> state of the art machine intelligence
> billions of dollars of funding
> PhD requirement to even enter the field
uses fucking python. is the entire ML field secretly composed of midwits?
why the fuck does ML use python
> state of the art machine intelligence
> billions of dollars of funding
> PhD requirement to even enter the field
uses fucking python. is the entire ML field secretly composed of midwits?
What language would you use?
ASSEMBLY
rust
C++
typescript
go
anything that provides some ability to incorporate it into a useable product
Go, Rust - none or limited ML backend library support
typescript - just no
C++ is usable, sometimes ML frontend is even written in it, but C++ is just so good damnt hard to debug.
Python - excelent write speed, easy to change code, excelent binding to ML backends
>C++
>hard to debug
what? the debugging tools for it are excellent and anything else is a chair-keyboard interface error
woah you hecking owned him bro, bazinga, dude!
he's right.
Oh come on he got his ass
>nooo but he didnt say nìgger in his post he's not a real heckin BOTnerrino
For
t. underageb&
>C++
>the debugging tools for it are excellent
lmao bro this is Stockholm's syndrome. Please try to phone a friend or family.
please, in detail, explain how they arent
surely you wouldnt just be saying "C++ is le bad LOL!" just because its funny and everyone else does it, would you?
Hold on a sec bro I have a problem with my
std::map<std::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::less<std::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::basic_string<char, std::char_traits<char>, std::allocator<char> > const, std::basic_string<char, std::char_traits<char>, std::allocator<char> > > > >
aha XD bro templates have long names thats so funny dude LOL
so basically you have no point and you just repeat what every other nocoder says. people like you should not have access to this board
>sees unreadable log because C++ doesn't have reflection and debuggers have to hack something togethertogether so they just fully expand types
>haha XD it's le funny joke and not a fundamental flaw of the language
1. works on my machine
2. not a flaw for me
Do yourself a favour and try another language.
i used many and i only like c, c++, lua and flavors of assembly similar to 8086
shut the fuck up
>typescript
>rust
only a microsoft chud would even propose this.
There are a few programming languages that could be used for ML, as much as I love javascript it isn't one of them, although with the amount of code written in it it could be possible.
real programming languages for ML
>ruby / crystal lang
>go
>C/C++
>There are a few programming languages that could be used for ML
it needs optimized libraries for linear algebra, statistics, and calculus. it also needs to be able to use GPUs or TPUs at the moment. otherwise your titty jiggle CNN is gonna take months to train
>real programming languages for ML
>>ruby
Are we living in the same universe?
ruby is not more suited to ML than JS
Braindead retard
good luck getting anything done
LMAO, castrate yourself, my friend. You don't know what you are talking about. Like, seriously, consider suicide.
Ruby
Identical except no libs.
No, ruby is red. Python is green.
Assmebly
common lisp, the actual answer
fuck low level langs
fuck webshit langs
we need something actually powerful and not primitive.
> common lisp, the actual answer
kek, they ditched it even at in AI for a reason
>for a reason
let's hear it anon, ill wait
ask the MIT AI department, they would explained better than me
What do you think? Do you think the government made lisp illegal you retard?
>yeah i actually have no idea what I'm talking about, ill just post something condescending and pretend like I'm right
every time
That posts tell a lot more about you than me
Have you ever used PDDL?
Have you ever implemented a ML algorithm?
anon i haven't forgotten about that """reason""" you mentioned.
do you actually have one or were you talking out of your ass?
The reasons can be inferred from the post if you weren't a pea brain retard
lisp and functional style in general have no advantage in AI algorithms (especially in ML algos, being supervised, unsupervised or reinforcement)
Natural framework of those problems is linear (and non) algebra and statistics, and they don't fit better in a non typed functional language (quite the opposite)
The only fit for language like lisp are the niche were **fully** declarative language (like prolog) were the best anyway
Adding
spoonfed, lisp was always slow as molasses
Go eat your daily toejam and stop posting please
Again, do you ever wrote a ML algorithm
are you actually stupid? Meme networks are literally a chain of functions being composed.
>reasons to ditch lisp
1.) Ugly syntax
2.) Too many dialects
3.) No standard implementation
4.) Low user libraries
5.) Hard to teach
6.) Niche advantages that are rarely used by the average lisp programmer
>reasons to use python
1.) Simple syntax
2.) 2 dialects, but one is deprecated and is slowly being erased
3.) 1 accepted implementation with other options for optimization
4.) Plentiful user libraries
5.) Easy to teach
6.) C bindings means python doesn't need to pull niche tricks to be fast.
It's amazing how retarded nuBOT is. Posts like these make me want to find another ib because clearly this one is lost for good.
>pure syntax
Theoretically pure, but that theory means nothing for the average programmer who has to deal with the ugly syntax.
>common lisp
Sure that's your favorite, but there are plenty more competing and still used variants.
Guile
Emacs lisp
Closure
Clisp
Racket
(The giant scheme gap)
>implementations
Same as above
>low 3rd party libraries
Unpopular because it is split between all the different dialects. Also popular is a valid metric to measure the value of a language.
>hard to teach
Its easy at first until you start dealing with all the theory behind meta programming that no one wants to talk about.
>niche advantages that are rarely used
Well for one everyone wants to talk about how they can in theory treat functions as a list, but I have rarely seen anyone do this.
>but that theory means nothing
no. it is not just "theoretically pure".
it's unironically easier to learn than python or C syntax since it's literally just (function arg1 arg2 ...)
that's an advantage since the average programmer can worry less about syntax and more about their actual project. the fact that you're not used to it yourself isn't an argument or a fault of the language
>Sure that's your favorite
no. in my initial post, i specifically said common lisp would be more suitable than python for ML. i never mentioned clojure or scheme or anything else. you're adding that in on your own.
>implementations
same as above
>Unpopular because it is split
maybe it's true, but like i said, that's not a fault of the language.
