it was then maintained by at least 20 different people that had no idea what they were doing

very little of the altered logic makes any sense and 95% of the time consists of trying to reverse engineer what they were trying to do

then fix it up without breaking other parts of the codebase as many of the logic bugs interact with each other and end up cancelling out

with the garbage generated by LLMs... I suspect all development is going to turn into this

Sounds awful. I was hoping for the future where we would all be out of a job.

we had two new starters and one long term employee quit after being asked to work on that codebase

(and ended up having to do it myself at extremely short notice... worked though)

on edit: https://www.zippia.com/cobol-programmer-jobs/salary/#

It also used to pay better, although at this point I'm pretty sure operating COBOL is just a sign that you're unwilling or unable to invest in upgrading your <1990s IT. The remaining companies operating COBOL systems likely aren't even able to pay top dollar.

on edit: amusingly, there is a relevant front page discussion right now https://news.ycombinator.com/item?id=36055986

Not really. We're comparing a programming language that has been legacy since the 1990s and has been gradually phased out over a period of 30 years and a process that accelerates the creation of legacy code in any language.

COBOL was always going to die eventually. I think, if anything, the fact you can still net over 100k writing it in 2023 kinda proves my point.

the logic is simply broken

AI lay-person, web developer here.

This hits at something that's been nagging at me for a few weeks. LLMs get labelled "intelligent", but they're "just" spitting out word/symbol patterns. They don't comprehend the meaning of the patterns (as best I have come to understand it from reading news).

Deeper than this, "intelligence" isn't rooted in words. It's rooted in experience, logic and their interaction. The words are a side-effect; they're a way to label the experiences.

We're not training LLMs with experiences, only a reference to experience had by others, which may or may not map to the experience of the interlocutor. The LLM has no experience and limited/no logic to wrestle those experiences into wisdom.

Debugging is a "pursuit of wisdom" activity. Statistics, which underpin LLMs often identify truths (we wouldn't use them in science if they didn't), but they also hide truths. You need to clean data, take unusual slices of it, recombine it, and otherwise process it in order to uncover the true truths. I'm not sure LLMs can do that yet (if at all), and strongly doubt they can without _experience_ (sensory input, intention, failure etc).

It is not obvious to me that this is true.

The argument against this is that a model has to develop an abstract understanding of the world in order to get better at predicting the next token.

You'll find different opinions as to wether this happens with todays llm's, but it's definitely not a clear cut issue.

I think they only develop abstract understanding about the relationships of the words/group of words.

I think the word you are looking is knowledgeable. They have huge knowledge repository and can remix it in whatever way they please, but I don't think if you gave them a novel problem (outside their realm of knowledge) they could solve it.

As another example, ImageBlind, https://ai.facebook.com/blog/imagebind-six-modalities-bindin..., learned modality translations explicitly not in its training set such as text -> sound.

If a technical product manager was able to generate the code needed for a feature using some combination of LLMs and CoPilots—even for proof of concepts, for example—the business would need to hire fewer developers and saves significant operating costs.

Extending it to your everyday person. Aunt Jane bought a new washing machine and wants to add it into her smart house. Either she's gotta call the niece to hook it all up, write the integration to Home Assistant, write automations for her iOS devices, etc, or, she asks her favorite LLM / Co-Pilot for help and it step-by-step walks her through it directly, with code, with individual troubleshooting.

There's business in both use cases, but the total addressable market is quite a bit bigger for one of them.

Your average person's ideal amount of exposure to code is zero (as in, most would prefer to not even think about code). If the device isn't automated from the getgo in an easy way, then Aunt Jane has no smart device and no matter how easy it is to write the connection, she's not going to have it unless her niece shows up.

You're severely overestimating the desire for an everyday person to automate their home. Most people are perfectly fine manually flicking light switches and unless you live on a very specific schedule, often automation ends up falling short of practical reality, so even the majority of people who "automate" just end up moving the light switch to their phone.

I know a cartographer who makes maps in some specialized software and then also has to occasionally interact with some official federally managed database. This person is not a software engineer and the database interaction has been historically painful in the office. This year ChatGPT just helped whip up some Python scripts.

Two hypotheses: 1) the “in house tooling” market is huge and will make great use of this tech; 2) that will contribute to pushing the floor lower and lower where more and more roles and people will take advantage of it.

This is exactly where it completely breaks down. The AI might as well show her some hieroglyphics.

Yes, today, developers are people who know programming languages, that's something non-developers don't know, so teach an AI do code and you can spare a developer, right?

