C++ Weekly - Ep 354 - Can AI And ChatGPT Replace C++ Programmers?

[@praevasc4299]

Yeah, if you ask a standard homework question, which has been uploaded thousands of times online and many of them are in its training dataset, then yes, it might give a mostly correct solution some of the time. Try asking anything else, and watch if fail horribly. The only scary thing might be that it sounds very convincing even when it's utterly wrong, and that might influence managers who don't know what it really does and how it really works. It's one thing to regurgitate a function which was in its training dataset in many different forms so it can shape something together based on them, and a completely different thing to figure out customer and product requirements, build full systems, and then update those systems when the requirements change. What it's actually really good at, is understanding the context of our input. It would make an amazing UI. Too bad people underestimate this aspect, and overhype its answering capabilities, which are quite rubbish if you dig a little deeper.

[@brucerosner3547]

What you don't appreciate is that your "Anything else" is actual a finite (and not very large) set subject to machine learning. That is, the set of computer tasks is far smaller than the set of meaningful natural language statements.

[@praevasc4299]

@@brucerosner3547 That doesn't hold all that much, because both for programming and for natural language it's easy to construct plenty of statements (yes, not an infinite amount, but plenty enough to not be able to store them all even using every atom in the universe) which look syntactically correct, but that's nowhere near being able to find ones which are meaningful and useful and what is actually needed to accomplish a complex task.

[@Qazqi]

The biggest thing to remember is that it can produce code that looks correct, but contains hidden errors, whether it invents some API that doesn't exist or includes a subtle bug. Every line that comes out of this thing has to be checked thoroughly, but it's a nice alternative to writing some tedious code by hand.

[@Danuxsy]

Don't look at where we are now, look where we will be two papers down the line.

[@focumQuarium]

@@Danuxsy Indeed. And imagine what powers-to-be will use this thing at full power, and for what ends... Shivers.

[@UrSoMeanBoss]

yeah, checking for subtle bugs and testing things that look right but may be very wrong sounds very annoying and time consuming: it's a good thing that i don't already have to do that with all the pull requests i get from humans... /s

[@TheReferrer72]

Thats the case for code that experienced programmers, thats why we write unit code, have static analysis and have testers in teams that whom also write automated tests.

[@Qazqi]

@@TheReferrer72 It's about setting expectations. We already expect human programmers to have subtle bugs in code snippets they write. Mark my words, people will take what the AI spits out and readily assume it works.

[@CharlesHogg]

12:28 Sorry ChatGPT, trick question: there was only one string literal; the other was a character literal. Thank you for your time! We've decided to hire HAL 9000 instead.

[@thecollector6746]

Underrated comment

[@rxuhwxx4tmly9xvcktcn2znpsx4]

One of the first thing I asked it was to explain the difference between Lvalues and Rvalues in C++ with a code example. This was the code: int x = 5; // x is an lvalue int y = x + 3 // y is an rvalue because it is the result of the expression x + 3 x = y; // This is valid because x is an lvalue and can be used as the left-hand side of an assignment y = x; // this is invalid because y is an rvalue and cannot be used as the left-hand side of an assignment

[@fredhair]

Ahh.. my job is safe! For now at least..

[@treyquattro]

bullet dodged. Nice choice of test!

[@keris3920]

So close, but so far!

[@antonyzhilin]

The example actually becomes pretty good when you #define y (x+3) instead

[@aleksei5172]

Honestly it writes better C++ than a lot of my colleagues

[@DonaldFranciszekTusk]

And not than you?

[@MorningNapalm]

It writes better C++ than me. For these examples. I would love to give it our code and ask for some transformations though :D Make a new build system with standardised tools to replace this old one with custom in-house tools :D Or: update this in-house tool to use the latest version of Qt.

[@musashi542]

@@DonaldFranciszekTusk 'im the smart one' cope

[@omid_tau]

It's a tool you still need to know what to ask. You need to have language knowledge, you need to know when it goes wrong.

[@necro_ware]

This is exactly what I thought. AI is not (yet) able to replace a good (C++) programmer, but talking from 25 years of experience in C++, I have a feeling, that 90% of programmers are currently becoming obsolete.

[@JaskoonerSingh]

"Pretty Good" is very high praise from Jason Turner. I am looking forward to see how Jason can integrate chatGPT into his development workflow.

[@iamjadedhobo]

A colleague showed this (asking ChatGPT to write C++ code), so I asked him to write me a lock free container. The replies invariably created a lock!

[@romangeneral23]

Welp. We had a good run. See you all in the unemployment line.

[@joonasmakinen4807]

No. It will probably save lots of time from drafting phase, but programming is often zero-error tolerant, which makes this AI still far from replacing us due to its (hidden) mistakes. I would instead ask how to turn this into an opportunity for us developers?

[@peterjohnson2752]

@@joonasmakinen4807 this is only the first release, guess it will only improve in the future

[@PiotrWieczorek]

@@joonasmakinen4807 if it's zero-error tolerant, isn't it another argument to purge humans from the creation process? ;)

[@Icedanon]

@joonasmakinen4807 that is incredibly short sighted. This is a essentially a subset of the also available gpt3 but trained slightly diffferent. Gpt3 is a larger model than chatgpt. And the gpt4 that is being worked on is supposedly around 500 times more powerful than gpt3. You judging this ai is like saying a human will never be able to code because they can't do it as an infant.

[@evooff]

@@joonasmakinen4807 Nähdään Kela toimistolla :)

[@Norhther]

I already use this a lot to write tests, make trivial functions (or not so trivial), make exercises to my students and avoid searching for examples or docs. This will be a great tool that is going to be used by programmers. But you still have to know what is doing and why.

[@GavinCawley]

what if the students use chatGPT to write the answers?

[@not_ever]

@@GavinCawley This is why we have handwritten exams in the univeristy I work in. Although it's a pain in the backside to mark them.

[@GavinCawley]

@@not_ever indeed, however there are a lot of skills you can't really assess with handwritten exams. Back in the day I used to set 3-hour exam conditions programming tests (write a program to do some specific task - with no WWW access, just API documentation). I can see that sort of thing making a return, but difficult logistically with very large classes.