>Also popular is a valid metric to measure the value of a language
no it's not.
this essay addresses the popularity of some languages very well
http://paulgraham.com/icad.html
>Its easy at first until you start dealing with all the theory behind meta programming that no one wants to talk about.
depends on the complexity of the macros you're writing. your average programmer won't write big complex macros often, if at all. simple macros are actually rather easy. try it out sometime
>Well for one everyone wants to talk about how they can in theory treat functions as a list, but I have rarely seen anyone do this.
i mentioned the advantages of lisp above. those should be enough to show that lisp does have tangible advantages for the average programmer. about that specific one, it is used in every lisp program under the sun. when people talk about "treat functions as a list", they mean macros. while you may not write many macros yourself, you most certainly will use them often in any lisp program. built-in or from libraries or whatever, but all it means is that you can use lisp code to modify lisp code like if it was any regular old list, because it is
>unironically easier
It is about as easy as python at best. You still need to learn about all the subtle shit and the standard libs that you would have to learn in python. Also people generally have a harder time grasping functional programming than imperative programming and this can be shown by which is more popular.
>muh post
Sure, but justify that to the google engineers that wrote the tensor libraries or to the data engineers that need to write in it.
>split is not the fault of the language
It kind of is. The spec is too hands off. When you look at most other programming languages they don't have this problem.
>popularity is not a valid metric
Yes it is. This essay depends on the mythic hacker. This titan stands above the industry and commands the directions that people go. This man doesn't exist. Languages get popular because of work. The language has to do something better than everyone else. The language has to justify its use. People have to sell it to their companies to use it. Languages that remain popular were sold to their companies and did what the company expected of them. Languages that were only temporarily popular did not. The languages mentioned in essays didn't remain popular. Perl, lisp, tcl, tk are all dead or on the decline. Ruby another smash hit is on the decline as well. Python has been around a long time and hasn't wavered in popularity. I would say that, that says leagues about the language. The mythic hackerman liked it and the shareholders and people who actually have to maintain the code after hackerman fucks off to google like it.
>macros
They are part of the reason lisp died. It turned every lisp project into its own subdialect of a dialect.
Delusional has a face and it's yours.
>delusions
Kek
>The google engineers should have used common lisp because I said so
>Lisp is easier to learn and use that's why its so niche
>language split and drift is not the fault of a language that barely had a spec
>I worship mythic hackerman, but when mythic hackerman chooses python to write the ML libraries I just assume mythic hackerman was wrong and that me Joe lisp shelp knows better
>The best languages are the ones that died after a few years
>every lisp guide says to avoid macros like the plague if you can, but they are the best part of the language
Forgot your meds again. Go get 'em, they're good for you.
>t. copegays when they actually have to back their arguments instead of throwing out 4chan buzzwords
Kek
ML is a scientific problem. You don't need state of the art theoretic CS discoveries to solve a optimization problem.
You need a good language, fast and simple enough
That's why lisp is not used in ML industry
>ML is a scientific problem.
True
>You don't need state of the art theoretic CS discoveries to solve a optimization problem.
False, a large part of the field is concerned with discovering new optimization methods and several parts of the field work squarely in the theoretical CS arena.
>That's why lisp is not used in ML industry
>industry
No backepdaling please.
It is patently false that lisp was not used for ML. Until 2009-2010, there was https://lush.sourceforge.net/ for example. Theano got the ball going and the rest was purely momentum.
Lisp is a superior tool than python for ml in all areas that matter except libraries, but the libraries didn't exist in python at the time it began being used either so even that's not a real argument.
The problem is that nobody knew lisp anymore at the time, pure and simple.
> False, a large part of the field is concerned with discovering new optimization methods
in math/algos (and parallalelization), not in programming language feature you idiot
if you think that find a speedup in matrix product is a theoretical cs problem you're irremediably hopeless
> No backepdaling please.
take your meds, no one is backpedaling. You are imagining things
btw
>lisp is not used
>well lisp was used
I think you need a nap
> Lisp is a superior tool than python for ml in all areas
just no. Programming language is almost irrelevant to ML, with rare exception
>retard doesn't know what optimization is or does
>thinks ML finds solution by magic or something
>calls others retarded
>thinks it's OK for data processing to take so long it dwarfs epoch time
OK schizo, whatever you say.
>retard doesn't know what optimization is or does
You are the one which is confused at what level the optimisation speedup is done.
The optimization is done at math level most of the time (approximation of proper gradient recursion, truncation of gradient, assuming adaptive weights uncorrelated with past data, synchronization of time and iteration in recursion, matrix inversion lemma, convenient matrix decomposition, etc)
And on top of that, parallelization Whenever is possible
>thinks ML finds solution by magic or something
I never sad that. ML finds solution by crunching numbers and whatever crunch it fast while remaining simple is the best
You don't need hygienic macros, TCO, cons cell, metacircular evaluation or whatever you deranged lispgay claim
>calls others retarded
I called you because you are one
>thinks it's OK for data processing to take so long it dwarfs epoch time
Lisp is no faster than python. Data (pre)processing is outside the ML core concepts btw
>OK schizo, whatever you say.
Kys retard
nta but lithp was once popular as a machine learning language. part of the reason it nearly died was because the AI winter destroyed the lisp machine market.
No, it was used in symbolic strong ai/robotics (not even lisp, but lisp based languages)
For ml it was never that popular, because ml at those times were mostly algorithms for controllers and sensors
> controllers
*control problem is a better definition
>It is about as easy as python at best.
no. it's easier. you're just having baby duck syndrome. also you can do imperative programming in common lisp just fine.
>muh post
wtf? you're the one replying to my post which was specifically about common lisp kek
and about the implementations, you're really trying hard to make it sound like it's worse than it is.
>muh popularity
did you read the essay i linked?
>macros are part of the problem
i was just explaining the feature you asked about.
unironically, learn common lisp. then form real opinions about it instead of parroting BOT memes. I'm really not trying to be condescending, but i'm honestly tired of people coming in and saying the same retarded things when they obviously have not looked into it at all
>baby duck
Why did the industry centralize around imperative? Especially when lisp came out in the 50's? Thus most would have baby ducked to it. Its because functional languages are too closely related to calculus and set theory and not enough to language. Functional language patterns expect verbs to memorize information about nouns because that makes sense in set theory when mapping between sets. When talking about what is intuitive to people we don't think about state like that. Mapping between sets is already an abstract concept and then imagining problems like that is another layer of abstraction.
>it can be imperative
Almost no one uses lisp like that.