But no, what developers really do is tell the computer precisely what it should do, a programming language simply the best way we have found to do it. Using GUIs, natural languages, etc... may fell less obscure to the layman, but the problem doesn't go away, in fact, it is often made worse because of the fuzziness of natural languages. There is a reason developer have been using programming languages for so long despite countless attempts to find alternatives, they simply haven't been matched for the task of telling a computer what to do precisely. And if something better is found, then developers will use that, and besides some syntactic technicalities, the job will be essentially the same.

    bool
    files_read(const char *path, char **data, size_t *size) {
        FILE *f = fopen(path, "rb");
        if (!f) {
            return false;
        }
        fseek(f, 0, SEEK_END);
        size_t n = (size_t)ftell(f);
        rewind(f);
        if (size) {
            *size = n;
        }
        if (!data) {
            goto ok;
        }
        *data = (char*)malloc(sizeof(char)*(n + 1));
        if (fread(*data, 1, n, f) != n) {
            free(*data);
            return false;
        }
        *data[n] = '\0';
     ok:
        fclose(f);
        return true;
    }

> The issue in this code is related to memory access. The line data[n] = '\0'; should be (data)[n] = '\0';.

And then went on to generate corrected code and this explanation:

> The original expression, data[n], first applies the array subscript operation, which is equivalent to (data + n), and then tries to dereference the resulting pointer, which is not the intention here. This will likely lead to a segmentation fault because data is a pointer to pointer, and data[n] will try to access memory that has not been allocated.

> The corrected expression, (*data)[n], first dereferences data to get the allocated char array, and then applies the array subscript operation to set the null terminator at the correct location.

I'm not trying to say this isn't impressive, but seems fairly obvious and more like something a linter / analyzer would pickup ?

Actually, no, let's talk about the bug. What's with people's obsession of eliminating local temporaries and working straight with pointer-to-result?

        char *buff = malloc(n + 1);
        if (fread(buff, 1, n, f) != n) {
            free(buff);
            fclose(f);
            return false;
        }
        buff[n] = '\0';
        *data = buff;
      ok:
        fclose(f);
        return true;

It kinda works tho. In my anecdote Copilot works much better with C# than Python simply because I can write the signature of a function and that it generates the content.

(I know Python has type annotations too, but Copilot just isn't as smart as with C#. Perhaps because there isn't enough training data in typed Python?)

But yes, what you described matches my experience. I'd say Copilot benefits those "tedious but relatively strict" languages the most.

But it obviously doesn't have knowledge of your domain, and I don't think any LLM could do a perfect job without that.

It happened so fast that so many people are in denial and many people aren't even asking the obvious questions. Simply ASK chatGPT to refactor your code to be more unit testable and IT Can do it.

Of course it clearly has your typical LLM problems but it is roughly 50% there. The game has changed. If AI gets to 100%, not only will the game be changed, but humans will no longer be part of the game.

See what I asked chatGPT:

   ME:
   def (x: int):
      for i in range(x):
         print(i)

   Can you refactor the code so it can be more unit testable?


   CHATGPT:
   Certainly! To make the code more unit testable, we can refactor it into a function that returns a list of integers instead of directly printing them. Here's the refactored code:

   python

   def generate_number_list(x: int) -> List[int]:
       number_list = []
       for i in range(x):
           number_list.append(i)
       return number_list

   In this refactored code, we've created a function called generate_number_list that takes an integer x as input. Instead of printing each number, we use a number_list variable to store the generated numbers by appending them in the loop. Finally, we return the number_list.

   By returning the list instead of printing the numbers directly, we make the code more testable because we can now write unit tests that assert the correctness of the generated list.

IME GPT-4 can't write a bug-free 10 line shell script. It's particularly poor at inferring unstated requirements - or the need to elicit the same.

There's a general problem with LLMs: they're too eager to please. It shows up as confirmation bias. Embed a perspective in your prompt, and LLMs continue in the same vein.

You can, with careful prompting, try to provoke and prod the text generation into a more correct shape, but often it feels to me more like a game than productivity. I have to know the answer already to know how to ask the right questions and make the right corrections. So it feels like I'm supervising a child, and that I should be amazed it can do anything at all. And it is amazing; but for productivity outside tightly constrained environments (e.g. converting freeform dialogue into filling out a bureaucratic form - I think this is a close to ideal use case), I struggle to see it scaling up much, from what I've seen so far.

For creativity - e.g. making up a story for a child - it's not bad. One of my favourite use cases, after discovering how bad it is at writing code.

So that means a 10 line bash script at best would have 5 lines of bugs.

But the AI is actually better then that. You've just increased the bar.

The game is designing software to requirements. It's writing literature for a new era. It's creating X for A audience with N vauge unspecified needs -- where X is a complex product made of many parts, involving many people, with shifting and changing problems/solutions/requirements.

The game was never writing the stack overflow answer -- that was already written.

People say, "oh but clients will have to get good at listing specs, that'll never happen". Like bruh the clients will obviously be using LLMs to make the specs too. Eventually the whole B2B workflow will just be LLMs talking to each other or something of the sort.

The problem is this was never a stackoverflow question and there was never an answer for it.

Try finding it. The LLM is already playing the game because it came up with that answer which is Fully Correct, Out of Thin Air.

Look, clearly the LLM can't play the game as well as a trained adept human, but it's definitely playing the game.

>The game is designing software to requirements. It's writing literature for a new era. It's creating X for A audience with N vauge unspecified needs -- where X is a complex product made of many parts, involving many people, with shifting and changing problems/solutions/requirements.