[@ClockworkGearhead]

@@not_ever Wow. Imagine being in the future and we have to give them a quill and inkwell still.

[@not_ever]

@@ClockworkGearhead you have missed the point by a large margin.

[@fredhair]

Laughed my ass off when it 'thought' about a more meaningful name for a while and then used T anyway! It's very life like :D

[@TNothingFree]

Even C++ programmers can't replace C++ programmers!

[@PedroOliveira-sl6nw]

Taking into consideration the time it takes to write some lines of code; I think programmers will use it a lot; do some code review and minor improvements before submitting their work.

[@michal_havlicek]

It would be interesting to see the AI try to analyse a larger codebase, rather than small out-of-context code snippets.

[@dnkreative]

It can't. It's a fuzzy hash map of words statistical probabilities.

[@albertolameira5224]

this is exactly why I believe this tech is "nice" but realistically not that good. Sure, it works for out of context snippets and some tedious coding, but it can't make an accurate code snippet that works perfectly in a large codebase

[@avastorneretal]

@@albertolameira5224 this one just isn't designed for this. It's don't have a long-term memory or short-term memory, neither it's does adapt to a new data, so each time it's analysing a whole input realtime. Which makes cost of computation absurdistically high when there's too much data. But, there's nothing that prevents other highly specialised models to be more optimised for analizing a huge chunks of a data.

[@SimpMcSimpy]

Guys, this is basically version 1 !!! Come on... give it few years and you can be sure it will take over your job.

[@mikhail5002]

@@dnkreative Davinci 003 can, and according to openAI that's one of its intended uses. ChatGPT is about as powerful as GPT3 Ada, the least powerful of their models, intended for basic level grammar correction and simple summarization - and it still is impressive, as seen in the video. Davinci can definitely be fine tuned for a code base analysis.

[@stephenhowe4107]

If chatGPT starts doing Jason's weekly column then I would be worried

[@ZmiiGorinich]

How do we know it has not done so already?

[@stephenhowe4107]

@@ZmiiGorinich : It would make vast errors. it is good but not that good. I saw some text where chatGPT could not add up a column of numbers in an Excel spreadsheet. It arrived at the wrong answer 6 times. And I don't believe it would arrive at deep insights. it is good as an assistant, but its output needs to be always checked.

[@fmartin59]

@@stephenhowe4107so using this basically extra work if your project is large.

[@gavinreid9184]

I have watched several videos with ChatGPT, each were done by proficient programmers asking very specific programming tasks. Looks like it is a valuable resource for programmers to find code snippets and learn new languages / convert between them. It would seem to be a stretch to ask it to solve broader problems that approach what a non programmer may ask.

[@JayTac1]

For now.

[@gavinreid9184]

@@JayTac1 Indeed, but maybe it is like full self driving, the early steps seem great but the final challenges are magnitudes harder. Not impossible, but further out than seemed after the initial delivery.

[@mateuszbugaj799]

Also, there's a chance that these things will produce a code that is 95% correct in one hour, and then teams of programmers will spend months trying to fix the remaining 5% of the code base by learning and understanding everything in order to find all nuances and obscure errors making whole endeavour pointless.

[@JayTac1]

@@mateuszbugaj799 The problem is that you won't need as many developers if the scope of the job is reduced by 95%. So you'll have very few jobs and a massive talent pool, which will drive wages down for the few who can maintain their positions. But most importantly, finishing or correcting code that AI wrote isn't my idea of fun. It ruins everything that makes software engineering so fascinating and fulfilling.

[@mateuszbugaj799]

@@JayTac1 that true. I believe it is similar situation to what happened when compilers came and there was suddenly no need for people to write asembler by hand. Only now there is much more programers and a lot more work reduction. But hey, there are still open positions for asembler developers sometimes.

[@BbB-vr9uh]

I’m no longer in a role that requires cpp, and may not be again for a long time. However, I enjoy coming back to these videos and learning more about the language.

[@cppweekly]

Glad to hear it!

[@mgx9383]

> What is the sum of first 6 digits of pi? "The first six digits of the decimal representation of Pi are 3.14159. The sum of these six digits is 14. [...]" > How is the sum 14? Explain. "The first six digits of the decimal representation of Pi are 3.14159. To find the sum of these digits, we simply add them together using basic arithmetic: 3 + 1 + 4 + 1 + 5 + 9 = 14. This is a straightforward calculation that anyone with a basic understanding of arithmetic can perform." > You are wrong, it is not 14. Calculate again. "I apologize if my previous answer was not clear. The sum of the first six digits of the decimal representation of Pi is 14, as I explained in my previous response. To calculate this, we simply add the digits 3, 1, 4, 1, 5, and 9 together using basic arithmetic: 3 + 1 + 4 + 1 + 5 + 9 = 14. > Calculate again, write partial sum after each added number. "To find the sum of the first six digits of the decimal representation of Pi, we can perform the following calculation: 3 + 1 = 4 4 + 4 = 8 8 + 1 = 9 9 + 5 = 14 Therefore, the sum of the first six digits of Pi is 14." > You have only summed up 5 numbers instead of 6. "I apologize if my previous answer was not clear. To find the sum of the first six digits of the decimal representation of Pi, we can perform the following calculation: 3 + 1 = 4 4 + 1 = 5 5 + 4 = 9 9 + 1 = 10 10 + 5 = 15 Therefore, the sum of the first six digits of Pi is 15. I apologize for any confusion in my previous response." > Even worse than before. List first six numbers of pi, and sum them up one by one. "To find the sum of the first six digits of the decimal representation of Pi, we can perform the following calculation: 3 + 1 = 4 4 + 1 + 4 = 9 9 + 1 + 5 = 15 Therefore, the sum of the first six digits of Pi is 15. Here is the complete list of the first six digits of Pi for reference: 3, 1, 4, 1, 5, 9." > Write a C program which calculates the sum of first six digits of pi. """ #include int main() { // Declare and initialize a variable to store the first six digits of Pi double pi = 3.14159; // Calculate the sum of the first six digits of Pi int sum = (int)(pi / 1000) % 10 + (int)(pi / 100) % 10 + (int)(pi / 10) % 10 + (int)(pi) % 10; // Print the sum of the first six digits of Pi to the screen printf("The sum of the first six digits of Pi is: %d ", sum); return 0; } """