>you are responding
Yes because you argue it was a mistake not to implement it in lisp, but then ignore the fact that lisp is less popular and more fragmented than python. Implementing in common lisp would only allow a subset of the lisp community to use the ML libraries while implementing in python would allow all of the python community to use it.
>did you read the essay I linked
Yes, but his argument fails to materialize in real life. The mythic 20 to 1 ratio only really applies in C and more verbose languages and isn't as big as a lead in python. If you were to use his approach for something like ML you would require people to learn both lisp and the DSL that would come with tensor. Also his approach has a fragmenting effect on reusability as every program or library that you would write would end up with its own DSL. This again is an issue of sub fragmenting a language.
>learn common lisp instead of parroting BOT memes
What BOT memes? Who even talks about common lisp here except people on the lisp generals. Who are all pro lisp by the way. This was my take on lisp after learning guile. When reading the docs on that the whole take was don't use macros unless you absolutely have to.
It's a waste of time
Lispgay are like Jehovah's witnesses
They are all schizo who think lisp is the best programming language ever conceived and the fight about if that particular language is really a lisp or they fight their imaginary enemy language
I learned lisp, it was cool but nothing out of this world honestly
Honestly enjoyed Haskell more than lisp. Lisp just becomes a mess of parenthesis too quickly even when you have an editor or an IDE to manage it, it is a pain.
I never touched Haskell but I felt the same for ocaml. It was nicer to me
>Lisp just becomes a mess of parenthesis too quickly even when you have an editor or an IDE to manage it
and the truth finally surfaces
you could have just said "i cant into parens" that's literally the real gripe gays have with lisp, everything else is just cope
people could just press a button on the editor and activate a form of m-exps without too many parens to lisp, but they are so limited and without creativity that they don't even consider that possibility.
It takes a week to teach lisp fully so that students become productive.
For python, it takes at least 6 months.
You are delusional.
>muh favorite
>muh flavor
>unironically lists elisp
b8/8 I responded
>For python, it takes at least 6 months.
lol
That's right, normal students aren't as slow as you and don't typically require the 10 years and counting you've been at it. But don't give up, timmy, one day you'll be able to program something more interesting than fizzbuzz! Ganbatte~
Wow you sure told him anon
anon, you're responding to someone with the same position as you.
he's saying lol because python can be learned very quickly.
>t. has never touched lisp
>Ugly syntax
pure syntax
>Too many dialects
i specifically mentioned common lisp above.
>No standard implementation
SBCL is what everyone uses nowadays. ANSI CL is the standard all the implementations must follow
>Low user libraries
less than python, but that's just because it's unpopular. nothing to do with the language itself being bad.
>Hard to teach
false. in fact it's probably one of the easiest languages to teach, precisely due to it's pure and simple syntax.
>Niche advantages that are rarely used by the average lisp programmer
what? can you elaborate on this one? I think the advantages of lisp seem "niche" to the average blub programmer since they don't have them in their language, but the average lisp programmer definitely appreciates these features (interactive programming, macros, higher order functions, closures, etc)
>bro just let me post bullshit and infer some good arguments from that
wtf?
>do you ever wrote a ML algorithm
yes. multiple. in python, and CL would be better unironically
also common lisp is not purely functional, it is strongly typed and way faster than python. go look up what common lisp is before being condescending about it online.
it might just be bait honestly. specifically the ESL anon hasn't posted anything of actual value yet
>in fact it's probably one of the easiest languages to teach, precisely due to it's pure and simple syntax.
lisp syntax is yet another example of the fact that simple doesn't mean easy. nobody wants to sit there conforming his thoughts to an alien structure regardless of it's conceptual purity and simplicity.
and this is leaving aside paren complaints; postfix has the same problem. it's just not how people think so you have to do extra translation steps to tell the computer what to do.
you people are retarded, there's nothing wrong with lisp syntax, it's just (blah blah (blah blah)), it's easy
Yes, I am aware what the syntax is and how easy it is to describe.
>nobody wants to sit there conforming his thoughts to an alien structure regardless of it's conceptual purity and simplicity.
How do you explain mathematicians? They’ve been developing a complicated difficult and extremely terse notation for thousands of years. We are at the point now where no one cares if you are filtered by mathematical notation.
Code monkey seething is ridiculous, just because you want to import a library and get something done as easily as possible doesn’t mean there is no value in other “esoteric” or “alien” languages.
>They’ve been developing a complicated difficult and extremely terse notation for thousands of years.
more density = easier to read more information, this is scientifically proven
That is my point. And it doesn’t stop millions of people from bashing their head against a wall trying to learn all sorts of mathematics every day. No idea why people hate on languages with syntax / notation that is designed with a purpose. Of those, forth and APLjk come to mind first, but Lisp’s syntax is clean and easy. Don’t get the parens hate.
>Don’t get the parens hate
indoctrination in schools unironically, if we teached apl in school or uni, people would find C ugly
Lisp isn't hard to learn because of the core syntax. It's hard to learn because of all the other idiosyncratic details of any specific Lisp that you have to learn before you can do simple things. Common Lisp being a lisp-2, for example. Clojure has some wierd javalike scaffolding, understanding how entrypoints work can be a little painful for a new programmer. Racket is full of autustic academic shit that you hit the moment you try to do something you think would be simple like load some data from json or yaml.
None of those issues have anything to do with syntax. You can even use Hy, which is basically python with lisp syntax, to get a feel for how little the parens really matter.
>Racket is full of autustic academic shit that you hit the moment you try to do something you think would be simple like load some data from json or yaml.
While I haven't dealt with json or yaml in racket, your claim that it's full of academic crap is false. Go to docs.racket-lang.org and type json or yaml and you will get what you want. Racket is by far the most practical Scheme, they have a lot of libraries and their documentation is one of the best because they all use the same documentation tool unlike Python.
I also assume from your post that Lisps are harder than Python in which I disagree. Where does this nonsense with Python = easy comes from? Is it the one line hello world? The only easy thing in Python is the ability to write shit code because language features are poorly implemented (lambda, for comprehensions, conditional expressions, syntax). That there is a library for almost everything is true but nobody talks about the quality of them. They all use different documentation layouts, all are incompatible with each other and aren't even compatible with a lot of the language's feature. Look at numpy, like why can't I do math operations on their shitty built in types? Does the language not have interfaces, contracts, traits or whatever they call them?
