AI is creating a generation of illiterate programmers
nmn.gl
nmn.gl
It took 3 times longer than if I had done it by hand because I just kept copy and pasting each error back to ChatGPT ending up in a dead end having to start over from scratch. As soon as you have a requirement that is not the 90% of cases you can quickly end up in a dead end especialy if you have more than one such requirement.
Once the LLM starts hallucinating you can easily get into corner where you will need to reset and start from scratch.
Set Cursor in "agent" mode (with Claude as a model).
Ask what you need.
It will automatically test your Docker deployment and configuration.
Then you become like me, a monkey clicking "Accept".
But a happy monkey.
You framed that as a criticism of ChatGPT, I think that is indeed the point the OP is making. IMO if this is something you used to do you could get a nice starting point with a single prompt and then go from there, then that would take 3 times less compared to reading the documentation from get-go.
I think LLMs are great for reducing the friction with just starting a task, but they end up being more of a nuisance when you need to dig into more nuanced problems.
It seems like a lot of the models in ChatGPT and Copilot were trained on that content and in turn tends to produce a lot of dead end solutions for anything that isn't the 90% cases and often leads to more pain than reading documentation and building a solution through iteration/experimentation.
> It took 3 times longer than if I had done it by hand because I just kept copy and pasting each error back to ChatGPT
You know that if you already know what you're doing you can tweak the output directly instead of mindlessly pasting it back into ChatGPT, right?
I spent a year just coasting at a job and not really writing any code.
It was quite difficult to get to my same level after starting my new job.
Now that I’m back there I’m weary of losing my touch again.
Re-learning a large codebase would take longer, of course, but I'm talking about just getting up to speed like if I was starting a new project.
But surely you can tell it's not like riding a bike.
> just getting up to speed like if I was starting a new project.
This is the best case scenario for getting back up to speed.
It may sound harsh, but there's not a lot of skill involved in starting new coding projects (it's one of the best things about the craft imo). Of course you'd get back up to speed quickly, especially if you have 10+ years of programming under your belt.
I generally feel the opposite. Starting a new project, not just a prototype, requires a lot of decisions and knowledge that mostly only has to be applied at the beginning of a project. Language choice, core dependency choices, build system, deployment, general project structure, etc. Each one can be changed later on, but the longer a project goes on, the more of an impact these early decisions have, and the more important it is to get them mostly right from the start.
On the other hand, when you join an existing project, many decisions of the project are already made and you just write code that resembles existing code.
But even so, many web projects are as easy as `npx create-next-app` and you have a solid foundation.
It’s much harder to start working on an existing codebase.
> you just write code that resembles existing code.
This is so incredibly wrong.
Every line of code you write as an IC on an existing project should take into consideration the existing code, the teams patterns and code style, and the reasoning behind existing systems.
The choices that were made that you’re unaware of makes it much harder to write any code, let alone code that also resembles existing code.
Decisions don’t require skill, necessarily. Anyone can just choose a web framework or a database.
Working within the restrictions of an existing system has a much higher skill floor.
User: { AI generated 1600 line handler "getUsers" with sql injection, n+1 queries, an exposed secret, plaintext passwords and credit cards, and lots of unused variables }
I've seen in some companies, relatively trivial tickets are bounced from developer to developer for months and management believing these are hard to do.
But on the other side does society also need highly educated experts, with deep understanding of things and the ability to find and prevent the s** which will harm us. This is not limited to IT, it's the same in every area.
Education prevents disaster. But AI prevents education, maybe. We will see how this will play out for us. Maybe the AI-Overlord won't be just a joke anymore at some point, and benevolent AGIs can replace the necessary experts.
Edit: Arvind Naranyan of AI Snake Oil uses LLMs to create "one-time-use" apps, mostly little games or educational things for his kids. As an AI skeptic I think this is cool and should be celebrated. But there is a serious downside when using it blindly for real work.
I have never liked the term "can program" anyway, programming is the easy part....
I used to be a non-programmer (although extremely technically advanced). I was very good with the terminal because I always played around with ADB and other things. I even wrote Windows batch scripts (but looking at them now is embarrassing. I used a lot of repetition instead of loops).
When ChatGPT became available, and my friend taught me how to use it to write Python scripts, I went all in. I didn't write a word of my own code. I just copy /pasted Chatgpt's output into notepad++. It was really annoying sometimes to get ChatGPT to change a simple thing. It hallucinated often and "fixed" stupid things I didn't ask for help with. I used all these scripts for various personal things. For example, it made me a GUI to edit the metadata of every .mp3 in a folder).
