Hacker Newsnew | past | comments | ask | show | jobs | submit | emil-lp's commentslogin

Related. 31 days ago

Reverting the incremental GC in Python 3.14 and 3.15

265 points, 130 comments

https://news.ycombinator.com/item?id=48077924


Using LaTeX makes sense because that's what all journals and conferences expect.

I'm writing two books, both in LaTeX.

I really don't get what the problem is.


Using LaTeX is mostly fine, except for the endless compile times, useless error messages, lack of unicode, etc. like the GP said.

I'm maintaining an internally used LaTeX document class and the development experience is even worse. TeX has no concept of such avant-garde ideas like lists, dictionaries, or namespaces. Things break all the time, and sometimes only when you load three specific packages in a specific order because they all patch each other's routines. I still haven't completely groked the idea of fragile commands and expanding macros. Characters can change meaning depending on context, even the `comment` character (%) or the `escape` character (\), (and I believe even the curly braces) for example when used inside `\path{}` or `\url{}` [1]. It makes a difference whether you comment out line endings or not. The LaTeX3 syntax looks like a bad joke. I mean, look at it:

\ExplSyntaxOn \tl_set:Nn \l_tmpa_tl {A} \group_begin: \tl_set:Nn \l_tmpa_tl {B} \par value~inside~group:~\tl_use:N \l_tmpa_tl \group_end: \par value~outside~group:~\tl_use:N \l_tmpa_tl

\tl_set:Nn \l_tmpb_tl {A} \group_begin: \tl_gset:Nn \l_tmpb_tl {B} \par value~inside~group:~\tl_use:N \l_tmpb_tl \group_end: \par value~outside~group:~\tl_use:N \l_tmpb_tl \ExplSyntaxOff

????

Let's just let it retire and focus our efforts on Typst and pushing publishers to accept Typst.

[1] Just look at all these poor souls trying to achieve something as exotic as putting a URL with a percent sign inside a footnote: https://tex.stackexchange.com/questions/12230/getting-percen...


> except for the endless compile times, useless error messages, lack of unicode, etc.

Some of these have been fixed; see my sibling comment [0] for more details.

> TeX has no concept of such avant-garde ideas like lists, dictionaries, or namespaces. […]. The LaTeX3 syntax looks like a bad joke.

But that is in fact the entire purpose of LaTeX3. I agree that the syntax looks intimidating, but it's actually quite nice once you learn it, and it's written that way to provide namespacing in TeX. Similarly, LaTeX3 defines lists, dictionaries, and most other conventional datastructures.

> Things break all the time, and sometimes only when you load three specific packages in a specific order because they all patch each other's routines.

Hmm, well it depends. The LaTeX kernel and the TeX engines are more stable than nearly all other software, but the third-party packages do indeed break occasionally. But you see similar dynamics play out in most other ecosystems: JavaScript the language is incredibly stable and has excellent backwards compatibility, but if you use 50+ third-party packages, then things do indeed break occasionally.

> Characters can change meaning depending on context

Much like operator overloading in other languages, catcode changes in TeX can indeed be misused and are sometimes confusing, but they're also a pretty useful solution to problems that would otherwise be tricky to solve.

All this isn't to say that TeX doesn't have issues—I criticize LaTeX myself fairly frequently—but most of these are due to the fact that LaTeX is 40-year-old software built on a 50-year-old engine, and has remained backwards-compatible with documents throughout that entire time. And La(TeX) is slowly modernizing, so I'm fairly hopeful that things will continue to improve.

[0]: https://news.ycombinator.com/item?id=48515090


Thanks for your insight, much appreciated!

However, regarding this:

> Much like operator overloading in other languages, catcode changes in TeX can indeed be misused and are sometimes confusing, but they're also a pretty useful solution to problems that would otherwise be tricky to solve.

