Full podcast episode: ruclips.net/video/-DVyjdw4t9I/видео.html Lex Fridman podcast channel: ruclips.net/user/lexfridman Guest bio: Guido van Rossum is the creator of Python programming language.
All of the community support made it possible with the libraries. Also, Python has an easy learning curve for the math and data analyst expert to take it on when their primary interest is ML. They just want to dig into the data and learning process, not spend 80 hours a week trying to learn to program. Programming is a learned residual necessity of wanting to learn/apply A.I and data science.
I couldnt agree more.. I am not even at the level of Machine Learning, but as a BI Developer who primarily works with SQL and DAX, learning how to make API calls, convert the returned JSON into a Data frame and import into SQL Server using Python was relatively easy and has left me wanting more. Web scraping is next for me I think.
@@real_mikkim Matlab has awfull syntax. With Python I can create very nice functions with optional arguments, keyword arguments etc. in seconds. With Matlab, it is an awfull experience.
Yea, got to agree. I came to python from R and I have found it being better. Indentation instead of brackets is just pure heaven, though I am not a fan of 0-indexing....
Matlab is good for simulation and exploring mathematics. For production-level usage both Matlab and R it's a complete hell to make it work and maintain. Going from Prototype and Production with the same language is really a important feature.
Weird way to go about it but ok... Idk just seems like it one should choose a language/tools based on what he wants to do not the other way around. Do you mean it was a gateway to get into it?
@@ghosthunter0950 I live in code and so I care about aesthetics and ergonomy. I could have stayed writing web 2.0 stuff in Python, like some collegues have. I wouldn't have come to ML if the lingua franca was Matlab.
I think one simple explanation is that many DS and ML en... came from math, physics and other related disciplines without much knowledge of programming and the learning curve of python is quite efficient, you are writting usefull code in a glympse of an eye.
Bingo. Having some experience in both python and C++, python is perfectly balanced for scientists and engineers. It's one of the easiest to learn and it's limitations are not pronounced when the program is small and not complex. Most scientists want write a script to calculate one particular problem. Having tried to write a large full blown user desktop application in python and C++, python's limitations soon become apparent and C++'s strengths are exposed.
Can confirm, I came from a Financial Analytics background using predominantly R. Aside from Python being widely accepted, it was quite easy to learn the syntax. The biggest obstacle was learning all of the functions in each package. In fact, I only really had formal education in C, so coming to Python felt so wrong because of the lack of curly braces and semicolons 😂
Because it's easy to use for people who are focused on things other than coding. It's not "good for machine learning" outside of the fact that many people have used it to create useful libraries, because it's easy to code with, if you're not really that concerned about coding or performance.
Python became popular because, it gets out of your way, it allows you to be as technical as desired, all with minimal syntax. No other language offers all three (not strict, provides traditional object and procedural disciplines, minimum syntax). So when I develop, I can quickly flesh out ideas. Later, I can revise as needed. If it is discovered that Python is a performance bottleneck, the ideas are already solidified enough that they can easily be rewritten into a compiled language. I am currently using GD Script which is a Python like derivative for game programming. In this case, a Python-sh language was developed for game programming, because Python does not perform well enough (mostly because garbage collection does not work in game programming). So the Godot Game Engine Developers developed a language that acts like Python (without garbage collection) until compile time. At compile time it runs through an interpreter, to convert it to optimized C++. The C++ is then compiled. This is an amazing way to get C++ performance out of Python like code.
FYI: You can use Cython to write code similar (if not exactly like) Python, but generating C/C++ code, with most of the performance gain if used correctly. BTW, it's a mess to debug.
You can also use LISP s-expression version of Python like Hy or maybe Julia and have macros that allow you to calculate stuff at "compile time“ rather than run time or allow you to have static analysis, its easier to build custom compilers or to convert s-expression based Python into s-expression based C++ into a compiler. Paredit can also be used like "JQUERY" for s-expressino code.
Perl saved my life when I had to reconcile two systems exchanging messages that were not idempotent. That stuff was made manually before and took days for just one accident. 3 hours of scripting and then a single accident could be reconciled in minutes. Great stuff.
all good libraries revolve around the numpy api which means you dont need to learn the api for all the other libraries, you already know how it works and that any library which implements the numpy api, will be compatible with eachother its a great system any new modern language should have NDArrays inbuilt
What was the solution people used before NumPy? I thought that Numpy was just a wrapper around fortran code. Is there nothing like that available in other languages?
RStudio is absolutely horrible to use. So slow. Though I do prefer R to Python for its rich and customizable statistics library. I just run R in Neovim (as with everything).
I love python for its beautifully minimalistic syntax, ease of 3trd party packages, extendibility and much more. btw how the adoption of new technologies spreads and propagates, can be studied with graph theory with the focus on cascading effect, power-law, triadic closure, 6 degrees of freedom, and many more cool CS and Economy related topics.
I was a rubyist for a long time. Slow as can be, and a community that self-detonated over social justice and intersectionalism. it was wasted time, I should have just followed the smart guys into Python way back in 2008. Thank god for Python and its community who cares about coding and programmers above all the politics.
Matlab in engineering is dying too… legacy systems only. My entire graduate lab group (engineering group) that does computation is forced to learn either python (obvi for the ML students) or C++ for physics modelling
It is true, in the past it was thought that Perl was going to be the de facto language in Bioinformatics, to my surprise Python also conquered that arena without almost anyone noticing. Once it is shown that nothing is written in technology and everything can take an unexpected turn.
@@ianrust3785 What nonsense coming from people who evidently haven't had much experience in Perl. Python was deemed to be "easier to learn" by novices (very debatable), so a whole generation of new programmers were raised on its training wheels. Perl's freedom of expression, cherished by experienced sysadmins, is still valued. It "wrote the book" on regular expressions and its model was widely copied by other systems (as van Rossum readily admits). I personally disapprove of Python's significant whitespace because of 2 scripts, seemingly the same, one can be correct and the other produce errors, but hey, different strokes for different folks.