After several months of exposure to code, I knew the basic syntax and tried to code myself when I could, and Stack Overflow was a big help. I still had very basic skills and didnt even know what a `class` was.
Fast forward to now, I consider myself extremely good at Python, and decent in other languages. I now use classes and dataclasses all the time. I always add type hints to my code. I follow Python PEPs and try to design my code to be short, maintainable, and to the point. If a library lacks documentation, I just read the source code. I started using an IDE, and the intellisense is really a step up from using notepad++.
I still use copilot, but I refuse to paste large code blocks from any llm. It makes the code so much harder to maintain. If I really do like the code chatgpt gave me when I get stuck, I write it myself, instead of copy/paste.
I don't know if everyone has the same experience as me, but I would certainly say that ChatGPT helped me become a successful programmer.
I shudder at the idea of what I would have become if life made it so that I would only read RD (bc of a lack of time, of support, of curiosity) for the rest of my life. A shallow half-knowledge I would have been contented with, oblivious to what literature is really about.
Another similar idea I am culpable of: watching videos of people playing games instead of playing for myself. Not the same experience at all.
I remember all those devs who could only do tutorials using visual tools like VB6 and ASP forms.
Once the requirements went beyond the basics the project always got handed off to a real dev.
Of course not. The ultimate goal of AI is to get rid of the developers altogether and reduce costs (more profits!). You think BigTech are spending hundreds of billions just to make developers more productive while retaining them? That's not a profitable strategy.
The copy/paste with LLM interactive stage is just a transitional stage as the LLM improves. We'll be past that in 5 years time.
No, AIs won't replace all developers -- you still need people doing the systems designs that the AI can implement code for. But it could easily reduce their numbers by some large percentage (50%? 80%?).
Edit: I would no longer advise my kids to get a degree in Computer Science.
Given the rate of improvement up to that point, how much longer do you think this would be the case?
> We’re becoming 10x dependent on AI. There’s a difference.
This is true, but I also don't need 10x the ability to write cursive any more. I used to have great hand writing, now it's a very poor. That said, my ability to communicate has only increased throughout my life. Similarly, my ability to spell has probably diminished with auto-correct.
Yes, you will become dependent on AI if you're a developer, because those who use AI will take your job (if you don't) and be significantly more productive than you. It sucks, but that's reality. My grandfather started programing on physical cards, he knew a lot of stuff I had no idea about. That said, I'd be able to run circles around him in the programming (and literal) sense today.
The question is really what skills do you need to know to still execute the job with AI effectively. I'd argue that doesn't change between AI and not having AI.
As a seasoned engineer, I spend probably 60% of my time designing, 30% managing/mentoring/reviewing and 10% coding. That probably wont change much with AI. I'll just be able to get a lot more done with the coding and design in the same timeframe. My skillsets likely will remain the same, though writing small functions may diminish slightly.
>...and 10% coding.
Disagree here:
>AI will take your job (if you don't) and be significantly more productive than you
Physically typing out the code is one of the last steps. Before that can happen the problem which is to be solved must be identified. An appropriate approach to the problem must be formulated.
Look around at some of the deranged hype in this AI cycle or look back on previous hype cycles. Many of the proposed "solutions" do not solve the problems of consumers. Solutions in search of a problem are as abundant as novel tech. These things dominate the zeitgeist. Perhaps they help inspire us, but they are not guaranteed to solve immediate problems.
There's an important distinction between problem solving skills, tools and the most probable token. To innovate is to do something new. AI, on the other hand, does the most probable thing.
The "no AI without understanding the solution" rule is a start here.
That being said, one of my top google searches are things like "kubernetes mount configmap as files" because I don't quite do them often enough to have them totally internalized yet.
I continue to believe that these tools would be more reliable if rather than merely doing next-token-prediction on code, we trained on the co-occurrence of human-written code, resulting IR, and execution traces, so the model must understand at least one level deeper of what the code does. If the prompt is not just a prose description but e.g. a skeleton of a trace, or a set of inputs/outputs/observable side effects, then simultaneously our generative tools only allow outputs which meet the constraints, with the natural and reasonable cost that the engineer needs to think a little more deeply about what they want. I think if done right, this could make many of us better engineers.
What you want - what everybody wants - is a GPT plus. In this case, GPT plus a compiler's parsing ability. I don't think that's what we have, though. What we have is a GPT that was trained on more and better code snippets. That's not good enough.
The problem with people sneering at LLM assisted coding is that they miss the developments and advancements in tooling and integration.
Copy pasting from the ChatGPT browser tab and hoping for the best is not where it's at anymore.
I mean, maybe there's a magic tool out there? The only one I've used (briefly, before turning it off in annoyance) is Copilot, which appears to pretty much just do that.