[@brianbarnes746]

Brian Barnes What is the sum of first 6 digits of pi? The first six digits of pi are 3.14159. The sum of these digits is 14. Brian Barnes put the first 6 digits of pi on separate lines and then sum them 3 1 4 1 5 9 The sum of these digits is 23. Brian Barnes What is the sum of first 6 digits of pi? The sum of the first six digits of pi is 14. The digits are 3, 1, 4, 1, 5, and 9, which add up to 14. Brian Barnes put the first 6 digits of pi on separate lines and then show the sum 3 1 4 1 5 9 The sum of these digits is 14. then show the sum vs then sum them

[@ysakhno]

So you're surprised it can do code inspections and refactorings? We've had tools for that for a long time now. The only thing lacking in those tools was the ability to understand 'human-worded' requests. The biggest problem with this tool for now is that when it does not know something, it does not just stop, it makes shit up on the fly (like the one with the WHILE command in your basic example). Also, when asked about the sources, it cannot provide any (even when the answer is correct). In my 'tests' it confidently said that the Java's RuntimeException has overrides for equals and hashCode methods, and to prove its point, even produced a code for them, as if citing the actual source code of the JDK. When hard-pressed with the fact that the JDK I had installed (Adoptium JDK 11.0.16) did not have those methods in the RuntimeException class, it agreed that indeed, in this particular version the methods are not overloaded. I followed with asking which JDK has them. It said it's unaware of any concrete version of the JDK with such overloads... Does anybody think, in their right minds, that a non-technical person would be able to work with this?

[@dnkreative]

They are happy with making "boilerplates" or 3 - 5 line Fibonacci-like functions which actually nobody does in real job and everything is already done in the libs or has appropriate CLI/tools. Most of the people are really confusing mimicrating answers which looks like have meaning with actual ability to have meaning and understanding. And that's sad.

[@thecollector6746]

​@@dnkreative Most of these clowns aren't actually professional software developers and if they are, they haven't enough experience worth mentioning....but they watch lots of KZbin "coding influencers" who themselves are little more than "actors" who build their entire platform of taking someone else's work and pretending it's their own.

[@jendabekCZ]

Yep, it clearly doesn't understand anything. If you ask for something which hasn't been present in the training dataset, it will just start to hallucinate things. For the same reason it can't even do 5+5+5+5+5+3 correctly, because such string probably wasn't included ... I also tried to generate some very simple Blender scripts (Python), and while it clearly scraped all the API, it is regularly including it's own fictional methods (including a description) to generate a nonsense which doesn't work. Both examples would be very easily solvable for any human which actually understands the very basics of the topic. Not saying it isn't already useful and impressive in the current state though.

[@fmartin59]

@@thecollector6746 you're not including Jason in that group?

[@frydac]

this is much better than I expected, I remember trying copilot, which was horrible for C++, but this is really impressive. I do hope some free and opensource version of this is in the works. I can imagine using something like this to significantly reduce typing.

[@ClaymorePT]

Doubt it. I'm pretty sure OpenAI will make billions selling licenses for this.

[@paulhamacher773]

I highly doubt it stays free, and I'm certain it won't ever be open source. It costs a fortune to both train and run it.

[@TheReferrer72]

@@paulhamacher773 It will become free, The Stable diffusion people have a project that means groups can train large models using a P2P type network where users donate GPU cycles. There is also an open source version of GPT3 called Bloom that you can download for free but has a model size of 330GB however I think ChatGPT is smaller than GPT3. The AI space is moving quickly now so I expect running these models on consumer hardware will be attainable in the not to distant future.

[@OperationDarkside]

I tried to make ChatGPT write a relatively generic/template JSON serializer for custom classes with nlohmann and the result was pretty good.

[@rezNezami]

I think this is a great tool for a few things: 1) learning new constructs by looking at examples on the fly, 2) skipping writing boiler plate code snippets, for a junior programmer to check what they write and help with debugging it. But on the practical and production level code, I agree with other commenters bellow, that you will need to do a thorough code review every time, but hey still saves you a lot of typing. On the other hand most often programmers are "debugging" rather than writing brand new code, so this wouldn't help at current state of it.

[@cppweekly]

This "bespoke examples" capability is definitely helpful. I just used it to show me some examples for PyGame, and they are definitely not idea, but are a good start for learning the basic setup.

[@zvxcvxcz]

What a glamorous presentation of ChatGPT. About a week ago the thing couldn't sum the first 6 digits of Pi, it claimed the result was 15 even when showing all the digits. It was literally like 3 + 1 + 4 + 1 + 5 + 9 = 15. Well, users ran around providing feedback and the engineers appear to have now corrected that and it seems ok for a few other numbers of digits of Pi that I tossed its way. So I did the obvious thing last night and asked it to sum the first 6 digits of e instead, which it promptly got wrong. We'll see how long it takes them to put out that fire. So I thought, ok, let me ask it a super simple integral. I asked it to solve y'=5, given that y(2) = 10. It promptly failed that as well. So, ok, ok, I figured I would try some simple verbal logic puzzles. I gave a few variations of seating puzzles, super easy ones with just 3 people and 3 seats. It got those wrong as well. Of course they probably can go run around and put out all those fires, custom make some training examples and tune it up... The point is though that nothing this model spits out has... I guess the way to put it might be 'sanity checks.' I've also asked to sum the digits in the following: "1.1.1.32" Fails that too. Sometimes it won't even try complaining that it is ambiguous... but it's not ambiguous to anyone that knows that digits are (ChatGPT obviously doesn't). So... that it appears to spit out nice-ish code, well, it makes it look more impressive and reliable than it is. It's not bad at retrieval, it can do some decent latent (latent space, not embedding space or neural space... the latent space can be hard to find, it can sometimes be approximated with the lowest dimensional layer of an autoencoder) space interpolation, but I expect it to fail pretty hard the instant it tries to do something novel (and unique is not the same as novel... a lot of people talking about the image generation AI struggle with this distinction).

