Hacker News Reader

Comment by hiAndrewQuinn | Hacker News Reader

I have the extremely fortunate scenario of watching my wife just begin a computer science degree now, with no prior experience programming whatsoever. This is not hyperbole: ChatGPT is an absolute game-changer when it comes to getting through the basics.

It is simply far, far easier for a newbie to see a computer generate 1 / 3 / 5 slightly-different attempts to solve a specific problem they have, and then to pattern match sufficiently to be able to solve the problem themselves, than it is to muddle around by yourself with it for hours and hours with no end in sight.

I thought she would have to ask me, someone who has been slinging Python in some form since 2008, a question at least once a week. In reality it's been about 3 times over the last 9 months.

ChatGPT doesn't do much for experienced SWEs, but it demolishes the difficulty curve for newbies. I wish I had this when I was learning.

View on HN

> It is simply far, far easier for a newbie to see a computer generate 1 / 3 / 5 slightly-different attempts to solve a specific problem they have, and then to pattern match sufficiently to be able to solve the problem themselves, than it is to muddle around by yourself with it for hours and hours with no end in sight.

That's because it is easier. Multiple choice is easier than open questions.

It's also a good way to not really grasp anything deeply.

But you don’t need to grasp anything deeply. You want to create something small and get a little bit of feedback quickly.

When we teach math to children we don’t start with algebraic structures and proofs. We give them simple recipes.

Later we show them problems that the recipes can’t solve and dig deeper. „By the way, the thing we did last year, it actually works like this…“

> When we teach math to children we don’t start with algebraic structures and proofs. We give them simple recipes.

This is not mathematics, but arithmetic/calculating. When one starts teaching not-children-anymore mathematics (typically at the university), one indeed starts with something of the kind of algebraic structures and proofs.

The GP mentioned that their wife is entirely new to programming but starting a CS degree.

She will certainly learn the deeper concepts in university, but catching up with coding quickly, even in a shallow manner, is super beneficial. I think it's a huge plus.

I don't mean to neg anyone, but will someone pick up the fundamentals, if they can keep leaning on a chatbot?

If they want to

Anyone can learn anything if they put their mind to it. Sure. But the question is: does it scale?

School has always been exactly the same way. Students who want to learn will learn. Students who do not will not. ChatGPT changes nothing for the latter, it is a tool for the former.

As a teacher, I have to tell you this is 100% wrong.

Well it’s right, but you can turn a student from one who doesn’t want to learn into one who does.

You can absolutely motivate unwilling or uncooperative students to learn. I have done it many times.

I'm not saying you can't switch between being one of those students, that's a whole other thing.

It's like learning from docs examples (guides), tutorials, stack overflow questions etc. But on steroids. We all have learned this way and still do when we quickly want to grasp the surface level. There's nothing wrong with that.

I've been studying AI by leaning heavily on both YT videos (in particular the Karpathy ones) and chatbots.

It has greatly accelerated the process of building prototypes in multiple languages and building various permutations of solutions.

Doing that has allowed me to study various problems both from the bottom up and the top down. I think that has been useful for understanding concepts.

I could have done the same things without AI, but it would have taken a lot longer.

I agree with this. I was a bit blunt. Each situation is different. Superficial learning might be all you need depending on the circumstances.

I still think it’s better to learn without too much tooling of any kind, but that’s a topic for another day.

Arithmetic is very much a part of mathematics, it seems bizarre to me to argue otherwise.

It may very well be a trivial and uninteresting part of mathematics compared to the rest, but it’s still mathematics nonetheless.

Their disagreement is that arithmetic is a part of mathematics, rather than being all of mathematics.

> When we teach math to children we don’t start with algebraic structures and proofs.

Depends. In my high school, math was tought not by telling us the proofs, but by asking us to provide them.

True, but that’s not a good comparison. If we would compare this with arithmetic it would be like handing out a calculator and claiming this is way faster without “muddling around” (also known as “learning things”).

You teach programming by giving simple assignments, not by handing out “calculators”.

> It's also a good way to not really grasp anything deeply.

Having recently done a screenshare with a fresh CS grad as they developed a simple program and seen how much Copilot was doing for them, I was both impressed and appalled. Yes, it's able to do a ton of that simple stuff, but this new grad has no idea _why_ it's doing what it's doing. They are more concerned with getting it done quickly due to not working on the problem early enough (that's a whole other issue).

It genuinely concerns me for the next generation of CS grads and how well they are going to understand the systems they are building. They are already coming out of school unprepared for the professional development world, this is just going to highly exacerbate the problems.