It can do all of this. It can talk like you and parrot exactly what your saying and also go into more detail and re-frame your words more eloquently.

What you're not getting is that all the things you mentioned the LLM can do in actuality to varying degrees to the point where it is in the "game." and at times it does better than us. Likely, you haven't even tried asking it yet.

I think if you're an expert in an area, this effect is easier to see through. You know where the github repo is, where the library example is, which ebooks there area -- etc. and you're mostly at-ease not using them and just writing the solution yourself.

These systems are not "fully correct" and not "out of thin area". They are trained on everything ever digitised, including the entire internet. They, in effect, find similar historical cases to your query and merge them. In many cases, for specific enough queries, the text is verbatim from an original source.

This is less revolutionary than the spreadsheet; it's less than google search. It's a speed boost to what was always the most wrote element to what we do. Yes, that often took us the longest -- and so some might be afraid that's what labour is -- but it isnt.

We never "added value" to products via what may be automated. Value is always a matter of the desire of the buyer of the products of our labour (vs. the supply) -- and making those products for those buyers was always what they wanted.

This will be clear to everyone pretty quickly, as with all tech, it's "magic" on the first encounter -- until the limitations are exposed.

I actually work in an area where what took 3mo last year, I can now do in maybe 3 days due to ChatGPT. But when it comes to providing my customers with that content, the value was always in how I provided it and what it did for them.

I think this makes my skills more valuable, not less. Since the quality of products will be even more stratified by experts who can quickly assemble what the customer needs from non-experts who have to fight through AI dialogue to get something generic.

The user needs to keep in mind that it can give plenty of false information. To make good use of it, the user needs to be able to verify if the returned information is useful, makes sense, compare with first hand sources, etc. In the hands of expert that is really powerful. In the hands of a layman (on the subject in question), they can generate a lot of crap and misunderstand what it is saying. It is similar to the idea that Democracy can be a great tool, but it needs an educated and participatory populous or it may generate a lot of headaches.

Of course you shouldn't think of it as magic. But, the experts self admit they don't fully understand how LLMs can produce such output. It's definitely emergent behavior. We've built something we don't understand, and although it's not magic, it's one of the closest things to it that can exist. Think about it. What is the closest thing in reality to magic? Literally, building something we can't understand is it.

It's one thing to think of something as magic, it's another thing to try to simplify a highly complex concept into a box. When elon musk got his rockets to space why were people so floored by decades old technology that he simply made cheaper?

But when someone makes AI that can literally do almost anything you ask it to everyone just suddenly says it's a simple stochastic parrot that can't do much?

I think it's obvious. It's because a rocket can't replace your job or your identity. If part of your skillset and identity is "master programmer" and suddenly there's a machine that can do better than you, the easiest thing to stop that machine is to first deny reality.

Well I take myself to be an expert in this area, and I think it's fairly obvious how they work. Many of these so-called "Experts" are sitting on the boards of commercial companies with vested interests in presenting this technology as revolutionary. Indeed, much of what has been said recently in the media is little more than political and economic power plays disguised as philosophical musings.

A statistical AI system is a function `answer = f(question; weights)`. The `answer` obtains apparent "emergent" properties such as "suitability for basic reasoning tasks" when used by human operators.

But the function does not actually have those properties. It's a trick -- the weights are summaries of unimaginable number of similar cases, and the function is little more than "sample from those cases and merge".

Properties of the output of this function obtain trivially in the way that all statistical functions generate increasingly useful output: by having increasingly relevant weights.

If you model linear data with just y = ax then as soon as you shift to "y = ax + b" you'll see the "emergent property" that the output is now sensitive to a background bias, b.

Emergence is an ontological phenomenon concerning how `f` would be reaslised by a physical system. In this case any physical system implementing `f` shows no such emergence.

Rather the output of `f` has a "shift in utility" as the properties of the data its training on, as summarised by the weights, "shifts in utilty".

In other words, if you train a statistical system on everything ever written by billions of people over decades, then you will in fact see "domains of applicability" increases, just as much as when you shift from a y=ax model to a y=ax+b.

To make this as simple as I can: statistical AI is just a funnel. ChatGPT is a slightly better funnel, but moreso, it's had the ocean pass through it.

Much of its apparent properties are illusary, and much of the press around it puts in cases where it appears to work and claims "look it works!". This is pseudoscience -- if you want to test a hypothesis of ChatGPT, find all the cases where it doesnt work -- and you will find that in the cases where it does there was some "statistical shortcut" taken

And frankly, any interaction with even GPT 3.5 should demonstrate this. It's not hard to make the network produce output that was never in the training set at all, in any form. Even just the fact that its skills generalize across languages should already disprove this claim.

Honest request because I am a bit skeptical, can you give an example of something it is not trained in any form and can give output for? And can it output something meaningful?

Because I have run a few experiments on ChatGPT for two spoken languages with standard written forms but without much of a presence on the internet and it just makes stuff up.