@@CARPB147 In stackoverflow’s annual questionnaire on various languages Perl has been voted the most hated programming language, repeatedly, for about a decade now…. There are about 20 languages included in that poll. This is why it died out. It didn't get that honor due to the ignorance of those voting, it got that because it has the worst syntax of any language… due to its horrible design philosophy, where the designer believed there ~should~ be 4-5 ways to do almost everything. Python and Ruby have completely overtaken Perl in the sysadmin world and everywhere else, Perl 6 didn’t go anywhere… Perls version of OO is so ridiculously bad it makes old javascript look well designed. Congrats to Perl for creating regex’s, every language has those now. Why should I bother today in 2022...? And why should an organization use the most hated language? That's just going to cause people to leave the org.
@@ianrust3785 I don’t expect you “to bother.” I expect you to go by annual questionnaires. I’ll go by experience. Perl is actually quite alive and working in systems worldwide. It’s even included by default in UNIX, Linux distributions, and in the latest macOS. No surprise given its affinity with UNIX commands and tools. Your favored questionnaires are “fashion statements” meant to steer students and the inexperienced emotionally into specific segments of the job market. For ML, the existing libraries are compelling arguments for Python. Real programmers are not that fazed by one or another language. Or say this one or that one “is dead” 😆 They know that languages don’t really “die”. Witness the venerable 1950’s FORTRAN. Real programmers are not “one-finger pianists playing on a one key piano”. They even know how to write Perl programs as readable as they can be in virtually any language. 😆
@@CARPB147 A serious programmer actually does need to differentiate between bad, dying technology and new, promising technology. Otherwise they'll just be pigeonholed into a crap technology, the way that you are. It would be a waste of braincells to learn every other dying technology where there are better modern alternatives that are widely used. I'm sorry that you spent your adult life obsessing over Larry Wall's poorly-conceived rats nest of a language... but your pain doesn't have any meaning, I'm afraid. You'd be hard pressed to find a population of more experienced devs than those who frequent stackoverflow. Perl must be really unfashionable if it's been voted most disliked 10 years in a row. How did it get so out of fashion...? Did the world conspire to demonize poor old Perl for some reason? No, it was voted that because people used it and found that it was a piece of crap language where it takes hours of in depth study to even understand how to initialize an array.. it is a bad implementation of a bad set of ideas on language design, it was only ever popular because people needed a good way to manipulate text in CGI during the internet boom... one of the very few things it is good for, but which Python is much better for... And it is indeed dying... we can be thankful for that. Carry onward!
Will be good to see a talk about Julia an MIT language that connects software engineering with scientists, looks like a a good replacement for math lab closing the gap between engineering and science.
Julia is certainly faster than Python for Scientific computation and machine learning, but is far to replace python just because the community and support that python already has will be hard to replace
@@EigenCharlie yeah but for a lot of people it's not about replacing, is about complementing. Julia can call Python, R, C++ code and more. I use a lot of pure Julia for tasks that are just very slow in Python, and at the same time I use PyTorch in Python.
Because writing stochastic gradient is much simpler in python due to high-level syntax and dynamic type setting along with automated garbage collection!
Easiest answer of my life, the crossover between programming and applied mathematics/engineering is already small enough. It's assumed that programmers are good at math but it's rare to have true polymaths. To be able to include people who have more pure math skills with an easier programming language is a big deal. I would not consider myself a good programmer at all but I am actually considering contributing to some open source Electrical Engineering libraries this summer in Python to bolster my resume, just because the language is so inclusive.
It's "readability" in my opinion. It's clean and human readable. There are no layers of complexity to peel back to understand what's really going on. Or rather Python hides the complexity so that you can focus on what really matters. Your work, not your code.
Its pretty easy to translate matlab code into python code in many instances, so if I find matlab code in the matlab community that would be useful I first check if there is a python implimentation, and if there isnt I write one
The most important reason is that mathematicians/computer scientists usually are not that good at coding. They will find/modify an algorithm from a research paper (written as pseudocode), and in Python you can almost just copy that pseudocode in and it will just work.
I use python because it lets me skip OOP, using only modules, and I get the same results as OOP. Using modules allows me to organize my code efficiently so I know what I'm doing.
Two reasons. First: Data Scientists are not programmers (at least that's not an implicit skill of data science). Python allows these people that don't program to write stuff in a way that is almost how it would be written on paper. Second: Execution speed is usually not important when it comes to data science. Data science doesn't really matter much about optimisation, and in some cases would be a bad idea because you would want to make sure that what is happening while the script is running is accurate to the task, and not necessarily fast.
I'd disagree or rather, put it in other words. Data science cares a lot about speed, as it has to deal with a lot of data. But all those speed critical parts are well handled in python using specialized libraries like numpy, numba and so on. So the program speed is comparable to compiled languages, while the programming time is drastically reduced due to well maintained high level libraries.
@@hansdietrich1496 Yeah I completely agree. I guess I was more meaning in terms of the kinds of people using it. Like you say, the speed critical parts are sorted by people that know both the math and the programming, but aren't necessarily data scientists by trade. Data scientists are more worried about the math and the second priority would be the speed.
A somehow marginal reason : for old AI guys, Python was the best replacement for LISP. R is in fact closer to LISP, but it is not as versatile as Python.
Now that I know that variadic numerical arrays are so important to machine learning and natural language processing I will be picking up a language especially designed for this; APL : and I will do this as soon as possable.
Because groups of people did the hard work of creating fundamental abstract tools(libraries) usable by free by other who then made more tools for free usage.
Python won I think because of readability. Not all programmers are not Software Engineers, there is a large range of skill. With Python, a novice can more easily understand code that an expert coder would write.
yes, for experienced coders its good to look at others codes and understand easily, you dont need too much, same thing can be tiresome for C++ and java for example.
