> I fear LLMs have frozen programming language advancement and adoption for anything past 2021.
Why would that be the case? Many models have knowledge cutoffs in this calendar year. Furthermore I’ve found that LLMs are generally pretty good at picking up new (or just obscure) languages as long as you have a few examples. As wide and varied as programming languages are, syntactically and ideologically they can only be so different.
There's a flywheel where programmers choose languages that LLMs already understand, but LLMs can only learn languages that programmers write a sufficient amount of code in.
Because LLMs make it that much faster to develop software, any potential advantage you may get from adopting a very niche language is overshadowed by the fact that you can't use it with an LLM. This makes it that much harder for your new language to gain traction. If your new language doesn't gain enough traction, it'll never end up in LLM datasets, so programmers are never going to pick it up.
> but LLMs can only learn languages that programmers write a sufficient amount of code in
i wrote my own language, LLMs have been able to work with it at a good level for over a year. I don't do anything special to enable that - just front load some key examples of the syntax before giving the task. I don't need to explain concepts like iteration.
Also llm's can work with languages with unconventional paradigms - kdb comes up fairly often in my world (array language but also written right to left).
> Because LLMs make it that much faster to develop software
I feel as though "facts" such as this are presented to me all the time on HN, but in my every day job I encounter devs creating piles of slop that even the most die-hard AI enthusiasts in my office can't stand and have started to push against.
I know, I know "they just don't know how to use LLMs the right way!!!", but all of the better engineers I know, the ones capable of quickly assessing the output of an LLM, tend to use LLMs much more sparingly in their code. Meanwhile the ones that never really understood software that well in the first place are the ones building agent-based Rube Goldberg machines that ultimately slow everyone down
If we can continue living in the this AI hallucination for 5 more years, I think the only people capable of producing anything of use or value will be devs that continued to devote some of their free time to coding in languages like Gleam, and continued to maintain and sharpen their ability to understand and reason about code.
* One developer tried to refactor a bunch of graph ql with an LLM and ended up checking in a bunch of completely broken code. Thankfully there were api tests.
* One developer has an LLM making his PRs. He slurped up my unfinished branch, PRed it, and merged (!) it. One can only guess that the approved was also using an LLM. When I asked him why he did it, he was completely baffled and assured me he would never. Source control tells a different story.
* And I forgot to turn off LLM auto complete after setting up my new machine. The LLM wouldn't stop hallucinating non-existent constructors for non-existent classes. Bog standard intellisense did in seconds what I needed after turning off LLM auto complete.
LLMs sometimes save me some time. But overall I'm sitting at a pretty big amount of time wasted by them that the savings have not yet offset.
> One developer tried to refactor a bunch of graph ql with an LLM and ended up checking in a bunch of completely broken code. Thankfully there were api tests.
So the LLM was not told how to run the tests? Without that they cannot know if what they did works, and they are a bit like humans, they try something and then they need to check if that does the right thing. Without a test cycle you definitely don’t get a lot out of LLMs.
I once toured a dairy farm that had been a pioneer test site for Lasix. Like all good hippies, everyone I knew shunned additives. This farmer claimed that Lasix wasn't a cheat because it only worked on really healthy cows. Best practices, and then add Lasix.
I nearly dropped out of Harvard's mathematics PhD program. Sticking around and finishing a thesis was the hardest thing I've ever done. It didn't take smarts. It took being the kind of person who doesn't die on a mountain.
There's a legendary Philadelphia cook who does pop-up meals, and keeps talking about the restaurant he plans to open. Professional chefs roll their eyes; being a good cook is a small part of the enterprise of engineering a successful restaurant.
(These are three stool legs. Neurodivergents have an advantage using AI. A stool is more stable when its legs are further apart. AI is an association engine. Humans find my sense of analogy tedious, but spreading out analogies defines more accurate planes in AI's association space. One doesn't simply "tell AI what to do".)
Learning how to use AI effectively was the hardest thing I've done recently, many brutal months of experiment, test projects with a dozen languages. One maintains several levels of planning, as if a corporate CTO. One tears apart all code in many iterations of code review. Just as a genius manager makes best use of flawed human talent, one learns to make best use of flawed AI talent.
My guess is that programmers who write bad code with AI were already writing bad code before AI.
I don't think this is actually true. LLMs have an impressive amount of ability to do knowledge-transfer between domains, it only makes sense that that would also apply to programming languages, since the basic underlying concepts (functions, data structures, etc.) exist nearly everywhere.
If this does appear to become a problem, is it not hard to apply the same RLHF infrastructure that's used to get LLMs effective at writing syntactically-correct code that accomplishes sets of goals in existing programming languages to new ones.
> LLMs have an impressive amount of ability to do knowledge-transfer between domains, it only makes sense that that would also apply to programming languages, since the basic underlying concepts (functions, data structures, etc.) exist nearly everywhere.
That would make sense if LLMs understood the domains and the concepts. They don't. They need a lot of training data to "map" the "knowledge transfer".
Personal anecdote: Claude stopped writing Java-like Elixir only some time around summer this year (Elixir is 13 years old), and is still incapable of writing "modern HEEX" which changed some of the templaring syntax in Phoenix almost two years ago.
I would argue it's more important than ever to make new languages with new ideas as we move towards new programming paradigms. I think the existence of modern LLMs encourages designing a language with all of the following attributes:
- Simple semantics (e.g. easy to understand for developers + LLMs, code is "obviously" correct)
- Very strongly typed, so you can model even very complex domains in a way the compiler can verify
- Really good error messages, to make agent loops more productive
- [Maybe] Easily integrates with existing languages, or at least makes it easy to port from existing languages
We may get to a point where humans don't need to look at the code at all, but we aren't there yet, so making the code easy to vet is important. Plus, there's also a few bajillion lines of legacy code that we need to deal with, wouldn't it be cool if you could port (or at least extend it) it into some standardized, performant, LLM-friendly language for future development?
I think that LLMs will be complemented best with a declarative language, as inserting new conditions/effects in them can be done without modifying much (if any!) of the existing code. Especially if the declarative language is a logic and/or constraint-based language.
We're still in early days with LLMs! I don't think we're anywhere near the global optimum yet.
> It’d be like inventing a new assembly language when everyone is writing code in higher level languages that compile to assembly.
Isn't that what WASM is? Or more or less what is going on when people devise a new intermediate representation for a new virtual machine? Creating new assembly languages is a useful thing that people continue to do!
We may end up using AI to create simplified bespoke subset languages that fit our preferences. Like a DSL of sorts but with better performance characteristics than a traditional DSL and a small enough surface area.
It does further than non-determinism. LLM output is chaotic. 2 nearly identical prompts with a single minor difference can result in 2 radically different outputs.
> For those that don't know its also built upon OTP, the erlang vm
This isn't correct. It can compile to run on the BEAM: that is the Erlang VM. OTP isn't the Erlang VM; rather, "OTP is set of Erlang libraries and design principles providing middle-ware to develop [concurrent/distributed/fault tolerant] systems."
Importantly: "Gleam has its own version of OTP which is type safe, but has a smaller feature set. [vs. Elixir, another BEAM language with OTP support]"
The comment you are replying to is correct, and you are incorrect.
All OTP APIs are usable as normal within Gleam, the language is designed with it in mind, and there’s an additional set of Gleam specific additions to OTP (which you have linked there).
Gleam does not have access to only a subset of OTP, and it does not have its own distinct OTP inspired OTP. It uses the OTP framework.
What's the state of Gleam's JSON parsing / serialization capabilities right now?
I find it to be a lovely little language, but having to essentially write every type three times (once for the type definition, once for the serializer, once for the deserializer) isn't something I'm looking forward to.
A functional language that can run both on the backend (Beam) and frontend (JS) lets one do a lot of cool stuff, like optimistic updates, server reconciliation, easy rollback on failure etc, but that requires making actions (and likely also states) easily serializable and deserializable.
You can generate those conversions, most people do.
But also, you shouldn’t think of it as writing the same type twice! If you couple your external API and your internal data model you are greatly restricting your domain modelling cability. Even in languages where JSON serialisation works with reflection I would recommend having a distinct definition for the internal and external structure so you can have the optimal structure for each context, dodging the “lowest common decimator” problem.
I understand your point, and I agree with it in most contexts! However, for the specific use case where one assumes that the client and server are running the exact same code (and the client auto-refreshes if this isn't the case), and where serialization is only used for synchronizing between the two, decoupling the state from it's representation on the wire doesn't really make sense.
This is also what really annoyed me when I tried out Gleam.
I'm waiting for something similar to serde in Rust, where you simply tag your type and it'll generate type-safe serialization and deserialization for you.
Gleam has some feature to generate the code for you via the LSP, but it's just not good enough IMHO.
Rust has macros that make serde very convenient, which Gleam doesn't have.
Could you point to a solution that provides serde level of convenience?