And here we go! https://poe.com/s/UJxaAK9aVN8G7DLUko87 Note that it took me a long time, because GPT 3.5 really really wanted to misunderstand what I was saying; there is a strong bias to default to its training samples, especially if it's a common idea. But eventually, with only moderate pushing, its code did work.

What's interesting to me here is that after I threw the whole "step by step" shebang at it, it got code that was almost right. Surprisingly often, GPT will end up with code that's clever in methodology, but wrong in a very pedestrian way. IMO this means there has to be something wrong with the way we're training these networks.

edit: https://poe.com/s/gZW5ZGgiomWzabKJCUcA I gave it a more complete prompt because I only have one completion per day, but GPT-4 got it in one shot.

edit: https://poe.com/s/2lS8rjbGqHrzSkpEvLzr GPT 3.5 flubbed it given the same prompt.

https://www.thedailybeast.com/the-radical-movement-to-worshi...

I disagree, it's not.

But to keep it with the religious theme... which side sounds more similar to religion? The side that claims it's absolutely impossible for LLMs to be anything more then a statistical operation or the side that claims they don't know? One side seems to be making a claim based on faith while another side is saying we don't know enough to make a claim... So which side sounds more religious?

We can also say we understand chemistry but we don't understand how consciousness comes out of chemistry.

You can also say that humans are "just" physical processes, but that word "just" is doing a lot of heavy lifting.

A child touches a fireplace once -- not a thousand times. Because they are in direct causal contact with the world and their body has a whole-organism biochemical reaction to that stimulus which radically conditions their bodies in all sorts of ways

This is a world apart from statistical learning wherein P(A|A causes B) and P(A|B) are indistinguishable -- and the bridge of "big data" merely illusory

Bro if you are an expert you'd already know that most of the exclamations that they don't fully understand LLMs is coming from researchers at universities. Hinton was my example on an "expert" as well and he literally quit google just so he can say his piece. You know who Hinton is right? The person who repopularized backprop.

>A statistical AI system is a function `answer = f(question; weights)`. The `answer` obtains apparent "emergent" properties such as "suitability for basic reasoning tasks" when used by human operators.

Every layman gets its a multidimensional curve fitting process. The analogy your using here to apply properties of lower dimensional and lower degree equations to things that are millions of dimensions in size on a complex curve simply doesn't apply because nobody fully understands the macro details of the curve and how that maps to the output it's producing.

The properties of a 2d circle don't map one to one to 3d let alone 500000000d.

>Much of its apparent properties are illusary, and much of the press around it puts in cases where it appears to work and claims "look it works!". This is pseudoscience -- if you want to test a hypothesis of ChatGPT, find all the cases where it doesnt work -- and you will find that in the cases where it does there was some "statistical shortcut" taken

You don't even know what science is. Most of software engineering from design patterns to language choice to architecture is not science at all. There's no hypothesis testing or any of that. An expert (aka scientist) would be clear that ML is mostly mathematical theory with a huge dose of art layered on top.

The hypothesis for the AI in this case is, and I'm parroting the real experts here,: "we don't understand what's going on." That's the hypothesis. How is that even testable? It's not so none of this is "science". ML never was a science, it's an art with some theoretical origins.

But your "hypothesis" is it's just "statistical parroting" which is also untestable. But your claim is way more ludicrous because you made a claim and you can't prove it while I made a claim that basically says "we can't make any claims because we don't understand". See the difference?

Experts in the area, including Hinton, the father of modern AI, self admit they don't fully understand what's going on but they think that LLMs know what they are talking about.

>These systems are not "fully correct" and not "out of thin area". They are trained on everything ever digitised, including the entire internet. They, in effect, find similar historical cases to your query and merge them. In many cases, for specific enough queries, the text is verbatim from an original source.

I never said the systems are fully correct. I said that for my specific example the answer is fully correct and out of thin air. No such question and answer pair exists on the internet. Find it and prove me wrong.

>This will be clear to everyone pretty quickly, as with all tech, it's "magic" on the first encounter -- until the limitations are exposed.

Except many experts are saying the exact opposite of what you're saying. I'm just parroting the experts..

>I actually work in an area where what took 3mo last year, I can now do in maybe 3 days due to ChatGPT. But when it comes to providing my customers with that content, the value was always in how I provided it and what it did for them.

So if they knew you were just copying and pasting their queries to chatgpt would they still care about the "how"? I doubt it.

However, in its current state users still have to know how to program in order to make good use of it. It will still give you lots of errors, but being able to get something close to your goal can save you a lot of time. Someone who does not know how to program will not be able to use these to put together a complex, useful and reliable system. It might change in the future, but these things are hard to predict.

Don't worry about this. You can get over the fear. I'm in the top 10% of stackoverflow users in terms of points and it's all because my stupidest questions from decades back gathered thousands of points from other stupid idiots like me. Who cares. Literally the line graph keeps climbing with no effort from me all from my dumbest questions. Just ask and don't worry about the criticism, you'll get a bit, but not too much.