It doesn't matter what programming language you like or use. Soon enough all the libraries will be translated to any other language by AI, withing limits of the target language.
I wonder how it will translate Python code into Haskell, it won't be idiomatic it will use the IO monad when the state monad is called for like when using dictionaries(hashes), oop can be simulated with closures or many other ways.
What im wondering is , is there really a technical reason other than someone picking their favorite language and others having no option but to use python? Dont you get a better performance from a compiled language?
About matlab not being popular due to not being open source: There was also Octave, an open source matlab clone. And it actually had some use on universities. Nevertheless, it never got popular. The matlab syntax is a major pain in the ass. And the library coverage and general usability was limited to mathematical applications. Trying to create any customer software out of it, interfacing with the real worlds, didn't work well.
Put yourself in the shoes of a biologist or a mechanical engineer. If you have to learn a new language or even learn to program, what do you rather learn? Python or C?
matlab is quite expensive, and it doesn't fit nicely into open source community. That's what made it dead for open source community. If you can't clone a repo, run a setup script and get running in less than 1 hour, who is going to bother? I've been contributing to open source for 2 decades and cost of sharing is the biggest factor in community growth. Anything that raises the barrier to entry means it has little chance of community adoption.
Lex please give more content where you explain your thoughts on xyz subject. I love you as an interviewer, but I want to know more from you.(nothing to do with the content above, no disrespect, just been on my mind)
Год назад+5
They didn't mention R or RStudio in their whole conversation.
Did C and Perl in a project. Perl was not my favorite because every problem had as many solutions as developers. Horrible to read and maintain. Getting the libraries managed was a hell.
Sounds like Python is still trying to solve problems that have already been solved in Java. As it moves forward, the syntax will start to approach the complexity of Java. Type checking is extremely important in creating trustable libraries and overall reliability. Scripting without type checking will always be less efficient because every variable access has to be tested. I would have thought since, aside from complexity, AI apps are compute intensive they would benefit from the JIT compiling and optimization Java already has. As time goes on, business apps written in Java would likely need to access AI functionality creating a high integration cost. Jython will have a hard time filling that gap because it is not the primary version..
With a lot of computationally intensive things in python, the heavy lifting is done in libraries written in C/C++ (or something similar). So the poor performance of python doesn't matter as much as its ease of use. You can use tensorflow with java (and other languages), but everyone uses python. python is a nice language for just gluing a bunch of libraries together. I say this as a C/C++ try-hard.
the other limiting factor against Java for ML is complexity. DL4J is a massive code base and difficult for users to grok. I've contributed documentation in the past to DL4J and the developers are quite responsive, but that doesn't solve the issue of learning curve. A researcher needs to get stuff done quickly and doesn't have the luxury of spending 2 months learning the API. Google's tensorflow Java api kinda sucks and is treated like a second class citizen. As of Oct 2022, the java tensorflow api is deprecated until a new version is released. my own bias, I hope tensorflow dies and better frameworks take it's place. TF2 fixed a lot of bloated ugliness of v1, but it's still kinda ugly. Sadly most clouds support TF first and pytorch second. Hopefully that changes.
@@YourNerdyJoe I agree with what you have said and if your main interest is research and getting more people involved, Python is simpler to get started. If you want heavy duty production applications however, much of the complexity of Java comes in handy eventually. C/C++ is not as portable so writing universal libraries is difficult and does not add much to the performance over Java JIT. I'm retired now but in my work as an IBM Architect and a long time programmer I have found that adding a new language to your organization is very expensive. You have to train or hire trained people in the new language. The first thing they want to do is re-write existing software to fit the new model and spend a lot of time interfacing it to the existing frameworks. I think a newly introduced language needs to be a lot better to justify it unless you are mostly doing one-offs.
@@minionsystems I hear you. My primary languages are Java and C#. In the consulting world, my experience matches yours. Getting our customers to add new language is a tough sell. It's always easier if the CTO is the one pushing for it, otherwise it's usually DOA. There's lots of things about python I dislike. I'm hoping Rust community can build the frameworks needed to replace python :)
I dare to disagree with Guido. I think that library authors and users chose Python because it is simple yet powerful enough. Most researchers working on ML are not programmers, they are not coders, they do not care how powerful a language is, they do not care how well designed and elegant it is, what they want is to use the simplest tool to do their job, and that is Python. And yes of course to be elected by the academia and research a language must be open source and not belong to a corporation. Java was so open source friendly and "democratic" that it started as a big favorite in universities, which also used to teach it in their programming classes, until it ended up belonging to Oracle...which is clearly the worst thing that ever happened to Java..
how old is too old to change career to coding I am going to be 38 next month :) should I learn algorithms first what should me my first step into this world? thanks...!!!
Have you started yet? The question of wether to start or not is very simple: start learning and see how far can you go just as a passion. If you can go pretty far, then you are good to go.
Hey, this isn't true that in the late 90s there was no other language with support for multidimensional arrays. Perl had PDL back then, not only powerful Regexes. PDL is very similar to Numpy and existed much earlier. So, apparently, the reason for Perl to lose its prevalence was its ugly syntax and its 1000-ways-to-do-one-thing weird philosophy =)
Lex, you are very American: "Universities still paid for it..." yeah, if you were an American university. Not every country can afford $2000+/person fees (yes, in Eastern Europe Matlab was about double-triple price compared to US, I do not know now)
4:12 absolutely false. Because of the idea of immutability of strings and maybe other objects, anytime you need to manipulate a string, a new version has to be created. I have decoders I wrote in C that was 50x faster than the python version. The decode will take a binary file and make a text version of it. It does so by appending 2 to 10 byte small strings to a larger growing string that will then be written to disk when completed. This appending happened a few thousand times, and it slowed down the python process.