Edit: The difference with generating code (like with Gleam) and having macros generate the code from a few tags is quite big. Small tweaks are immediately obvious in serde in Rust, but they drown in the noise in the complete serialization code like with the Gleam tools.
> Rust has macros that make serde very convenient, which Gleam doesn't have.
To be fair, Rust's proc macros are only locally optimal:
While they're great to use, they're only okay to program.
Your proc-macro needs to live in another crate, and writing proc macros is difficult.
Compare this to dependently typed languages og Zig's comptime: It should be easier to make derive(Serialize, Deserialize) as compile-time features inside the host language.
When Gleam doesn't have Rust's derivation, it leaves for a future where this is solved even better.
"Elixir has better support for the OTP actor framework. Gleam has its own version of OTP which is type safe, but has a smaller feature set."
At least on the surface, "but has a smaller feature set" suggests that there are features left of the table: which I think it would be fair to read as a subset of support.
If I look at this statement from the Gleam OTP Library `readme.md`:
"Not all Erlang/OTP functionality is included in this library. Some is not possible to represent in a type safe way, so it is not included. Other features are still in development, such as further process supervision strategies."
That quote leaves the impression that OTP is not fully supported and therefore only a subset is. It doesn't expound further to say unsupported OTP functionality is alternatively available by accessing the Erlang modules/functions directly or through other mechanisms.
In all of this I'll take your word for it over the website and readme files; these things are often not written directly by the principals and are often not kept as up-to-date as you'd probably like. Still even taking that at face value, I think it leaves some questions open. What is meant by supporting all of OTP? Where the documentation and library readme equivocates to full OTP support, are there trade-offs? Is "usable as normal" usable as normal for Erlang or as normal for Gleam? For example, are the parts left out of the library available via directly accessing the Erlang modules/functions, but only at the cost of abandoning the Gleam type safety guarantees for those of Erlang? How does this hold for Gleam's JavaScript compilation target?
As you know, Elixir also provides for much OTP functionality via direct access to the Erlang libraries. However, there I expect the distinction between Elixir support and the Erlang functionality to be substantially more seamless than with Gleam: Elixir integrates the Erlang concepts of typing (etc.) much more directly than does Gleam. If, however, we're really talking about full OTP support in Gleam while not losing the reasons you might choose Gleam over Elixir or Erlang, which I think is mostly going to be about the static typing... then yes, I'm very wrong. If not... I could see how strictly speaking I'm wrong, but perhaps not completely wrong in spirit.
Ah, that’s good feedback. I agree, that documentation is misleading. I’ll fix them ASAP.
> Elixir also provides for much OTP functionality via direct access to the Erlang libraries.
This is the norm in Gleam too! Gleam’s primary design constraint is interop with Erlang code, so using these libraries is straightforward and commonplace.
Thanks for the clarification. I've read about Gleam here and there, and played with it a bit, and thought there was no way to directly access OTP through the Erlang libraries.
This can be just my lack of familiarity with the ecosystem though.
Gleam looks lovely and IMO is the most readable language that runs on the BEAM VM. Good job!
I wonder why so many have got this wrong across this thread? Was it true once upon a time or something, or have people just misunderstood your docs or similar?
OTP is a very complex subject and quite unusual in its scope, and it’s not even overly clear what it even is. Even in Erlang and Elixir it’s commonly confused, so I think it’s understandable that Gleam has the same problem further still with its more distinct programming style.
i just implemented a project in elixir with LLM support and would never have considered that before. (i had never used elixir before) - So who knows maybe it will help adoption?
They are working towards "the real thing", whatever your definition of real is.
BTW in the 90s people tried to come up with a type system for Erlang, and failed:
--- start quote ---
Phil Wadler[1] and Simon Marlow [2] worked on a type system for over a year and the results were published in [3]. The results of the project were somewhat disappointing. To start with, only a subset of the language was type-checkable, the major omission being the lack of process types and of type checking inter-process mes-sages. Although their type system was never put into production, it did result in a notation for types which is still in use today for informally annotating types.
Several other projects to type check Erlang also failed to produce results that could be put into production. It was not until the advent of the Dialyzer [4] that realistic type analysis of Erlang programs became possible.
I don’t mean to minimize the huge effort by the Gleam team; however, Elixir cannot become Gleam without breaking OTP/BEAM in the same ways Gleam does. As it stands now, Elixir is the superior language between the two, if using the full Erlang VM is your goal.
Gleam is really quite a nice language. I did AoC in it this year as well and came away with the following: (incomplete list for both positive and negative, these are mainly things that come to mind immediately)
Positive:
- It can be pretty performant if you do it right. For example, with some thought I got many days down to double digit microseconds. That said, you do need to be careful how you write it and many patterns that work well in other languages fall flat in Gleam.
- The language server is incredibly good. It autoformats, autocompletes even with functions from not-yet-imported-but-known-to-the-compiler packages, shows hints with regarding to code style and can autofix many of these, autofills missing patterns in pattern matches, automatically imports new packages when you start using them and much much more. It has definitely redefined my view of what an LSP can do for a language.
- The language is generally a joy to work with. The core team has put a lot of effort into devex and it shows. The pipe operator is nice as always, the type system is no haskell but is expressive enough, and in general it has a lot of well-thought out interactions that you only notice after using it for a while.
Negative:
- The autoformatter can be a bit overly aggressive in rewriting (for example) a single line function call with many arguments to a function call with each argument on a different line. I get that not using "too much" horizontal space is important, but using up all my vertical space instead is not always better.
- The language (on purpose) focuses a lot on simplicity over terseness, but sometimes it gets a little bit much. Having to type `list.map` instead of `map` or `dict.Dict` instead `Dict` a hundred times does add up over the course of a few weeks, and does not really add a lot of extra readability. OTOH, I have also seen people who really really like this part of Gleam so YMMV.
- Sometimes the libraries are a bit lacking. There are no matrix libraries as far as I could find. One memoisation library had a mid-AoC update to fix it after the v1.0 release had broken it but nobody noticed for months. The maintainer did push out a fix within a day of realizing it was broken though. The ones that exist and are maintained are great though.
I can live with these negatives. What irritates me the most is the lack of if/else or guards or some kind of dedicated case-distinction on booleans. Pattern matching is great but for booleans it can be kinda verbose. E.g.
case x < 0 {
True -> ...
False ->
case x > 10 {
True -> ...
False ->
case x <= 10 {
True -> ...
False -> ...
}
}
}
You most likely asked an AI for this. They always think there is an `if` keyword in case statements in Gleam. There isn't one, sadly.
EDIT: I am wrong. Apparently there are, but it's a bit of a strange thing where they can only be used as clauses in `if` statements, and without doing any calculations.
> Having to type `list.map` instead of `map` or `dict.Dict` instead `Dict` a hundred times does add up over the course of a few weeks, and does not really add a lot of extra readability.
I did it in F# this year and this was my feeling as well. All of the List.map and Seq.filter would have just been better to be called off of the actual list or Seq. Not having the functions attached to the objects really hurts discoverability too.
Re argument formatting, I'd guess it's because it uses the Prettier algorithm which works like that.
However in my experience it's much better than the alternative - e.g. clang-format's default "binpack"ing of arguments (lay them out like prose). That just makes them hard to read and leads to horrible diffs and horrible merge conflicts.
I like Gleam, but I am somewhat annoyed by the fact, that I don't have the full functional freedom in calling recursive (inner) functions wherever I want. I don't know, why new functional languages do not get this right all the way, straight from some rnrs document or implementation. Another thing is the separate operators like .> and .< and so on. What I liked were of course pipes and pattern matching.
To me it felt less elegant than Scheme (GNU Guile) which I usually use (with nice parallelism if I want to, pipelines, and also pattern matching), and, aside from syntax, conceptually perhaps less elegant than Erlang. On the other hand it has static typing.
I also tried OCaml this year, but there are issues in the ecosystem making a truly reproducible environment/setup, because opam doesn't produce proper lock files (only version numbers) and it seemed silly to not be able to even include another file, without reaching for dune, or having to specify every single file/module on command line for the OCaml compiler. So I was left unsatisfied, even though the language is elegant and I like its ML-ness. I wish there was a large ecosystem around SML, but oh well ...
Might be I should take another look at Erlang soon, or just finally get started with Haskell. Erlang using rebar3 should have proper lock files, has pattern matching, and if I remember correctly no such limitations for calling functions recursively. No longer sure how or whether Erlang did inner functions though.
Heh, similar thoughts! The main difference that I only used Scheme for SICP, and I've used a bit of Haskell.
I like Haskell in theory, but: just to get a hello world takes a lot of CPU and disk space. The standard library is full of exceptions (you can use a different prelude, that opens a whole different can of worms). The ergonomics of converting between the thousand different string types are awful.
