> When an AttributeError on a builtin type has no close match via Levenshtein distance, the error message now checks a static table of common method names from other languages (JavaScript, Java, Ruby, C#) and suggests the Python equivalent
What benefit does the lazy import have here - if we use the value in a type hint at module scope anyway? Would that require Deferred evaluation of annotations -- which I don't think are enabled by default?
Python is such a weird language. Lazy imports are a bandaid for AI code base monstrosities with 1000 imports (1% of which are probably Shai Hulud now).
And now even type imports are apparently so slow that you have to disable them if unused during the normal untyped execution.
If Instagram or others wants a professional language, they should switch to Go or PHP instead of shoehorning strange features into a language that wasn't built for their use cases.
I think this is just a natural consequence of an easy-to-use package system. The exact same story as with node. If you don't want lots of imports, don't make it so damn easy to pile them into projects. I'm frankly surprised we still see so few supply chain attacks, even though they picked up their cadence dramatically.
This seems a lot more due to an import running arbitrary code because stuff can happen in the top-level of a module rather than only happening in functions. From what I can tell, it seems pretty common for dynamically typed languages and pretty much entirely absent from statically typed ones, which tend to have a main function that everything else happens inside transitively. I guess this makes it easy if what you're writing is something that runs with no dependencies, but it's a pretty terrible experience as soon as you try to introduce the concept of a library.
Interesting, I had no idea that existed! I still think there's a a difference between "here's a hook you can use to run stuff earlier" and "importing any module is fundamentally the same as running it as a script unless the module happens to use a special conditional to wrap stuff inside of" though (and I say this as someone who doesn't go out of his way to defend Go design decisions)
True, but this is yet another code path that isn't exercised until specific conditions happen. That means even more latent application behaviour can go undetected by unit testing and security profiling until the moon is in the right phase, which is a boon for submarine attacks.
Empirically, I have used the current accepted way to do lazy imports (import statement inside a function) before AI coding was even a mainstream thing, for personal code that sometimes needs a heavy import and sometimes doesn’t.
The lazy statement would be an improvement as it allows one to see all the imports at the top where you expect them to be.
As a now deleted comment pointed out, lazy imports had been requested forever. They were rejected forever and were accepted just when BigCorps wanted them.
Python-dev now is paid to shore up the failed Instagram stack.
Is the biggest user larger than the combined set of individual users who had asked for (or would benefit from) the same thing? I honestly don't know, but I don't think that things are always as simple as you're implying in a world where we have the collective action problem.
If you’re asking some some kind of abstract moral value sense, I have idea.
If you’re asking whether project leads give more weight to a single, tangible, vocal stakeholder than they do to unknown numbers of anonymous and lightly-engaged stakeholders? Yes.
I mean, yes, demonstrably, the phenomenon you're describing happens. Your previous comment seems pretty sarcastically dismissing the idea that someone could disagree with this being a good thing though, and I was making a counterargument against the underlying opinion that seemed apparent.
But also great for speed. Larger libraries can take a measurable amount of time to import (even if they have no transitive dependencies). If only some of your code paths actually need the large library then it makes sense to import it lazily. Without lazy you have to do it conditionally which can lead to the imports happening in strange places rather than all being listed out at the top of the file.
> I've left this one to the bonus section because I've never used set operations on Counters and I'm finding it extremely hard to think of a use case for xor specifically. But I do appreciate the devs adding it for completeness.
Yeah, but applied to counters it would be the symmetric difference between multisets, which doesn't have a natural definition. If I understood the proposal they'd be defining it as absolute value of the difference of the counts, which isn't even associative.
If they only considered parities it could be interpreted as addition in F_2, which is more natural, but I'd still agree that it's hard to see how you'd use something like this in practice.
I was so into Python for 10 years, was enjoyable to work in. But have deleted 100k+ lines this year already moving them to faster languages in a post AI codebot world. Mostly moving to go these days.
This is straightforward in the first instance, but how do you see maintenance of those projects going forward - especially adding more complex features ?
I can see one way forward being to prototype them in python and convert.
1. The very first steps are quite simple. Hello world is literally just `print("hello world")`. In other languages it can be a lot more complex.
2. It got a reputation as a beginner-friendly language as a result.
3. It has a "REPL" which means you can type code into a prompt and it will execute it interactively. This is very helpful for research (think AI) where you're trying stuff out and want to plot graphs and so on.