If you want a real response, try again to make an intelligent post and not a totally retarded one.
The reason being that symbolic AI fell out of favor and it was used mostly for its symbolic manipulation capabilities. That AI winter singlehandedly killed lisp. That's literally the only reason.
because the idea that intelligence was MODEL-PLAN-ACTION were the kind of retardation you could only see at MIT
I wouldn't call it a "winter", but rather a spring. The intuition of Rodney Brooks on behavioral-ism stop defining "intelligence" as an intrinsic property of an agent but rather a property attributed from an external observer of an "emergent intelligent behavior" emerging from hard-wired stimulus and inhibitors were a savior for the field
you'll free to bring your uneducated ass away from here, and please do it
How did it save anything when it didn't have any impact whatsoever on deep learning or its development (not to mention it already existed at the time)?
deep learning is only a subset of ML which is a subset of AI
There is a lot more to AI that deep learning
> yes. multiple. in python, and CL would be better unironically
so do I, for a living. And never once in my life I thought that lisp were better suited for a deep reinforcement learning algorithm
I'm sure you did in your fantasy
>deep learning is only a subset of ML which is a subset of AI
And all useful ML is DL, just as all AI that works is ML in the exclusive sense of DL (RL also is just DL nowadays).
>There is a lot more to AI that deep learning
For all intents and purposes, DL is 100% of AI.
Also, stop moving the goalposts. You claimed that moving away from intrinsic to extrinsic view of intelligence helped the field, yet the field is entirely DL, there has been no real advances in any other direction, and you implied DL isn't what you mean. Clearly huge contradictions here.
>And all useful ML is DL, just as all AI that works is ML in the exclusive sense of DL
not remotely true
>RL also is just DL nowadays)
not true again. There are a lot of soft-actor critic algorithm that work without necessary using dl
> You claimed that moving away from intrinsic to extrinsic view of intelligence helped the field, yet the field is entirely DL
you are still reiterating your wrong thesis. I don't know what kind of discussion you're used to
> here has been no real advances in any other direction,
In the behaviorism-soft ai approach there were big advantages, first of all the speed of action avoided divergence in the modelling and most importantly, robots worked without the closed-world assumption, which was a big chore at the time
sure thing bud. Whatever makes you sleep at night
>facts aren't real because... they just aren't, OK?
Take your meds, schizo.
Worst larp post I've ever seen on BOT, well done. Want a medal? You'll have to imagine it, but you seem quite good at doing that.
>and never once in my life I thought that lisp were better suited for a deep reinforcement learning algorithm
lol maybe because you have no idea what lisp is anon
try looking it up
Unironically common lisp would be just about perfect for deep learning.
>compiles to very fast code
>best in class development velocity
>won't lose your work for a typo, not only because you will know about typos in advance but because worst case scenario you get the restart handler to fix and continue your work
wolfram language already exists
>What language would you use?
bash
OP here, Lisp of course
>What language would you use?
Kotlin
Lua, the language torch and subsequently waifu2x was written in
python is only used because universities teach everything in it, and universities use it only because students straight up demand it (and will leave poor ratings for your course otherwise), universities listen because students pay handsomely, and students demand python because they're all midwits who barely know anything about technology other than the six digit income they're about to make. For a more realistic answer at this point I'd pick julia
Java, sir
Lua or js
Python is like a car with good ergonomics for the ML people. The actual math and algorithmics are much more nuanced and involve a lot of JIT compilation, C++ code, computation graphs and linear algebra.
Python just happens to be the tool of choice for comfortable manipulation of high-level concepts for high IQ people.
because ML's most important applications are in science and engineering and ML researchers know that lab scientists and engineers don't have the fucking time to work through structure and interpretation of computer programs. python just werks
ML uses c/c++/rust python is just glue.
Requirements for ML:
- Quick iteration iteration time (lots of parameters to check for any arch choice)
- Ease of modifying the overall arch/workflow
- Accelerated single-block compute (all compute happens at once, repeatedly, with no need to move back to python runtime for a long long time)
- Good tool facilities (tensor manipulation, text and stats/prob toolkits, etc.)
Things nobody gives a shit about:
- Fast non-library speed (virtually no time is spent in the python runtime)
- Verbosity
- Low level memes (makes modifications harder, too verbose, greatly reduces turnaround time, error risk requires far more time and effort to work out, no room for mistakes due to long runtimes)
A ML algorithm was trained to quantify the brain size of posters based on their posts. It has an average error rate of 0.001 cubic centimeters. I used it on your post. Here is the result:
>Error: brain not found
It's the first time it ever did that. Strange!
> Things nobody gives a shit about:
they literally have no idea if they should give a shit about it. Nvidia has really cucked these researchers into thinking chunky GPU ops are the only way
Show another way.
ML is accelerated by using vector and matrix ops, which is exactly what the GPU does. what would you do instead?
show me good research in dynamic sparsity or networks with tight control flow
doesn't happen cuz paper farms know that scaling up the latest LLM with nvidias hardware is the way to go for an easy publication in one of like 300 annual conferences
Scaling papers barely get published. Nowadays the popular paper scams are either: copy some technique from 5 years ago and rename it (UNet, cyclegan, latent space diffusion models for recent examples), bribe (coordconv, for example), or dazzle with (wrong) math (basically all the mathy papers in the past 10 years).
Goddamn this truly is the lowest IQ board
Nvidia shill confirmed
CPU AMX instructions (intel, apple, arm) gonna crush the GPU "muh flops" watt wasters
>Requirements for ML:
>- Quick iteration iteration time (lots of parameters to check for any arch choice)
>- Ease of modifying the overall arch/workflow
>- Accelerated single-block compute (all compute happens at once, repeatedly, with no need to move back to python runtime for a long long time)
>- Good tool facilities (tensor manipulation, text and stats/prob toolkits, etc.)