I'm sorry, but I've never seen overloading of such fundamental characters like the comment character or escape character anywhere. Or at least if you use these characters inside a string, it's pretty clear that the string context is special. In LaTeX I have no way of knowing which catcodes a macro has modified without essentially parsing the entire thing, which breaks syntax highlighters and language servers (something that increases quality of life in other languages substantially), because the compile times are prohibitive. The decision to let users redefine %, \ and literally every character seems like a really, really bad idea to me.

Other languages and syntaxes seem to do just fine, so I'm not sure what you mean by tricky to solve.

> most of these are due to the fact that LaTeX is 40-year-old software built on a 50-year-old engine, and has remained backwards-compatible with documents throughout that entire time

I realize that, and I appreciate what LaTeX (and by extension TeX) has done. It's a giant in sciences and the software world, of absolutely critical importance, but still. We learned a lot of lessons about writing software in the last 50 years, and Typst is applying these from the ground up. Unfortunately I don't have a lot of confidence that LaTeX can be modernized.


> I'm sorry, but I've never seen overloading of such fundamental characters like the comment character or escape character anywhere.

I believe that Racket [0], Mathematica [1], Raku [2], and Rust [3] let you assign arbitrary meanings to most symbols, but these are indeed much more restricted than TeX is (and for good reason). But the issue is really more that TeX barely supports lexical scoping, and that it lets you change the global catcodes at any point in the document, since changing the meanings of characters before any code runs or only in the middle of a scope is pretty useful.

> Or at least if you use these characters inside a string, it's pretty clear that the string context is special.

Verbatim is essentially equivalent to strings in other languages, and it mostly works pretty well, aside from the huge problem that it's impossible to nest it or pass it as an argument to most macros.

> In LaTeX I have no way of knowing which catcodes a macro has modified without essentially parsing the entire thing

Agreed, this is pretty annoying, but the only consolation is that most documents don't change their catcodes very often (since it's usually a pretty terrible idea).

> Other languages and syntaxes seem to do just fine, so I'm not sure what you mean by tricky to solve.

Texinfo [4] is what I was mainly thinking of, since that is able to completely redefine TeX's syntax without needing to manually implement parsing itself (which would be the best strategy today, but was less feasible back when computers were much slower). Similarly, active characters are pretty useful (this is how ~ is defined to insert a non-breaking space, and is also useful for faux-Markdown [5]).

> We learned a lot of lessons about writing software in the last 50 years, and Typst is applying these from the ground up.

Yup, I first learned about Typst on the first day that it was released, and even then I thought that it had a good chance of suceeding, since it solved the problems that most users had (bad error messages and slow compile times), whereas the other TeX competitors focused on things like better typesetting quality, better extensibility, or easier programability, which most users don't care about at all.

I would be personally a little disappointed if Typst replaced LaTeX, but until that happens, I definitely hope that it continues to do well.

> Unfortunately I don't have a lot of confidence that LaTeX can be modernized.

There are lots of other non-LaTeX TeX formats that are quite modern (ConTeXt [6] is my personal favourite, but OpTeX [7] is nice too), and even LaTeX itself has improved quite a bit over the last 5 years or so [8]. But yeah, it's modernization process is still much slower than nearly any other piece of software, so I am also worried that this may end up being too little too late.

[0]: https://docs.racket-lang.org/guide/language-get-info.html

[1]: https://reference.wolfram.com/language/Notation/tutorial/Not...

[2]: https://docs.raku.org/language/slangs

[3]: https://doc.rust-lang.org/book/ch20-05-macros.html#function-...

[4]: https://en.wikipedia.org/wiki/Texinfo#Texinfo_source_file

[5]: https://tex.stackexchange.com/a/236457/270600

[6]: https://www.ctan.org/pkg/context

[7]: https://petr.olsak.net/optex/

[8]: https://www.latex-project.org/news/latex2e-news/ltnews.pdf#s...


> I would be personally a little disappointed if Typst replaced LaTeX, but until that happens, I definitely hope that it continues to do well.

I'm curious about why you'd be disappointed. Is it because you think typst is an irremediably inferior technology to you or is it because you are invested in LaTex?

I do not mean to ask this in accusatory way. I'm not very well informed about typst or tex (I've only used typst once and I thought it was easy and nice).