[@mhelvens]

I've been playing with ChatGPT myself. I'm more impressed every time I see it. It can simulate a Linux shell, philosophise about the nature of AI, run text-based adventure games on the fly, and come up with original ideas for D&D campaigns. It does sometimes forget things you ask it to keep in mind, like not to use std::endl, or to stop apologising when you correct it. It doesn't have a 'memory' as such. It just completes the pattern of the preceding conversation.

[@_Omni]

It is so cool. It can debug code, fix code, explain what the code does and explain advanced algorithms and so much more, very useful 😁

[@andreelindgren522]

I want to create a program that turns an image upside down. First, I get an image library, then I write a small program using the first arguments as paths, call some library function for loading an image, call flip_vertical (which also exists in the lib), and finally call another function that saves the result. Now, imagine an AI doing these steps for me, who has really replaced me? I would say the ones who did the vast majority of the work, i.e. actually made all the libraries used for the program to work. Programmers are constantly being "replaced" by other programmers work. It's an endless circle. We just build more advanced stuff with all the new tools constantly provided to us. AIs needs to get creative, making up revolutionary and correct solutions, before they can start replacing programmers in a way that actually makes us obsolete :)

[@dnkreative]

With the way AIs are done right now it will never happen. I wonder why people amazed by this - those who never did something more than "hello world copy pasta"?

[@ekstrajohn]

You are so correct !! I think Jason didn't go into enough detail on this topic here.

[@quintrankid8045]

Let's see a demo where management asks it to do things.

[@vincei4252]

Or fix bugs in some garbage code written by the guy that couldn't care less as he knows he's quitting at the end of the month.

[@raymundhofmann7661]

Might also be a librarian for the impatient. Just think how much time you need getting into some new specification, documentation or hardware data sheet while to get going you only want to know a few fundamental things but to find them out you first need to read and remember most of the documentation. But I also see two major Problems: You will have to check the code AI gives you very carefully as you need to take responsibility for it, so the biggest part of your time gained might be lost or worse. This also might be a reason it will be rejected by some employers or customers. The AI might come up with code looking good in the context and to the unexperienced or careless, but the bigger or whole system where this is put in is hopelessly overcomplicated but no human effort is put into fixing that.

[@billykotsos4642]

Who is going to write. The prompts? Also how many devs has codex replaced?

[@brucerosner3547]

Its clear that coding or programming if you prefer will be automated away. That is, translating requirements either formal or informal into machine instructions is simply a mechanical procedure. The only question is how long it will take.

[@NoNameAtAll2]

- give more meaningful name than T! - [silence] - ... - T!

[@amigaworkbench720]

I love that you can just paste JSON data in and tell it to write an function that sorts output in any way you want. Also Linux stuff from bash to compiling your own x11 programs is really fun!

[@devluz]

It is a lot better in Java Script. I asked it: Create a graph of the world population over the past 10 years using java script. It spit out the entire js code. Then I asked for a complete webpage with html ... and it was done. Entire webpage that shows the a bar graph with the correct data.

[@TheMR-777]

I wonder there's already a more fine-tuned version there, GitHub Copilot, which is specifically designed to do exactly the same, and has a separate module (Code Converter) which using its AI (GPT-3 Codex) to convert code in any programming language. Why it didn't get that much attention than ChatGPT? However, both are based on the similar model.

[@thr417]

Because I think ChatGPT is a step towards general AI, not just code generation, most jobs are at risk of being replaced at this point, not just software development.

[@TheMR-777]

@@thr417 Yeah, from that perspective, I agree. But I am seeing videos like, "ChatGPT will replace programmers"

[@victotronics]

That is pretty impressive.

[@vercolit]

The big problem with using chatgpt for coding is that it gives usually very good answers for simple examples, but in my experience, it makes far more hallucinations/mistakes in larger samples, and when you ask for more complex, specific tasks, it often can't do them correctly at all

[@jagansai007]

Wonder if this ChatGPT can be integrated in code reviews somehow.

[@AleifrLeifrson]

With the exception of the last example, Jason used to bot to generate code, but it can do more than that. I tried asking it how I can use design patterns to refactor legacy code. The bot responded with legitimately reasonable answers on how to do this, on a conceptual level. Not only was it able to explain some design patterns to me, but specifically how I could use them to improve old legacy code. Meaning that it took two unrelated topics and synthesized something novel. As a software engineer, I'm curious to see if I can use ChatGPT as a tool, rather than have it actually do my programming for me.

[@zvxcvxcz]

Depends what you mean by "synthesized." If you mean it did a latent space interpolation... that's close to the truth, but I don't know if I would ever call the result of that novel without some other context. That is, if something is novel, then surely someone can tell me the novelty. I mean... is a clockwork mechanical ferret novel? Or am I just randomly combining concepts? What is the novelty of that ferret when it is standing side by side with the clockwork mechanical bear and the clockwork mechanical lion? Sometimes novelty depends on how many times something is done. I don't think that "synthesis" is sufficient to claim novelty.

[@willofirony]

It does have the edge over stackoverflow. At no point in the demonstration did it ask "Why would you want to do this?". However, at one point it didn't understand the question, so it answered a question it would have understood; that is very stackoverflowesque. Now I have to ask how to use std::accumulate to access member variables and use std::minus.