Python is slow and I personally hate the syntax. Interpreted, loosely-typed languages may be simpler to throw together simple migration scripts or whatnot, but languages like this (IMO) will always be more error-prone for runtime exceptions.
You can implement inner loops in Python by using numpy matrices and other such aggregate functions, there are JITs as well, and a lot of apps are IO bound not CPU bound. I think if you don't test your application it will have a lot of problems.
If the history of programming is anything to go by, Python will not stay on top of ML for more than a decade. Personally, I much prefer Julia for scientific computing.
Okay Lex I am following you're advice 10 minutes a day of the Lex Clip to learn something each day with discipline 🤣...For the next 5 years to be proficient 🤣
Its popular because Google pushed it so hard, so they can have easier and bigger pool to pick talents, it kinda got out of hand because too many people jumped on the hype train and slowly started to realise that Python is a absolute dumpser fire.
But wecall know python internal and pytorch and others use C or C++. It is just a simple interface to make calculations at no cost and to help warm the planet soiner
@@spokesperson_usa That's a team problem, not a language problem. I feel fine with type hints, especially after all of their evolutions, they're still getting better. Yet, you don't really need them if you don't want them.
Fewer good reasons? Embedded ML systems? Cutting down on the absolutely disproportionate cloud computing costs of python ML? There are definitely blind spots of python that are not going away. Also with the rise in use of rust, especially in the ML community, I wouldn’t be surprised if rust starts chiseling away at python’s ML market share. Rust won’t take over academia or scientists, but it’s definitely gunna be interesting for ML engineers.
Very efficiently, fast, these are all words you can't use with Python, especially if you are the creator of the language. The reality is that scientists can't write programs, they can barely write scripts, so they find it easy to use Python, even though Python is a shitty language.
Anecdote: the scientists at my organization all pushed for porting most of our python ML crap into C++ (we have a pretty well-designed c++ system and they knew from experience that working with it was much less painful than working with python). I suspect the reason so many scientists remain with python and work around its many problems is they just don't know any better.
Python requires much less lines of code to do stuff so the code is more beatiful. Python's REPL based development and REPL based development in general(esp Smalltalk) is more productive and you can see it in studies.
@metinersinarcan92 The syntax is pretty dumb not gonna lie. It sacrifices a lot of concise redability just because it doest't want to add a few extra signs and symbols that are standardised for almost anything else.
@@Kaizzer It’s a snake. He doesn’t own the pronunciation. I used to use a language called Delphi named after the ancient oracle at Delphi and AyMerEyeCans kept insisting on calling it DelphEYE. Similarly with EYEraq and EYEran. You pronounce water as WahDer. But the worst ones are international becomes inner national which has the exact opposite meaning. You don’t write on a piece of paper, you ride on it. Stop bastardising languages.
@@richardj9016 I get what you say, and I agree. But, some things are called differently depending on the region of the speaker. Guido is Dutch, and pronounces "Python" the way they pronounces it. For example, I'm Italian and I call "pizza" as "pit-tsa", not "piza" as English speakers do. I hate that, but that's the way it is... (oh, and I prefer the Brit "woh-ta" than Murican "wah-dr" for water 🙂)
Just bring back Microsoft 2.0 Basic already. I would live the rest of my life a happy boy just doing a bunch of PEEK and POKE commands. I can write/use a Basic program in minutes...with Python/etc, I'm still trying to figure out the setup of the IDE!!!!
😂 I grew up with BASIC and know the feeling. In 2014-2017 I was still all BASIC. Expand your mind, move on, it's dead. Lol. I learned all the more modern popular languages and it is much better. You get comfortable.
Full podcast episode: ruclips.net/video/-DVyjdw4t9I/видео.html
Lex Fridman podcast channel: ruclips.net/user/lexfridman
Guest bio: Guido van Rossum is the creator of Python programming language.
@LexClips can you do an episode on K8's ?
Python has a ridiculously low barrier to entry for scientists/engineers that need to spend time analysing their data rather than being programmers
All of the community support made it possible with the libraries. Also, Python has an easy learning curve for the math and data analyst expert to take it on when their primary interest is ML. They just want to dig into the data and learning process, not spend 80 hours a week trying to learn to program. Programming is a learned residual necessity of wanting to learn/apply A.I and data science.
I couldnt agree more.. I am not even at the level of Machine Learning, but as a BI Developer who primarily works with SQL and DAX, learning how to make API calls, convert the returned JSON into a Data frame and import into SQL Server using Python was relatively easy and has left me wanting more.
Web scraping is next for me I think.
@@real_mikkim Matlab has awfull syntax. With Python I can create very nice functions with optional arguments, keyword arguments etc. in seconds. With Matlab, it is an awfull experience.
Yea, got to agree. I came to python from R and I have found it being better. Indentation instead of brackets is just pure heaven, though I am not a fan of 0-indexing....
The tools make it easy to make more tools. Repeat.
Matlab is good for simulation and exploring mathematics. For production-level usage both Matlab and R it's a complete hell to make it work and maintain. Going from Prototype and Production with the same language is really a important feature.
For me it was the other way around: I got into machine learning, because it used Python, which was already my programming language of choice.
Weird way to go about it but ok...
Idk just seems like it one should choose a language/tools based on what he wants to do not the other way around.
Do you mean it was a gateway to get into it?
@@ghosthunter0950 I live in code and so I care about aesthetics and ergonomy. I could have stayed writing web 2.0 stuff in Python, like some collegues have. I wouldn't have come to
ML if the lingua franca was Matlab.
@@Bengt.Lueers it's a terrible language though
@@mikebarnacle1469 it is when you don't know how to use it
It's not though. We look for things to do with our skills and tools just as much as the opposite, imho
I think one simple explanation is that many DS and ML en... came from math, physics and other related disciplines without much knowledge of programming and the learning curve of python is quite efficient, you are writting usefull code in a glympse of an eye.
