AI can spew code, but kids should still suffer like we did, says Raspberry Pi Raspberry Pi, a company started with the aim of democratizing computing and recreating the programming frenzy of the 1980s and 1990s, is warning that "vibe coding" cannot replace the skills picked up during the process of learning to code. In a position paper titled "Why kids still need to learn to code in the age of AI," the …
It's going to be a long time (if it happens at all) before AI is truly capable of understanding what the meatbag writing the prompts actually wants and be able to solve novel problems it can't just crib from code it's seen before. AI still has no understanding of anything it does, how it works etc, that still needs a human.
You & I understand that. What worries me is that managers will not and force the use of AI generated code "better productivity innit".
All nice & good until it is not.
It's a tool that you have to learn to use.
Getting copilot to write the outline framework or to generate boiler plate for a typical operation is better than searching stackoverflow coding it from memory
You have to learn how to phrase the requests, just like you had to learn the correct magic incantations to the C++ compiler
I don't code a lot, but i found that talking to people on a forum helped understand how to do it, in context,.
I have not used AI, but just getting the answer is not a learning experience, and will not show the underlying concept.
As others have said, you will not know if the AI generated code is right or not.
And that only proves that boilerplate is just moronic.
I never write
function main () :
#... blahblah
if (main):
main()
because it's effing wasting my time.
Plaintext code is an inherent waste of time. The line that you "need to understand" ignores the fact people don't actually understand the boilerplate -- they just copy and paste it.
Copilot can indeed be useful for boilerplate code, but so is a code snippet library utility, and a snippet library has the benefit of being easily updated to new language features, and less likely to generate random mistakes.
The most useful thing I have found AI to help with is in understanding a language I am not familiar with. That’s something that doesn’t require the answer to be completely correct, as i can figure out what it means even if there are minor mistakes. For situations where I have found a code solution in another less familiar language and need to translate it, it’s genuinely useful.
The biggest hinderance to AI adoption is the vastly overblown promises of AI companies that are billions in the red and are desperate to start making profit, of at least meaningful growth. Negative reaction to the actual limits of AI after exaggerated claims inevitably leads to disappointment and rejection.
Unfortunately I think a crash is now the only thing that will reset that situation to reasonable expectations and put us in a situation where widespread acceptance of AI adoption as useful but limited tool can happen.
>” You have to learn how to phrase the requests…”
Reads like we need a new high-level,programming language designed to help developers or perhaps more correctly AI prompt writers, to create the framework and boiler plate needed to productively driver tools like Copilot.
The laugh to some extent is that such tools already exist and are used in some environments for the creation of system and software specifications. I suggest if Copilot can’t read and process a BPMN diagram and produce working output it’s definitely got a long way to go before it is going to be useful for real business applications.
My esteemed employer has migrated to Github, and Copilot has made itself known so I tried it and asked it to find all the commits to a company repo which had something to do with a feature I wanted to take a look at. It said it couldn't give me an answer because there was some problem talking to the Github API and could I rephrase my question.
So having seen the level of help that Copilot offers, I went to my profile and disabled that as much of that as I could.
"AI still has no understanding of anything it does, how it works etc, that still needs a human."
Plenty of software, such as accounting programs, need minor updates from time to time to adhere to changing laws, tax reporting requirements, etc. Does one go back to an AI and completely re-code the application? Something such as a tax table changes often enough that it would be considered from the outset as something that can receive updates, so that's a known issue, but government likes to spring nonsensical crap on companies often enough. I've seen applications with nomenclature that applies to a particular thing, but with a renaming of fields and some slight modifications be useful someplace else. I spent some time at trade show talking with a company that built a system that kept track of something for schools and thought how well it would work for the tool crib at a large company. Perhaps I should have only hinted and angled for a consulting contract, but then I'd have to do real work and write stinkin' reports. Would an AI get the connection?
Even with "vibe" coding you need to describe the problem you are trying to solve meticulously.
To solve a programming problem with a computer, you must specify the problem and all its conditions. This description is the specification and needs to be complete and correct. Only then you have a program that will function correctly. But when you have a complete and correct specification, then the ONLY remaining step required is compilation into an executable format. In essence, the specification is the program.