(I suspect possibly the reason that this does _not_ seem to be commonly done is that reliably producing _something_ (even if it's nonsense) is seen as more appealing to users than potentially not producing anything. And of course passing type checking is no guarantee of correctness.)
No, it's not. It's a guarantee that it should at least compile, though. That is, it means certain kinds of incorrectness are ruled out.
However just like learning to drive, I think there is much scope for forcing the basics once in a while.
The old programming blue shell, as it were.
Please find one. I highly doubt there are any at all.
AI isn’t translating idea to code. It’s guessing as to what the most likely next code is, given a context.
Compiling is exactly translating code to machine code, but hiding some (or a ton of) technical details.
Those are two wildly different mechanisms to interact with.
Assembly and compiled languages require the same precision of thought, just at different levels of technical detail.
LLMs can, in effect, translate some ideas into code. Under the hood, though, they are not. They are predicting next tokens based on context.
There are impressive and obviously useful emergent properties of how good they are at predicting tokens, but that is what they are doing. There's no question there. That's a fact.
It's not a direct translation, however, as is the case with compiled languages. Compiled languaged directly translate the idea you communicate to them into machine/byte code in a deterministic, predictable, and importantly tractable way.
LLMs do not do that. They aren't abstracting what the computer can do so that you can more effectively write instructions.
as you say
> It might not always be good or maintainable code
What if your idea _was_ good and/or maintainable machine/byte code? What if you needed to examine the output code and ensure some quality of it. With a compiler, if you can communicate in a programming language what exactly what you want, it will _always_ produce machine/byte code that matches that.
LLMs just don't do that. It's just not how statistical models work.
For me, AI is like pairing with a very knowledgeable and experienced developer that has infinite time, infinite patience, and zero judgements about my stupid questions or knowledge gaps. Together, we tackle roadblocks, dive deep into concepts, and as a result, I accept less compromises in my work. Hell, I've managed to learn and become productive in Vim for the first time in my career with the help of AI.
a Typescript programmer who doesn't understand JavaScript well?
Or a Python programmer who doesn't know C?
Or a C programmer who doesn't know assembly code?
Or an assembly coder who doesn't make his own PCBs?
The shame is not in outsourcing the skill to, e.g. make a pencil[1]; rather, the shame is not retaining one's major skill. In IT, that is actually thinking.
What most people who would call themselves programmers are working on is far beyond what AI can do today, or probably in the foreseeable future.
that would be a generous invented statistic if it only was addressing the inherent stochastic nature of llm output, but you also have to factor in the training data being poisoned and out of date
in my experience the error rate on llm output is MUCH higher than 5%
While it is important to understand the fundamentals of coding, if we expected every software engineer to be well versed in assembly that wouldn't necessarily result in increased productivity.
LLMs are just the next rung up on the abstraction ladder.
There will always be people interested in the gritty details of low level languages like Assembly, C, that give you a lot more granular control over memory. While large enterprises and codebases, as well as niche use cases can absolutely benefit from these low level abstraction specialists, the avg. org doesn't need an engineer with these skills. Especially startups, where getting the 80% done ASAP is critical to growth.
If someone who can't read, write, or debug code can still be called a programmer, which unique skills do they still possess?
But I think everyone who isn't at least a standard deviation above the average programmer (like myself) shouldn't be focused on being able to read, write, and debug code. For this cohort the important ability is to see the bigger picture, understand the end goal of the code, and then match those needs to appropriate technology stacks. Essentially just moving more and more towards a product manager managing AI programming agents.
Someone who delegates programming tasks to others through natural language instructions isn't a programmer, but a manager.
That's not enough. You will inevitably hit a problem which you will leave for "tomorrow" because tomorrow is not a No-AI day.
Anecdotally, I wouldn't be surprised if there was some truth to it: in the early 2010s I saw tons of people saying "defiantly" that meant to type "definitely", and I couldn't figure out why. Then I misspelled "definitely" once and saw the spell checker suggested "defiantly", and it finally clicked.
Spell checkers have definitely improved since then (I can't remember the last time I saw the definitely/defiantly mix-up), but I can't help but wonder how bad suggestions have affected the way people understand languages.
A similar phenomenon exists on TikTok with words taking on contradictory or even opposite meanings, though this doesn't seem to be caused by spell check so much as constant incorrect usage rewriting the definition in people's minds (like "POV: you're getting mugged" when it should actually be "POV: you're mugging someone" or "POV: you're a bystander watching someone get mugged").