IMO it is undeservedly popular, or at least was. Wind back 10 years to when it was rapidly gaining mindshare:
1. While "hello world" is simple, if you went further to more complex programs you would hit two roadblocks: a) the lack of static type checking means large programs are difficult to maintain, and b) it's really really slow.
2. While the language is reasonable, the tooling (how you install packages, manage the code and so on) was eye-bleedingly abysmal.
3. While the REPL did technically exist, it was really bare bones. It couldn't even handle things like pasting code into it if the code contained blank lines (which it usually does).
However since it has become arguably the most popular language in the world, a lot of people have been forced to use it and so it is actually getting quite decent now. It has decent static types (even if lots of people still don't use them), the REPL is actually decent now (this changed very recently), and there's a new third party tool called `uv` to manage your code that is actually good.
The biggest issue with it now is that it's still horrifically slow (around 50-200x slower than "fast" languages like C++, Rust etc). It is pretty unlikely that that will ever change. People always try to excuse this by saying Python is a "glue" language and you just use it to connect components written in faster languages, but a) that's pure "you're holding it wrong", and b) that only works in some cases where there are nicely separated "slow bits" that can be moved to another language. That's the case for AI for example, where it's all numerical, but for lots of things it isn't. Mercurial was a competitor to Git that was written in Python and lost partly because it was way too slow. They've started writing parts in Rust but it took them 10 years to even start doing that and by then it was far too late.
> what would you suggest?
It really depends on what you want to make. I would pick something to make first and then pick the language based on that. Something like:
* AI: Python for sure. Make sure you use uv and Pyright.
* Web-based games: Typescript
* Web sites: Typescript, or maybe Go.
* Desktop GUI: Tbh I'd still use C++ with QtWidgets. Getting a bit old-school now tbf.
Also Rust is the best language of them all, but I dunno if I'd pick it as a beginner unless you really know you want to get into programming.
the funny thing is that everyone, including myself, posited that python would be the winner of the ai coding wars, because of how much training data there is for it. My experience has been the opposite.
I felt the opposite, because Python isn’t a great language. It won because of Google, fast prototyping, and its ML interop (e.g. pandas, numpy), but as a language it’s always been subpar.
Indentation is a horrible decision (there’s a reason no other language went this way), which led to simple concepts like blocks/lambdas having pretty wild constraints (only one line??)
Type decoration has been a welcome addition, but too slowly iterated on and the native implementations (mypy) are horribly slow at any meaningful size.
Concurrency was never good and its GIL+FFI story has boxed it into a long-term pit of sadness.
I’ve used it for years, but I’m happy to see it go. It didn’t win because it was the best language.
Lambdas are intentionally kneecapped in python because Guido van Robson doesn't want to make a functional language. (As in "functional programming", not that it doesn't work.)
AI benefits from tools to verify its halucinations. That's much easier in a typed and compiled language. Then have a language that can't be monkey patched at runtime and the confidence increases even more.
If you mean "easy to get something out of it" then yeah, it's great.
Typescript wins in terms of training data IMHO, by which I mean that the training data is large enough that AI does great with TS, and the language is (IMHO) superior to Python in many ways.
I personally now use a mixture of Typescript and Rust for most things, including AI coding. Its been working quite well. (AI doesn't handle Rust as well as TS, in that the code isn't quite idiomatic, but it does ok)
It turns out that volume of training data isn't the most important thing. Elixir beats Kotlin and C#, which beat pretty much everything else. Kotlin is probably the sweet spot for most things.
The versioning issue I've seen across libraries that version change in many languages.
I don't tend to hit Python 2 issues using LLMs with it, but I do hit library things (e.g. Pydantic likes to make changes between libraries - or loads of the libraries used a lot by AI companies).
I’ve found recent Claude to be much better in this regard. I think a lot rests on the quality of the harness and the work behind the scenes done to RAG up to date docs or search for docs proactively rather than guessing.
I also don’t have issues with quality of Python generated. It takes a bit of nudging to use list comps and generators rather than imperative forms but it tends to mimic code already in context. So if the codebase is ok, it does do better.
That could be it. I still see LLMs fail a set of static typing challenges that I created a couple years ago as a benchmark. Google models still fail it. I wonder if the lack of typing in a lot of the training data makes python harder to reason about?