For an AI to create that program, you need to create a prompt that is the complete and correct specification. Therefore, the AI is nothing more than a compiler. The actual work has been done writing that prompt.
The AI is completely superfluous. You have already written the program when you wrote the promptspecification.
No you don't need to write a precise specification. A specification, as in a URS or even a functional spec is not pseudocode, nor is is UML or similar. There are general best ways to do things. AI knows them, and codes them.
The programmer needs to understand how to use this output in the bigger system, how to hook up all the ins and outs etc. A non programmer will struggle with that. AI is very good at helping people along by supplying pre-built chunks of code. I know this, because I do it every day and no amount of whinging on here (or pointless downvotes) changes the truth.
Unpopular thing to say amongst this groupthink site, but tough shit. Suck it up.
Quote
"AI is very good at helping people along by supplying pre-built chunks of code."
In my line of programming , thats called "A library" or "A macro" and its hardly needs AI to generate the code needed to run the robot loaders. the rest of the code is generated by the programmers interacting with the CAD/CAM system.
And we've tried AI for our stuff.... the word coming back is NOPE!. the code to be honest does look ok to a beginner or to a less experienced hand and may just work. sadly in high speed manufacturing the word 'may' can leave a trail of destruction or worse.... injury
Labeling people who prefer modern tools and pragmatism as 'keyboard monkeys' is a great way to make yourself look like a pompous ass. Exactly the kind of outdated elitism that holds teams back.
To return to a serious software engineering discussion. I’ve always found UML to be more ceremony than substance in most projects I’ve worked on.
AI makes me faster, more consistent, more effective and allow me to focus on solving actual problems.
Drawing diagrams that no one maintains and date the moment you write them doesn't and never will.
Unpalatable fact, (in anticipation of a deluge of downvotes): companies who cling to stuffy, theoretical processes as a measure of competence are going to be left behind. Period.
The industry is moving on whether you like it or not.
Those who can’t adapt are welcome to keep perfecting their diagrams while the rest of us ship working software.
>” To return to a serious software engineering discussion. I’ve always found UML to be more ceremony than substance in most projects I’ve worked on.
AI makes me faster, more consistent, more effective and allow me to focus on solving actual problems.”
Softwares engineering is about precision, ie. It is an attempt to make software development an engineering discipline more than a craft or an art. AI, certainly it is current form is a tool for the hobbyist, not an aspiring engineer.
Remember at the point a software developer whose function is to produce working code, joins a project (especially a business transformation programme), much has already been decided.
>” Drawing diagrams that no one maintains and date the moment you write them doesn't and never will.”
I expect the same applies to the AI prompt scripts…
The maintenance problem isn’t going to go away just because AI has been used. A big part of being an engineer is actually reading and maintaining the documentation…
This post has been deleted by its author
I work for a company that writing pretty critical code. No, not lives-at-stake-level but livelihoods-at-stake-level. My mind boggles at your statement.
Stop telling me (and the world in general) "I don't need to write a precise specification". This is just plain wrong - and dangerous. Same thing with testing edge cases or actually documenting your code. And unless your code has almost zero impact, these three things, however much we all like to avoid them[1], have to be done in a professional environment.
Unpopular thing to tell people, but tough shit. Suck it up.
[1] Yeah, I have been guilty as charged on all three points in the past. We all have. Why else would all those memes exist with "my code is self documenting" et cetera?
Part of my job is writing code for pharmaceuticals production. So potentially lives-at-stake level if done wrong. A precise specification, continually maintained with any changes, is an absolute requirement. So is testing edge cases and documenting the code. (Not "just" edge cases; we are required to test both branches of every IF statement!) All of this is for good reason - to make sure we have code that does exactly what we need it to do, so the product is good quality, so the patients get their lifesaving medicine.
Having an AI write code in this environment? If that's a joke, it's not a funny one.
>” There are general best ways to do things. AI knows them, and codes them.
The programmer needs to understand how to use this output in the bigger system, how to hook up all the ins and outs etc.”
Exactly the same as using boilerplate, library and model code.
Ie. Having assessed and grabbed something as being relevant and potentially useful, a good programmer will adapt and hone it into the flow of logic they are developing.
Hence the value of AI in software development isn’t as great as some would like us to believe.