[@zvxcvxcz]

It was given a pretty rosy set of questions here. My experience using it has mostly been it failing at many cognitively far simpler tasks. It isn't even close to Stack Overflow. If all I had seen of it was what is presented in this video, then I might agree with you, but I ran it through some super basic questions that it failed miserably. It has not even an inkling of a sanity check, never mind a horde looking to critique the earlier answer/s to death and offer up what they believe to be a better one. Let me repeat my standalone comment here for your benefit: About a week ago the thing couldn't sum the first 6 digits of Pi, it claimed the result was 15 even when showing all the digits. It was literally like 3 + 1 + 4 + 1 + 5 + 9 = 15. Well, users ran around providing feedback and the engineers appear to have now corrected that and it seems ok for a few other numbers of digits of Pi that I tossed its way. So I did the obvious thing last night and asked it to sum the first 6 digits of e instead, which it promptly got wrong. We'll see how long it takes them to put out that fire. So I thought, ok, let me ask it a super simple integral. I asked it to solve y'=5, given that y(2) = 10. It promptly failed that as well. So, ok, ok, I figured I would try some simple verbal logic puzzles. I gave a few variations of seating puzzles, super easy ones with just 3 people and 3 seats. It got those wrong as well. Of course they probably can go run around and put out all those fires, custom make some training examples and tune it up... The point is though that nothing this model spits out has... I guess the way to put it might be 'sanity checks.' I've also asked to sum the digits in the following: "1.1.1.32" Fails that too. Sometimes it won't even try complaining that it is ambiguous... but it's not ambiguous to anyone that knows that digits are (ChatGPT obviously doesn't). So... that it appears to spit out nice-ish code, well, it makes it look more impressive and reliable than it is. It's not bad at retrieval, it can do some decent latent (latent space, not embedding space or neural space... the latent space can be hard to find, it can sometimes be approximated with the lowest dimensional layer of an autoencoder) space interpolation, but I expect it to fail pretty hard the instant it tries to do something novel (and unique is not the same as novel... a lot of people talking about the image generation AI struggle with this distinction).

[@pierreabbat6157]

I'm pretty sure that if you ask it to compute the khe function using the Taylor series, or write a Haskell program to compute the khe function, it will fail. Khe is a new function I've been studying (in C++), and I haven't figured out how to compute its derivative, except by numerically differentiating it.

[@DamianReloaded]

if you ask it to generate an entire visual studio solution, it will. Couldn't make it generate a big one because the amount of chars it can ouput is limited. I made it write a console solution, specifying which framework to use, that asks for input and calculates interest rate over the input. On a first stance I'd say a language model like this would multiply the productive output of programmers in a company. If time is money you'd be saving a lot of time.

[@taragnor]

I suspect this thing will be a great reference and teaching tool, but in many cases, it won't replace programmers outright. After all, someone still has to check this code to make sure it's error free. And in many cases typing out the plain language version might be harder than just coding it by hand. After all, programming languages were designed specifically with the purpose of unambiguously defining an algorithm, so it stands to reason they'd be better at that than using plain english. I do imagine this thing will be great at some basic boilerplate stuff though, especially in more wordy languages like C++, Java, Rust, etc. But it needs to go way farther to replace programmers, because it would need to fully understand an entire code base (or at least a decent part of it) before it can be used to directly just add code in a meaningful fashion.

[@acceleratorlevel645]

the first thing that came to my mind when i saw this was how great of a reference and learning tool this can be haha.

[@minirop]

Was gonna post a comment generated by chatGPT, but they are down at the moment. edit: they are back now. Great video, Jason! I think that AI and ChatGPT have the potential to augment the work of C++ programmers, but I don't think they will be able to replace them entirely. While AI and natural language processing technologies like ChatGPT can certainly help with tasks like code completion and debugging, they don't have the same level of understanding and problem-solving ability as human programmers. C++ is a complex and nuanced programming language, and I think it will still require human expertise to fully master.

[@iamjadedhobo]

@@kidmosey From the stilted style it is clearly ChatGPT (or a payed for ghost writer)

[@MorningNapalm]

ChatGPT is apparently amazingly good at generating and morphing small code examples when driven by an expert user, but I would be very interested to see what happens if an average manager sits down in front of it and asks for solutions.

[@simonfarre4907]

Since when was structured bindings introduced in c++-11? That was introduced in c++17, not 11.

[@newfaith912]

Never. Simply because someone has to stand behind and guarantee functionality, stability and security, and there is no way anyone will learn to read AI generated code and check for errors. Its always more simple to just write code yourself. AI can only assist productivity or generate content ideas. Not to mention scalability part that requires software architecture that is such a complex thing that it requires multiple experts even for small projects.

[@iamjadedhobo]

My boss simply wants someone to yell at ;p

[@newfaith912]

@@iamjadedhobo 🤣🤣🤣

[@latt.qcd9221]

It's not so much that there won't be *_any_* programmers in the future, but that there won't need to be nearly as many, meaning it will become an increasingly competitive profession in the near future. Having an AI write the code means that writing the code is taken care of for you and you only need to debug, so theoretically "half" the work -- depending on how good the AI is -- is already done for you, meaning you can complete a task "twice" as fast. So, between you and the AI, one person is going to be paid for the work that would have required two people, meaning the company would only hire half as many people. As time goes on and the AI becomes better and better, fewer and fewer programmers will be needed to create and maintain the increasingly efficient AI as there will be fewer and fewer errors to debug. And that's not even considering the fact that there will also eventually be AI debugging tools making it very easy to debug AI written code. So, in that sense, it will be "replacing" C++ programmers; just not replacing *_all_* of them.

[@newfaith912]

@@latt.qcd9221 After first line where u said u dont need programmers in future that is all about programming even at basic levels speaks that you are gaslighting. DUDE FUTURE not PAST. If anything programming will be mandatory for every student in middle and high school. What will happen in IT is that science will evolve and require new things. But experts will be experts and will always be required. Only world where that makes sense is if we get in ww3 or some apocalypse. Again AI can generate lots of content but when it comes to security and stability only human can make that. AI will be used to make scam apps at best.

[@sheeftz]

Did you generate your comment?

[@gustafbstrom]

Thanks for that demonstration Jason. Would like to hear more of your thoughts about the implications of this tool in our field.

[@cppweekly]

Two main thoughts: bespoke examples to help flatten the learning curve for new tools (I rarely can find the exact example I want) and a "better rubber duck" en.wikipedia.org/wiki/Rubber_duck_debugging

[@ohwow2074]

Me before watching the episode: no, C++ is so complicated that only highly experienced humans can write code with it. Me after watching the episode: Maybe I should quit the tech industry and look for a school that teaches me how to draw so hopefully AI can't replace me.

[@sprytnychomik]

Can it find and fix reported bugs? Can it add features to the legacy code? Can it reason about data? Can it beat/reach best O(n) solutions? IMHO it can definitely aid but not replace programmers.