The tons of python code would be great training data if there was any consistency across the ecosystem. Yet every project I've touched required me to learn it's unique style.
Then I'd imagine they practically poisoned half the training set because python2 is subtly different.
100k lines is tiny what are you on about, especially in the monolithic app sass world where many Fyll stack apps that handle all business ops are probably written with Django.
Our entire business runs on 300k lines of Ruby (on Rails) and I can keep most of the business logic in my head. I would say our codebase is not exactly “tiny” and just cracking the ceiling into “smal” territory. And comparatively, people probably write even less code in equivalent rails apps to django ones. 100k lines of C++ is miniscule.
Obviously “deleting code” in this context doesn’t mean purging version control history but the current state of the codebase.
No, no, it is not, or at least not in my experience (I do not and never have done web development - medium performance C++ code - I don't see how I could write, understand and support 100K lines of code in this area).
Your experience doesn't match mine. I have, mostly solo, and part time, written multiple codebases that on that kind of magnitude (it is about the level where it still will fit in one person's head pretty easily IMO). It doesn't take much to reach that kind of size. Now, if all of it was super dense and subtle code, then yeah, that would be a lot, but in my experience that's usually a pretty small part of any given codebase.
> in my experience that's usually a pretty small part of any given codebase
Our experiences differ then. Mine is that almost all of the code I write is directly targeted on the usually quite complex problem I am trying to solve. I don't do boilerplate, for example.
I tend not to have much boilerplate (and write abstractions to avoid it), but I do still find there tends to be a lot of supporting code around the 'difficult bits' (TBH, most of the code I write is supporting a small amount of relatively simple but subtle operations, but such is the nature of embedded software). But different codebases are quite different in this regard: this is why such different scales shouldn't be too surprising in different domains.
I am not a python dev but have the utmost respect for the ecosystem.
But damn, with all the supply chain attacks now in the news, could they just make a simple way (for non python insiders) to install python apps without fearing to be infected by a vermin with full access to my $HOME ...
There is little that they can do short of running the programs in a VM. Linux distros aren't engineered to consider applications as something different from the user running them. You need a completely different security model to achieve that and the Python runtime isn't tackle that.
From this example:
Do we finally have "lazy imports" in Python? I think I missed this change. Is this also something from Python 3.15 or earlier?3.15: https://docs.python.org/3.15/whatsnew/3.15.html#whatsnew315-...
> When an AttributeError on a builtin type has no close match via Levenshtein distance, the error message now checks a static table of common method names from other languages (JavaScript, Java, Ruby, C#) and suggests the Python equivalent
Oh, that is such a nice thing.
What benefit does the lazy import have here - if we use the value in a type hint at module scope anyway? Would that require Deferred evaluation of annotations -- which I don't think are enabled by default?
Annotations were made lazy by default in 3.14:
https://peps.python.org/pep-0649/
https://docs.python.org/3/reference/compound_stmts.html#anno...
With 3.15, using lazy typing imports is more or less an alternative to putting such imports behind an "if TYPE_CHECKING" guard.
[dead]
Note that you can work around it by implementing `def __getattr__(name: str) -> object:` at the module level on earlier Python versions
Somehow I have no trouble imagining this being used as a rationale to avoid unnecessary "magic" to the language for years
[dead]
Yes, 3.15+
Python is such a weird language. Lazy imports are a bandaid for AI code base monstrosities with 1000 imports (1% of which are probably Shai Hulud now).
And now even type imports are apparently so slow that you have to disable them if unused during the normal untyped execution.
If Instagram or others wants a professional language, they should switch to Go or PHP instead of shoehorning strange features into a language that wasn't built for their use cases.
> Python is such a weird language. Lazy imports are a bandaid for AI code base monstrosities with 1000 imports
Just because you don’t like a feature doesn’t mean it’s because of AI and bad code.
Too much syntactic sugar causes cancer of the semicolon.
I think this is just a natural consequence of an easy-to-use package system. The exact same story as with node. If you don't want lots of imports, don't make it so damn easy to pile them into projects. I'm frankly surprised we still see so few supply chain attacks, even though they picked up their cadence dramatically.