>However, in its current state users still have to know how to program in order to make good use of it. It will still give you lots of errors, but being able to get something close to your goal can save you a lot of time. Someone who does not know how to program will not be able to use these to put together a complex, useful and reliable system. It might change in the future, but these things are hard to predict.

Of course. I think the thing I was trying to point out is the breadth of what chatgpt can do. So if you ask it to do a really in depth and detailed task it's likely to do it with flaws. That's not the point I was trying to emphasize, not the fact that it can't do any task with great depth but the fact that it can do ANY task. It has huge breadth.

So to bring it line with the direction of this thread. People were thinking about making special LLMs that refactor code to be unit testable. I mean we don't have to make special LLMs to do that because you can already ask chatgpt to do it already. That's the point.

I've had a few flamed and closed but that's just 1 or 2 out of I'd say around 13 or 14 questions. It's a low percentage for me.

And I absolutely assure you much of my questions are stupid af.

Your example is so trivial, that there are definitely similar code examples. Maybe not word for word, but similar enough, that this is not really mindblowing "making things out of thin air" for me. It seems like a standard coding class example, so not surprising, that it also can make the unit tests.

Find one. Dated before 2021. In fact, according to the theory that it's statistical parroting there should be multiple examples of for loops printing out numbers being converted to unit testable functions because AI needs multiple examples of it to form the correct model.

Find one. And it doesn't have to be from stack overflow either. Just a question and answer data point.

It's not though. It doesn't print the values anymore, so the behavior isn't the same.

Refactoring isn't allowed to change behavior.

By definition all unit testable functions have to return data that can be asserted. You throw that data to IO it's not unit testable.

IO is testable via an integration tests. But not unit tests. Which is what my query exactly specified. I specified unit tests.

Replacing a function that does `print("hello world")` with a function that does `return "hello world"` isn't a valid way to make it unit testable.

instead it generates a load of redundant boilerplate

if I saw a developer check that in I'd think they were incompetent

I'm not good at prompting (if I believe what others say they can do with ChatGPT), but that's one thing that bother me with this system. They will do anything you ask them to without questioning it (in the limit given by their creators).

Is it possible to set it up in a way that they will challenge you instead of blindly doing what you ask? In this particular case, is it possible to ask it to do a code review in addition to performing the task?

I've tried various time (with the v3.5) to "tune" it so that each answer will follow a specific format, with links and recommended resources, with several alternatives, etc. The goal is to have it to broad my perspectives as opposed to focus too much on what I'm asking. But it never worked for more than a couple of questions.

Are there ways to do that? What am I doing wrong?

As it stands right now, chatGPT can actually challenge you.

I'm not sure where your side track is coming from. Who in their right mind would ever check in code that prints a range of numbers from 0 to x?

The example wasn't about writing good code or realistic code. It's about an LLM knowing and understanding what I asked it to do. It did this by literally creating a correct answer that doesn't exist. Sorry it doesn't satisfy your code quality standards but that's not part of the task is it? Why don't you ask it to make the code quality better? It can likely do it Maybe that will stop the subtle insults (please don't subtly imply I'm incompetent that's fucking rude)

Like why even get into code quality about some toy example? What's the objective? To fulfill some agenda against AI? I think that's literally a lot of what's going on in this thread.

it was your example

plus: toy examples are all it really does

It's my example sure, but it's your reaction to it that makes no sense.

plus: that's not true.

(and it is true I'm afraid)

It is pure pablum. It is almost the perfect example of LLMs producing fluent vapid bilge.

For loops in python that return non lazy evaluated lists are fine. Python was never suppose to be an efficient language anyways, grading python based off of this criteria is pointless.

It doesn't matter how snobbish you are on language syntax though. I fed it code and regardless of whether you think it's garbage it did what I asked it to do and nothing else.

Would you prefer the AI say, "this code is garbage, here's not only how to make it unit testable but how to improve your garbage code." Actually we can make the output more unpredictable as LLMs do have a non deterministic seed that can increase the creativity of the answer.

.

Please, take the code it has produced and integrate it into the original function. All it does is replace the range call. That's it. It has absolutely and totally failed at the given task whilst outputting plausible garbage about why it has succeeded.

By definition a unit test can only test functions that return data. So there's no other option here.

A mutable object is functionally identical to a return value if you control the initial state and lifetime of the object. Like you can do in a unit test.

And as I demonstrated in my other comment I 100% retained the semantic structure of the function whilst making it 100% unit testable.

How the hell do I test a print function? I take the print function and match it with what? It has no output so how can I test it printed the correct thing? I can't.

I can test a list. I just match it with another list. Making your code unit testable is about segregating IO from logic. Write pure logic where all functions have inputs and outputs and those things can be tested. Your io prints should be small because all functions that do io cannot be fully tested.

IO is pollution. Any output to IO is the program exiting the logical mathematical universe of the program and that output can be verified only by an external entity. Either your eyes for stdout or another process or files or a bunch of other ways.

Unit tests are about internal local tests that touch local functionality and logic. If you want something unit testable it needs a local output and an input and it shouldn't rely on io in it's data path.