So, you being basically me, I have some recommendations:
Idris (2): good stdlib, has dependent types. A beautiful language. The compiler is self-hosted and bootstrapped by lisp - very elegant! The ecosystem is basically nonexistent though.
PureScript: also improves on Haskell in many ways. But, it's more of a frontend language, and though you can do backend, you're stuck with JavaScript runtime. Oh well.
It seems like it should be a common feature to be able to view between each array operation in a debugger without having to augment the code with `echo`
The out of bounds handling didn't seem all that good to me. Sure you can filter out undefined. You could also just make a function that returns an empty array if out of bounds, or array of 1 element if not.
// JS
function getElemFromGrid(grid, x, y) {
return (x < 0 || x >= grid.width ||
y < 0 || y >= grid.height)
? []
: [grid.elems[y][x]]
}
...
neighbors = [
...getElemFromGrid(grid, x + 1, y + 0),
...getElemFromGrid(grid, x + 1, y + 1),
...getElemFromGrid(grid, x + 0, y + 1),
...getElemFromGrid(grid, x - 1, y + 1),
...getElemFromGrid(grid, x - 1, y + 0),
...getElemFromGrid(grid, x - 1, y - 1),
...getElemFromGrid(grid, x + 0, y - 1),
...getElemFromGrid(grid, x + 1, y - 1),
]
I also find grids made of 2 dimensional array to be code small. An array of arrays is NOT A GRID as there is nothing enforcing the inner arrays to be the same length as each other. Also, in efficienct code it would be better to provide a 1 dimensional array and bounds. Now, out of bounds checks based on accessing outside the array won't work. You need to check actual bounds. Again giving preference using a helper
You're right, but loads of times I just left that there because I probably did something more involved in the map that I ended up deleting later without realising.
This sounds like the kind of situation where the LSP could suggest the simpler code, I'll see if there's an issue for it already and suggest it if not.
Elixir has one opinionated formatter -- Quokka -- that will rewrite the code above properly. It can also reuse linting rules as rewrite policies. Love using it.
Just so no one misunderstands this. The creator (Evan) didn't get into, or start, any drama himself that I ever noticed. I'd argue he's a very chill and nice dude.
I've been on the edges of the community for probably a decade now (lurker), and all of the drama came from other people who simply didn't like the BDFL and slow releases strategy.
I'm a backend dev mostly and use Elm for all my frontend needs. Yes there are some things compiler-side that could be improved, but basically it's fine.
I appreciate not having to keep up with new releases!
Yeah I'm not a front end dev but I do kind of keep track of what's going on in that space. From what I saw it seemed Elm was all but dead. Maybe that's not true, but that was the impression from the outside looking in.
Not many people use elm directly but it has influenced the design of so many frameworks and libraries.
Right now I’m toying with the idea of building a GNOME application in rust, and the framework I’m using is relm4 which provides elm like abstractions over gtk-rs.
Previously I’ve built web applications with F# and elmish, which again provides elm like abstractions for building F# applications.
I recently used Gleam + Lustre for a small app that I normally would have built with Elm + PostgREST. It went very well, and I'm now planning to use it for a larger rewrite (of a rails app).
I've considered playing around with it. Every now and then I tell myself I'm gonna learn front end dev, but as soon as I start looking at it (FE in general not Lustre specifically) I get depressed and give up lol
I looked at lustre for a recent project and it seems very nice. But the ecosystem is pretty small yet (I could find no examples of auth, for one), so I ended up going with liveview.
I'm hoping it succeeds and gets bigger because I really like its ergonomics.
One thing im wondering with the LLM age we seem to be entering: is there value in picking up a language like this if theres not going to be a corpus of training data for an LLM to learn from? Id like to invest the time to learn Gleam, but I treat a language as a tool, or a means to an end. I feel like more and more I'm reaching for the tool to get the job done most easily, which are languages that LLMs seem to gel with.
In the medium to long term, if LLMs are unable to easily learn new languages and remap the knowledge they gained from training on different languages, then they will have failed in their mission of becoming a general intelligence.
I feel that was more true 1-2 years ago. These days I find Claude Code write almost as good (or as bad depending on your perspective) Elixir code as JavaScript code and there must be less Elixir code in the training data.
I personally find it much more painful to generate valid Rust code that compiles and does what I want than e.g. valid python code that runs and does what I want.
i think it's pretty clear that some of "the things you expect to make an LLM good at a language" (like strong typing) are not actually the case. other things like "don't indirect your code by jumping to something unexpected" might be more important.
If anything llms would be poorer in codegen for static languages because they are more verbose - More tokens to generate and use limited context windows parsing code.
The advantage rather for llms in strongly typed languages is that compilers can catch errors early and give the model early automated feedback so you don’t have to.
With weakly typed (and typically interpreted) languages they will need to run the code which maybe quite slow to do so or not realistic.
Simply put agentic coding loops prefer stronger static analysis capabilities.
not necessarily. if your refactoring loop requires too many cycles you'll fall off the attention context window.
also, some nonstatic languages have a habit of having least surprise in their codebases -- it's often possible to effectively guess the types flowing through at the callsite. zero refactoring feedback necessary is better than even one.
I’ve found Claude (at least until Opus 4) would routinely fail at writing a bash script. For example it would end an if block with }. Or get completely lost with environment variables and subshells.
But those are exactly the same mistakes most humans make when writing bash scripts, which makes them inherently flaky.
Ask it to write code in a language with types, a “logical” syntax where there are no tricky gotchas, with strict types, and a compiler which enforces those rules, and while LLMs struggle to begin with, they eventually produce code which is nearly clean and bug free. Works much better if there is an existing codebase where they can observe and learn from existing patterns.
On the other hand asking them to write JavaScript and Python, sure they fly, but they confidently implement code full of hidden bugs.
The whole “amount of training data” is completely overblown. I’ve seen code do well even with my own made up DSL. If the rules are logical and you explain the rules to it and show it existing patterns, the can mostly do alright. Conversely there is so much bad JavaScript and Python code in their training data that I struggle to get them to produce code in my style in these languages.
the flip side of that, right now ai coding agents try to generate code, not software.
it seems semi intuitive to me that a typesafe, functional programming language with referential transparency would be ideal if you could decompose a program to small components and code those.
once you have a referentially transparent function with input/out tests you can spin on that forever until its solved and then be sure that it works.
Claude reads and writes Gleam just fine. I think as long as the language syntax is well documented (with examples) and has meaningful diagnostics, LLMs can be useful. Gleam has both brilliant docs and diagnostics rivalling Rust. Gleam is also very well designed language wise, not many reserved words, very explicit APIs… also something that helps LLMs.
Contrast with the likes of Swift - been around for years but it’s so bloated and obscure that coding agents (not just humans) have problems using it fully.
If you just see language as a tool, unless you’re self employed or working in open source, wouldn’t the lack of job market demand for it be the first blocker?
If you're fortunate, you can find companies with a passion for good code who use lesser-known languages. Picking Erlang, or Haskell, or OCaml generally filters out candidates who don't share your interest in seeing what can be done outside the mainstream.
I’m more interested in what happens when a language is designed specifically for llms? When doing vibe coding a lot of code is a lot more verbose than I’d do normally. Do we drop down the abstraction level because llms are just so good a churning out boilerplate?
Llms are already good at churning boilerplate, so the next step really is making them good as so they develop taste and architectural consistency imho.
I have similar concerns to you - how well a language works with LLMs is indeed an issue we have to consider. But why do you assume that it's the volume of training data that drives this advantage? Another assumption, equally if not more valid IMO, is that languages which have fewer, well-defined, simpler constructs are easier for LLMs to generate.
Languages with sprawling complexity, where edge cases dominate dev time, all but require PBs of training data to be feasible.
Languages that are simple (objectively), with a solid unwavering mental model, can match LLMs strengths - and completely leap-frog the competition in accurate code gen.
Its pretty much the same thing as in every previous age, where not having a community of experience and the supporting materials they produce has been a disadvantage to early adopters of a new language, so the people that used it first were people with a particular need that it seemed to address that offset that for them, or that had a particular interest in being in the vanguard.
And those people are the people that develop the body of material that later people (and now LLMs) learn from.
I don't know Gleam at all so I can't comment on that specifically, but I think everyone has the experience of a coworker who writes C++ as if it's C or Python as if its Java or whatever else.
A language doesn't have to be unique to still have a particular taste associated with its patterns and idioms, and it would unfortunate if LLM influence had the effect of suppressing the ability for that new style to develop.
I recently built something in Hare (a very niche new language), and Claude Code was helpful. No where near as good as it is with TypeScript. But it was good enough that I don’t LLMs being in the top 5 reasons a language would fail to get adopted.