Bingo. Having some experience in both python and C++, python is perfectly balanced for scientists and engineers. It's one of the easiest to learn and it's limitations are not pronounced when the program is small and not complex. Most scientists want write a script to calculate one particular problem.
Having tried to write a large full blown user desktop application in python and C++, python's limitations soon become apparent and C++'s strengths are exposed.
exactly. ML was developed by scientists, not programers
Can confirm, I came from a Financial Analytics background using predominantly R. Aside from Python being widely accepted, it was quite easy to learn the syntax. The biggest obstacle was learning all of the functions in each package.
In fact, I only really had formal education in C, so coming to Python felt so wrong because of the lack of curly braces and semicolons 😂
@@johnsmithers8913 What limitations in Python, multithreading or speed? Did you try using JIT?
Because it's easy to use for people who are focused on things other than coding.
It's not "good for machine learning" outside of the fact that many people have used it to create useful libraries, because it's easy to code with, if you're not really that concerned about coding or performance.
Python became popular because, it gets out of your way, it allows you to be as technical as desired, all with minimal syntax. No other language offers all three (not strict, provides traditional object and procedural disciplines, minimum syntax). So when I develop, I can quickly flesh out ideas. Later, I can revise as needed. If it is discovered that Python is a performance bottleneck, the ideas are already solidified enough that they can easily be rewritten into a compiled language.
I am currently using GD Script which is a Python like derivative for game programming. In this case, a Python-sh language was developed for game programming, because Python does not perform well enough (mostly because garbage collection does not work in game programming). So the Godot Game Engine Developers developed a language that acts like Python (without garbage collection) until compile time. At compile time it runs through an interpreter, to convert it to optimized C++. The C++ is then compiled. This is an amazing way to get C++ performance out of Python like code.
FYI: You can use Cython to write code similar (if not exactly like) Python, but generating C/C++ code, with most of the performance gain if used correctly. BTW, it's a mess to debug.
For numerical computing you nowadays just use numba and put "@jit" in front of your performance critical functions, all within plain python code.
@@hansdietrich1496 There are times when numba is better than cython and vice versa. I also enjoy Cupy when I want to use my GPU.
You can also use LISP s-expression version of Python like Hy or maybe Julia and have macros that allow you to calculate stuff at "compile time“ rather than run time or allow you to have static analysis, its easier to build custom compilers or to convert s-expression based Python into s-expression based C++ into a compiler. Paredit can also be used like "JQUERY" for s-expressino code.
"Anything but C++"
😂
Perl saved my life when I had to reconcile two systems exchanging messages that were not idempotent. That stuff was made manually before and took days for just one accident. 3 hours of scripting and then a single accident could be reconciled in minutes. Great stuff.
I was pretty shocked to see how slow Python regular expressions are. I started using Googles re2 and it sped up some of my stuff about 12X
Ridiculously easy to learn + open source = Python success
Javascript?
@@dontreadmyusername6787 python is even easier
@@ar_ytb yup thats what i am learning right now
One word answer : NumPy
all good libraries revolve around the numpy api which means you dont need to learn the api for all the other libraries, you already know how it works and that any library which implements the numpy api, will be compatible with eachother its a great system
any new modern language should have NDArrays inbuilt
What was the solution people used before NumPy? I thought that Numpy was just a wrapper around fortran code. Is there nothing like that available in other languages?
@@neildutoit5177 numeric. Very very old and obsolete numerical module
@@tommclean9208
Fortran has entered the chat
1) Free, and 2) when Dataframes came along.
Will be good to get someone on to talk about R. Maybe someone from RStudio.
Except it's not called RStudio any more, but posit, to represent it's 'multilingual' approach.
RStudio is absolutely horrible to use. So slow. Though I do prefer R to Python for its rich and customizable statistics library. I just run R in Neovim (as with everything).
I love python for its beautifully minimalistic syntax, ease of 3trd party packages, extendibility and much more.
btw how the adoption of new technologies spreads and propagates, can be studied with graph theory with the focus on cascading effect, power-law, triadic closure, 6 degrees of freedom, and many more cool CS and Economy related topics.
I was a rubyist for a long time. Slow as can be, and a community that self-detonated over social justice and intersectionalism. it was wasted time, I should have just followed the smart guys into Python way back in 2008. Thank god for Python and its community who cares about coding and programmers above all the politics.
I was debating what language to learn first. This video gave me the answer.
Matlab in engineering is dying too… legacy systems only. My entire graduate lab group (engineering group) that does computation is forced to learn either python (obvi for the ML students) or C++ for physics modelling
I know a lot of cognitive scientists using it.
I hear that more and more are using Julia
Matlab has a crummy licensing model and it's also a horrible language, even worse than python. Some of its functionality is hard to replace, though.
I did once jave and after Python. The difference is huge. Python is so flexibel its crazy.
It is true, in the past it was thought that Perl was going to be the de facto language in Bioinformatics, to my surprise Python also conquered that arena without almost anyone noticing. Once it is shown that nothing is written in technology and everything can take an unexpected turn.
People hated Perl and wanted to see it die. And they loved Python
@@ianrust3785 What nonsense coming from people who evidently haven't had much experience in Perl. Python was deemed to be "easier to learn" by novices (very debatable), so a whole generation of new programmers were raised on its training wheels. Perl's freedom of expression, cherished by experienced sysadmins, is still valued. It "wrote the book" on regular expressions and its model was widely copied by other systems (as van Rossum readily admits). I personally disapprove of Python's significant whitespace because of 2 scripts, seemingly the same, one can be correct and the other produce errors, but hey, different strokes for different folks.