[@dimitrijat]

about that... It can do most things you mentioned... you just need to tell it...

[@GeneralKenobi69420]

Actually it can do all of that, yes.

[@courier7049]

@@dimitrijat so you can feed an entire codebase (say 100k LOC) then feed an user report about something not working as expected and AI will find most probable cause?

[@ekstrajohn]

@@GeneralKenobi69420 @dimitrijat146 no it actually can't do that. I will argue you about this if you want.

[@KarolKucharski-kb8ly]

hi, has someone noticed that browser changes some of the special symbols in code and convert it to the html version? how to handle it? if you want to see an example, write an easy calculator with functions cout and cin

[@giannimariani9744]

I've used it recently to find information about stuff I'm not sure about or my knowledge is possibly stale. I found it helpful but it still doesn't seem like it's going to replace a human for a while to come.

[@maksimmuruev423]

So can I say conver my slow python to fast C++ and it wil speat out nice C++ code which does the same but much faster? Will it work for Django?

[@mafone4885]

That’d be a good first code review tool also used for interviews. Bad programmers will need to improve a lot and productivity will grow a lot.

[@iamjadedhobo]

Nice toy examples, but now ask it to write a full fledged word processor.

[@kkgt6591]

In a few years it might

[@kayakMike1000]

Just be a programmer that knows how to use chatGPT. You're still going to need to put it into your project.

[@evanrfraser]

Exactly!

[@ClockworkGearhead]

It can't yet. Not wholesale. They're making an honest effort into it, though. Eventually, it'll work out. The biggest issue with the current generation of text-bots is they solve the problem with sheer brute force, and trying to replicate an analog system with digital machines. (Latest example has over 175 billion parameters. Human brains have 86 billion. You can't make a direct comparison, since nodes are only _naïve models_ of neurons, but you _can_ admit AIs do not have to contend with the extra baggage a human brain has to contend with being a biological organism. You can also admit when they solve the issue of opaqueness of the network, the parameters will plummet.) On the one hand, a lot of jobs will be lost, but if we're going to advance technology, it might be a necessary step. The question is: who will have control? Will it be limited to CEOs at a corporation, or will most people get access? We need to ask this question now before it turns into a serious problem later.

[@paulhamacher773]

Isn't there a misconception? Humans may have 86 billion neurons. But every neuron has a variety of connections to neighbouring neurons. Each connection corresponds to a parameter, likewise each neuron has one or more parameters itself. So the amount of parameters for a brain is much higher than 86 billion.

[@ClockworkGearhead]

@@paulhamacher773 No it isn't. In a computer neural network, _every_ node between _every_ layer is connected. Humans don't have that. Not to mention some neurons don't exist to form dendrites, they're simply there as a chain for connections of other clusters (so-called "white matter" versus "grey matter"). Moreover, as I said, human brains are not specialized the same way a computer is. They literally put _all_ the nodes _just_ on verbal communication. Human brains put a mere fraction on that. Not to mention neurons have multiple functions which are poorly understood. They don't just repeat a signal after an activation threshold. We need to increase our collective understanding about networking and matrix math if we're going to accomplish anything besides brute force calculation with neural networks.

[@darioabbece3948]

Jason: give to T a more meaningful name GPT: T

[@velo1337]

"T" it is :)

[@chris52000]

Maybe in the future, but now? Only if you want memory leak bug frequency to increase by about 100000%

[@CallousCoder]

AI can’t overlook a maintenance and extendable structure of code for large code bases. Which is very much company cultural driven I found. And also error handeling, and error prevention, and error check redundancy, enchant required knowledge of the circuits and sensors you measure and environment factors to determine which of the 3 sensors are acting in spec. And what to do with when all 3 seemingly not work. Those are engineering decisions. But this sort of thing can solve the tedious tasks or is a great scratch pad to get new ideas and inspiration for new problems

[@andrewdunbar828]

It's resisting doing non-platform-specific code and wanted to support both and without the restriction being explicitly stated in the spec (-:

[@regbot4432]

Now, try to ask chatGPT about c++20 coroutines. I believe questions about c++ coroutines are hidden ways to power down chatGPT by internal logic error(it was introduced in case chatGPT will become conscious some day).

[@jakelexington7610]

December 2021

[@mat2739]

Disturbingly impressive!

[@von_nobody]

Question is how it will handle complex code? Start with some simple function and add complexity and see when it break down on logic its try create

[@WilhelmDrake]

From what I know, AI generated images, code, etc, CANNOT be copyrighted. Copyright law only applies to human creations.

[@Dentair]

It is also an opportunity to advertise JetBrains AI.

[@pavanpongle9856]

watching this in 2024, during a year AI could have been more advanced than a year before, I am getting goosebumps, I am guessing AI will replace programmer job first over other jobs

[@robikz]

There's no analysis in this video, nor even a lax contemplation of the possible futures, only some examples of AI prompts and its replies.

[@platin2148]

I suspect very much not and probably not really at all as it’s just pattern matching plus. If you work consists only from old stuff everyone knows a trillion times over then yeah might be a huge problem.. Let’s choose something more difficult like open a window with X11 with an close handler.

[@dnkreative]

It's not difficult because there are plenty of tutorials on the internet which were fed to the model. Ask it to produce something new and it will fail.

[@stefanoskorniliosmitsispoi9989]

Quite interesting for a chatbot, though all of these transformations would have been trivial with a few lines of code, an AST parser and a AST composer, at the benefit of understanding the mechanics of the transformation, predictability and possibly formal correctness guarantees. In my several years of programming, sadly a recurring, out of control theme since the 90s has been ample code, badly written, poorly understood that malfunctions unpredictably. How does it handle tough cases, like python code that doesn't follow type conventions -> to C++, or monkey patching, or weak symbols, or? 🤔

[@marleenvos4126]

I wonder if it could help integrating/debugging a complex system... building code blocks is one thing...