On the other hand, if you write a substantial amount of code in a niche languages, the LLMs will pick up your coding style as it's in a sizable chunk of the training corpus.
claude is really good at elixir. IME, It's really really good with a few "unofficial" tweaks to the language/frameworks, but this could be my bias. the LLM cutoff was a fear of mine, but i think it's actually the opposite. we know that as few as 250 documents can "poison" an LLM, i suspect that (for now) a small language with very higg quality examples can "poison" LLMs for the better.
Seems like you are not target audience for these new languages and that is OK. But I guess there are still many people that want to try new things (on their own even).
Yes, because LLMs don't change the fact that different programming languages have different expressive capabilities. It's easier to say some things in some languages over others. That doesn't change if it's an LLM writing the code; LLMs have finite context windows and limited attention. If you can express an algorithm in 3000 loc in one language but 30 loc in another, the more expressive language is still preferred, even if the LLM can spit out the 3000 lines in 1s. The reason being if the resulting codebase is 10 - 100x larger than it needs to be, that has real costs that are not mitigated by LLMs or agents. All things being equal, you'd still prefer the right tool for the job, which does not imply we should use Python for everything because it dominates the training set, it means we should make sure LLMs have capabilities to write other programming languages equally well before we rely on them too much.
I've looked at Gleam before but it didn't seem to have any mechanism for dynamic dispatch like interfaces or type classes. Did it change in the meantime?
The answer I’ve seen is “just pass structs of functions around”, which is just one step more explicit than the implicit version we’re all use to, but honestly I kinda like it to free ourselves of all the ceremony around generics.
It’s discouraged to pass around structs of functions to replicate type classes in Gleam. Instead the preference is to not type class style patterns in your projects, favouring a concrete style instead.
Gleam has first class functions, so it has dynamic dispatch.
Both of type classes and interfaces desugar to high order functions, so anything you write with them can be written with first class functions, though with a less concise API.
It's not a "hack" because many language DO NOT let you store functions with state. Gleam does, I write PHP, and that does as well.
PHP has interfaces and whatnot, but a lot of the time I do polymorphism by just having a class that has Closure members. When you can arbitrarily pass around functions like that, it's basically equivalent to an interface or abstract class, with a bit more flexibility.
I keep running into Gleam posts and podcasts and videos. I think it could be an especially attractive alternative to JS for UI with the Lustre library. Haven't tried it yet, my backlog is ever growing...
It’s really good. But it needs generics. This is a huge downside. It’s a typed and clean functional programming language but it arbitrarily followed golangs early philosophy of no generics. Ironically golang is one of the most hated languages among many fp advocates.
By the developers own action of adding generics ultimately the golang team admits they were wrong or that generics are better. If gleam gets popular I think much of the same will occur.
There’s simply too much repeated code without generics. I tried writing a parser combinator in gleam and it wasn’t pretty.
I think he means maybe AI can get around languages lacking features - like how codegen was used for a long time.
Codegen is more and more rare these days, because languages have so many tools to help you write less code - like generics. LLMs could, theoretically, help you crank out similar repetitive implementations of things.
There are multiple levels of possible generics at play here: the container type and the element type. You can have a map/reduce/filter that operate on any element type (generic elements) while still being specialized to linked lists. On the other hand, you might prefer a generic map that can operate on any container type and any element type, so that you can use the same map method and the same function to map over arrays of numbers, sets of numbers, lists of numbers, trees of numbers, or even functions of numbers!
Haskell allows both sorts of generics. In Haskell parlance they call this higher-kinded polymorphism and the generic version of map they call fmap (as a method of the class Functor).
I saw your other comment that you meant interface. But an example of a language that went without a feature people thought the language desperately needed was Go with generics. They only added them more than ten years later, when they figured out the best way to implement them.
It might be the same with gleam, with first version in 2019 and 1.0 in 2024. The language authors might think they are either uneeded and lead to anti patterns, or are waiting to see the best way to implement them.
Why does it need generics? There's a great blog post about how you can replace a lot of trait behaviour with just functions. Maybe something like that can be done for generics
Gleam is a beautiful language, and what I wish Elixir would become (re:typing).
For those that don't know its also built upon OTP, the erlang vm that makes concurrency and queues a trivial problem in my opinion.
Absolutely wonderful ecosystem.
I've been wanting to make Gleam my primary language, but I fear LLMs have frozen programming language advancement and adoption for anything past 2021.
But I am hopeful that Gleam has slid just under the closing door and LLMs will get up to speed on it fast.
> I fear LLMs have frozen programming language advancement and adoption for anything past 2021.
Why would that be the case? Many models have knowledge cutoffs in this calendar year. Furthermore I’ve found that LLMs are generally pretty good at picking up new (or just obscure) languages as long as you have a few examples. As wide and varied as programming languages are, syntactically and ideologically they can only be so different.
There's a flywheel where programmers choose languages that LLMs already understand, but LLMs can only learn languages that programmers write a sufficient amount of code in.
Because LLMs make it that much faster to develop software, any potential advantage you may get from adopting a very niche language is overshadowed by the fact that you can't use it with an LLM. This makes it that much harder for your new language to gain traction. If your new language doesn't gain enough traction, it'll never end up in LLM datasets, so programmers are never going to pick it up.
> but LLMs can only learn languages that programmers write a sufficient amount of code in
i wrote my own language, LLMs have been able to work with it at a good level for over a year. I don't do anything special to enable that - just front load some key examples of the syntax before giving the task. I don't need to explain concepts like iteration.
Also llm's can work with languages with unconventional paradigms - kdb comes up fairly often in my world (array language but also written right to left).
LLMs still struggle with lisp parens though
> Because LLMs make it that much faster to develop software
I feel as though "facts" such as this are presented to me all the time on HN, but in my every day job I encounter devs creating piles of slop that even the most die-hard AI enthusiasts in my office can't stand and have started to push against.
I know, I know "they just don't know how to use LLMs the right way!!!", but all of the better engineers I know, the ones capable of quickly assessing the output of an LLM, tend to use LLMs much more sparingly in their code. Meanwhile the ones that never really understood software that well in the first place are the ones building agent-based Rube Goldberg machines that ultimately slow everyone down
If we can continue living in the this AI hallucination for 5 more years, I think the only people capable of producing anything of use or value will be devs that continued to devote some of their free time to coding in languages like Gleam, and continued to maintain and sharpen their ability to understand and reason about code.
This last week:
* One developer tried to refactor a bunch of graph ql with an LLM and ended up checking in a bunch of completely broken code. Thankfully there were api tests.
* One developer has an LLM making his PRs. He slurped up my unfinished branch, PRed it, and merged (!) it. One can only guess that the approved was also using an LLM. When I asked him why he did it, he was completely baffled and assured me he would never. Source control tells a different story.
* And I forgot to turn off LLM auto complete after setting up my new machine. The LLM wouldn't stop hallucinating non-existent constructors for non-existent classes. Bog standard intellisense did in seconds what I needed after turning off LLM auto complete.
LLMs sometimes save me some time. But overall I'm sitting at a pretty big amount of time wasted by them that the savings have not yet offset.
> One developer tried to refactor a bunch of graph ql with an LLM and ended up checking in a bunch of completely broken code. Thankfully there were api tests.
So the LLM was not told how to run the tests? Without that they cannot know if what they did works, and they are a bit like humans, they try something and then they need to check if that does the right thing. Without a test cycle you definitely don’t get a lot out of LLMs.
Using AI to write good code faster is hard work.
I once toured a dairy farm that had been a pioneer test site for Lasix. Like all good hippies, everyone I knew shunned additives. This farmer claimed that Lasix wasn't a cheat because it only worked on really healthy cows. Best practices, and then add Lasix.
I nearly dropped out of Harvard's mathematics PhD program. Sticking around and finishing a thesis was the hardest thing I've ever done. It didn't take smarts. It took being the kind of person who doesn't die on a mountain.
There's a legendary Philadelphia cook who does pop-up meals, and keeps talking about the restaurant he plans to open. Professional chefs roll their eyes; being a good cook is a small part of the enterprise of engineering a successful restaurant.
(These are three stool legs. Neurodivergents have an advantage using AI. A stool is more stable when its legs are further apart. AI is an association engine. Humans find my sense of analogy tedious, but spreading out analogies defines more accurate planes in AI's association space. One doesn't simply "tell AI what to do".)
Learning how to use AI effectively was the hardest thing I've done recently, many brutal months of experiment, test projects with a dozen languages. One maintains several levels of planning, as if a corporate CTO. One tears apart all code in many iterations of code review. Just as a genius manager makes best use of flawed human talent, one learns to make best use of flawed AI talent.
My guess is that programmers who write bad code with AI were already writing bad code before AI.
Best practices, and then add AI.
I don't think this is actually true. LLMs have an impressive amount of ability to do knowledge-transfer between domains, it only makes sense that that would also apply to programming languages, since the basic underlying concepts (functions, data structures, etc.) exist nearly everywhere.
If this does appear to become a problem, is it not hard to apply the same RLHF infrastructure that's used to get LLMs effective at writing syntactically-correct code that accomplishes sets of goals in existing programming languages to new ones.