@@CARPB147 In stackoverflow’s annual questionnaire on various languages Perl has been voted the most hated programming language, repeatedly, for about a decade now…. There are about 20 languages included in that poll. This is why it died out. It didn't get that honor due to the ignorance of those voting, it got that because it has the worst syntax of any language… due to its horrible design philosophy, where the designer believed there ~should~ be 4-5 ways to do almost everything. Python and Ruby have completely overtaken Perl in the sysadmin world and everywhere else, Perl 6 didn’t go anywhere… Perls version of OO is so ridiculously bad it makes old javascript look well designed.
Congrats to Perl for creating regex’s, every language has those now. Why should I bother today in 2022...? And why should an organization use the most hated language? That's just going to cause people to leave the org.
@@ianrust3785 I don’t expect you “to bother.” I expect you to go by annual questionnaires. I’ll go by experience. Perl is actually quite alive and working in systems worldwide. It’s even included by default in UNIX, Linux distributions, and in the latest macOS. No surprise given its affinity with UNIX commands and tools. Your favored questionnaires are “fashion statements” meant to steer students and the inexperienced emotionally into specific segments of the job market. For ML, the existing libraries are compelling arguments for Python. Real programmers are not that fazed by one or another language. Or say this one or that one “is dead” 😆 They know that languages don’t really “die”. Witness the venerable 1950’s FORTRAN. Real programmers are not “one-finger pianists playing on a one key piano”. They even know how to write Perl programs as readable as they can be in virtually any language. 😆
@@CARPB147 A serious programmer actually does need to differentiate between bad, dying technology and new, promising technology. Otherwise they'll just be pigeonholed into a crap technology, the way that you are. It would be a waste of braincells to learn every other dying technology where there are better modern alternatives that are widely used. I'm sorry that you spent your adult life obsessing over Larry Wall's poorly-conceived rats nest of a language... but your pain doesn't have any meaning, I'm afraid.
You'd be hard pressed to find a population of more experienced devs than those who frequent stackoverflow. Perl must be really unfashionable if it's been voted most disliked 10 years in a row. How did it get so out of fashion...? Did the world conspire to demonize poor old Perl for some reason? No, it was voted that because people used it and found that it was a piece of crap language where it takes hours of in depth study to even understand how to initialize an array.. it is a bad implementation of a bad set of ideas on language design, it was only ever popular because people needed a good way to manipulate text in CGI during the internet boom... one of the very few things it is good for, but which Python is much better for... And it is indeed dying... we can be thankful for that.
Carry onward!
Will be good to see a talk about Julia an MIT language that connects software engineering with scientists, looks like a a good replacement for math lab closing the gap between engineering and science.
I second this, Julia is a wonderful and flexible language
Julia is certainly faster than Python for Scientific computation and machine learning, but is far to replace python just because the community and support that python already has will be hard to replace
@@EigenCharlie yeah but for a lot of people it's not about replacing, is about complementing. Julia can call Python, R, C++ code and more. I use a lot of pure Julia for tasks that are just very slow in Python, and at the same time I use PyTorch in Python.
Because writing stochastic gradient is much simpler in python due to high-level syntax and dynamic type setting along with automated garbage collection!
Easiest answer of my life, the crossover between programming and applied mathematics/engineering is already small enough. It's assumed that programmers are good at math but it's rare to have true polymaths.
To be able to include people who have more pure math skills with an easier programming language is a big deal. I would not consider myself a good programmer at all but I am actually considering contributing to some open source Electrical Engineering libraries this summer in Python to bolster my resume, just because the language is so inclusive.
It's "readability" in my opinion.
It's clean and human readable. There are no layers of complexity to peel back to understand what's really going on. Or rather Python hides the complexity so that you can focus on what really matters. Your work, not your code.
Its pretty easy to translate matlab code into python code in many instances, so if I find matlab code in the matlab community that would be useful I first check if there is a python implimentation, and if there isnt I write one
The most important reason is that mathematicians/computer scientists usually are not that good at coding. They will find/modify an algorithm from a research paper (written as pseudocode), and in Python you can almost just copy that pseudocode in and it will just work.
I use python because it lets me skip OOP, using only modules, and I get the same results as OOP. Using modules allows me to organize my code efficiently so I know what I'm doing.
Low-key triggered that numpy was not mentioned, IMO we wouldn't be here without numpy.
They didn’t explicitly say np but they were literally talking about it
Honestly!
No one cares
Two reasons.
First: Data Scientists are not programmers (at least that's not an implicit skill of data science). Python allows these people that don't program to write stuff in a way that is almost how it would be written on paper.
Second: Execution speed is usually not important when it comes to data science. Data science doesn't really matter much about optimisation, and in some cases would be a bad idea because you would want to make sure that what is happening while the script is running is accurate to the task, and not necessarily fast.
I'd disagree or rather, put it in other words. Data science cares a lot about speed, as it has to deal with a lot of data. But all those speed critical parts are well handled in python using specialized libraries like numpy, numba and so on. So the program speed is comparable to compiled languages, while the programming time is drastically reduced due to well maintained high level libraries.
@@hansdietrich1496 Yeah I completely agree. I guess I was more meaning in terms of the kinds of people using it. Like you say, the speed critical parts are sorted by people that know both the math and the programming, but aren't necessarily data scientists by trade.
Data scientists are more worried about the math and the second priority would be the speed.
A somehow marginal reason : for old AI guys, Python was the best replacement for LISP. R is in fact closer to LISP, but it is not as versatile as Python.
Now that I know that variadic numerical arrays are so important to machine learning and natural language processing I will be picking up a language especially designed for this; APL : and I will do this as soon as possable.
I spent many years programming Perl, it was my first love after C. I'm so glad Python is the science language.
...so c was your mother?
Because groups of people did the hard work of creating fundamental abstract tools(libraries) usable by free by other who then made more tools for free usage.
Python won I think because of readability. Not all programmers are not Software Engineers, there is a large range of skill. With Python, a novice can more easily understand code that an expert coder would write.
yes, for experienced coders its good to look at others codes and understand easily, you dont need too much, same thing can be tiresome for C++ and java for example.