This seems a lot more due to an import running arbitrary code because stuff can happen in the top-level of a module rather than only happening in functions. From what I can tell, it seems pretty common for dynamically typed languages and pretty much entirely absent from statically typed ones, which tend to have a main function that everything else happens inside transitively. I guess this makes it easy if what you're writing is something that runs with no dependencies, but it's a pretty terrible experience as soon as you try to introduce the concept of a library.
> it seems pretty common for dynamically typed languages and pretty much entirely absent from statically typed ones
Counter-example is Go and init() function.
Interesting, I had no idea that existed! I still think there's a a difference between "here's a hook you can use to run stuff earlier" and "importing any module is fundamentally the same as running it as a script unless the module happens to use a special conditional to wrap stuff inside of" though (and I say this as someone who doesn't go out of his way to defend Go design decisions)
What would your alternative look like?
True, but this is yet another code path that isn't exercised until specific conditions happen. That means even more latent application behaviour can go undetected by unit testing and security profiling until the moon is in the right phase, which is a boon for submarine attacks.
Empirically, I have used the current accepted way to do lazy imports (import statement inside a function) before AI coding was even a mainstream thing, for personal code that sometimes needs a heavy import and sometimes doesn’t.
The lazy statement would be an improvement as it allows one to see all the imports at the top where you expect them to be.
As a now deleted comment pointed out, lazy imports had been requested forever. They were rejected forever and were accepted just when BigCorps wanted them.
Python-dev now is paid to shore up the failed Instagram stack.
I too am outraged that a product would prioritize its biggest users.
Is the biggest user larger than the combined set of individual users who had asked for (or would benefit from) the same thing? I honestly don't know, but I don't think that things are always as simple as you're implying in a world where we have the collective action problem.
If you’re asking some some kind of abstract moral value sense, I have idea.
If you’re asking whether project leads give more weight to a single, tangible, vocal stakeholder than they do to unknown numbers of anonymous and lightly-engaged stakeholders? Yes.
I mean, yes, demonstrably, the phenomenon you're describing happens. Your previous comment seems pretty sarcastically dismissing the idea that someone could disagree with this being a good thing though, and I was making a counterargument against the underlying opinion that seemed apparent.
On most unix-likes all "imports" via shared libraries (e.g. in C / C++) are lazy by default.
But also great for speed. Larger libraries can take a measurable amount of time to import (even if they have no transitive dependencies). If only some of your code paths actually need the large library then it makes sense to import it lazily. Without lazy you have to do it conditionally which can lead to the imports happening in strange places rather than all being listed out at the top of the file.
> I've left this one to the bonus section because I've never used set operations on Counters and I'm finding it extremely hard to think of a use case for xor specifically. But I do appreciate the devs adding it for completeness.
Check out symmetric difference
https://en.wikipedia.org/wiki/Symmetric_difference
Yeah, but applied to counters it would be the symmetric difference between multisets, which doesn't have a natural definition. If I understood the proposal they'd be defining it as absolute value of the difference of the counts, which isn't even associative.
If they only considered parities it could be interpreted as addition in F_2, which is more natural, but I'd still agree that it's hard to see how you'd use something like this in practice.
It's nice that python 3.15 added Iterator synchronization primitives: https://docs.python.org/3.15/library/threading.html#iterator.... These will nicely complement my threaded-generator package which is doing just this but using a thread/process+generator+queue: https://pypi.org/project/threaded-generator/
I was so into Python for 10 years, was enjoyable to work in. But have deleted 100k+ lines this year already moving them to faster languages in a post AI codebot world. Mostly moving to go these days.
This is straightforward in the first instance, but how do you see maintenance of those projects going forward - especially adding more complex features ?
I can see one way forward being to prototype them in python and convert.
Interested in why you'd use Python in the first place? Advice for someone who knows nothing about programming - what would you suggest?
IMO the main reasons people use Python are:
1. The very first steps are quite simple. Hello world is literally just `print("hello world")`. In other languages it can be a lot more complex.
2. It got a reputation as a beginner-friendly language as a result.
3. It has a "REPL" which means you can type code into a prompt and it will execute it interactively. This is very helpful for research (think AI) where you're trying stuff out and want to plot graphs and so on.
IMO it is undeservedly popular, or at least was. Wind back 10 years to when it was rapidly gaining mindshare:
1. While "hello world" is simple, if you went further to more complex programs you would hit two roadblocks: a) the lack of static type checking means large programs are difficult to maintain, and b) it's really really slow.
