On Learning to Program with LLMs
se-radio.net
se-radio.net
I love that I could compile my program and try it without needing to check it with a teacher or superior. It cost literally nothing to try, and if it failed the computer would tell you it was wrong.
In a sense you are constantly bouncing off of reality (or the limits of the compiler) and that allows you to learn really quickly by experiment. LLMs with their tendency to hallucinate or just provide you with the solution seem like they could be both a blessing and a curse (for learning).
LLMs could be very useful for learning and probably are, but they also allow you to ask the wrong questions and circumvent the learning you were supposed to be doing.
I wouldn't be surprised to see people using it to simply solve exercises for them, and those people will have a hard time trying to do things on their own.
Recently I was trying to manipulate data in Google Sheets and was using ChatGPT to help. In the beginning is was fantastic, I was very productive because I didn't have to stop and think about formulas, read crappy documentation, or analyze data transformation. ChatGPT just gave me the right answer in a split second, as long as I kept asking the right questions.
Then I stumbled upon a particular issue that wasn't really too complicated but ChatGPT could not give me a correct answer. Unknowingly I spent 3 days trying to fix problems with the solution and every time I got an answer that was slightly wrong, not in subtle ways.
I have 20 years experience as a software engineer and still, I continued to waste time in this loop. After 3 days I decided to apply my engineering skills and solved the matter in 30 min. Now I know the solution and it was way simpler than I thought.
What surprised me was how dumb the whole process was. My questions certainly weren't the problem - as bad as they possibly were, the solution wasn't too far off. Not only ChatGPT had become a crutch but it put me in a situation that no human-tutor would ever put me.
So removing the pain of quick interactions with a tutor has benefits but the technology is not quite ready to be considered as true guidance in forming (or even helping) someone's understanding of a subject.
I've been using ChatGPT a lot for general language but when it requires logical thinking it falls pretty flat.
I've been using GPT4 as a tool to interrogate lesson transcripts for a language I'm learning and mention in the prompt to specifically focus on things mentioned in the transcript, if it's not in the transcript, check the helper script I update as I move on through the process (which does sadly take up more and more context window) and figure out if my answer is in one of those previous lessons, and to not guess. Hallucinations are quite rare, I don't think I can name an egregious instance of it in the 25 lesson I've done of approximately 20 minutes in length each, though I'm sure it's happened.
It's also pretty good at suggesting drills based on the contents of the lesson, there are probably a whole bunch of lesson plans in the training data.
The end result has been progress at a pace I could only dream of previously, and it doesn't matter if a question is too basic because I'm asking a computer. There is zero concern of any question being embarrassing because it's only between me, GPT4, and the OAI engineer who happens across the conversaiton.
But when the task at hand involves logical thinking, that's when I believe the LLMs of today are still a very much work in progress.
So I'm skeptical of trying to use them as tutors for now. I'm sure things will evolve quite quickly from now.
"I spent days trying to solve X by doing Y, then it turned out I could have solved it by doing Z instead" is an experience I've had countless times before LLMs were a thing. Sometimes you really do need these three days of stumbling before you can build up the confidence to do the easy solution.
(Then again, I don't know the specifics of your case.)
I have no evidence other than hundreds of hours using ChatGPT.
Well, yeah, but the central insight here is that LLMs enable a worse-is-better approach with a tighter feedback loop. They're not as good as a regular tutor, but a regular tutor costs money, gets impatient, is only available at set hours, etc.
Part of the appeal of learning-by-LLM is that you can get a flash of motivation at 2 AM and go "hey, I should totally learn about X, that would be cool!", open up ChatGPT, ask some very naive questions, and get just enough to get started.
This is a more recent version of the argument against most machines and similar automation tools. There is usually some validity to it, but we saw the same arguments against MOOCs, calculators, computers, type writers, smart phones, spell checkers, etc.
Like the commenter for this thread was saying, there's no fear of asking a stupid question. Or going to slowly or too quickly.
If only more people would focus on cultivating learning instead of locking down tech like this because they're afraid (and perhaps lazy to create new teaching methods).
For learning to code I suspect that with a strong enough version of GPT a new user won't hit the limits, and if it goes wrong having to debug isn't a bad thing ;)
I wish more interfaces supported intake of PDFs/URLs/HTML content as domain material. Then at least we'd be able to point the bot at the most current docs and have it most-heavily weight that information.
On the other hand: that's happening today already en masse, so I am not sure if it can actually get worse.
I agree, that LLMs as an addition to the established resources (teachers, books, youtubers, etc.) will have lots of value.
If you believe the future of coding sits in AI/LLMs/ML, this next generation of learners might just be using the 'beta' tools-of-tomorrow, rather than just being people that need to learn to manually code.
(My personal view is that manual coding will be changed almost unrecognisably within 5-10 years, but lets see! So much of learning to program is about learning syntax, and I can see a world of AI-powered-future-programming-languages where syntax is less important than clarity)
Meaning that there won't be coders that sit around punching code all day, but rather just generate diagrams and/or feed models with descriptions of systems, and in return get fully functional systems.
5 years? 10 years? 15 years? 20?
You: Here is a bug: [explanation]
GPT: Ok, fixed it!
You: It is still there.
Gpt: Ok, now fixed for real!
You: It is still there...
There's a whole field emerging of prompt engineers, but I think what's desperately missing is good quality integrated UX engineering.
Being able to write a comment and/or a method signature and having an LLM spit out a generated piece of code is kinda cool, but the interface for showing that the code is buggy needs refining. I think that's likely to be quickly showing a test case: given these inputs I'm expecting this output. LLM refactors against all tests, and you iterate together towards the correct solution.
That's one small example of how we can start to think about the future UX that isn't about having a "chat" about your code.
Similar UX thought is needed in other disciplines, it'll be interesting to see how this domain evolves.
It's even more effective when you can limit the domain somewhat to things that can fit in the context window, which eliminates out of date API information.
We will now shortly be able to add visual feedback of the applications into the loop.
Frankly I feel like we are looking at this the wrong way. In the future we might not even have a way to program some things manually. There might be a llm for each device that generates necessary hardware instructions on the fly like today's jit based on natural language or environment. May be there would be a inter llm spec/language that would felicitate cross llm functionality. All short circuiting out the programmers between the user and the utility.
The phase-2 LLM converts the designs to code in either react/flutter (both of which current LLMs are already quite good at) and deploy for a certain set of users. Basically same flow as today minus the engineers. Most of this is even possible today. The only thing saving engineering necks is the context window size which prevents fitting the whole repo in memory. Hopefully we (the programmers) will build our savings to a sustainable level in the next 5 years before signing off on manual coding or finding another thing to do than typing on keyboards to make computers dance.
This is basically science fiction
> This is basically science fiction
Product owners describing requirements in detail! More like fantasy, I think.
If you're learning to code now, using an LLM, who's to say that the LLMs won't improve faster than you do? In 10 years will there be a recognisable programming job for most people learning programming now? We're standing on a cliff watching the tide rise, wondering if it'll reach us. People learning programming are down in the sea, trying to climb up out of it.
Years ago taxi drivers actually needed to know their way around their local area. Then we got GPS and mapping apps, meaning Uber could exist and any random dude could be a taxi driver instantly. In future* the drivers will be unnecessary and the cars will be the drivers too. [*timeline 5-50 years]
In the early 2000s there were plenty of jobs for mediocre developers making websites for small companies. Now we've got Facebook, Squarespace & Shopify, that role is gone.
Lots of observations around here recently about how kids are less tech-savvy now, they know how to use appliances - not general purpose computers. Make stuff easy enough that anyone can do it, and nobody will know how it works.
Up till now software was mostly an arms race between competitors because of software’s low marginal cost and high utility. Now with ZIRP over and many of parts of the industry consolidating, I fear enshittification is a sign that the good times are over and we will now be milked to make up for all the profit lost the industrialists lost to the plebs.
That's the fundamental reason why there was such a falling out between Stack Overflow executives and the moderators.
People will say it was just the way they handled the interactions, which was part of it, but I don't believe there was any way to do it in a way that the users wouldn't hate.
Because the basic task for Stack Overflow IS to replace the users answering the questions with AI. They don't have a choice. Because now they are already competing with ChatGPT etc. doing exactly that.
It's true that taking away human answers from Stack Overflow will make it impossible for AI to answer many of those types of questions. But it will still retain fundamental knowledge and skills from other sources. And can be trained on its own debugging experiences as it gets deployed for building software. So although that presents a challenge, it's not insurmountable at all.
By going in that direction, I feel like they put the nail in the coffin. Why would I use their website when I can just ask ChatGPT directly?
I think they had one other option, which was to identify the fundamental limitations of LLMs, and find a way to demonstrate value over and above.
I don't believe there is any such fundamental limitation, especially when you look say two years down the road and don't limit yourself to language-only models or the AI models of today. Already GPT-4 has audio and image input.
Good human programmers today have better abilities and more current knowledge overall than GPT-4. But it's a matter of degree, and a matter of time.
What types of limitations do you suppose? And again, we have to anticipate that these technologies evolve rapidly. LLM is a misleading term in my opinion for a fundamentally multimodal model. And totally new types of models or model integrations will be likely be coming out over the next few years that will make current systems obsolete. Such as combining AlphaZero-type systems with language models.
I don't see the capacity for LLMs to perform novel and creative reasoning tasks. They're limited by virtue of being a pattern-matching system without the ability to make genuine causal or logical inferences.
I wrote here about my recent experiments with ChatGPT - https://news.ycombinator.com/item?id=37583333
I'm vaugely interested to know how GPT4 deals with questions like these, but I'm not willing to pay to find out.
Anyway, it will only be a year or two before things like tree search over problems spaces are explicitly integrated into models or systems similar to this. Deep Mind may have already done something like that.
That really hasn't been my experience with StackOverflow for the last 10 years.
If an inexperienced engineer asks for code and the LLM produces something that has sub-par performance, is overly complex, and botches various edge-cases, the engineer may not realise until years later if at all that they could have produced something much better.
Combined with the fact that there are plenty of things that mostly only junior engineers will ask about, and I think you have a problem.
Code generation AI systems will continue to get better rapidly.
They are not just language models anymore but now can include image input and output.
There will be systems trained (or context "trained") for software platforms that put a wide variety of task configurations within easy reach by translating to relatively small amounts of DSL code.
The next revolution may come from applying things like Monte Carlo search to concrete domains defined on the fly by LLMs. Or just a tight integration between implementation and testing within the platform and AI loop.
Also consider the programming ability increases when models are trained to be able to unroll state very accurately within their platform domains as part of the software design process.
Looking anywhere from a few months out to say five years, we should anticipate increasingly capable and more general AI programming systems.
Who knows how long it will take for these systems to be fully deployed in all areas. But beyond five years or so, it is kind of hard to imagine what the leading-edge capabilities of these types of systems might be.
Maybe no one will want to buy a particular computer game or application anymore. Instead, you buy an AI programmer that comes with a kit for configuring and even coding some custom parts of software specifically tailored for you.
Maybe the only thing that really sells at some point are AI asset streaming services which allow your existing AI swarms to on demand pull in new neural radiance models, behavior models, neural world state models, neural avatar models, domain experts, document drafters, etc.
These will be integrated with mixed reality. So you are looking at basically portable Holodeck providers. The commodification of all types of skills and knowledge, manifested in person on demand.
An LLM is never going to be able to figure out what you need written, and bug-free code is a big ask, putting it mildly. Even structurally correct code is a bug if it doesn't actually do what you want. Someone still has to understand the problem domain, design an architecture to solve it, and verify that it does what it's supposed to do, this is a professional position, no way around it.
Where we are now, Copilot and friends are a nice assist for drudge work, a way to learn new libraries/APIs/languages, but a danger to beginners (you have to be able to proof the code it produces, it will still frequently make straightforward errors) and for experts it's less useful. As they improve I expect them to become an essential part of the profession, but they aren't actually able to replace it.
We've heard variations on this argument for decades and I see no difference why LLMs, despite their immense capabilities, will be massively different. They still take guidance and produce garbage output which requires knowledge and expertise to turn into a cohesive product. Will they replace some script writing, or some business logic translation? Of course. But so did SQL, no-code, and plain-English testing frameworks. LLMs are an incredibly powerful tool but they're not a silver bullet.