>Things nobody gives a shit about:
>- Fast non-library speed (virtually no time is spent in the python runtime)
>- Verbosity
>- Low level memes (makes modifications harder, too verbose, greatly reduces turnaround time, error risk requires far more time and effort to work out, no room for mistakes due to long runtimes)
unironically, just use forth
https://mind.sourceforge.net/mind_faq.html
So what library, written in Forth, is equivalent to Keras or PyTorch again?
forth did not had any investment mady by big corporations, but is entirely possible to do modern ai software with it, forth was an ai language in the old days, but people just remembers lisp.
I’d very much like a faster runtime and especially something statically typed. It feels pretty retarded to coerce a simple loop into a matrix operation just so it doesn’t take ages. And it’s significantly nicer to get an error up front rather than four hours into a run on a compute cluster.
t phd student doing ML
OK you surely know better than people in the field that have actually done valuable work. Fucking midwit
so basically you're looking for nim
If you're getting python-level runtime errors after 4 hours it probably means you didn't test on sufficiently good dummy data. (Other reasons usually have nothing to do with the language or runtime, like corrupt data or storage flaking out, etc).
As for coercing the matrix, typically the tradeoff is a few awkward optimization points vs dealing with constant tedium of declaring types that are obvious or writing extra boilerplate to deal with user input.
> feels pretty retarded to coerce a simple loop into a matrix operation
this is the crux of it.
"it's written in C++" posters don't understand that you have to coerce so much shit into pre-defined python APIs and have no easily accessed escape hatch
You meant to say C/C++/CUDA
Forgot Fortran.
Nothing is actually written in Fortran anymore.
BLAS implementations that are used by e.g. numpy (they're pluggable) are often ancient fortran libraries. Nobody uses it for new stuff though, that's true.
I bet there are still oldgays writing MPI jobs in fortran.
FORTRAN programmers divide in two groups: the ones that understand the language is amazing as a backend for many number crunching APIs and the old farts (and their students) running simulations with the most unmaintainable code you will ever see, or maybe not but is still pretty bad.
tbqh nothing is really there to rival Fortran when you have to develop HPC code for clusters where you can't log into the worker nodes to debug.
Not true. OpenBLAS uses lapack which is Fortran. That's just off the top of my head, I believe there are several other commonly used python libraries that use fortran.
This thread reads like a bunch of underage wannabe programmers/MLers talking out of their asses. Thread should have ended with
Today you've realized that programming languages are tools and specializing in deep "technical knowledge" of a languages is not as valuable
Would you rather ML to use assembly and take years to come up with a dataset that recognizes the text "Hello World"?
ML is supposed to be easy to code. Python is just the interface.
Python is the programming language of everyone who isn't a developer but needs some way of getting the computer to do whatever he needs to achieve some other goal.
Before python, these people likely used Autohotkey, Excel formulas, VBA, Bash scripts or maybe PHP
or just hire programmers instead??
Eh. I'm not talking about hiring.
I'm talking about the people that have an interest in for example Data Science or ML.
They don't necessarily also have to be programmers but they'll need something.
>hire two people to do one job
>each must be paid more than the average CEO
>one of them isn't supposed to do anything 99% of the time as he waits for the other to program the specified solution
>must wait hours to change a single parameters instead of seconds because of the "two body" problem
Get a brain, cletus.
>implying 1:1 ratio
1 programmer could satisfy multiple ai researcher at the same time
Wow, what a slut.
But the one waiting is not the dev, it's the researcher. You can't modify your shit if you don't have the results from the previous run to decide what to change next, but here you're waiting forever for the dev to implement what you described (maybe he's busy with something else, maybe he's out to lunch, maybe he has trouble understanding your request and you have to meet him to explain, etc.), just for you to run it because only you are well-placed to judge if the numbers you're getting show the training is worth continuing or not. And if the dev is incompetent, ho lee shiet.
All this for literally 0 advantage because the training time will be the same anyway since it's all time spent in cuda, not in the python interpreter.
Programmers aren't mathematicians or AI researchers
Perfect, you now need to teach a programmer how do neural networks work. You better pray HR chose you a person with a proper math background and not some random bootcamper.
The the first mistake if you actually want to make a product.
>paying professionals when gradserfs will work for free
wow, it's like you want it done properly or something
Before python, they used bespoke solutions like LUSH for the most part. It was common for labs to have their own special snowflake language based on their focus (some had probabilistic languages, some used lisp variants or languages with logic programming support if they were working with symbolic stuff, etc.)
Another common choice was matlab. The more engineery places used fortran. C was also used, but rarely.
I want the usual BOT shitposter to see some of that old "scientist taking up programming" code from before the mass adoption of Python just to see how many cases of sudden heart attacks are reported in the news
It's still a huge problem. Any physicunt or mathtard uses C nowadays and the code literally doesn't relies on "running on linux having been compiled with gcc in debug mode and no optimization" to not segfault, if you're lucky enough that it even compiles at all. It's also always slow and completely unreadable (single-letter names everywhere, but importantly even pajeet seems to have a better grasp of programming concepts). Truly dreadful.
literally relies*
or literally doesn't run without relying on*
lmao
some things don't change as much, after all
>is the entire ML field secretly composed of midwits?
They're doing statistical analysis and pretending it's something new, so you tell me.
ML uses numpy, numpy is written in C++ (?
ML does not use numpy lol
Yes it does, scikit-learn takes advantage of pandas and numpy.
https://scikit-learn.org/stable/install.html
Are you fucking retarded
Just about any data loading and prep that's done uses Numpy, and if it's done with Pandas it's still backed by Numpy
Python is an incredibly well-designed language, and the libraries are all written in C(++) anyways, so it isn't like the 250ms python takes to call the C library matters when gradient descent takes days/weeks/months
PyTorch and TensorFlow are both implemented in C++. Your post was correct in spirit I suppose.
>Python is an incredibly well-designed language
My fucking sides!
> >python le badly designed language
>brainlet detected
>One of the most popular starting language
>Used by almost everyone in IT
>Used by tons of people outside of the computer science field
>used by a ton of developers
If python is a poorly designed language then there is no well designed language.
You must have an IQ of 18 or above to browse this site.
Then why are you here?