2. While the language is reasonable, the tooling (how you install packages, manage the code and so on) was eye-bleedingly abysmal.
3. While the REPL did technically exist, it was really bare bones. It couldn't even handle things like pasting code into it if the code contained blank lines (which it usually does).
However since it has become arguably the most popular language in the world, a lot of people have been forced to use it and so it is actually getting quite decent now. It has decent static types (even if lots of people still don't use them), the REPL is actually decent now (this changed very recently), and there's a new third party tool called `uv` to manage your code that is actually good.
The biggest issue with it now is that it's still horrifically slow (around 50-200x slower than "fast" languages like C++, Rust etc). It is pretty unlikely that that will ever change. People always try to excuse this by saying Python is a "glue" language and you just use it to connect components written in faster languages, but a) that's pure "you're holding it wrong", and b) that only works in some cases where there are nicely separated "slow bits" that can be moved to another language. That's the case for AI for example, where it's all numerical, but for lots of things it isn't. Mercurial was a competitor to Git that was written in Python and lost partly because it was way too slow. They've started writing parts in Rust but it took them 10 years to even start doing that and by then it was far too late.
> what would you suggest?
It really depends on what you want to make. I would pick something to make first and then pick the language based on that. Something like:
* AI: Python for sure. Make sure you use uv and Pyright.
* Web-based games: Typescript
* Web sites: Typescript, or maybe Go.
* Desktop GUI: Tbh I'd still use C++ with QtWidgets. Getting a bit old-school now tbf.
Also Rust is the best language of them all, but I dunno if I'd pick it as a beginner unless you really know you want to get into programming.
I'm still on the lookout for a comprehensive Django-like web framework for go. That would be an instant hit for me.
Same here. Django is my last holdout for Python. Everything new is go.
Go is terrible for scientific/ML work though, the libraries just aren't there. The wrapping C API story is weak too even with LLMs to assist.
Try and write a signal processing thing with filters, windowing, overlap, etc. - there's no easy way to do it at all with the libraries that exist.
I think the purpose of go is to write CRUD. Stray from that and you're on your own.
i don’t really see it this way. the value of a token in Python is much higher than it is in lower-level language
Thats very intersting, If I may ask was it from professional projects or personal projects?
Same, I’m not sure how Python survives this outside of machine learning.
All of our services we were our are significantly faster and more reliable. We used Rust, it wasn’t hard to do
the funny thing is that everyone, including myself, posited that python would be the winner of the ai coding wars, because of how much training data there is for it. My experience has been the opposite.
I felt the opposite, because Python isn’t a great language. It won because of Google, fast prototyping, and its ML interop (e.g. pandas, numpy), but as a language it’s always been subpar.
Indentation is a horrible decision (there’s a reason no other language went this way), which led to simple concepts like blocks/lambdas having pretty wild constraints (only one line??)
Type decoration has been a welcome addition, but too slowly iterated on and the native implementations (mypy) are horribly slow at any meaningful size.
Concurrency was never good and its GIL+FFI story has boxed it into a long-term pit of sadness.
I’ve used it for years, but I’m happy to see it go. It didn’t win because it was the best language.
> there’s a reason no other language went this way)
Except of course for those that did, Haskell, Fortran for example.
Lambdas are intentionally kneecapped in python because Guido van Robson doesn't want to make a functional language. (As in "functional programming", not that it doesn't work.)
Guido van Rossum didn't oppose functional programming, but he wanted to keep the language (and the interpreter) simple.
I’m always baffled when language complaints come down to syntax
AI benefits from tools to verify its halucinations. That's much easier in a typed and compiled language. Then have a language that can't be monkey patched at runtime and the confidence increases even more.
If you mean "easy to get something out of it" then yeah, it's great.
Typescript wins in terms of training data IMHO, by which I mean that the training data is large enough that AI does great with TS, and the language is (IMHO) superior to Python in many ways.
I personally now use a mixture of Typescript and Rust for most things, including AI coding. Its been working quite well. (AI doesn't handle Rust as well as TS, in that the code isn't quite idiomatic, but it does ok)
It turns out that volume of training data isn't the most important thing. Elixir beats Kotlin and C#, which beat pretty much everything else. Kotlin is probably the sweet spot for most things.
a lot of the training data is either for python 2 or just generally very low quality
The quality issue doesn't seem unique to Python.