When hiring is done via arcane coding exams, ya, that’s what they are concerned with.

Did I say anything about hiring or arcane code exams?

This fresh grad happens to already be working for us and I was mentoring him.

I didn't say "seems" easier, I also said it is easier. There is no disagreement on this fact.

The good thing about ChatGPT is that you can keep asking it to explain something in a different way until it explains it in a way that “clicks” with you.

Of course you then have to compare this explanation against all others and see if it fits or if it’s not a valid explanation.

It’s also great for exploring topics which have polluted namespaces on search engines.

Overall though I think the majour benefit of ChatGPT is it just teaches people to clearly define problems as natural language questions. Developing this ability helps the subconscious mind solve problems when the user is away from screens.

+1 for this, I think this is one of the most useful and underrated things enabled by ChatGPT for effective learning.

One of the types of teaching video that I've found very helpful is where an expert, usually a professor explaining a concept (e.g zero knowledge proof) to five different audiences starting from school kids up to fellow researchers or professors. You can basically ask many different targeted levels of questions to ChatGPT for example, explain the concept as I am five years old, etc.

Yes, but my impression is that chatGPT returns are really linked to the intelligence of the user. How good are to get things out of it. For a clever person starting to code you can easily jumps all the technical hops of the base knowdledge needed to start creating. In every discipline you need base knowdledge that allow you to be able to make the connections.

When you want to compare software or library X versus Y ChatGPT is excellent. Google is a morass of spam on those queries.

> Overall though I think the majour benefit of ChatGPT is it just teaches people to clearly define problems as natural language questions. Developing this ability helps the subconscious mind solve problems when the user is away from screens.

Yeah we've had them for a while, it was called the rubber duck, ;)

> ChatGPT doesn't do much for experienced SWEs, but it demolishes the difficulty curve for newbies.

Interesting, my impression was that ChatGPT helps experienced SWEs by filling out large amounts of boilerplate-ish code. It sounds like that's not your experience, though?

Which might make you, what, maybe 10% more productive overall? The more experienced you are, the less time you spend writing boilerplate code, so it's not that huge of a win in the grand scheme of things compared to what it provides when you're learning something new.

Not writing any unit tests are you?!?!? ;)

It makes us far more productive; all of us, so we just do more as a team rather than fire someone. For me, copilot writing out if / then / else stuff like it's reading my mind, so tab -> 15 lines, all correct. Or mappings (.map .filter etc) which it completely does within milliseconds. As an experienced programmer, I cannot avoid these things and they do save me a lot of time and debugging. Oh yes, and tests. Going from a jira task to a bunch of jest tests for code that doesn't exist yet is a pretty big win. Very much more than 10% of the average day. You have to be sharp, be a good code reader and writer and it cannot do anything 'alone', but it makes us vastly more efficient.

> Which might make you, what, maybe 10% more productive overall? The more experienced you are...

But the more exeperienced you are, more valuable your output is (in most cases). So the 10% of your productivity could be more valuable than a novice's 50%.

The more experience you are, the less the value is in the output objective productivity. The ability to write self descriptive, maintainable and bitrot resistant code in one pass goes way above feature delivery.

that would be copilot not chatgpt

I think it works well in both ways. I've been developing software professionally for the past ~20 years and there's no doubt that the combination of ChatGPT and CoPilot greatly reduces the amount of time I would usually spend on boilerplate code. Now I can more or less just review the suggestions, review tests and progress with my solutions a lot faster.

At the same time, I see my 9 year old girl using ChatGPT to create Minecraft extensions; she has no prerequisites for actually developing software, but she manages to make some small extensions and have them work either way.

Acknowledging that she doesn't actually learn much from this, I think one can assume that everything in between (e.g. using ChatGPT for pair learning) is definitely also possible.

But whatever AI is outputting, there's no way these qualified guesses can beat experience from a professional developer. I know that the output will never be better than the prompt, but sometimes the output - no matter how much effort you put into the prompt - is visibly just a qualified guess.

While I am not the most experienced, I did recently finish a CS a degree and find chatgpt a great resource for getting a general understanding to go off of when approaching something outside my domain of experience. Having a core understanding of things in general makes it easier to identify when it's wrong about things, and the majority of time I am not using it to generate code, outside of asking for an example. Being able to ask a few direct question on a topic "demolishes" the time spent to become familiar with something, as most of the time I am just looking for answer about some specifics or to clear up some ambiguity in my understanding of a new concept.

I would say with someone new to coding it can be bad and good, as a lot of times it glosses over things, or can be slightly incorrect as it makes assumptions (or more so just answers in a more general context, and when asked to elaborate, or challenged on specifics it will reformat/improve it's answer, but without knowing you need to do so, I could see it easily see it providing half-baked foundational knowledge. You can ask it "x" and it will give a answer, but then if you ask it I am trying to do "y" with "x" and isn't "z" an issue or area of concern with its answer it will reformulate the information provided as its original response was flawed, but if you don't know exactly what the "y" you want to do is, or the "z" being foundational knowledge to challenge it on, you can easily get a whole wall of text that is out of context with what you are actually trying to learn.

> find chatgpt a great resource for getting a general understanding to go off of when approaching something outside my domain of experience.

This is my hang up with LLMs … I don’t trust them.

Frankly it seems just as easy if not easier to just google keywords and read sites.

Google vs LLM is like asking random people on the internet (some are brilliant, some don’t know anything, some are nuts, … etc.) vs asking random people on the internet but all of them have a history of suffering from hallucinations and are routine liars with a compulsive need to answer confidently even when they know nothing.

Especially with an Idea of how they work, I find it very hard to trust them too. Doesn't help that the couple of times I asked them about something subtle, GPT gave me code with the same bugs I was trying to fix. I can see how a beginner would find it useful though.

So it's really just a refinement of Wikipedia, but with custom answers to more explicit questions.

I don't know about you but I trust Wikipedia a lot more than I trust ChatGPT.

Plus, with Wikipedia, you can clearly see the sources or the lack of them. I tried asking ChatGPT for a source twice. One time, it gave a source. Another time, it gave the "as a language model I can't" spiel.

Wikipedia actually has a system for creating knowledge and verifying it, while ChatGPT was trained to convince humans.

> suffering from hallucinations and are routine liars with a compulsive need to answer confidently even when they know nothing.

Sounds like a description of narcissistic personality disorder or schizoaffective. Of course proto ai would have a personality disorder, go figure.

In the future, the job of an ai psychologist will be to certify the personality of ai products. Gotta make sure you’re not shipping a shrink wrapped psychopath.

They're already doing this. That's why Chatgpt is so successful. They sanded off the offensive edges by showing some Nigerians CSAM.

I tried that a few times, but by the time I've explained enough for the code to be close to what I want i could've just implemented it myself twice over... And that's ignoring the fact that it's always generating subtly wrong or just poorly performing code.

It might change eventually as ide integrations improve, but for now it's a novelty for me.

If you have boilerplate anywhere in your code, then you have a design problem.

Yeah, my design problem is the language I'm using, the language version I'm using, the outdated framework I'm using, and the 3! frontend frameworks I'm using on my job.

Stop using them?

A good craftsman crafts his own tools.

Both this post and the one two levels back honestly sound like someone who's never worked in a SW company. Craft my own tools? When? Who's paying for that? Too much boilerplate? Tell that to the 20+ people who built this thing and are long gone. You really have to work for a specific kind of company to be able to change any of that.

I have about 15 years experience professionally as a software engineer, in all kinds of companies, at all kinds of different seniority levels.

I call bullshit. Your task is to deliver a project that does what it's supposed to, all in a timely fashion. You do whatever it takes, including getting rid of old code and rewriting stuff if it makes you more productive overall.

Writing code that does what you want is much faster than taking someone else's code and twist it to do something it was not written to do. You should always be very careful about what dependencies you use, as most often you will have to deal with their limitations and bugs, and they could kill your project.

If you had 20 separate people who contributed to the same thing and left then you most likely have huge code debt. Managing that sort of thing is a daily task for any software engineer. Apparently your strategy is to just add debt onto the pile and put your head in the sand pretending it's not your problem. Instead you should plan tactical refactoring moves to make the codebase leaner and better able to adapt to the business needs, all without breaking any existing functionality and workflows.

You're asking "when". That's the wrong question. It's up to you to manage your time and invest it into tasks that help you save time later on. A software engineer is usually not paid by the hour either, so you just do the work until you're satisfied with the effort you've put in.

My father-in-law is a woodworker who makes chairs that cost five digits, and canoes that run into six digits. He is the epitome of a good craftsman.

He does not make his own tools. He buys them, because there are already extremely talented tool-makers who are very good at their craft.

How many hours does it take him to make a canoe? I imagine it must run into the high hundreds, if not more.

A full canoe can take him up to a year, depending on the wood species. He made one once from ebony, which is extremely difficult to work. Also if it wasn’t clear, these aren’t intended for use. They all do (except for the ebony - he said the heat absorption from the dark color would delaminate the fiberglass coat within hours) float and work extremely well if you want, but given the cost, most people use them as art.

He also makes paddles, which are still expensive (I think the cheapest starts at about $1000), but some people actually use those.

But I'm sure he carefully reviews what tools are available and most likely still modifies them to fit his workflow.

Certainly a lot of youtube woodworkers do that, can't say I work wood myself.

That depends what your definition of tool is. He custom-makes jigs all the time, as those are bespoke for whatever particular shape of thing he’s making.

But saws, blades, chisels… maybe he could make them, but why? Metalworking is its own subject of mastery.

AFAIK, he doesn’t often modify any tools of that sort. He spends time selecting the one he wants, yes.

Making a tool doesn't mean making every component of the tool.

Don’t hesitate to be more specific on your boiler plate problem, because after 19 years career I’m yet to have or see, or even being counted a case where it has any significance on productivity loss. It is at worse questionable yet somewhat healthy warm up.

I worry that our reliance on LLMs (or rather, their reliance on these languages) will make it harder to adopt new technology

Imagine, a vision of the future... A thousand million chatbots, all outputting 2018 JavaScript

It’s boilerplate all the way down.

Not necessarily. Trying to make code over-DRY is also an anti-pattern.

In my experience, the key is that it can also provide good explanations for the basics as well. You can fairly easily overwhelm it, but with some practise of prompting, learning new language/domain basics it is pretty good, and much better than googling to learn the same. The best of all is that you can query your understanding and get decent feedback.

As an SWE, my problem isn’t boilerplate. It’s architecture and orchestration. I cancelled my Copilot subscription because it just doesn’t help with this. Let alone ChatGPT which has zero context about the application.

I would like an AI to tell me if I’m reinventing the wheel or if something similar can be abstracted. Whether what I’m doing aligns with existing patterns in the codebase. How the application as a whole could be refactored to improve maintainability, performance, or both.

This can all already be done by a human with experience and context, so it must be possible with AI. This will be the biggest game changer for me.

I can’t believe you canceled copilot, at $10 a month it pays for itself every single day for me. I’m just baffled how it can be so useful for me and not useful for you.

Once you get over the novelty it gets pretty annoying.

I’ve been using it for a full year now, that’s just not been my experience.

I've found it extremely similar to IntelliCode for C#, which is not very impressive. Surprisingly I've found it does a better job of writing accurate comments than generating useful code.

Maybe it's good for some languages, or for code that's doing the exact stuff a hundred people have already written. Copilot X certainly looks cool, but who knows when that'll be ready.

I use it the other way around. I write the comments first and then I find it can often fill in the code.

I find it especially helpful with SQL queries, and with boilerplate type code. It’s also very helpful when I’d need to lookup the usage of various methods. I just start with the comments and most of the time it saves me the context switch to the browser to lookup the usage.

Experienced SWE here. ChatGPT is awesome for me. It's great for learning new technologies and it's great for writing programs that I'm too lazy to write myself.

Hey, also an SSE here, do you find that writing code is becoming a grind now that GPT can generally handle most changes within a single context?

I'm having a harder and harder time justifying writing a lot of code myself when faster tools are right at my fingertips

Would love to hear how other SWE's who are using GPT feel about this

I mostly use ChatGPT to write code while I'm playing video games. Cutscene/ load screen comes up and I review what it sent and have it iterate. Then back to games.

It's much slower at producing good code than I am but it's much faster for me to write a prompt than a program, so ultimately it allows me to do two things at once because 'prompt time' is so short I can fit it in between other things.

I don't think of programming as a grind unless I'm working in a shitty codebase, which does happen, certainly.

Also as a learner - what i've found incredible is the ability to get AN ANSWER...without context switching. 1 window, 1 result location, no fishing...its trivial to test that given solution...as a beginner...my problems are fairly one dimensional.

Not to mention...things like "cool want to learn python"..."wait - wtf..how do i setup a venv in python!!"

Graying SWE here. ChatGPT is quite helpful, and so is Copilot. I mostly use ChatGPT for well defined functions that I need to implement that don’t rely on too much external context. I love it for writing unittests. It frequently makes mistakes, but it’s easier than starting from zero.

> ChatGPT doesn't do much for experienced SWEs, but it demolishes the difficulty curve for newbies. I wish I had this when I was learning.

Yes and no. It doesn't help me that much with technologies that I'm senior within, but it has allowed me to work with a wider tech-stack. For example, I can now confidently write any SQL querys I want (so far anyway) although my prior knowledge in SQL isn't that deep.

If you are an experienced SWE and you aren’t getting something out of it, you are holding it wrong.

Reply on news.ycombinator.com