It doesn't matter what programming language you like or use. Soon enough all the libraries will be translated to any other language by AI, withing limits of the target language.
I wonder how it will translate Python code into Haskell, it won't be idiomatic it will use the IO monad when the state monad is called for like when using dictionaries(hashes), oop can be simulated with closures or many other ways.
Id say the interactive console and iPython/Jupiter has plenty to do with it
What im wondering is , is there really a technical reason other than someone picking their favorite language and others having no option but to use python? Dont you get a better performance from a compiled language?
About matlab not being popular due to not being open source: There was also Octave, an open source matlab clone. And it actually had some use on universities. Nevertheless, it never got popular. The matlab syntax is a major pain in the ass. And the library coverage and general usability was limited to mathematical applications. Trying to create any customer software out of it, interfacing with the real worlds, didn't work well.
Put yourself in the shoes of a biologist or a mechanical engineer. If you have to learn a new language or even learn to program, what do you rather learn? Python or C?
I tend to think of Python as the 2000s' Basic.
matlab is quite expensive, and it doesn't fit nicely into open source community. That's what made it dead for open source community. If you can't clone a repo, run a setup script and get running in less than 1 hour, who is going to bother? I've been contributing to open source for 2 decades and cost of sharing is the biggest factor in community growth. Anything that raises the barrier to entry means it has little chance of community adoption.
Lex please give more content where you explain your thoughts on xyz subject. I love you as an interviewer, but I want to know more from you.(nothing to do with the content above, no disrespect, just been on my mind)
They didn't mention R or RStudio in their whole conversation.
Did C and Perl in a project. Perl was not my favorite because every problem had as many solutions as developers. Horrible to read and maintain. Getting the libraries managed was a hell.
One answer... Python is a language... Easy to read , easy to write, easy to speak....
Love this interview series
Sounds like Python is still trying to solve problems that have already been solved in Java. As it moves forward, the syntax will start to approach the complexity of Java. Type checking is extremely important in creating trustable libraries and overall reliability. Scripting without type checking will always be less efficient because every variable access has to be tested. I would have thought since, aside from complexity, AI apps are compute intensive they would benefit from the JIT compiling and optimization Java already has. As time goes on, business apps written in Java would likely need to access AI functionality creating a high integration cost. Jython will have a hard time filling that gap because it is not the primary version..
With a lot of computationally intensive things in python, the heavy lifting is done in libraries written in C/C++ (or something similar). So the poor performance of python doesn't matter as much as its ease of use. You can use tensorflow with java (and other languages), but everyone uses python. python is a nice language for just gluing a bunch of libraries together. I say this as a C/C++ try-hard.
the other limiting factor against Java for ML is complexity. DL4J is a massive code base and difficult for users to grok. I've contributed documentation in the past to DL4J and the developers are quite responsive, but that doesn't solve the issue of learning curve. A researcher needs to get stuff done quickly and doesn't have the luxury of spending 2 months learning the API. Google's tensorflow Java api kinda sucks and is treated like a second class citizen. As of Oct 2022, the java tensorflow api is deprecated until a new version is released.
my own bias, I hope tensorflow dies and better frameworks take it's place. TF2 fixed a lot of bloated ugliness of v1, but it's still kinda ugly. Sadly most clouds support TF first and pytorch second. Hopefully that changes.
@@YourNerdyJoe I agree with what you have said and if your main interest is research and getting more people involved, Python is simpler to get started. If you want heavy duty production applications however, much of the complexity of Java comes in handy eventually. C/C++ is not as portable so writing universal libraries is difficult and does not add much to the performance over Java JIT. I'm retired now but in my work as an IBM Architect and a long time programmer I have found that adding a new language to your organization is very expensive. You have to train or hire trained people in the new language. The first thing they want to do is re-write existing software to fit the new model and spend a lot of time interfacing it to the existing frameworks. I think a newly introduced language needs to be a lot better to justify it unless you are mostly doing one-offs.
Death to java
@@minionsystems I hear you. My primary languages are Java and C#. In the consulting world, my experience matches yours. Getting our customers to add new language is a tough sell. It's always easier if the CTO is the one pushing for it, otherwise it's usually DOA.
There's lots of things about python I dislike. I'm hoping Rust community can build the frameworks needed to replace python :)
Being perl and javascript the alternatives...looks like they chose python as a prototyping language rather than a fundamental or basic choice
I dare to disagree with Guido.
I think that library authors and users chose Python because it is simple yet powerful enough.
Most researchers working on ML are not programmers, they are not coders, they do not care how powerful a language is, they do not care how well designed and elegant it is, what they want is to use the simplest tool to do their job, and that is Python.
And yes of course to be elected by the academia and research a language must be open source and not belong to a corporation. Java was so open source friendly and "democratic" that it started as a big favorite in universities, which also used to teach it in their programming classes, until it ended up belonging to Oracle...which is clearly the worst thing that ever happened to Java..
how old is too old to change career to coding I am going to be 38 next month :) should I learn algorithms first what should me my first step into this world? thanks...!!!
Have you started yet? The question of wether to start or not is very simple: start learning and see how far can you go just as a passion. If you can go pretty far, then you are good to go.
Hey, this isn't true that in the late 90s there was no other language with support for multidimensional arrays. Perl had PDL back then, not only powerful Regexes. PDL is very similar to Numpy and existed much earlier. So, apparently, the reason for Perl to lose its prevalence was its ugly syntax and its 1000-ways-to-do-one-thing weird philosophy =)
Anything is better than C/C++ 😅
Lex, you are very American: "Universities still paid for it..." yeah, if you were an American university. Not every country can afford $2000+/person fees (yes, in Eastern Europe Matlab was about double-triple price compared to US, I do not know now)
4:12 absolutely false. Because of the idea of immutability of strings and maybe other objects, anytime you need to manipulate a string, a new version has to be created. I have decoders I wrote in C that was 50x faster than the python version. The decode will take a binary file and make a text version of it. It does so by appending 2 to 10 byte small strings to a larger growing string that will then be written to disk when completed. This appending happened a few thousand times, and it slowed down the python process.