> LLMs have an impressive amount of ability to do knowledge-transfer between domains, it only makes sense that that would also apply to programming languages, since the basic underlying concepts (functions, data structures, etc.) exist nearly everywhere.
That would make sense if LLMs understood the domains and the concepts. They don't. They need a lot of training data to "map" the "knowledge transfer".
Personal anecdote: Claude stopped writing Java-like Elixir only some time around summer this year (Elixir is 13 years old), and is still incapable of writing "modern HEEX" which changed some of the templaring syntax in Phoenix almost two years ago.
I bet LLMs create their version of Jevons paradox.
More trial and error because trial is cheap, in the end less typing but hardly faster end results
The motivation isn’t there to create new languages for humans when you’re programming at a higher level of abstraction now (AI prompting).
It’d be like inventing a new assembly language when everyone is writing code in higher level languages that compile to assembly.
I hope it’s not true, but I believe that’s what OP meant and I think the concern is valid!
I would argue it's more important than ever to make new languages with new ideas as we move towards new programming paradigms. I think the existence of modern LLMs encourages designing a language with all of the following attributes:
- Simple semantics (e.g. easy to understand for developers + LLMs, code is "obviously" correct)
- Very strongly typed, so you can model even very complex domains in a way the compiler can verify
- Really good error messages, to make agent loops more productive
- [Maybe] Easily integrates with existing languages, or at least makes it easy to port from existing languages
We may get to a point where humans don't need to look at the code at all, but we aren't there yet, so making the code easy to vet is important. Plus, there's also a few bajillion lines of legacy code that we need to deal with, wouldn't it be cool if you could port (or at least extend it) it into some standardized, performant, LLM-friendly language for future development?
I think that LLMs will be complemented best with a declarative language, as inserting new conditions/effects in them can be done without modifying much (if any!) of the existing code. Especially if the declarative language is a logic and/or constraint-based language.
We're still in early days with LLMs! I don't think we're anywhere near the global optimum yet.
This is why I use rust for everything practicable now. Llms make the tedious bits go away and I can just enjoy the fun bits.
> It’d be like inventing a new assembly language when everyone is writing code in higher level languages that compile to assembly.
Isn't that what WASM is? Or more or less what is going on when people devise a new intermediate representation for a new virtual machine? Creating new assembly languages is a useful thing that people continue to do!
We may end up using AI to create simplified bespoke subset languages that fit our preferences. Like a DSL of sorts but with better performance characteristics than a traditional DSL and a small enough surface area.
I believe prompting an AI is more like delegation than abstraction especially considering the non-deterministic nature of the results.
It does further than non-determinism. LLM output is chaotic. 2 nearly identical prompts with a single minor difference can result in 2 radically different outputs.
> For those that don't know its also built upon OTP, the erlang vm
This isn't correct. It can compile to run on the BEAM: that is the Erlang VM. OTP isn't the Erlang VM; rather, "OTP is set of Erlang libraries and design principles providing middle-ware to develop [concurrent/distributed/fault tolerant] systems."
Gleam itself provides what I believe is a substantial subset of OTP support via a library: https://github.com/gleam-lang/otp
Importantly: "Gleam has its own version of OTP which is type safe, but has a smaller feature set. [vs. Elixir, another BEAM language with OTP support]"
Hi, I’m the creator of Gleam!
The comment you are replying to is correct, and you are incorrect.
All OTP APIs are usable as normal within Gleam, the language is designed with it in mind, and there’s an additional set of Gleam specific additions to OTP (which you have linked there).
Gleam does not have access to only a subset of OTP, and it does not have its own distinct OTP inspired OTP. It uses the OTP framework.
(I know Erlang well, but haven't used Gleam)
The library the parent links to says this:
> Not all Erlang/OTP functionality is included in this library. Some is not possible to represent in a type safe way, so it is not included.
Does this mean in practice that you can use all parts of OTP, but you might lose type checking for the parts the library doesn't cover?
No, it means that one specific package only offers bindings to certain parts. It’s the documentation for one library, not the language.
> Hi, I’m the creator of Gleam!
What's the state of Gleam's JSON parsing / serialization capabilities right now?
I find it to be a lovely little language, but having to essentially write every type three times (once for the type definition, once for the serializer, once for the deserializer) isn't something I'm looking forward to.
A functional language that can run both on the backend (Beam) and frontend (JS) lets one do a lot of cool stuff, like optimistic updates, server reconciliation, easy rollback on failure etc, but that requires making actions (and likely also states) easily serializable and deserializable.
You can generate those conversions, most people do.
But also, you shouldn’t think of it as writing the same type twice! If you couple your external API and your internal data model you are greatly restricting your domain modelling cability. Even in languages where JSON serialisation works with reflection I would recommend having a distinct definition for the internal and external structure so you can have the optimal structure for each context, dodging the “lowest common decimator” problem.
I understand your point, and I agree with it in most contexts! However, for the specific use case where one assumes that the client and server are running the exact same code (and the client auto-refreshes if this isn't the case), and where serialization is only used for synchronizing between the two, decoupling the state from it's representation on the wire doesn't really make sense.
Totally. This is where I would generate them.
>You can generate those conversions, most people do.
Hi, what do people use to generate them, I found gserde (edit: and glerd-json)
There’s several options, depending on what you want. The most commonly used option is the language server.
The language server code action :)
Also. The lisp cam now generate serialisers and deserializers for some types iirc.
This is also what really annoyed me when I tried out Gleam.
I'm waiting for something similar to serde in Rust, where you simply tag your type and it'll generate type-safe serialization and deserialization for you.
Gleam has some feature to generate the code for you via the LSP, but it's just not good enough IMHO.
Multiple of such tools exist and have done for years. Serde isn’t a Rust-core project, and similarly the Gleam alternatives are not Gleam-core.
Rust has macros that make serde very convenient, which Gleam doesn't have.
Could you point to a solution that provides serde level of convenience?
Edit: The difference with generating code (like with Gleam) and having macros generate the code from a few tags is quite big. Small tweaks are immediately obvious in serde in Rust, but they drown in the noise in the complete serialization code like with the Gleam tools.
> Rust has macros that make serde very convenient, which Gleam doesn't have.
To be fair, Rust's proc macros are only locally optimal:
While they're great to use, they're only okay to program.
Your proc-macro needs to live in another crate, and writing proc macros is difficult.
Compare this to dependently typed languages og Zig's comptime: It should be easier to make derive(Serialize, Deserialize) as compile-time features inside the host language.
When Gleam doesn't have Rust's derivation, it leaves for a future where this is solved even better.
In Gleam code generators are most commonly used, similar to in C#, Go, or Elm.
Yes, my point is that it's not a good experience.
I would be very interested in hearing about your experience writing Gleam and the problems you have!
We regularly collect feedback and haven’t got problems reported here, so your feedback saying otherwise would be a useful data point.
I might try to collect my thoughts somewhere when I get some time over.
Thank you for Gleam btw, I do really like the rest of the language.
Fair enough, but to be fair to my statements, the quotes I chose were largely from gleam-lang.org or the Gleam OTP library.
Take for example this section of the Gleam website FAQ section:
https://gleam.run/frequently-asked-questions/#how-does-gleam...
"Elixir has better support for the OTP actor framework. Gleam has its own version of OTP which is type safe, but has a smaller feature set."
At least on the surface, "but has a smaller feature set" suggests that there are features left of the table: which I think it would be fair to read as a subset of support.
If I look at this statement from the Gleam OTP Library `readme.md`:
"Not all Erlang/OTP functionality is included in this library. Some is not possible to represent in a type safe way, so it is not included. Other features are still in development, such as further process supervision strategies."
That quote leaves the impression that OTP is not fully supported and therefore only a subset is. It doesn't expound further to say unsupported OTP functionality is alternatively available by accessing the Erlang modules/functions directly or through other mechanisms.
In all of this I'll take your word for it over the website and readme files; these things are often not written directly by the principals and are often not kept as up-to-date as you'd probably like. Still even taking that at face value, I think it leaves some questions open. What is meant by supporting all of OTP? Where the documentation and library readme equivocates to full OTP support, are there trade-offs? Is "usable as normal" usable as normal for Erlang or as normal for Gleam? For example, are the parts left out of the library available via directly accessing the Erlang modules/functions, but only at the cost of abandoning the Gleam type safety guarantees for those of Erlang? How does this hold for Gleam's JavaScript compilation target?
As you know, Elixir also provides for much OTP functionality via direct access to the Erlang libraries. However, there I expect the distinction between Elixir support and the Erlang functionality to be substantially more seamless than with Gleam: Elixir integrates the Erlang concepts of typing (etc.) much more directly than does Gleam. If, however, we're really talking about full OTP support in Gleam while not losing the reasons you might choose Gleam over Elixir or Erlang, which I think is mostly going to be about the static typing... then yes, I'm very wrong. If not... I could see how strictly speaking I'm wrong, but perhaps not completely wrong in spirit.