The versioning issue I've seen across libraries that version change in many languages.
I don't tend to hit Python 2 issues using LLMs with it, but I do hit library things (e.g. Pydantic likes to make changes between libraries - or loads of the libraries used a lot by AI companies).
I’ve found recent Claude to be much better in this regard. I think a lot rests on the quality of the harness and the work behind the scenes done to RAG up to date docs or search for docs proactively rather than guessing.
I also don’t have issues with quality of Python generated. It takes a bit of nudging to use list comps and generators rather than imperative forms but it tends to mimic code already in context. So if the codebase is ok, it does do better.
That could be it. I still see LLMs fail a set of static typing challenges that I created a couple years ago as a benchmark. Google models still fail it. I wonder if the lack of typing in a lot of the training data makes python harder to reason about?
The tons of python code would be great training data if there was any consistency across the ecosystem. Yet every project I've touched required me to learn it's unique style. Then I'd imagine they practically poisoned half the training set because python2 is subtly different.
You can test on the device directly, without needing to recompile to try something.
Three things I find unlikely about this:
- You wrote 100K lines of code (I've worked on several large C++ projects that were far smaller)
- You wrote those lines in Python (surely the whole point of Python is to write less code)
- You deleted them (never delete anything, isn't this what modern VCS is all about?)
But whatever floats your boat.
> You deleted them (never delete anything, isn't this what modern VCS is all about?)
The person said: "deleted 100k+ lines this year already moving them to faster languages"
Are you saying that when you move code to another language/rewrite in another language, you leave the original languages code in your repo?
They didn't say they deleted it from their git history. I delete code all the time (doesn't mean its "gone", just that its not in my git head).
Well, they deleted it from somewhere. As I assumed they were using a VCS I assumed they deleted it from that. Or are they really short of disk space?
In this context I would assume deleting code to mean deleting it from the current version of the software, not removing from the VCS history entirely.
100k lines is tiny what are you on about, especially in the monolithic app sass world where many Fyll stack apps that handle all business ops are probably written with Django.
Our entire business runs on 300k lines of Ruby (on Rails) and I can keep most of the business logic in my head. I would say our codebase is not exactly “tiny” and just cracking the ceiling into “smal” territory. And comparatively, people probably write even less code in equivalent rails apps to django ones. 100k lines of C++ is miniscule.
Obviously “deleting code” in this context doesn’t mean purging version control history but the current state of the codebase.
> 100k lines is tiny
No, no, it is not, or at least not in my experience (I do not and never have done web development - medium performance C++ code - I don't see how I could write, understand and support 100K lines of code in this area).
And so, what does your Ruby code actually do?
Your experience doesn't match mine. I have, mostly solo, and part time, written multiple codebases that on that kind of magnitude (it is about the level where it still will fit in one person's head pretty easily IMO). It doesn't take much to reach that kind of size. Now, if all of it was super dense and subtle code, then yeah, that would be a lot, but in my experience that's usually a pretty small part of any given codebase.
> in my experience that's usually a pretty small part of any given codebase
Our experiences differ then. Mine is that almost all of the code I write is directly targeted on the usually quite complex problem I am trying to solve. I don't do boilerplate, for example.
I tend not to have much boilerplate (and write abstractions to avoid it), but I do still find there tends to be a lot of supporting code around the 'difficult bits' (TBH, most of the code I write is supporting a small amount of relatively simple but subtle operations, but such is the nature of embedded software). But different codebases are quite different in this regard: this is why such different scales shouldn't be too surprising in different domains.
Uhm what? All of those things are totally ordinary.
> All of those things are totally ordinary. reply
I would need some evidence of that.
funny how we may have to wait even longer for llms to pick up this update in their pre-training
I am not a python dev but have the utmost respect for the ecosystem.
But damn, with all the supply chain attacks now in the news, could they just make a simple way (for non python insiders) to install python apps without fearing to be infected by a vermin with full access to my $HOME ...
There is little that they can do short of running the programs in a VM. Linux distros aren't engineered to consider applications as something different from the user running them. You need a completely different security model to achieve that and the Python runtime isn't tackle that.