I think your complaint here is an example of chatGPT superiority. It understood something you didn't. Well now you know.

Removing the print function from the logic and returning the data is 100 percent the correct move. Do you understand?

Of course you can make the function with a print statement more unit testable without completely changing it's semantics!

You pass in an outputstream and use that as the target for print.

Then your unit test can create its own stream and test the content of the stream whilst production code can pass in standard out.

That way you don't completely change the semantic meaning of the code.

And once again that GPT function is useless. It is identical to list(range()) and it doesn't do what the first function does. Anyone can make anything more unit testable if it doesn't have to do the same thing.

For clarity I reproduce the original function you gave and then I present what the change I am suggesting is

  def cool_function(x):
    for i in range(x):
      print(i)

  def cool_function(x, output_stream=sys.stdout):
    for i in range(x):
      print(i, file=output_stream)

But now we can write a unit test like

  def test_output():
    output = io.StringIO()  
    cool_function(1, output)   
    contents = output.getvalue()
    assert contents=="1\n"

No need to apologize. This is a discussion. No one did anything wrong.

>For clarity I reproduce the original function you gave and then I present what the change I am suggesting is

This is called dependency injection and it's a valid way of segregating IO away from pure logic. Although this pattern is popular among old school OOP programmers it's getting out of vogue due to the complexity of it all. You used a python trick here of default values, but typically dependency injection changes the function signature and ups the complexity of the code by a lot. Let me show you the full output of the code that chatgpt was implying:

   #unit testable code (without using dependency injection tricks)

   def cool_function(x: int) -> None:
       IO_function(logic_function(x))

   def logic_function(x: int) -> List[int]:  
       return [i for i in range(x)]

   def IO_function(x: Any) -> None:
       print(x)
       
   def test_output():
       assert logic_function(4) == [i for i in range(4)]

As a side note you will see it's actually an improvement to the code. It's simpler, no dependency injection, no confusing function type signature and a much simpler test case. The other thing that must be noted is the modularity.

Making tests unit testable in this way allows for your logic to be portable. What if I want to repurpose cool_function to output it's logic to another function? In your example you don't have the components to do that, it's harder for your case as you'd have to create another component for injection.

In short not only did chatGPT produce A correct answer. But it produced the better answer compared with your dependency injection. That being said your dependency injection is valid BUT you were not correct in saying that chatGPT's answer was worse or incorrect.

3 functions is better. Think about it. Do people write all their stuff in one big function? No. Better to compose higher level functions with smaller ones rather then write one big monolith like you did. The more modular something is the better.

Also IO_function is there for illustration purposes. Technically it's just wrapping print with a name so you can understand the intent. In reality you just use the regular print here without a wrapper, so in actuality only two functions are defined.

>The job of ChatGPT was to make cool_function unit testable. You haven't done it.

It did. By giving it a return value. Just like you did by giving it a new input value.

>You still have cool_function using side effect generating code hitting the actual IO system.

Yeah but one component of cool_function is pure and you can unit test that. Cool function itself can never be tested because it generates no output, you test the unit components of cool function. That's the point of unit tests.

>Genuinely the worst unit test I have ever seen written, on a poor form per line basis, absolute bananas. If you don't understand why [i for i in range(4)] is bad in a unit test and [0,1,2,3] is correct then I need you to walk away from the computer.

Let's just talk about it like adults. Just tell me what exactly about it makes you think it's bad?

Most likely it's some pedantic stylistic philosophy you have? I'm thinking you only want to test literals? Perhaps you prefer [0,1,2,3]? Am I right on the money?

Logic potentially has errors so you don't put logic in your test code. Makes sense, but who cares. For trivial shit it's fine. While in this case the logic in the test is identical to the function, typically 'logic_function' represents something significantly more complex and the list comprehension so I could care less if I'm not following the strictest form of testing. The comprehension is just something akin to an alias shortcut I prefer to use over writing out a massive literal. For the toy example the test is pointless because the logic is identical but typically it's fine to use range as an alias to represent a sequence of numbers.

Someone who strictly follows these stylistic rules without seeing intent or having the ability to bend the rules is just an inflexible pedantic programmer. It's not good to boast about it either by telling other people to walk away from a computer. That's just rude.

Output formatting touches io. In this case it is no longer a unit test that touches these things. Unit tests by definition test ONLY internal logic and transformations.

It is literally the definition of unit tests.

When you test things like stdout that becomes an integration test and Not a unit test. It requires some external thing or some global black magic monkey patch that changes what print does to do integration testing.

(Btw making print formatting unit testable means segregating the formatting from the print. Produce the string first, test that, then print, because print can never be unit tested by definition)

Typically programmers segregate these levels of testing because unit tests are easier to write. But to write unit tests your code has to be written in a way to cater to it. Often this style of coding actually improves your code it makes it much more modular. The reason is because pure functions that output data can be composed with all kinds of io functions. You can move it all over the place and to different platforms with different forms of IO. Print has no meaning in certain embedded systems so it can't be moved... By segregating the logic out it makes it so I can move the logic without the io baggage.