There is a dumbing down effect, for sure.
also we didn't have those discussion before, everyone forgets about abacuses and log rulers
Do they just copy paste emails and office menus into an agent and send responses? Do they actually make any decisions themselves, or do they just go with whatever?
So I guess office jobs are now grinding machines, where one side has keywords, the other side extracts keywords, and in the middle there is an enormous waste of resources to fit some manager's metrics.
There isn't any such dependency (yet).
Despite strong corporate push I'd say the average person I work with uses it maybe once a week - usually research or summarization.
The inroads LLMs are making on programming aren't translating to other office jobs cleanly. Think it's largely because other areas don't have an equivalent source of training data like github
It'll no doubt come, but for now all programmers seem to have mostly automated themselves out of a job not others from what I can tell.
>Do they actually make any decisions themselves
There is little to no GPT decision making happening despite all the media chat about AI CEOs and similar bullshit. Inspiration for brainstorming is about as close as it gets
my assumption was that these office people were secretly using it to automate dull parts of their work without the approval or knowledge of their coworkers
Yup, because I'm one of these office workers. Financial controller stuff in PE space & coding as hobby and interest in AI space.
There just isn't a direct equivalent thing that captures the knowledge in a machine readable way the way a code base does. It's all relationships, phone calls, judgement calls, institutional knowledge, meetings, coordination, navigating egos and personal agendas etc. There is nothing there that you can copy paste into an LLM like you can with say a compile error.
Even the accounting parts that conventional wisdom says should be susceptible to this...it's just not anywhere close to useful yet. Think about how these LLMs routinely fail test like "is 9.90 or 9.11 bigger" or miscount how many Rs are in strawberry. You really want to hand decisions about large amounts of money to that? And maybe send a couple million to the wrong person cause the LLM hallucinated a digit wrong. It's just not a thing.
Maybe with some breakthroughs on hallucinations it could be but we all know that's a hard nut to crack.
>automate dull parts
I've been trying hard to find applications given my enthusiasm for coding & AI. No luck yet. Even things where I thought this would do super well...like digging through the thousands of emails via RAG+LLM is proving oddly mediocre. Maybe that's an implementation flaw, not sure.
Because AI doesn't have the latest documentation.
i mean maybe it's just me but i quite like this new normal where i'm an "intermediary". i review every line that comes out and research things if they don't make sense or i'm confused. and a lot of times it helps me figure out how to approach a problem or refine an existing one. i'm still architecting everything and diving in.
i guess if you liken progamming to like i dunno, chopping your own tree down, turning it into 2x4s to build a house then yeah, it seems like your skills are atrophying.
i do not miss the "joy" of writing random boilerplate 100s of times or variations of the same function. i do however fear for the the new generation who maybe didn't cut their teeth on having minimal guidance. i can totally see people just coasting and blindly pasting things, but that's always sort of existed on stack exchange, etc.
I think AI is just giving cover to all the incompetent people in your org, and in fact will create more incompetence via the effect you yourself noticed.
Eventually you’ll find many if not most of your coworkers are just wrappers around an LLM and provide little additional benefit—as I am finding.
I am curious, do you think it will stay this way? AI couldn't code anything 2.5 years ago, why should it stop with glue code?
Yesterday I fed the first 3 questions from the 2024 Putnam Exam [0] into a new open source LLM and it got 3/3. The progress is incredible and I don't expect it to suddenly stop where it is.
The open source model as worked through the problem step by step (it was R1 so I could read its chain of thought). That could be an elaborate ruse, but also unlikely.
LLMs literally grew out of autocomplete, and aside from having a huge amount of information they still don't understand, and certainly don't learn during inference.
Do you have any podcast/blog/videos that you recommend which make your case clearly?
Then you have the project manager / architect skill of shaping things and having an understanding of the tech stack - but does not need to be a great coder. Maybe they were in the past, but they're not grinding silly LeetCode questions and cannot solve them, and are smart enough to see that as the hamster wheel surrogate goal that it is.
I think we need less of the former, much less. I don't see it as a bad thing. I myself have severe carpal tunnel, RSI, and back problems, from sitting at a computer.
Almost all of the students that jumped into programming courses did so for money. They have no passion (less than me by far) for coding. Its a lie perpetuated by toxic positivity and invested groups that coding is some magical, super cool, super smart activity.
Its very overdue at this point. DeepSeek R1 and V3 are competing with o1, open source and local hostable. NVidia is coming out with small AI GPUs for consumers.
The winners of this AI age is everyone. EXCEPT 'coders'. Many getting 6 figures to maintain a CRUD app. I don't feel sorry for them! Cooking up garbage apps in PHP - I cannot wait for it to be overturned by energy efficient AI made apps in Rust.