The fact that it isn't well designed is why it's popular. Software that is better in principle comes with everything you need. software that is inferior does not, and someone has to fill it in. They do it by building tools. Every third-party “tool” is actually a weakness: it’s doing something that either didn’t need to have been done at all, or could have been done better in the first place. however, it creates a community. There’s a way for people to contribute. It’s like a game that leaves little bonuses lying around for people to pocket. If even a novice can make a small contribution, so much better. They feel good about themselves, and now they’ve made a commitment to that ecosystem. The switching costs suddenly became significantly higher. In the realm of programming languages, this is especially true. And in fact, to be “successful” in terms of size, perhaps every language should start out somewhat flawed.. it should have bad scope rules, weird modules, funky overloading, floating points as the only numbers, and so on. To a rational observer, this might seem an awful idea. but programmers, as a species, have gotten acculturated to salt mines as a natural habitat. They will think nothing of it. Of course, you need something to draw them in: “one weird trick” that lets them do entirely unnecessary and perhaps wholly unwise things that let them demonstrate virtuosity. The latter draws them in, and the former seals the commitment. One could view C++ templates as a form of the latter.
Customers regard more highly an organization that has recovered well from failure than one that had not yet failed. Good designs are like the latter: they do not fail, and therefore cannot pleasantly surprise you. Inferior designs fail, but when the community comes to your rescue, you are left happier than you were before.
>Software that is better in principle comes with everything you need.
So python? The standard library is pretty massive. Is this a botpost or something it's borderline nonsensical.
>One could view C++ templates as a form of the latter.
Template metaprogramming was a mistake.
Are you shit posting or have you never programmed? A language should not include everything and every library is not a weakness. You don't know the first thing about maintaining code. When you add a feature you are making a commitment to it. The features you add increase scope. The point of a language is to build a tool to solve problems. A good language is flexible enough to solve any or most problems while being well optimized and easy to use. That or it is designed to solve a specific problem very well. 3rd party libraries are recognition of a good language. If someone with no connection to your language wants to maintain a 3rd party library in it for free that means you have done a good job. They view your language as being worthy of their free time. 3rd party libraries are important to a language because they allow you to solve more problems out of the box while not expanding the language maintainers scope. It means someone else has made a promise to maintain that problem domain. If they are good enough they may actually get merged into the standard library.
python posters can just use a low level fallacy to defend their shit lang
>le python posters
Says the lisp poster. You are pushing a 70's meme. There are barely any real world applications of lisp and thousands of python applications. Your post is syntactically pure cope.
>into parens
I gave detailed reason why I didn't like lisp the parens were just the cherry on top of the shit sandwich. Being able to deal with them doesn't me I like them. You have so much lisp Stockholm syndrome that you can't even admit how annoying they are.
>thousands of python applications
another low level fallacy!
How is it a fallacy to cite repeated past success as evidence of something being good vs the alleged benefits of something else that while existing longer has produced far less successful products. The quality of a tool is defined by how well it gets a job done, Python has plenty of evidence to show that it more than meets this standard while Lisp's record is underwhelming.
>How is it a fallacy to cite repeated past success as evidence of something being good vs the alleged benefits of something else that while existing longer has produced far less successful products. The quality of a tool is defined by how well it gets a job done, Python has plenty of evidence to show that it more than meets this standard while Lisp's record is underwhelming.
average python defender
>There are barely any real world applications of lisp
Apart from Emacs, all lisp applications are either very expensive or so domain specific you'll never hear of them if you don't work at that one weird place just outside of Boston.
Fwiw I agree with most of the points wrt Python over Lisp but complaining about parens does make you look like an entry level gay who doesn't know shit about programming.
>Python is an incredibly well-designed language
I use Python... and it really isn't well-designed. It's got some pretty clothes that it throws on over the top, but that's it.
You've just not seen what good languages actually are.
> written in C++
https://pytorch.org/get-started/pytorch-2.0/
> we announce torch.compile, a feature that pushes PyTorch performance to new heights and starts the move for parts of PyTorch from C++ back into Python
out of touch and btfo;d
script kiddies who take themselves too seriously always make attempt like this but in the end these attempts fail. you best use different languages for maxing out performance, for maxing out computational complexity and program depth, and also for manchildren too low iq to learn more abstract concepts but want to feel like adults (hence why python and javascript also exist)
Pretty sure python is still the only way to program for TPUs
https://github.com/pytorch/pytorch
https://github.com/tensorflow/tensorflow
did you not see this part?
> starts the move for parts of PyTorch from C++ back into Python
it's not done yet lol
>why the fuck does ML use python
Because "data scientists" just need something to glue libraries together.
I mean it works doesn't it?
Why should they pain themself with learning c, c++, rust, etc when they can use python that has a way easier language to learn and understand just to get a little better performance?
Just because Python is an "easy" language, it does not mean that ML is easy
Cause they're not dumb. The Python calls the C libraries and just passes all the parameters. Why spend time fucking with memory when you need to do a lot of testing and fiddling with something. Python removes an entire class of errors through it's garbage collector which is useful when your dealing with ML.
numpy, tensorflow, pandas, and whatever other python libraries that do intensive numerical stuff are all really precompiled binaries with python interfaces. As long as the programmer isn't retarded and doesn't reinvent some time-complex procedure in interpreted Python code (non-cs researchers do odd things) then it doesn't really matter. I dislike Python for being weakly typed, but would rather it for ML over the clusterfuck that is C++ ten times out of ten.
Python is not weakly typed. There's just no static compile-time type checking. It's dynamically typed. C is weakly typed.
It used to be Lua. See luatorch.
It was just called torch back then but yeah. Torch got a python port they didn't care about too much, but as theano died, and everyone was using python because of theano (plus tensorflow was also python also for this reason), they fully reprioritized to python sadly.
It's because data scientists aren't programmers and need something that will work without too much bullshit.
Some are, some aren't. Both choose python for good reasons.
Python's just used to call libraries written in C/C++.
ML is hard enough without some neckbeard helicoptering his dick about curly braces
Python is high IQ. Choosing something in between Python and C is literally the midwit route.
>"I'm too smart for Python and therefor why does anyone use it?"
>made up curve, artificially made to look smooth, unlabelled y axis
go read the actual paper. the real result doesn't look like that.
The Dunning-Kruger effect has failed replication. So ironically, unironically referencing the effect to try to make a point is Dunning-Kruger in and of itself.
unironically quite ironic indeed
Because Python is easy to type and people don't want to waste more time than they have to writing retarded syntax for no reason.
use nim retards
I never understood why Lua didn't take off for ML.