Chatgpt 100 percent gets the difference that's why it did what it did. I think you and the OP don't fully understand the meaning of unit testing.

Don't take this the wrong way, but just because you don't know this doesn't say anything about your skills as a programmer. But just recognize that this concept is basic and is pretty much something universal among testing.

Output formatting is still a type of transformation! The function explicitly takes the numbers and prints them as decimal integers with newlines between each. A test to confirm that it IS in that format is still a unit test.

BTW I gave ChatGPT the prompt I would give, and I have to say the answer looks pretty good, even if I'm not a Python programmer and it's not the way I'd do it (which would be to change the function to allow passing in an output stream):

    class MyFunctionTestCase(unittest.TestCase):
      def test_my_function(self):
        expected_output = "0\n1\n2\n"
        with patch('sys.stdout', new=StringIO()) as fake_out:
          my_function(3)
          self.assertEqual(fake_out.getvalue(), expected_output)

  def my_function(x: int) -> str:
    output = ""
    for i in range(x):
        output += str(i) + "\n"
    return output

I'll quote part of my reply (which you missed):

   (Btw making print formatting unit testable means segregating the formatting from the print. Produce the string first, test that, then print, because print can never be unit tested by definition)

>BTW I gave ChatGPT the prompt I would give, and I have to say the answer looks pretty good, even if I'm not a Python programmer and it's not the way I'd do it (which would be to change the function to allow passing in an output stream):

It's wrong in this case. Unless you specifically asked it to write unit testable code, what it did here is write a hack that monkey patches the program. It's a huge hack. It didn't write unit testable code, but rather it wrote a integration test that monkey patches stdout, negating any need to make your code "unit testable" no refactoring needed using this method. The entire concept of refactoring code to be unit testable flies out the door in this case as you're just using integration tests to get around everything.

I mean yeah you use the unit test library but is not technically a unit test. It's fine I'm not a stichler for what style of testing is used in practice but what I am saying is that what chatgpt did previously was literally follow my instructions to the letter. It did it exactly 100% correctly. Think about it. I asked chatgpt to make the Code more unit testable. You didn't have chatgpt do anything to the code. You simply changed the test from a unit test to integration test. Huge difference. I mean if your case was the "proper" way then what does it even mean to make code "unit testable" if you're not even touching the code? Like why does the concept of "making code more unit testable" even exist if we're just changing tests to make everything unit testable? Ask yourself this and you'll realize that the only answer is basically what I just told you previously.

   ME:
   def (x: int):
      for i in range(x):
         print(i)

I didn't say don't bother extending it's capabilities either. You're just projecting that from your imagination. An hallucination so to speak not so far off from what LLMs do. I find your similarity to LLMs quite fascinating.

What I said is, the capability of doing the "extension" you want is already in the LLM. Sure go extend it but what you're not getting is that we've already arrived at the destination.

What I would like to do is feed in my entire 50k line program and get something out.

Also writing code, in 2023? With your hands? Pshh, that's so 2010s.

When you ask for large output or give it large input you are increasing the sample size. Which means more likely that part of the answer are wrong. That's it. Simple statistic that are inline with my initial point. With AI we are roughly half way there at producing answers.

If you keep the answers and questions short you will have a much higher probability of being correct.

So that 50k line program? My claim is roughly 25k of those lines are usable. But that's a fuzzy claim because I LLMs can do much better than 25k. Maybe 75% is more realistic but I'll leave it at 50% so there's a lower bar for the nay sayers to attack.

    > g(x) = sin(x)/x ; what is g(exp(-200)) ?
    > ChatGPT
    > 
    > To find the value of g(x) = sin(x)/x at the point g(exp(-200)), we 
    > substitute x = exp(-200) into the function:
    > g(exp(-200)) = sin(exp(-200))/exp(-200)
    >
    > Now, let's calculate this value using numerical methods:
    >
    > sin(exp(-200)) ≈ 
    > 0.0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000

    > (there it breaks off, running out of tokens )

Taking Nair's viewpoint, coding AI will be useful only to the extent that it assists programmers to build and employ their theory of the program. Or to itself develop a useful theory of a specific program, not just programs in general.

[1] https://gist.github.com/onlurking/fc5c81d18cfce9ff81bc968a7f...

I asked it to write a YAML parser in pure Rust. The first thing it did was write this:

    pub enum YamlNode {
       Scalar(String),
       Sequence(Vec<YamlNode>),
       Map(HashMap<String, YamlNode>),
    }

Literally unusable as a starting point.

I did try to correct it but it tripped up and somehow managed to fuck that up to.

YAML allows YamlNodes as keys in a map (aka complex keys). E.g.

    ? [ a, b]
    : "complex key, scalar value"

Also this is the most banal thing about YAML, something that should be obvious just browsing the examples. And I gave it test suite and link to YAML spec.

Most work in YAML is around super obscure corner cases.

You also get the bonus of nasty bugs like crash bugs and exploitable memory problems become almost entirely a thing of the past. You can ship new software you know is very reliable.

> But, AI is particularly bad at formal methods.

AI? Or GPT? Earlier, reasoning-based attempts at AI seem like they would be a natural fit for formal methods.

The one thing worse than debugging code is debugging "types". Those 30-page C++ compiler errors are definitely worse than whatever I had to do with (for example) Python.

That's definitely in the "programming language sucks" category. Good languages make it easy to debug problems with your types.

I was responding to a comment that was suggesting letting an AI to do that.

> The one thing worse than debugging code is debugging "types".

It depends on a language. In languages with first-class types, we have the same tools to debug types. Classical C++ templates are more akin to a macro system than a type system.

C++ is sort of the canonical example of a language with a terrible type system, so it's not really the right place to look for this comparison.

Everyone complains how slowly Rust compiles. But at least its error messages are much better.

We can actually try this now. Literally tell the LLM what you want to do and work with it. See how far you can take it. You will of course be asking the LLM for debug line inputs and you will be providing it with outputs as you run the debugger yourself.

Did I learn a bit of java and css and git?- sure, but I was up and running in about 4 hours with a mvp for my 1st one. There is NO way I could "learn" that in that timeframe. I just asked chatGPT 4 how to do it, and it told me. When I didn't know how to commit, it told me (actually I didn't even know the concept). It held my hand every step of the way.

I didn't need to learn something first, I just did it. And I have started doing it at work. "hmm 4 GB of fortinet logs in 20 files of gzip on mac.. how do I find a host name in that? - chatgpt.. oh- 1 line of zgrep.. never heard of it- hey it works.."

admittedly, I am bathed in tech, been hanging around folks talking about projects for years. But NOW I can execute- the problem? When it hits about 500 lines of java- maybe 10 functions, it is too big to drop into the prompt to debug and I don't know enough to fix myself. Solution, make smaller apps, get them working, create data files to reference in json, chain them together. eh, not perfect, but good enough for hobby.

Beware- fools like me who know nothing will be bringing code to production near you soon. Cool that you like to learn stuff, but syntax bores the crud out of me, each to their own, I'm just going to make. I find it more satisfying. Terrifying that code born like mine will end up in someone's prod, but it will.

If AI helps you, you'll emphasize on the overall benefit rather then nitpick at the details because of the clear conflict of interest that LLMs present to programmers.

Before you go on and write such a system it's better to test if the LLM can do debugging to an efficacy level that we require. I don't think anyone has tried this yet and we do know LLMs have certain issues.

But make no mistake, the possibility that an LLM knows how to debug programs is actually quite high. If it can do this: https://www.engraved.blog/building-a-virtual-machine-inside/ it can likely debug a program, but I can't say definitively because I'm too lazy to try.

I do wonder though whether the methods that the LLM provides are reflective of best practice or whether they are simply what happens to be most written in SO or blog posts.

"Write C++ code that sorts the following inputs"

versus

"Write (version) C++ code that sorts the following inputs, ensure the code is secure and uses best practices"

And you'll likely get a different answer.

My default assumption now, after watching dozens of post mortems, is that beyond a certain scale, nobody understands the code in prod. (edited added 2nd para)

Why does it always get side tracked into a comparison on how useful it is compared to human capability? Everyone already knows it has issues.

It always descends into a "it won't replace me it's not smart enough" or a "AI will only help me do my job better" direction. Guys, keep your emotions out of discussions. The only way of dealing with AI is to discuss the ramifications and future projections impartially.

Though i was surprised i can't easily run pdb instance via a python program, still have to use stdin/out apparently.

Next i want to implement automerge (or semiautomerge) between different outputs which e.g. contain variants of the same function to automatically resolve issues spawned from the model forgetting. That's so annoying

I also suspect a lot of issues are due to the training data being on old SW. I think we can automatically remap this with whitelisted functions and parameters (i recall inspect can do this), blacklisted ones from old version NOT present in the current, and maybe a transformation between the two -- or automatisch regenerate if it's wrong, maybe with a modification to the prompt.

Also talking to it in other languages generates massively different code (i used deepl) so i had the crazy idea of spawning Dockers and just letting this automatic/semiautomatic trouble Shooting+ just parallel generating lots of functions using wildly different inputs (and models) to brute force the problem of having to code

I do need to look into a nice terminal interface for N-way merges and parallel gen monitoring.

The most useful thing for me was making some vim keybinds and scripts to automatically grab Codeblocks, run them and quickly regenerate. You can literally just tell it "DF" and if fixes a pandas issue sometimes

The holy grail will probably be local fine tunes/LoRAs for specific issues or libraries, since it only costs a few $ for one. Sign me up for an expert plotly AI in a box for neat plots please

Edit : i also have literally no idea what I'm doing either, but linting and analyzing generated Code blocks could help expedite this whole process as well. And in principle you don't even have to run it if you know the type is wrong or something.

I don't know what this is called but computer science is ostensibly mathematics so i assume/hope there is some rigor here