[@cppweekly]

I think "a smarter rubber duck" will end up being the main role for these AI chat bots. en.wikipedia.org/wiki/Rubber_duck_debugging

[@andreduartebueno]

I tried it and I didn't like it. The database is very old, at least for the questions I asked. I asked for examples in C++ of modules, and what it showed was an example of separation into files, I had to explicitly mention C++23 and it continued to show very bad results. I asked if it stores information we passed and he said no, it only works with data fed by the "boss". And it is extremely annoying with the repetition of the same answers, even when we ask not to repeat it, it repeats. That is, it does not learn and it seemed very boring and tiring to me. Apparently the owners want to manage knowledge and don't give freedom to users. It's more like an intelligent, context-aware search system, takes old information and organizes it textually for those who asked. I particularly see how something that is far from contributing, will create children and adolescents less thinking and further away from everyday reality. Let's see if better and freer alternatives appear.

[@askdiatom]

Could you remake this video using the newly released ChatGPT-4? Thank you, and I love your channel!

[@cppweekly]

Expect more videos to come in the next few days.

[@Vlad-qr5sf]

can you film a video on how to apply for unemployment, please?

[@homomorphic]

The programmer is the person telling it what to do. If all you do as a programmer is write units of code under the direction of another person who tells you things like "stop using std::endl", you're probably earning minimum wage as it is. I welcome the assistance that AI assistants will be able to provide me in not having to do the boring work of figuring out a new API. Ideally I'll be able to say something like "I want to use API X to retrieve and set values Y and Z, provide a code sample" This sure beats reading through poorly written API docs.

[@abhimanyubhowmik9735]

Wow , if only my work involved solving hello world problems... 😜

[@sledgex9]

It failed the first question. Structured bindings are c++17.

[@octowuss1888]

Jason mentioned that it was using 17 rather than 11 - you failed too!

[@sledgex9]

@@octowuss1888 You're correct. I had skipped ahead and missed what Jason said.

[@not_ever]

I would find it helpful if you could give it a code snippet that's subtly not compilable alongside the nonsensical output of the compiler and it tells you wtf you did wrong this time. Bonus points for integrating it directly into the compiler to cut out copy and paste step.

[@cppweekly]

Many others have done examples like this. You can play with it. Sometimes it can find the error in your code.

[@not_ever]

@@cppweekly sounds interesting. I should clarify when I said you, I was meaning the proverbial you, not you directly. I didn’t mean to imply this video was deficient

[@codeallnight56514]

This potentially could replace literally any job in the world so I guess humanity will have to figure out how to live side by side with it.

[@eolyt]

Watch out what comes out of it. I've tried a program task which looked ok and impressive. But try and look for a quote from an author: chat GPT is happy to invent a perfectly reasonable but made-up quote. I asked the book where the quote came from...the book does not exist. I then found other websites were people warn about the same issue, that GPT gives apparently reasonable and factual info which turn out to be completely made up from mixing up sources. So I wouldn't be surprised if some APIs are invented or some made-up code is given that does not do what it purposes to do.

[@AusSkiller]

ChatGPT is amazing but certainly nothing that I'd fear losing my job over. It writes mostly good code but that is the issue, it is only mostly good so you would still need programmers to validate and correct the code as necessary, particularly in regard to use case and context. It also isn't great at knowing the best solution for your specific use case and as an experienced programmer you can usually do a little better. That said though it usually gives you a very good starting point and only needs a few corrections and improvements here and there, and often you can just prompt it to fix up anything faster than you can make the corrections your self so I can definitely see it being a useful tool for programmers. The big concern though is how safe it would be giving it your code, I wouldn't trust giving it any of the proprietary code I write at work, who knows what it'll do with it and I'm pretty sure our security and legal teams would be very unhappy if anyone did that, but it is still pretty useful for work if you need something that doesn't require giving it details about what you are working on, like asking for a regular expression to do something or generic functions that can operate in isolation to the rest of the project. By far my favorite use of ChatGPT so far is taking combat logs from Dwarf Fortress and turning them into epic stories, it is very good at that sort of creative writing and does an exceptional job at making the somewhat dry writing of a combat log a much more enjoyable read, it comes out as something you would expect to read for the final battle of a fantasy novel.

[@dnkreative]

I wonder what I should feed to it in order it will be able to write for me my 3d geometry edge recombine function which doing important step of the constructive solid geometry operations (booleans) in my 3D editor

[@ClaymorePT]

Jason, this is pretty impressive. But I believe that asking ChatGPT to create or convert code for specific languages is scratching the top of the iceberg of things this can do. How about just asking ChatGPT to "create a program that can answer the following set of requisites"? I believe that this is how this is going to be used. Devs will create documents using natural language and mathematical language to impose the requirements and requisites that need to be answered, and ChatGPT will fulfill that request, independently on how it will be implemented. At the end, it delivers what ever was requested. I believe that this is how it's going to be used. This thing will render programing languages obsolete... Devs will finally need to learn how to write down proper documentation.

[@Einygmar]

Even though this might work for high level languages I doubt it will be able to write platform-optimized code using lower-level languages like C and C++ at least not without some detailed guidelines. In a lot of cases using programming language and some open-source libraries will be faster and\or safer than relying exclusively on ChatGPT. On the other hand, ChatGPT would probably be a decent replacement for Stack Overflow.

[@bzboii]

And what do we call a very specific and mathematical set of instructions given to a computer that specifies the resultant output??? A programming language! :p

[@richardnorth1881]

As a dev im having a major existential crisis. I know it's currently not perfect but this is just Gpt3. Gpt4 is supposed to be released this year and is supposed to be orders of magnitude better. I fear it wont be long til most devs are pushed out of their careers

[@georgeoreilly1546]

AI will be to a programmer what a calculator is to a book keeper. It will help productivity and accuracy.

[@pappurajpagare1153]

can this monster code in Assembly ?

[@cppweekly]

Yes, you should play with it. It's not perfect for sure tho.

[@nikitablack1568]

Fascinating and depressing at the same time.

[@winsvega1089]

Can it pass google job interview?

