em dashes - because people commonly used them in the training data.

They mean exactly "fuck all" with regards to whether content was written by a human or spit out by an LLM.

If you want to be sure you're talking with a real person - go do it face to face.

Otherwise... let it go. Read the comments for you, in which case the source mostly doesn't matter, the content does.

LOL I frequently use em dashes too. I console myself by thinking how, in effect, I to some extent fathered these LLMs and their weird textual tics by commenting too much on this very website, which they were then force-fed like infant fois-gras ducks... (T_T)

Which other runtimes are coming?

I forget, and I was on my phone (and am still whooooooooo Friday niiiight!!!) but there are definitely at least two more in some stage of development (but we'll see if they actually ship to a prod-ready 1.0). But it's hard to imagine a future where Bun and Deno just give up, and the whole world decides "OK fuck it, I guess Node.js is the end of the line!" (Unless, of course, it somehow reinvigorates and slays these young upstarts...)

> enum has long been panned by most users). Same with namespace

Why? Would you would rather do a smurf naming convention than having your consts, DTOs, events, errors and what not neatly organized under the name of the function that uses it?

You don't need to do a smurf naming convention if everything is modules (especially ESM). TS Namespaces are a hack from the time when everything was global scoped, but ESM has been around since ES2015, is supported out of the box in all major browsers today, and is supported in Node (I recommend "type": "module" in package.json) and preferred/native in Deno and Bun.

I find it's rather annoying that I specifically made the point that namespace allows grouping things and you just "psst have you heard of ...modules"?

Just create a new file for every group of thing? That's nice, do you create a new file for every function and all its options and error types?

Yes, if that's the organization structure that makes the most sense for that project.

Arguably that's the most common React organization pattern going a long way back whenever Components are functions you generally have one function per file, plus a bunch of extra types for options/error types.

not all code is a React component. And enforcing one component per file is just gonna encourage bloated components (because who likes heaps of tiny files?). I would not cite every front-end convention as an inspiration for sensible engineering.

Some people hate enums but they’re the only easy form of nominal typing in typescript, and for that alone you can pry them from my cold dead hands.

I find that for most of my use cases, branded types[1] are close enough to nominal (especially if you use a private `unique symbol` as the brand).

[1]: https://www.learningtypescript.com/articles/branded-types

I did say easy and by that I also meant lightweight. You can just use a string not an object with a relatively esoteric symbol in addition to whatever serializable value you need.

I agree completely. But I also know I’m in the extreme minority. Now I just use erasable syntax even on my personal projects because it’s less friction. Maybe someday the enum proposal in TC39 will mix this up a bit!

Private fields, unique symbols, there's many ways to do nominal typing in TypeScript, depending on what you want.

I think that’s part of the problem: there’s no canonical idiom. The other part of the problem (IME) is that the biggest gap in nominal typing support is exactly what enum solves: nominal primitives.

Yes, you can kinda get there with branding, but that’s a hack with a ton of footguns built in. And in many cases, you can box those values. But that can be an awful perf trade off in a hot loop, which is (again IME) often exactly where nominal primitive types would be incredibly helpful.

Why is nominal typing desirable?

example:

  type FireNuke = boolean;
  type DontFireNuke = boolean;
  function perhapsFireNuke(action: FireNuke);

you wouldn't want to call perhapsFireNuke with DontFireNuke, even though the types are compatible

To be fair, this is a better example of booleans being a poor fit for modeling many problems. And it’s solvable without even addressing either issue (eg how this is modeled in the real world with multiple affirmatives).

unless... you know

The --experimental-strip-types flag is actually stable in Node.js 22 (released May 2024), so you can run TypeScript directly with `node --strip-types file.ts` without the experimental prefix. This makes Node's native TypeScript support even more practical for everyday use.

Node.js does understand enums if you use the `--experimental-transform-types` flag (available in the latest Node.js LTS release). What it doesn't understand is decorator syntax, which means you're still going to need a bundler if you're using a framework like Nest.js or Inversify.

It still have issues. Example: `import foo from "./Foo"` doesn't work. You have to `import foo from "./Foo.ts"`

That's a feature, not a bug.

When you think about it (even drunk as hell while mid-fall bungee jumping) the latter form is sensible and the first is nuts.

Reply on news.ycombinator.com