Lua is easier to bind to C/C++.
Lua is more simple than Python.
Also, mathgays largely prefer 1-based indexing.
MLgays are mostly cstards, not mathinbreds, so that wasn't very attractive (the ones with math degrees are those who graduated before CS degrees existed for the most part. On the other hand there is a decent chunk of mlgays with a stat phys background).
The reason python was the winner is also related: the new crop of people interested in ML were uniquely in the CS camp, and mostly knew python (or if not could pick it up easily). Lua, on the other hand, is a much rarer language in colleges.
Then the rest is the same momentum dynamics that C and unix enjoyed, for example. Because people knew python, tools were made in python, thus people learned python... for ML.
But it's true, a major defect of python for ML is that the C interop overhead is massive. This isn't a huge problem since back-and-forth are minimal in most ML applications, but lua doesn't have this problem, and the corollary is that it makes C libraries with bindings in lua far more usable (think numpy-for-lua). It also makes writing small acceleration code in C far easier (in python, you have to resort to numba (very limited), cython (python-flavored C, very complicated to build in non-trivial cases, let alone distribute, rarely actually that fast), python.h's horrid interface, or ctypes and its immense overhead, i.e. no good options at all).
no, it's because a top AI researcher makes roughly $100/hr. No company is going to waste their valuable time making them debug buffer overflows in C or fighting with the rust borrow checker.
Also, all of the number crunching happens with extremely optimized CUDA submodules. The python is basically just a convenient interface to low level libraries
its so idiots such as myself can get paid to use it for niche engineering applications
Programming is for people that doesn't really have anything important to do
people in want to do work, not fuck around with a language for making operating systems or some autistic shit from the 60s. python let's you use some 3rd party libraries and build and modify something quick and try it out. fucking around with low level code is the job of the library. go ahead and implement a convolutional neural network in python vs cpp see what happens
>PhD requirement to even enter the field
As a PhD student, let me let you in on a little secret... most CS PhDs can't fucking program to save their life. Time spent studying theory and advanced mathematics is time not spent learning how to properly design and maintain software, nor is it time spent using any remotely complicated languages. Python is common in AI, data science, and anything in between precisely because it is simple. Academics love Python because it allows them to focus on the task they're trying to accomplish, without regards to the implementation details.
What matters isn't that you write a good program. You don't have users. You have yourself. Your program will be run however many times it takes to get results for a paper, and then it is unlikely to be run again. Unless, of course, one intends to write a follow up paper. Speed is relevant only insofar as machine learning is very computationally expensive... so all of the code for training these things is written in C or C++, with bindings to use the library within Python.
I personally like programming. About a year or so ago, I worked on some tools for a project I was doing with some power systems folks. One of them could have been used in pretty much any language with a graph library, so I used Rust. Not because it needed to be fast, but because I wanted to write something in Rust. I'm in a minority in academia though. For most people, writing software is a means to an end and nothing more. If it can be done in Python, it will.
>uses fucking python. is the entire ML field secretly composed of midwits?
YOU are the midwit, since you don't even know what Python does in ML. The libraries are usually implemented in languages like C, and python acts as a glorified API to it.
Because pytorch and numpy are convenient to data science
>thanks to python i can do codemonkeys jobs but they could never do real engineering
you love to see it
retard:
because it is an amalgam of lisp and prolog
> ML uses Python
Do you forget that almost all libraries are made in C, C++, etc to make them fast.
Python is just a wrapper and makes things easy and fast to work.
this subhuman filth will advocate "C"
fuck off, ponytail
C has destroyed the planet
Why isn't Scala more popular?
brainlets cant into typing, thats why python/JS is so popular
Than what?
People programming for jvm will typically just write Java. Or maybe kotlin.
Due to Java and Kotlin already existing
The language design is pretty good. I love the type system and their fusion of OOP and functional works well but the tools are god awful. I used to syntax reverting tool from the compiler and I had to fix most of the syntax myself. Metals will now fix your imports when you move files (I think this feature came out this week) but it actually does a half assed job as well and renaming symbols sometimes doesn't work for some magical reason! Scalafmt makes code look ugly. Only VSCode seems to work well with Scala 3. Documentation for some libraries is lacking might be because a lot of them are Java wrappers.
I might have these issues because I am using Scala 3 though.
>Engineers (real) and scientists use any language for projects
>Programmers (fake engineers) spend hours of their day arguing over what language to use.
lol
because its easy to set up. You have a CNN in like 5 minutes instead of linking dependencies for an hour and trial&erroring
i'm finally going to learn how to program in this year, this thread is full of people that use python, just recommend me a book or a course please.
I just notice that practically all the lisp users magically disappear around halfway to 2/3rds through advent of code
that should say something
most people who do aoc are young students who have time to waste on december, lispgays are doing productive things with their families
I'm thinking python is the real filter for brainlets now
I haven't head a CompSci PhD complaining about it, it's always some midwit BOT user who make less than 200k a year
not like phds earn all that much
you do not really go the phd route for money
you do it to learn
>you do not really go the phd route for money
>you do it to learn
The point of a PhD is that it marks you as being THE world expert on something. Almost always something really off the beaten track and maybe only for a few months, but definitely the world expert on it.
The point of using python for ML is using python for easy access while making C do all of the hard work. If there ought to be a replacement it needs to be attached to C like python is. Something like lua.
explain this, then
https://jax.readthedocs.io/en/latest/jax-101/02-jitting.html
https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html
Python supporters BTFO'd by experts literally needing to shim the language with their own compilers
nothing says "shit language" like this
Nothing says shit poster like linking to posts he didn't read and can't even discuss in contexy, then making bold claims about straw men.
nim?
Machine learning is just a while loop u can do it in any language
in the end, the only real difference between programming languages in a practical sense, is how much money a big company is using to make it popular, it happened with C, C++, java, python, go, etc.
iterate in slow but easy language then export the model for inferencing in faster language
Cant tell if you're esl or AI
neither, im right
Python was always for retards and mouth breathers
>herp derp what the fuck are semicolons and braces moronman save meeeeee
what do you use to program ML you prick
from your post, you dont' program at a
what you using to develop your ML framework, genius
you fucking parasite. python is the only thing
python is ADVANCED you fucking parasite
> be one of the easiest languages to learn and use
> arguably has the best library support of any existing language
> "why the fuck does ML use python"
Hmmm.