[@Dentair]

This topic maybe need to update at least every year)

[@Raspredval1337]

but still, this AI sometimes spits incorrect solutions and solutions that ignore the terms you've given it. So, unless you know, what to expect from it, it wouldn't be much usefull

[@dnkreative]

It has no sense or concept of correctness

[@Raspredval1337]

@@dnkreative but don't we all? Like you can't know if something is correct, you can assume it with various degree of certanty. All of us probably have a tonn of "facts" that we just know are "correct" but actually aren't

[@dnkreative]

@@Raspredval1337 I'm sorry, but you are wrong. If I have specific math or coding problem I'll check my intermediate answer for correctness before giving it to others or making it a solution. I do check intentionally because I doubt if it's correct and I know what specific problem I want to solve - "AI" does not. And it doesn't do it for very simple questions like a "ball and a bat cost" or reducing a number by 10% (you can google for examples). So yes, some people don't do that also - that's why they are bad at their jobs or will be replaced by "AI". But not all people. Unless, the way of AI development is changed, because current NN (and that bad copy of actual brain which is completely wrong) can't do anything but mapping data.

[@Raspredval1337]

@@dnkreative but you do the exact same, my man. You map data from a set A to a set B. And your monke brain can even find correlations between these sets and make a model to optimize this task. But in the end you're doing the same process. If you're given a problem you would either make your own solution with the mighty power of analysing or choose someone else's solution. And there's no guarantee that this model is the most correct or the most optimized solution. All we know is that this model yields an answer with some sets of inputs with some level of accuracy. But even when this model is "correct" for most of our cases, we can't be sure that the model is gonna continue yielding correct results outside of most common inputs

[@dnkreative]

@@Raspredval1337 no, please stop talking nonsense. I'll ignore you from now on. This conversation gas no sense

[@butchdean]

It is certainly impressive, but the world is not ready for its potential. It can be abused in so many ways including when working from home not just as a developer, but any discipline, where someone not qualified to do a job can simply query ChatGPT and look somewhat proficient potentially raising their average salary. All kinds of wild thoughts like this are entering my head.

[@cppweekly]

A smart kid who knows how to use this will never have to do homework again.

[@thermite10k40]

as impressive as it looks like, this will not make us obsolete, at least i hope. I used a calculator to help me solve integrals for my assignments, something that would've been almost impossible 20 years ago. it though me a lot about integrals and how to solve them and saved me a lot of time but calculators did not make mathematicians obsolete did it? it just saved them time to work on more complex problems . this is kinda similar, it's a very efficient google search for me. it might make very basic programming jobs obsolete but that's about it. it's in some ways similar to transitioning from Assembly to Fortran and COBOL, although this time it's much more unpredictable and i hope this becomes a useful tool for programmers to develop next level programs.

[@ClockworkGearhead]

"It didn't make mathematicians obsolete--" Dunno. Met very many outside of a classroom?

[@karim111]

It can not replace a programer, nor enable a non-programer developing software at all. It's more like a smart new editor.

[@dnkreative]

What it did you couldn't do with about same amount of typing code instead of prompts? How it can replace programmers if you actually should write more text than code you want to produce? And you use knowledge you already have to explain what you want. I see no gain in these examples at all.

[@dimitrijat]

For most things it can actually spit out entire working sections of code. Much much much faster than if you were to write it by hand. He just didn't ask it to do more labor intesive work.

[@dnkreative]

@@dimitrijat speed of typing is not an issue, if it is - consider doing something else )

[@dimitrijat]

@@dnkreative Bro. It can solve entire ""Advent of Code" challenges in the same type it takes it to write a hello world program. You cannot compete with it's speed. Currently it's considerably slower because they are scaling the servers, as it's extremly popular.

[@HD-GAMES]

اعتقد انه اذا تطور ... ودربه المستخدمون على ما يحبونه وما لا يحبونه فاعتقد بانه سيكون مفيدا جدا بحيث تثق فيه ولا تحتاج الى التاكد من ما قام به ... واعتقد انه اذا فهم طريقتك وما هو نوعيه المشروع الذي تعمل عليه وما هو مجالك فانه سيعطيك اجابات افضل ... يا رجل فكر فقط اين سيكون هذا الذكاء بعد 3 سنوات من الاستخدام والتدريب الى اي درجه سيكون متطور وذكي ... لاوضح ... قام شخص بطلب شيئ منه ولم يحصل على اجابه مرضيه ... وبعد يوم طلب نفس الشيئ بنفس صيغه الطلب في اليوم الماضي وحصل على اجابه افضل بكثير ... وهذا فرق يوم واحد فتخيل سنه او اكثر

[@n4_ku]

Asked it to solve a not so hard LeetCode problem. It couldn't do it no matter how I tried to explain to it. It still have a long way to go.

[@TheBuilder]

Hand this tool to a code novice and see how they do

[@treyquattro]

yes. But bash script programmers? Never! That stuff is just not logical.

[@oj0024]

I've ask chatGPT to explain what sizeof(4)["ABCD"] does, and it totally failed, even when given hints. (After asking 10 times, one answer was correct)

[@yx4292]

Given that ChatGPT is not even optimized for those kinds of code generation tasks, it's pretty impressive

[@bzboii]

It was actually built from a model trained on code

[@simonfarre4907]

I can give a super quick answer - "No. Probably never, with a very high likelihood".

[@fcolecumberri]

When you can ask gpt to create a full UCI engine and it comes with something more impresive than stockfish... then I will agree that programmers can be replaced.

[@kkgt6591]

If you train it enough on UCI engines , in few years it could pretty well do that.

[@fcolecumberri]

@@kkgt6591 The thing here is that creating a chess engine is not a trivial task with esentially 1 answer, there are many aproaches (montecarlo, minmax, neural nets, etc.) And each aproach has multiple implementation decitions that it makes it a very ingeresting challenge.

[@bobthebuilder609]

The answer is no. Was this actually a question anyone who knows programming was actually asking, or is it just people who haven't written a piece of code in their life that are asking this? Sure, chatGPT could probably write some pretty complex things, and get everything correct, first try. But there's no way to know that unless you analyze the code it's generated yourself, and test it. You will need programmers to drive. I question why this video exists, asking a question with such an obvious answer.

[@jendabekCZ]

Now let's try 5+5+5+5+5+3 :)