Ah, that’s good feedback. I agree, that documentation is misleading. I’ll fix them ASAP.
> Elixir also provides for much OTP functionality via direct access to the Erlang libraries.
This is the norm in Gleam too! Gleam’s primary design constraint is interop with Erlang code, so using these libraries is straightforward and commonplace.
Thanks for the clarification. I've read about Gleam here and there, and played with it a bit, and thought there was no way to directly access OTP through the Erlang libraries.
This can be just my lack of familiarity with the ecosystem though.
Gleam looks lovely and IMO is the most readable language that runs on the BEAM VM. Good job!
the Erlang vm is called BEAM, not OTP. sadly, Gleam's implementation of OTP is not at the same level as Elixir's or Erlang.
Gleam uses regular OTP, it doesn’t have a distinct OTP inspired framework. Source: I’m the author of Gleam.
I wonder why so many have got this wrong across this thread? Was it true once upon a time or something, or have people just misunderstood your docs or similar?
OTP is a very complex subject and quite unusual in its scope, and it’s not even overly clear what it even is. Even in Erlang and Elixir it’s commonly confused, so I think it’s understandable that Gleam has the same problem further still with its more distinct programming style.
"Big Elixir" must be paying people to misunderstand Gleam today eh ;-)
i just implemented a project in elixir with LLM support and would never have considered that before. (i had never used elixir before) - So who knows maybe it will help adoption?
> what I wish Elixir would become (re:typing).
Elixir is slowly rolling out set-theoretic typing: https://hexdocs.pm/elixir/main/gradual-set-theoretic-types.h...
I dunno, my unfounded guess is that gradual type systems are super complex and very hard to get right.
Why use something complex and half working, when you can have the real thing?
They are working towards "the real thing", whatever your definition of real is.
BTW in the 90s people tried to come up with a type system for Erlang, and failed:
--- start quote ---
Phil Wadler[1] and Simon Marlow [2] worked on a type system for over a year and the results were published in [3]. The results of the project were somewhat disappointing. To start with, only a subset of the language was type-checkable, the major omission being the lack of process types and of type checking inter-process mes-sages. Although their type system was never put into production, it did result in a notation for types which is still in use today for informally annotating types.
Several other projects to type check Erlang also failed to produce results that could be put into production. It was not until the advent of the Dialyzer [4] that realistic type analysis of Erlang programs became possible.
https://lfe.io/papers/%5B2007%5D%20Armstrong%20-%20HOPL%20II...
--- end quote ---
[1] Yes, that Philip Wadler, https://en.wikipedia.org/wiki/Philip_Wadler
[2] Yes, that Simon Marlow, https://en.wikipedia.org/wiki/Simon_Marlow
[3] A practical subtyping system for Erlang https://dl.acm.org/doi/10.1145/258948.258962
[4] https://www.erlang.org/doc/apps/dialyzer/dialyzer.html
I don’t mean to minimize the huge effort by the Gleam team; however, Elixir cannot become Gleam without breaking OTP/BEAM in the same ways Gleam does. As it stands now, Elixir is the superior language between the two, if using the full Erlang VM is your goal.
I use many of the otp functions in gleam on thr regular, what functionality cant i call?
Gleam can call any erlang function, and can somewhat handle the idc types. [ im sure it has another name ].
Did i miss something that gleam fails on, because this is one of my concerns.
Gleam is really quite a nice language. I did AoC in it this year as well and came away with the following: (incomplete list for both positive and negative, these are mainly things that come to mind immediately)
Positive:
- It can be pretty performant if you do it right. For example, with some thought I got many days down to double digit microseconds. That said, you do need to be careful how you write it and many patterns that work well in other languages fall flat in Gleam.
- The language server is incredibly good. It autoformats, autocompletes even with functions from not-yet-imported-but-known-to-the-compiler packages, shows hints with regarding to code style and can autofix many of these, autofills missing patterns in pattern matches, automatically imports new packages when you start using them and much much more. It has definitely redefined my view of what an LSP can do for a language.
- The language is generally a joy to work with. The core team has put a lot of effort into devex and it shows. The pipe operator is nice as always, the type system is no haskell but is expressive enough, and in general it has a lot of well-thought out interactions that you only notice after using it for a while.
Negative:
- The autoformatter can be a bit overly aggressive in rewriting (for example) a single line function call with many arguments to a function call with each argument on a different line. I get that not using "too much" horizontal space is important, but using up all my vertical space instead is not always better.
- The language (on purpose) focuses a lot on simplicity over terseness, but sometimes it gets a little bit much. Having to type `list.map` instead of `map` or `dict.Dict` instead `Dict` a hundred times does add up over the course of a few weeks, and does not really add a lot of extra readability. OTOH, I have also seen people who really really like this part of Gleam so YMMV.
- Sometimes the libraries are a bit lacking. There are no matrix libraries as far as I could find. One memoisation library had a mid-AoC update to fix it after the v1.0 release had broken it but nobody noticed for months. The maintainer did push out a fix within a day of realizing it was broken though. The ones that exist and are maintained are great though.
I was also pretty surprised about the performance. It's not C, but it's so much faster than I'd guessed.
I do agree the language server is great. And it works in basically any IDE, which is another huge bonus.
With regards to having to type `list.map`, you actually don't need to! You can do this:
Some libraries just aren't there, and I do wonder how hard it would be to port C libraries over. Something I want to play with!double space before each line for code formatting
Thank you! TIL
I can live with these negatives. What irritates me the most is the lack of if/else or guards or some kind of dedicated case-distinction on booleans. Pattern matching is great but for booleans it can be kinda verbose. E.g.
There are (some) guards available though? You could rewrite your example as:
Guards are a bit limited in that they cannot contain function calls, but that's a problem of the BEAM and not something Gleam could control.Ah right, I remember now.
> Guards are a bit limited in that they cannot contain function calls,
I feel like that's not a small sacrifice.
> but that's a problem of the BEAM and not something Gleam could control.
Could Gleam desugar to a case expression like I wrote above?
Guards also eliminate checking for whether a pattern matching is comprehensive. Or do I remember wrong and the language does not check it at all?
You most likely asked an AI for this. They always think there is an `if` keyword in case statements in Gleam. There isn't one, sadly.
EDIT: I am wrong. Apparently there are, but it's a bit of a strange thing where they can only be used as clauses in `if` statements, and without doing any calculations.
There is though?
https://tour.gleam.run/flow-control/guards/
In Gleam we would typically write this code with `use`, which doesn’t introduce nesting.
> Having to type `list.map` instead of `map` or `dict.Dict` instead `Dict` a hundred times does add up over the course of a few weeks, and does not really add a lot of extra readability.
I did it in F# this year and this was my feeling as well. All of the List.map and Seq.filter would have just been better to be called off of the actual list or Seq. Not having the functions attached to the objects really hurts discoverability too.
Re argument formatting, I'd guess it's because it uses the Prettier algorithm which works like that.
However in my experience it's much better than the alternative - e.g. clang-format's default "binpack"ing of arguments (lay them out like prose). That just makes them hard to read and leads to horrible diffs and horrible merge conflicts.
I like Gleam, but I am somewhat annoyed by the fact, that I don't have the full functional freedom in calling recursive (inner) functions wherever I want. I don't know, why new functional languages do not get this right all the way, straight from some rnrs document or implementation. Another thing is the separate operators like .> and .< and so on. What I liked were of course pipes and pattern matching.
To me it felt less elegant than Scheme (GNU Guile) which I usually use (with nice parallelism if I want to, pipelines, and also pattern matching), and, aside from syntax, conceptually perhaps less elegant than Erlang. On the other hand it has static typing.
I also tried OCaml this year, but there are issues in the ecosystem making a truly reproducible environment/setup, because opam doesn't produce proper lock files (only version numbers) and it seemed silly to not be able to even include another file, without reaching for dune, or having to specify every single file/module on command line for the OCaml compiler. So I was left unsatisfied, even though the language is elegant and I like its ML-ness. I wish there was a large ecosystem around SML, but oh well ...
Might be I should take another look at Erlang soon, or just finally get started with Haskell. Erlang using rebar3 should have proper lock files, has pattern matching, and if I remember correctly no such limitations for calling functions recursively. No longer sure how or whether Erlang did inner functions though.
Heh, similar thoughts! The main difference that I only used Scheme for SICP, and I've used a bit of Haskell.
I like Haskell in theory, but: just to get a hello world takes a lot of CPU and disk space. The standard library is full of exceptions (you can use a different prelude, that opens a whole different can of worms). The ergonomics of converting between the thousand different string types are awful.
So, you being basically me, I have some recommendations:
Idris (2): good stdlib, has dependent types. A beautiful language. The compiler is self-hosted and bootstrapped by lisp - very elegant! The ecosystem is basically nonexistent though.
PureScript: also improves on Haskell in many ways. But, it's more of a frontend language, and though you can do backend, you're stuck with JavaScript runtime. Oh well.
> The standard library is full of exceptions
By the way, the number of partial functions is base that throw compiler warnings is increasing, for example:
https://hackage.haskell.org/package/base-4.21.0.0/docs/Prelu...
I hope it will increase further.
If you like ML-ness try ReScript v12. It occupies a nice middle-ground between OCaml and Gleam.
The `echo` part seemed interesting but it made me think, debuggers need this as a built in feature. If I have
It seems like it should be a common feature to be able to view between each array operation in a debugger without having to augment the code with `echo`The out of bounds handling didn't seem all that good to me. Sure you can filter out undefined. You could also just make a function that returns an empty array if out of bounds, or array of 1 element if not.
I also find grids made of 2 dimensional array to be code small. An array of arrays is NOT A GRID as there is nothing enforcing the inner arrays to be the same length as each other. Also, in efficienct code it would be better to provide a 1 dimensional array and bounds. Now, out of bounds checks based on accessing outside the array won't work. You need to check actual bounds. Again giving preference using a helperI don’t know gleam, but surely
Could be writtenYou're right, but loads of times I just left that there because I probably did something more involved in the map that I ended up deleting later without realising.
This sounds like the kind of situation where the LSP could suggest the simpler code, I'll see if there's an issue for it already and suggest it if not.
Elixir has one opinionated formatter -- Quokka -- that will rewrite the code above properly. It can also reuse linting rules as rewrite policies. Love using it.
That is correct.
Gleam is a great language. It didn't click for me when I was trying it out, but I'm glad to see more people enjoying it.
And I wonder if Gleam + Lustre could become the new Elm.
As a mostly back end dev Elm looked really nice but all the conflict with the creator and then the lack of compiler releases made me shy away a bit.
I have bumped into "the Elm architecture" in other projects though and it was nice.
> all the conflict with the creator
Just so no one misunderstands this. The creator (Evan) didn't get into, or start, any drama himself that I ever noticed. I'd argue he's a very chill and nice dude.
I've been on the edges of the community for probably a decade now (lurker), and all of the drama came from other people who simply didn't like the BDFL and slow releases strategy.
> the lack of compiler releases
I'm a backend dev mostly and use Elm for all my frontend needs. Yes there are some things compiler-side that could be improved, but basically it's fine.
I appreciate not having to keep up with new releases!
Yeah I'm not a front end dev but I do kind of keep track of what's going on in that space. From what I saw it seemed Elm was all but dead. Maybe that's not true, but that was the impression from the outside looking in.
https://iselmdead.info/
I can't believe this is still up tbh. And I can't believe there's still people defending Elm's lack of development
> It’s true that there hasn’t been a new release of the Elm compiler for some time. That’s on purpose: it’s essentially feature-complete.
Last talk I saw by Evan Czaplicki (from the 2025 Scala Days conf) he seemed to be working on some sort of database language https://www.youtube.com/watch?v=9OtN4iiFBsQ
> And I can't believe there's still people defending Elm's lack of development
Why? (I'm one such person defending Elm's lack of development)
I get the sentiment. He feels like everything he wants is there so it doesn't need much more development. But it still feels off.
Not many people use elm directly but it has influenced the design of so many frameworks and libraries.
Right now I’m toying with the idea of building a GNOME application in rust, and the framework I’m using is relm4 which provides elm like abstractions over gtk-rs.
Previously I’ve built web applications with F# and elmish, which again provides elm like abstractions for building F# applications.
The Roc programing language is being developed because Richard Feldman wanted an Elm like experience in more places then just the front end.
Which conflicts? Gleam seems to be released often?
The elm drama
I recently used Gleam + Lustre for a small app that I normally would have built with Elm + PostgREST. It went very well, and I'm now planning to use it for a larger rewrite (of a rails app).
Anecdotically I recently got into Ruby, via Sinatra, and enjoy the experience.
What are you lacking in ruby and rails, besides the types?
This Lustre? https://www.lustre.org/
Seems to be a filesystem, how would it replace a database?
This one is likely what they were referring to: https://github.com/lustre-labs/lustre
No, this one https://hexdocs.pm/lustre/index.html
I've considered playing around with it. Every now and then I tell myself I'm gonna learn front end dev, but as soon as I start looking at it (FE in general not Lustre specifically) I get depressed and give up lol
I looked at lustre for a recent project and it seems very nice. But the ecosystem is pretty small yet (I could find no examples of auth, for one), so I ended up going with liveview.
I'm hoping it succeeds and gets bigger because I really like its ergonomics.
One thing im wondering with the LLM age we seem to be entering: is there value in picking up a language like this if theres not going to be a corpus of training data for an LLM to learn from? Id like to invest the time to learn Gleam, but I treat a language as a tool, or a means to an end. I feel like more and more I'm reaching for the tool to get the job done most easily, which are languages that LLMs seem to gel with.
In the medium to long term, if LLMs are unable to easily learn new languages and remap the knowledge they gained from training on different languages, then they will have failed in their mission of becoming a general intelligence.
I feel that was more true 1-2 years ago. These days I find Claude Code write almost as good (or as bad depending on your perspective) Elixir code as JavaScript code and there must be less Elixir code in the training data.
There's certainly a lot more JS code out there to train on, but the quality of the Elixir code is likely overall much better.
I personally find it much more painful to generate valid Rust code that compiles and does what I want than e.g. valid python code that runs and does what I want.
i think it's pretty clear that some of "the things you expect to make an LLM good at a language" (like strong typing) are not actually the case. other things like "don't indirect your code by jumping to something unexpected" might be more important.
If anything llms would be poorer in codegen for static languages because they are more verbose - More tokens to generate and use limited context windows parsing code.
The advantage rather for llms in strongly typed languages is that compilers can catch errors early and give the model early automated feedback so you don’t have to.
With weakly typed (and typically interpreted) languages they will need to run the code which maybe quite slow to do so or not realistic.
Simply put agentic coding loops prefer stronger static analysis capabilities.
not necessarily. if your refactoring loop requires too many cycles you'll fall off the attention context window.
also, some nonstatic languages have a habit of having least surprise in their codebases -- it's often possible to effectively guess the types flowing through at the callsite. zero refactoring feedback necessary is better than even one.
in my daily experience Claude Code writes better Elixir code than JS (React). Surely this has to do with the quality of the training material
Can’t confirm or deny comparison with JS but I can second that it write decent elixir
The only problem I’ve ever had was on maybe 3 total occasions it’s added a return statement, I assume because of the syntax similarity with ruby
I’ve found Claude (at least until Opus 4) would routinely fail at writing a bash script. For example it would end an if block with }. Or get completely lost with environment variables and subshells.
But those are exactly the same mistakes most humans make when writing bash scripts, which makes them inherently flaky.
Ask it to write code in a language with types, a “logical” syntax where there are no tricky gotchas, with strict types, and a compiler which enforces those rules, and while LLMs struggle to begin with, they eventually produce code which is nearly clean and bug free. Works much better if there is an existing codebase where they can observe and learn from existing patterns.
On the other hand asking them to write JavaScript and Python, sure they fly, but they confidently implement code full of hidden bugs.
The whole “amount of training data” is completely overblown. I’ve seen code do well even with my own made up DSL. If the rules are logical and you explain the rules to it and show it existing patterns, the can mostly do alright. Conversely there is so much bad JavaScript and Python code in their training data that I struggle to get them to produce code in my style in these languages.
the flip side of that, right now ai coding agents try to generate code, not software.
it seems semi intuitive to me that a typesafe, functional programming language with referential transparency would be ideal if you could decompose a program to small components and code those.
once you have a referentially transparent function with input/out tests you can spin on that forever until its solved and then be sure that it works.
Claude reads and writes Gleam just fine. I think as long as the language syntax is well documented (with examples) and has meaningful diagnostics, LLMs can be useful. Gleam has both brilliant docs and diagnostics rivalling Rust. Gleam is also very well designed language wise, not many reserved words, very explicit APIs… also something that helps LLMs.
Contrast with the likes of Swift - been around for years but it’s so bloated and obscure that coding agents (not just humans) have problems using it fully.
If you just see language as a tool, unless you’re self employed or working in open source, wouldn’t the lack of job market demand for it be the first blocker?
If you're fortunate, you can find companies with a passion for good code who use lesser-known languages. Picking Erlang, or Haskell, or OCaml generally filters out candidates who don't share your interest in seeing what can be done outside the mainstream.
It's funny you mention Haskell, because it's one of those languages many love but can't find jobs in it even if they halved their salaries.
I hope this isn't the future of "new" languages. Hopefully newer AI tools can actually learn a language and they won't be the limiting factor.