"Therefore, the AI is nothing more than a compiler."
The problem is that when it screws up, you don't know what the input was that actually created the code. It would be bad to have invoices sent that say "GFY" on the bottom instead of "Thank you for your order", so how would you find out why it's doing that?
A Long Long Time Ago, 2006, the ever-brilliant David Brin (one of my favourite authors) wrote an article called "Why Johnny Can't Code"
It seems things are just going to get worse and Johnny will never code.
https://www.salon.com/2006/09/14/basic_2/
For three years -- ever since my son Ben was in fifth grade -- he and I have engaged in a quixotic but determined quest: We've searched for a simple and straightforward way to get the introductory programming language BASIC to run on either my Mac or my PC.Why on Earth would we want to do that, in an era of glossy animation-rendering engines, game-design ogres and sophisticated avatar worlds? Because if you want to give young students a grounding in how computers actually work, there's still nothing better than a little experience at line-by-line programming.
Only, quietly and without fanfare, or even any comment or notice by software pundits, we have drifted into a situation where almost none of the millions of personal computers in America offers a line-programming language simple enough for kids to pick up fast. Not even the one that was a software lingua franca on nearly all machines, only a decade or so ago. And that is not only a problem for Ben and me; it is a problem for our nation and civilization.
TBH, I think Python fits pretty well in the BASIC slot, and that's trivially available on every platform.
It's plenty powerful enough for the kinds of things I did with BASIC, enough sharp edges to make mistakes with, fast enough, and has sufficient bindings to make fun things before potentially moving up to something more serious.
From my involvement with the initial teaching of children to programme, I would regard Python as a second language.
Those first steps require the simplicity of Basic command line interaction. However,, for those who “click”, a rapid switch to Python is both possible and beneficial to the child’s development. Although, there is the viewpoint, expressed in the 1980s that we shouldn’t be exposing children to procedural programming languages as their first language and instead should be teaching them declarative languages such as Prolog.
Thanks for the article. I was unfamiliar with the article or it's author so thanks for both. I read it and I think the author is on the verge of suffering an overdose of nostalgia. I remember BASIC. I had to write some in Uni and even got paid to write some on the side: Can't say I ever liked it. C (which is what I ended up re-writing my got-paid-for-this BASIC code in a year later when I couldn't make any sense of it) provides at least as low-level view into "how the machine thinks" as C-64 BASIC, without having to use BASIC as the "shell" to interface with your O/S (something it's even worse for than application programming IMHO). The author himself admits that are many variants of BASIC he could download then, (and even now) if he really thought BASIC was the cat's meow, but that they were all too high-level. Really? He admits that his son was learning C++ but doesn't really say why that didn't accomplish the same purpose. Finally, he mentions some of the languages designed specifically to teach kids to code (I think he used Logo as an example but there was also Turtle Graphics and I think Scratch was already a thing...) but that they were too high level for his purposes: that point I concede. I totally agree with the underlying premise: someone who thinks the way to sort data is to use COBOL's "SORT" verb, or SQL's "ORDER BY" is missing a lot, but I have to admit that it's been a pretty long time since I had to write my own sort routines from scratch (it was in Perl as I recall -- that's how long ago!), but don't his specific examples don't resonate with me (personally) at all.
"Even in a world where AI can generate code, we will need skilled human programmers who can think critically, solve problems, and make ethical decisions."
First up, we are not in a world where AI can generate reliable code and a search of the multiverse would turn up rainbow unicorns far more often than a reliable AI.
Next: not just skilled human programmers. A bump to the proportion of humans capable of critical think would be wonderful.
Third: Critical thinking is not even on the checklist when looking for when it is sensible to unleash an LLM on the world.
"The friction introduced in the conversion of human reasoning into a rigid expression of logic is where the learning and development of computational thinking occurs."
We all hate legalese, or at least most of us do. Consider this just another way to cast human reasoning into a rigid expression of logic (when done correctly). I actually know a guy who did exactly that as a thesis for his law degree, mapping legal language to mathematic logical expressions to build a "reasoning machine". Lawyers have also obviously been doing obfuscated code longer than we did.
“So why bother with learning the skills necessary to program a computer?”