It's the most convenient way to create wrappers for C libraries.
Haskell. If it's good enough for Cardano, it's good enough for ML.
The user interface is python, the compute libs are all compiled langs.
The guys that wrote these ML applications are too busy building things to worry about your hateboner for python
lmao python script kiddie trying to justify his language of choice. real reason: you're too dumb to learn anything better and it's the same reason for ML PhDs. getting a phd in ML is literally nothing it's the most looked down upon post grad degree
What's Tensorflow written in?
turing-complete programming language
Mostly in C++, CUDA and a little bit of Python itself
The worst thing about python is the xdb forums pajeet tier versioning and lack of true dependency management, coming second is whitespace as syntax
ITT:
room temp iq pseudo-intellectuals that think harder to learn languages are always better.
>wahhh wahhh you have to manage memory yourself and you cant use anything dynamically typed
lisp is easier
No it isn't. And it's not the parens. See
Also, even once past the beginner stage lisp's advantages vs python tend to be minor and not that important. Homoiconicity starts working against you in little ways. Some operations are much more intuitive with infix, eg 'item in set' vs '(in? set item)' where you have to remember which operand comes first. Then you have python's array slice syntax which is often slightly more convenient than using function calls or special forms to access array sections. These are all really minor frictions and lisp has it's own advantages but that's partly the point: the syntax differences just aren't very important. And for a typical workflow glue, interactive notebook, etc the Python is probably going to win out based on big picture reasons like access to libraries, docs, and support.
>lisp's advantages vs python tend to be minor and not that important. Homoiconicity starts working against you in little ways. Some operations are much more intuitive with infix, eg 'item in set' vs '(in? set item)' where you have to remember which operand comes first. Then you have python's array slice syntax which is often slightly more convenient than using function calls or special forms to access array sections
are you dumb, don't you see that lisp being homoiconic everything can be implemented as macros? There are already infix lisps, but the community don't use because they like lisp notation, simple.
No, you are dumb. You seem to have forgotten the context of the comment, missed it's main point and more importantly you clearly have no experience actually programming to solve real problems.
cope
>why the fuck does ML use python
Because it's easy. Fantastic for training, experimenting and also prototyping. If I need to show board something, I'm going to do it all in jupyter notebook and run through the shit, or email them a PDF of the notebook. Python should however NOT be used for deployment.
To use myself as an example, I've spent over two decades programming. I transitioned into ML about 8 years ago. I do my data preprocessing and training in python. I do all my experiments and testing in python. I deploy all my shit in C++ and CUDA, whatever python I use for deployment is minimal and doesn't require dependencies past the version of python.
I can understand how someone that didn't have a programming background would just stick to python for deployment, but honestly, they shouldn't be deploying. The company should hire someone else to do that, or they should hire competent ML people that are also competent programmers.
It's a really easy to use language but fucking wasteful, and if you're working at scale, it just burns through resources.
You are looking at the front end, Python is ideal for this, trivial to change and does not require recompiling. That is why a simple git pull works for AUTOMATIC1111
The low level libraries are another issue, they are usually not written in Python.
manip layer: python/golang
autism layer: C/C++
just like any other useful bit of software
most engineering analysis software is in fortran you know.
fucking retard
Python is like the Kardashians. It's a retarded dumpster fire that's popular only because it's popular, but it has too much momentum to be stopped now, so you just have to accept it and move on.
It's not a dumpster fire, though. It has problems, but everything had problems and it's worth understanding what Python did right to get where it is. Instead of just jerking yourself off in ignorance.
yes
Because python abstracts the tedious shit so you can focus on the actual ML task. Nvidia provides C++ libraries for CUDA, no one's forcing you to use PyTorch.
Go be stupid somewhere else.
Python is C under the good. Built in functions and libraries are written in C. Python is all about using libraries and built in functions to perform a tssk. If you're using a for loop in python you're doing it wrong.
do I need to be good at math to enter DL/ML/AI?
I'm currently using ML for science purposes and I'm just importing packages and cleaning data without understanding any of what's going on like a total moron, I imagine that's what most people do except for the actual geniuses who write these algorithms. I'm working on my understanding, slowly thoughever.
Would love any input or tips from actual ML specialists or users.
Honestly, because of jupyter notebooks, numpy, matplotlib and similar libraries. The workflow for mucking around with data and then constructing a model and then training the model and then plotting performance is nicely streamlined.
It's especially useful to use a simple and well-known language given that people who specialize in ML may not spend as much time programming as a full-time software developer might. Often times these are domain experts who know some programming who then leave it to other teams to productionize their models.
python notebooks aren't even that good, wolfram notebooks are much better, wolfram lang is not used just because is proprietary
jupyter is based, it helped revive interactive programming but sucks hard enough to motivate competing projects
thanks pytards for tarding clojure clerk into existence
>why
Because python is superior.
>why
Because it makes (You) seethe, no other reason.
Mission Accomplished.
>make this efficient in Python using C++ libraries
You're missing the point. Making small amounts of setup code more efficient isn't worth it. You'll be saving less than 1% of 1% of the execution time and energy. Optimizing your own code is only really worth it once you get to about 5% faster (the break-even point is lower for common library code) and you only get much better when you go to 10 times faster; the ~10 times factor is where you change what sort of applications using the code can be conceived of.
“An idiot admires complexity, a genius admires simplicity, a physicist tries to make it simple, for an idiot anything the more complicated it is the more he will admire it"
primitive abd simple are different things schizo
Anyone who works with PhDs knows they only come in three flavors in descending frequency:
Midwits who cobble together semi-believable middling work on current topics but lack the IQ to understand their field in a complete manner.
Sycophants who are adept at getting their names thrown on papers and getting grant money, but everyone in their field secretly suspects they are retarded.
Autistic Rainmen who never get tenure anymore and advise everyone else. They are the only ones who actually understand anything, but are thought of as kooks. They usually end up as forever-postdocs.