I'm just curious about what is worse about typst that would make someone disappointed if that was the reason why you would be.


It's mostly that I'm quite personally invested in TeX, far more than most people are. I work on it nearly every day, and I'm even hosting an international TeX conference about a month from now [0].

I do have a few minor concerns about Typst, mainly that its parsing/compiling strategy allows less introspection and extensibility than TeX does (but this is necessary for its fast compile times), and the fact that it's backed by a for-profit company seems less robust than TeX's 50-year history of volunteer maintainership. And of course Typst is still missing quite a few features, but if it's successful I can't imagine that this will be an issue.

Also, I personally mostly use ConTeXt [1], which compiles much faster than LaTeX and has a very uniform syntax, so switching to Typst wouldn't have as many benefits compared to if I were switching from LaTeX.

So if Typst were to take over, I'd mainly be disappointed that my very deep knowledge of TeX would be no longer relevant. But I also wouldn't be surprised if Typst ultimately revitalized TeX, since it might motivate us to improve TeX even more, and Typst might be a good entrypoint for users interested in plain text typesetting (who might then switch to TeX if they find Typst too limiting).

[0]: https://tug.org/tug2026/

[1]: ConTeXt uses the same underlying TeX engine as LaTeX, but different macros. See https://tug.org/TUGboat/tb46-3/tb144berry-engines-formats.pd... for a brief overview.


Excuse me for going meta, but your series of comments here have been extraordinarily informative, wise, and a pleasure to read. Thank you for taking the time (and for your work on TeX).

> It's a lot more practical to navigate than Emacs' undo, too.

What the heck are you talking about?

Emacs undo is simply the State monad over a zipper into a persistent tree of buffer states.

I can't see how you could make it more practical?


I have a similar RCE for Midnight Commander.

1. Download my virus

2. Open Midnight Commander

3. Launch virus via MC.


Remember that OpenAI is subsidized from here to the highway.

A better way to model this, since you seem interested is the following:

How much would it cost you to start such a service for, say, 10k users?

Any other internet service has had virtually Zero cost, $0. Google, Facebook, youtube, Wikipedia, you name it. They all went into the dumpster to pick up a thrown away desktop computer, and they could serve up towards 100k if not a million users.

How much would it cost you to serve, say, 10k simultaneous users with a SOTA model? And if you wanted to go cash positive after a year, how much would each user have to pay?


> How much would it cost you to serve, say, 10k simultaneous users with a SOTA model? And if you wanted to go cash positive after a year, how much would each user have to pay?

My post has this same argument - we have multiple third party companies running open weight models. They are obviously not subsidised. And people are willing to pay for it. And these models are as good as the SOTA models from last year. So this kinda proves my point that SOTA is sustainable.


I didn't find the answer there, that's why I asked.

What hardware is needed, how much of it, cooling, and what does it all cost you?

Or are you saying I can take my old desktop and serve Deepseek v3.2 to 10k users simultaneously and it would cost me about $1 per megatoken?


I'm simply saying this: there are third party hosters of Open Weight models like deepseek and they have been doing this for a while.

Obviously they are not subsidised, do you disagree? If you agree, they have a way to price it at a point that people wanna pay for it and also they aren't losing money.

So there's nothing inherent about inference that makes it too costly or whatever.


> I'm simply saying this: there are third party hosters of Open Weight models like deepseek and they have been doing this for a while.

> Obviously they are not subsidised, do you disagree? If you agree, they have a way to price it at a point that people wanna pay for it and also they aren't losing money.

> So there's nothing inherent about inference that makes it too costly or whatever.

Do we have audited GAAP financial data for any of these companies? If we don't, all these are... vibes, man.


Good questions. The post you are replying to is also based on vibes. So how about we make a bet and come back after IPO filings?

But isn't it okay to suggest that random 3rd party hosting companies are ... not losing money? Why?