I’m more interested in what happens when a language is designed specifically for llms? When doing vibe coding a lot of code is a lot more verbose than I’d do normally. Do we drop down the abstraction level because llms are just so good a churning out boilerplate?
Llms are already good at churning boilerplate, so the next step really is making them good as so they develop taste and architectural consistency imho.
I think as ai tools actually learn languages that functional languages will win out as they are much easier to reason about.
it’s easy enough for them as soon as they have an RL environment for the language
This is the answer. We need online learning for our own code bases and macros.
The Gleam language, yes all of it, fits in a context window (https://tour.gleam.run/everything/)
I have similar concerns to you - how well a language works with LLMs is indeed an issue we have to consider. But why do you assume that it's the volume of training data that drives this advantage? Another assumption, equally if not more valid IMO, is that languages which have fewer, well-defined, simpler constructs are easier for LLMs to generate.
Languages with sprawling complexity, where edge cases dominate dev time, all but require PBs of training data to be feasible.
Languages that are simple (objectively), with a solid unwavering mental model, can match LLMs strengths - and completely leap-frog the competition in accurate code gen.
Its pretty much the same thing as in every previous age, where not having a community of experience and the supporting materials they produce has been a disadvantage to early adopters of a new language, so the people that used it first were people with a particular need that it seemed to address that offset that for them, or that had a particular interest in being in the vanguard.
And those people are the people that develop the body of material that later people (and now LLMs) learn from.
Gleam isn’t a very unique language. The loss from generalizing may be less than the improved ergonomics, if not now then as LLMs improve.
I don't know Gleam at all so I can't comment on that specifically, but I think everyone has the experience of a coworker who writes C++ as if it's C or Python as if its Java or whatever else.
A language doesn't have to be unique to still have a particular taste associated with its patterns and idioms, and it would unfortunate if LLM influence had the effect of suppressing the ability for that new style to develop.
This was one of my bigger worries for LLM coding: we might develop path dependence on the largest tools and languages.
I recently built something in Hare (a very niche new language), and Claude Code was helpful. No where near as good as it is with TypeScript. But it was good enough that I don’t LLMs being in the top 5 reasons a language would fail to get adopted.
On the other hand, if you write a substantial amount of code in a niche languages, the LLMs will pick up your coding style as it's in a sizable chunk of the training corpus.
claude is really good at elixir. IME, It's really really good with a few "unofficial" tweaks to the language/frameworks, but this could be my bias. the LLM cutoff was a fear of mine, but i think it's actually the opposite. we know that as few as 250 documents can "poison" an LLM, i suspect that (for now) a small language with very higg quality examples can "poison" LLMs for the better.
Of course there is, especially if you believe that LLMs further improve on reasoning.
Seems like you are not target audience for these new languages and that is OK. But I guess there are still many people that want to try new things (on their own even).
Yes, because LLMs don't change the fact that different programming languages have different expressive capabilities. It's easier to say some things in some languages over others. That doesn't change if it's an LLM writing the code; LLMs have finite context windows and limited attention. If you can express an algorithm in 3000 loc in one language but 30 loc in another, the more expressive language is still preferred, even if the LLM can spit out the 3000 lines in 1s. The reason being if the resulting codebase is 10 - 100x larger than it needs to be, that has real costs that are not mitigated by LLMs or agents. All things being equal, you'd still prefer the right tool for the job, which does not imply we should use Python for everything because it dominates the training set, it means we should make sure LLMs have capabilities to write other programming languages equally well before we rely on them too much.
where do you think the corpus of training data comes from?
I've looked at Gleam before but it didn't seem to have any mechanism for dynamic dispatch like interfaces or type classes. Did it change in the meantime?
The answer I’ve seen is “just pass structs of functions around”, which is just one step more explicit than the implicit version we’re all use to, but honestly I kinda like it to free ourselves of all the ceremony around generics.
It’s discouraged to pass around structs of functions to replicate type classes in Gleam. Instead the preference is to not type class style patterns in your projects, favouring a concrete style instead.
Does that mean pass every needed function as a parameter? Or just don’t write generic functionality?
Yes, one would use higher order functions, as is common in Erlang, Elixir, OCaml, Elm, etc.
Gleam has first class functions, so it has dynamic dispatch.
Both of type classes and interfaces desugar to high order functions, so anything you write with them can be written with first class functions, though with a less concise API.
What you are saying is: no, it doesn't.
Of course dynamic dispatch can be implemented in almost every language. The Linux kernel uses dynamic dispatch with C!
But that's a hack, not a language feature.
I think you might mean “ad hoc polymorphism” rather than “dynamic dispatch”. Gleam, C, Erlang, etc have the latter, not so much the former.
It's not a "hack" because many language DO NOT let you store functions with state. Gleam does, I write PHP, and that does as well.
PHP has interfaces and whatnot, but a lot of the time I do polymorphism by just having a class that has Closure members. When you can arbitrarily pass around functions like that, it's basically equivalent to an interface or abstract class, with a bit more flexibility.
> You can do [first, ..rest] and you can do [first, second].
> But you cannot do [first, ..middle, last].
I don't think you're supposed to do that! It's probably expensive!
fortunately the blog title police stepped in as soon as possible https://news.ycombinator.com/item?id=45426996
I keep running into Gleam posts and podcasts and videos. I think it could be an especially attractive alternative to JS for UI with the Lustre library. Haven't tried it yet, my backlog is ever growing...
Nice, thanks for the write-up! Any pros/cons that jump out to you for someone whose been trying to pick up Elixir on the side?
Thank you! I have not played around with Elixir just yet, sadly, so I can't help there, I'm afraid.
I don't know if I'm the only one experiencing this: It's too hard to find the download link from their website.
Ruined by ligatures. Use them in your own private coding time, not when trying to communicate with others. It obfuscates rather than clarifies.
'Ruined' is a bit strong, isn't it?
I actually agree. I don’t even use them in my private code. I will look into removing them from my blog!
For a quick fix: `font-variant-ligatures: none;` on body or similar.
For users: paste the following in console
I use NoScript and only got them after enabling JS for the blog page
I have this as !important in a global userStyle for code, pre, kbd, samp elements.
It’s really good. But it needs generics. This is a huge downside. It’s a typed and clean functional programming language but it arbitrarily followed golangs early philosophy of no generics. Ironically golang is one of the most hated languages among many fp advocates.
By the developers own action of adding generics ultimately the golang team admits they were wrong or that generics are better. If gleam gets popular I think much of the same will occur.
There’s simply too much repeated code without generics. I tried writing a parser combinator in gleam and it wasn’t pretty.
Gleam has always had generics! There’s no Gleam version without them
Go touted it's lack of features as simplicity. And it is: on the language writing side. Go is an incomplete language masquerading as a simple one.
That’s the worse is better philosophy which maybe doesn’t matter as much anymore especially with AI
What do you mean?
I think he means maybe AI can get around languages lacking features - like how codegen was used for a long time.
Codegen is more and more rare these days, because languages have so many tools to help you write less code - like generics. LLMs could, theoretically, help you crank out similar repetitive implementations of things.
Perhaps this is a silly question but how do you do functional with no generics? Arent they pretty much required for map/reduce/filter?
Sorry my comment was wrong. It’s been a while when I messed with gleam and I remember it was missing a critical thing but I misremembered what it was.
Gleam doesn’t support interfaces. Not generics. You are completely right.
There are multiple levels of possible generics at play here: the container type and the element type. You can have a map/reduce/filter that operate on any element type (generic elements) while still being specialized to linked lists. On the other hand, you might prefer a generic map that can operate on any container type and any element type, so that you can use the same map method and the same function to map over arrays of numbers, sets of numbers, lists of numbers, trees of numbers, or even functions of numbers!
Haskell allows both sorts of generics. In Haskell parlance they call this higher-kinded polymorphism and the generic version of map they call fmap (as a method of the class Functor).
Gleam does have generics.
Most Scheme implementations don't have generics, and you have to deal with a different map function for every data structure.
Gauche has a generic sequence interface which is great, and it's one of the reasons as a Python user I like Gauche as my "daily driver" Scheme.
I saw your other comment that you meant interface. But an example of a language that went without a feature people thought the language desperately needed was Go with generics. They only added them more than ten years later, when they figured out the best way to implement them.
It might be the same with gleam, with first version in 2019 and 1.0 in 2024. The language authors might think they are either uneeded and lead to anti patterns, or are waiting to see the best way to implement them.
Why does it need generics? There's a great blog post about how you can replace a lot of trait behaviour with just functions. Maybe something like that can be done for generics
the comment is wrong, what Gleam lacks is interfaces.
Which feels super strange, but doesn't seem to really be a problem, e.g. imagine a language where you'd write
Gleam wants you to write edit: found an article that explained this a bit better https://mckayla.blog/posts/all-you-need-is-data-and-function...