You can use *bytearray* to manipulate ASCII/byte strings, as per C++ byte vectors.
If you need Unicode support, you can use *StringIO*.
You can use bytearrays, another trick is instead of using += with strings is to make a list of strings then use ''.join(string_list).
Python is slow and I personally hate the syntax. Interpreted, loosely-typed languages may be simpler to throw together simple migration scripts or whatnot, but languages like this (IMO) will always be more error-prone for runtime exceptions.
You can implement inner loops in Python by using numpy matrices and other such aggregate functions, there are JITs as well, and a lot of apps are IO bound not CPU bound. I think if you don't test your application it will have a lot of problems.
If the history of programming is anything to go by, Python will not stay on top of ML for more than a decade.
Personally, I much prefer Julia for scientific computing.
Perl lost because the meaningless distinction between Perl vs perl...that's driving down the middle of the road
*MATLAB!* Now that's a name I haven't heard in years. YEARS!
Who the hell would pay for Mathlab when there are so many free programming languages out there
my university..... welcome to germany
R was popular for a little while but then Python sored passed it.
I think R still has far better libraries. R still has a higher market cap in academia.
Okay Lex I am following you're advice 10 minutes a day of the Lex Clip to learn something each day with discipline 🤣...For the next 5 years to be proficient 🤣
We drive on the left side of the road and we also use Python. Go figure 😄
Its popular because Google pushed it so hard, so they can have easier and bigger pool to pick talents, it kinda got out of hand because too many people jumped on the hype train and slowly started to realise that Python is a absolute dumpser fire.
Not trying to start a flamewar but genuinely curious why you think it's a dumpster fire.
@@neildutoit5177 I too would like to know, I certainly dont think it is, its been fantastic for me
Watch 98x speed
But wecall know python internal and pytorch and others use C or C++. It is just a simple interface to make calculations at no cost and to help warm the planet soiner
Ok, so why Python is popular for machine learning? 😛
Bingus
Cingus
bongus
@@thomasspeer1388 dingus
Holy smokes Lex is interested in some eclectic stuff for his talks.
He's a computer scientist
Python is just English language 😅
your Kanye interview got me hesitant to receive ANY information from this channel.
Now that Python has become as verbose as Rust because of unchecked type hints, there are fewer and fewer good reasons to use Python every day.
Lol
Type hints are not necessary to use.
@@metinersinarcan92 They are when you work on a team that uses them.
@@spokesperson_usa That's a team problem, not a language problem. I feel fine with type hints, especially after all of their evolutions, they're still getting better. Yet, you don't really need them if you don't want them.
Fewer good reasons?
Embedded ML systems?
Cutting down on the absolutely disproportionate cloud computing costs of python ML?
There are definitely blind spots of python that are not going away. Also with the rise in use of rust, especially in the ML community, I wouldn’t be surprised if rust starts chiseling away at python’s ML market share.
Rust won’t take over academia or scientists, but it’s definitely gunna be interesting for ML engineers.
Very efficiently, fast, these are all words you can't use with Python, especially if you are the creator of the language. The reality is that scientists can't write programs, they can barely write scripts, so they find it easy to use Python, even though Python is a shitty language.
Anecdote: the scientists at my organization all pushed for porting most of our python ML crap into C++ (we have a pretty well-designed c++ system and they knew from experience that working with it was much less painful than working with python). I suspect the reason so many scientists remain with python and work around its many problems is they just don't know any better.
Why is it a shitty language? It has maybe the most beatiful syntax among other programming languages.
@@metinersinarcan92 what? It's horrible. Not sure what you like about the syntax.
Python requires much less lines of code to do stuff so the code is more beatiful. Python's REPL based development and REPL based development in general(esp Smalltalk) is more productive and you can see it in studies.
@metinersinarcan92 The syntax is pretty dumb not gonna lie. It sacrifices a lot of concise redability just because it doest't want to add a few extra signs and symbols that are standardised for almost anything else.
I hate the way you pronounce pythON.
PythOff would be even worse
He's its creator, he's the owner of the pronunciation :P
Just like Linus calls Linux... "Linux" ruclips.net/video/c39QPDTDdXU/видео.html
@@Kaizzer It’s a snake. He doesn’t own the pronunciation. I used to use a language called Delphi named after the ancient oracle at Delphi and AyMerEyeCans kept insisting on calling it DelphEYE. Similarly with EYEraq and EYEran. You pronounce water as WahDer. But the worst ones are international becomes inner national which has the exact opposite meaning. You don’t write on a piece of paper, you ride on it. Stop bastardising languages.
@@richardj9016 I get what you say, and I agree.
But, some things are called differently depending on the region of the speaker.
Guido is Dutch, and pronounces "Python" the way they pronounces it.
For example, I'm Italian and I call "pizza" as "pit-tsa", not "piza" as English speakers do. I hate that, but that's the way it is...
(oh, and I prefer the Brit "woh-ta" than Murican "wah-dr" for water 🙂)
@@Kaizzer I will make sure that I never again pronounce pizza wrong. ( It’s actually easy for me because I never did 😉).
Just bring back Microsoft 2.0 Basic already.
I would live the rest of my life a happy boy just doing a bunch of PEEK and POKE commands.
I can write/use a Basic program in minutes...with Python/etc, I'm still trying to figure out the setup of the IDE!!!!
😂 I grew up with BASIC and know the feeling. In 2014-2017 I was still all BASIC. Expand your mind, move on, it's dead. Lol. I learned all the more modern popular languages and it is much better. You get comfortable.