I'll bet 50€ that big AI companies (OpenAI, Anthropic, xAI) have bad regarding their core financials. Excluding one time deals such as renting hardware, external cash infusions, the core AI model business (training models and selling inference) is unprofitable and will be be at the expected scales (50 billion € or more) for the next 5 years at least.

I bet that within 5 years they will be sold for scrap to bigger companies and will become divisions inside them.


This is not fair, lets just speak about Inference margins. You are bringing 100 other things to make it more ambgigous.

    Statement: margins API prices of all models are greater than 10% in Anthropic. 
Feel free to either agree to what I'm saying or bet otherwise.

1. You can't have a model without training. Training is part of the cost. A model trained 5 years ago is borderline useless. Also models cannot be retrained continuously, we do not have the technology (even if we had it, it would just increase the ongoing operational cost).

2. Rephrasing your statement:

  At June 2026 prices, profit margins for all Anthropic models are >10%.
That claim is super defensive. Serving a model can only be done if the model has been built and trained, can't have it any other way. Building it and training it costs lots of money.

Even so, fine, I'll take that bet. Anthropic inference prices are still marginally subsdizided. Once they're public they will hike their API prices several times over the next 24 months. Even that might not save them, because when we take all their expenses into account, they will probably need to raise prices 2-3x compared to their June 2026 prices.


> Even so, fine, I'll take that bet. Anthropic inference prices are still marginally subsdizided. Once they're public they will hike their API prices several times over the next 24 months. Even that might not save them, because when we take all their expenses into account, they will probably need to raise prices 2-3x compared to their June 2026 prices.

Ok lets state the bet like this: 5 months after their IPO, it will be clear that their API prices still have greater than 10% margin.


Since you couldn't answer, I asked ChatGPT.

It said: upfront investment: $3M to $6M.

Customers should pay $25k per month.

Checks out


I pasted everything you wrote and ChatGPT said

>The "$25k per month" figure is almost certainly the result of ChatGPT making assumptions, not a fact derived from any known business model.

https://chatgpt.com/share/6a28193b-6ec0-8333-a1af-d07e8d89ef...

Your whole calculation is also ridiculous - I think you assumed what revenue per month is required to pay off hardware within a year? Why would I use hardware within a year?

I would suggest consulting with ChatGPT and coming up with a better and more coherent argument.


I reran the numbers, break even after four years. 10k users.

Minimum upfront: about $15M

Comfortable upfront: $20M–$25M

Monthly revenue needed: $900k–$1.5M

Required price per 10k individual customers: $99–$149/month

API-equivalent output price: usually $8–$20/M output tokens, unless utilization is very high.


dawg -- what is your point? please write your statement clearly and we can discuss.

My apologies for not being clear about my point.

I try to do some napkin math of what it takes to start a company serving an LLM to customers. I thought maybe 10,000 users sounded like a reasonable number.

I'd like also to compare it to traditional Internet companies, e.g. Twitter. I'd guess Twitter with 10k users would cost me literally a dumpster dive, so essentially no cost.

If I were to start an LLM company, what would my initial investment look like?

Turns out around 15+ million dollars assuming I would get 10,000 users willing to pay 200$/month.

I just don't see how the numbers you show add up.


Did you read the paper? Is there something specifically you're missing? A proof? A theorem statement?

this is an empirical engineering paper with theoretical dressing, it would not need to be a theorem paper of course.

You don't think this is relevant to his perspective?

> In 2025 he left academia to become "Founding Mathematician" of Axiom Math in Palo Alto, California to research the application of artificial intelligence to fundamental mathematics.


No.

Understand first what Axiom Math does (https://www.youtube.com/watch?v=abYcV5LHMG4). It is only after he realized the possibilities of AI that he started Axiom Math with a student he mentored.


Not only is it not copyrightable, it's likely someone else's.

That said, you can still distribute it under a licence, it just means that it's not necessarily enforceable, but that's ultimately for the judge to decide.


Definitions ~ Types

Lemmas ~ Helper functions

Theorems ~ API


Tokens are their food, it's literally what keeps them alive.

Not feeding them tokens is neglect.

I try to feed them a healthy diet.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: