But it really feels like I didn't improve my skills by doing it.
The process of finding the right prompt to get the most out of the LLM is programming. You aren't doing anything different, you just switched languages.
The process of finding the right prompt to get the most out of the LLM is programming. You aren't doing anything different, you just switched languages.
The process is clearly shortened. I don't have to "go through the pain" of producing the code: it just pops up for free, and then I just have to tune it. It's the difference between writing a speech and proof-reading it. Or listening to a lot of music versus actually learning how to play an instrument. Because you can be critical about music does not mean at all that you can play an instrument.
I.e. because it was wrong to say X in the past doesn't mean it's wrong to say Y today, even if you find some similarities. Doesn't mean it's true either.
So let's ignore such worthless arguments and have a real discussion, shall we?
I'm the author of a package used for data acquisition from various bits and pieces of instrumentation hardware. It is used in some professional settings, but because it's a random collection of hacks, it's mostly useful to hobbyists with more free time than money. Retirees who are getting back into their old ham radio hobby, for example, like the guy I just (failed to) help. He was asking about how to adapt one of the larger programs in the package to a completely-different test instrument.
Although he wasn't a programmer, and although my code is in C++, he had "heard good things about Python" and was wondering if he should install it on his Windows box and give it a try. I didn't give him the full LART treatment, but I was probably a bit more patronizing than I should have been. I said something like, "That's not even the wrong way to do it. Try feeding the .cpp file to ChatGPT and see if it will help you understand what you need to do to modify it."
That was yesterday. This morning, I got an entirely undeserved thank-you note from him. He got a Python program back from ChatGPT, and after a few back-and-forth interactions, it actually worked, using code he couldn't read to communicate with an instrument I'd never heard of.
He wasn't a programmer yesterday, but today he is one. His first language is not assembly or C or BASIC or Python, but English.
This technology doesn't threaten people, it empowers them. It doesn't harm creativity, it enables it. I don't know about "worthless arguments," but my recommendation is that you change your outlook right fucking now if you intend to stay in the software business, or in the music business for that matter. You'll eventually thank me for the advice, and this time I'll deserve the gratitude.
But I believe that it is a different situation from what I was describing. For someone who is already a professional, I think that AIs risk to prevent them from improving. The industry will push us towards using the AI and tuning the output because that is more productive now, even if it means that we don't improve our skills.
> This technology doesn't threaten people, it empowers them.
AI threatens many people in many ways. For instance by enabling phishing to a level never seen before. By being a big copyright-laundering machine in favour of BigTech. By teaching people to believe an eloquent, all-knowing chat interface because most of the time it seems good enough to not fact-check. Etc.
> It doesn't harm creativity, it enables it.
Do you know any professional artist? I do. I have an example in the animation movies business. You know what genAI brings to their job? They are pushed to have a genAI generate images, select those that are good enough, photoshop them a bit and move to the next.
This is the definition of "harming creativity". If AI makes you twice as productive, it's okay to lower the quality of what you produce, right? You'll just sell more of lower-quality stuff. That's a tendency that has existed in software in the last decade, but it feels like LLMs will bring that to a whole new level.
> I don't know about "worthless arguments"
My apologies. Re-reading it out of context, it sounds harsh. It's not an excuse, but I wrote it after answering a few other comments that all said "people said that for technology X and they were wrong, which proves that if you say it now for technology Y you are most likely wrong as well".
> hits close to home, because in this case I was arguably the asshole
I wouldn't call that being an asshole. Not helping someone for free because they are stuck using the stuff you provided for free is completely fine.
I have had many people over the years telling me how my project sucked, how it was making them lose money, how they wouldn't use my project unless I added X and Y. When being part of a larger community, I've had people pressure me by telling everybody and their cat, on public channels, that "well you shouldn't use this project because it seems unmaintained" (last commit was some weeks before, issues had been answered days before). Or complaining about the license because it makes it harder for them to make money out of my free work.
Many times people don't even realise that they are assholes. They think that if I published my work for free, they are the customer, and they deserve free support.
Users of open source projects make me want to keep my stuff proprietary. Not because I don't want them to benefit, but just because I want them to stop bullying me.
As for copyright, that's a matter for the courts and legislature to decide. If copyright maximalism wins, such that training is not considered fair use, or that liability for infringement attaches to the trainer of the model rather than its users, it will really do a number on American competitiveness. Such a decision would transfer an enormous amount of power to countries that DGAF about copyright law.
The most competitive AI models are always going to rely on snarfing up as much data as possible, legally or otherwise, and I'm OK with that. Copyright is a recent invention, historically speaking; we got along fine without it for thousands of years. What happening in AI, on the other hand, can potentially take us to the next level as a species. If that means the end of copyright as a concept, so be it. My big concern is that like most futures (to paraphrase Gibson), this one won't exactly be equally distributed.
I kindly disagree :-). Management is about increasing profit, not improving the product. And we see the results: arguably most products are worse today than they were 10 years ago. Software is a great example of that: the hardware made huge progress, but the resulting software is slightly worse than it was with hardware orders of magnitudes slower.
> As for copyright, that's a matter for the courts and legislature to decide.
Unfortunately, I think the US are showing us differently. The billionaires decide and again, they optimise their profit.
> What happening in AI, on the other hand, can potentially take us to the next level as a species.
All the established science says that we are about to collapse (most species are, already: we are the cause of the current mass extinction that is the fastest in the history of the planet). We don't have a solution to the energy problem, and we're running out of time. We don't have a solution to the climate problem, and we're running out of time. And we don't have a solution to the biodiversity, that is well into a mass extinction.
There is one thing we know we can do: start doing less with less, adapt our society to handle the coming changes as well as possible and accept that it will be worse than today anyway.
The direction we are taking, though, is what you suggest: hope for a miracle, hope for what we don't know. "Maybe technology will save us". Sure. Maybe god will, too. Feels unlikely to me.
C compilers emit assembly as a direct result of the code written.
The same compiler given the same code will output the same assembly. This would be true for LLMs if you set the temperature to 1 and fix the seed, but you lose the benefits of language models at that point.
There are guarantees about the behavior of a compiler which can be verified by examining the code for the compiler.
The same is not true for language models. They are statistically based all the way down.
But prompt engineering isn't programming because it doesn't allow aggregation. It's not the activity of building a system from reliable components. Prompts don't allow reliable recursion, etc. At best, it is simply a different but valid activity. But AI companies no doubt aim to remove the mystery from prompts in the fashion they aim to remove the mystery from programming.
Yea... That Wall-E ship is our future isn't it?
Here’s the most obvious reason why that isn’t the case:
There is a direct cause effect relationship with the code you write and the resulting program.
That is not so with prompting, obviously.
Prompting is, therefore, not the same as programming.