I have read this type of thing so many times. Why learn assembler? Nobody uses it any more. Why learn C? I can write a couple of lines of Python that will do the job.
This thinking is like ignoring half the food chain. To get a C compiler, you need an assembler. To get a Python interpreter, you need C (or something similar). Ignoring and throwing away the lower levels of this chain is false thinking. We will ALWAYS need assembler (at least until there is a very major change in how most computers work). We will still need C and similar “low level” languages for many years to come. It is the very top level stuff (like Python or PHP or Perl) that are transitory and more likely to be discarded.
Assembly is forever (it will certainly outline me) and if you have any interest in software at all you really ought to learn at least one assembler. That way, you will start to understand why your 2 line Python program isn’t magic; that it does in fact make use of thousands of lines of C which in turn relies on many more assembler instructions under the bonnet.
AMEN! You can drive without understanding how brakes work, but when you get to that icy patch, you may reconsider that approach. Similarly, understanding the underlying code and architecture-- especially if you're using a debugger -- is vital to good use of even higher-level languages like Python.
Sounds like a luddite argument, flailing against something that might undermine their income stream.....
Why shouldn't programming languages be natural language based? Why should they be often some incomprehensible collection of symbols etc that only make sense after extended periods of study and still often result in people struggling agaibst the language.
I don't recall in star trek etc scenes of people bashing in code by hand, instead it's describe to the computer what you want and it spits it out
It’s not a Luddite argument at all. The computers we use today are fundamentally very stupid at heart. They need absolute, precise instructions to do anything vaguely useful. As of now, the best way (and I’m not saying this is desirable or not) is to use a classical text based computer language.
While it would be wonderful if we all had Star Trek levels of computer sophistication, we don’t. Obviously there are many people trying to change that; the current crop of ill conceived “AI” nonsense being one. But it’s not there yet, and I’m afraid complaining “why doesn’t it work like this?” Isn’t going to help
So, until such a Navana is found (and feel free to research and develop it if you like) you’ll have to make do with a keyboard, a display, and a brain
Although stating the problem to be solved and the goals for the solution should be done in a natural langue like English this language is far too imprecise to use instruct a computer. Computers are well meaning but terminally stupid, as the old saying goes "It never does what I want it to do, only what I tell it".
> don't recall in star trek etc scenes of people bashing in code by hand, instead it's describe to the computer what you want and it spits it out
That's the magic of screenwriting. Implied in the script is an accumulation of cultural experience that defines the overall course of the plot (the "six stories") and the efforts of set designers and special effects that make sure that the plot unfolds per the script. Even the movie "Galaxy Quest" that attempted to address what would happen if some advanced -- but naive -- civilization took Star Trek as literally true had to resort to dubious, unlikely (if rather fun) scenarios to get plot closure.
This says everything about why you cant use natural languages for programming
>” Why shouldn't programming languages be natural language based? “
They are, just that to get the machine to do exactly what you want, requires precision, which distorts natural language into a more formalised structure. You see this in policy and procedure manuals. The “collection of symbols” can largely be attributed to programmers not wishing to be verbose and minimising keying.
Hence why COBOL programmes being more verbose tend to be more readable than C and Rust.
But this is missing the point, you aren’t actually programming the AI, you are giving it an instruction to produce output that can be fed into a compiler, ie. It is translating your natural language instructions into a more formalised structure, namely a programming language.
A big assumption with AI is languages. I expect there to be a large resource for C, but other languages? So I doubt (general) AI will be of much use writing Algol68 or Ada programmes, even Rust programmes will without careful screening be contaminated by C or C++.
I don't see mentioned in the comments or article the fact that the company which makes Raspberry Pi computers is separate and independent from the Raspberry Pi Foundation being reported on. This may be obvious to many people, but I mention it anyway.
From my point of view the job market for programmers and software engineers has historically expanded since the 60's due to the everywhere increasing use of computers. Then suddenly AI makes writing software so efficient that even though the everywhere use computers still expands, the job market for programmers and software engineers contracts.
The question boils down to how fast the need for new software continues to increase versus how fast developer productivity increases due to AI. From my point of view, productivity increases will be so great the job market is likely to contact.
People make analogies about how the job market for skilled machinists didn't grow with the automobile market due to assembly-line productivity increases.
