Hacker News Reader

Comment by jeppester | Hacker News Reader

This very well sums up the anxiety I have about how this is going to affect programming.

That any creativity and understanding will be sucked out and replaced with prompt engineering.

I still think you will have to be a good developer to understand if the solution you get is good, but is that even going to matter when you can just ask again and try again? I think there will be lots of code bases that no-one really understands, but that it won't matter much.

All of this is powered by work done by real programmers who are not getting a dime for their effort. That just seems completely unfair to me, and even more so when already insanely rich corporations like MS only stand to gain from it.

I'm not sure I want to be part of it to be honest. It feels dirty and wrong as it works now. Laws might be able to change that, but with a whole industry of giant corporations ready to fight any such laws, I'm not optimistic.

View on HN

Even prompt engineering is just a short term problem, that'll get solved and largely disappear within a couple of years, if not months. Making sense of natural language is after all what LLM are already really good at, they just haven't made their way into image generators yet.

Another big issue that gets overlooked is that a lot of the problems that programming is used to solve today will disappear and no longer be considered problems to begin with. A whole lot of data conversion is no longer necessary when you have an AI that can directly interface with the raw data and manipulate it in whatever format you like.

People today are thinking of ChatGPT as making it easier to build websites, software and such, but real the power of AI is that it makes all of that obsolete. Why bother browsing the Web when ChatGPT can just give you what you are looking for in whatever format you desire, without ever touching a website?

AI is not just going to write software, AI is going to be the only software you'll ever need. BingChat today is already Photoshop, Shell and Webbrowser all in one. Won't be long until it learns to do most of the rest as well.

The interesting question is: What will be left after all this? If we automate away all the boring parts, what are going to be the bits that provide value? How will the Website of the future look like when AI will be the only one reading them? What's the value of a classic book when AI can write new books in real time? Will anybody bother with video games when they have a Holodeck at home?

> People today are thinking of ChatGPT as making it easier to build websites, software and such, but real the power of AI is that it makes all of that obsolete. Why bother browsing the Web when ChatGPT can just give you what you are looking for in whatever format you desire, without ever touching a website?

I don‘t think it will play out like this. Companies currently build systems to make the data robust/structured, to enforce a business process etc. Since they do not believe in answers any individual would give out of the unstructured data, they would also not trust an AI, and rightfully so. ChatGPT is not trustworthy at all and there is currently no path to make it so.

I found that companies are less interested in quality data than one might expect.

I worked in ontologies, focusing on making rigorous schemas for better data quality and easier integration. Our lunch was eaten by schemaless databases and big data. We protested that they wouldn't like what happened years down the line, but it was cheaper to start.

I don't mean to shake my tiny fist or overgeneralize my experience. But I think if AI gives good enough answers enough of the time, with a tiny fraction of the effort, people might just adapt around that model. People would rather have good now than great later. Or even mediocre now.

There is nothing stopping companies from offering AI oriented apis. If they no longer have to develop a frontend but just have a "here is our logo" etc endpoint, then... goodbye react! Goodbye FE developers! And at that point AI will probably be developing the very endpoints that it consumes.

Honestly it doesn't sound crazy to me.

For frontend - yeah, maybe. Or maybe the frontend becomes much more fancy. I‘ve built complex frontends that took a year or two but could imagine it all in weeks. I would also like to be able to build it in weeks.

I think that’s a simplistic view. Sure thing AI will make a bunch of problems disappear… but do you really think that with all the power of AI, we are going to keep publishing react apps? Golang web servers? Rest APIs? No. We’ll be publishing more sophisticated stuff. And no, AI simply cannot magically do everything (AI today is not the AI we will have in 100 years… unless you think the AI of today is the most we can get from AI)… so in between we (humans) will be working to use AI tooling to fulfill user requirements.

A lot of what we do as programmers, writers or artists IS 'prompt engineering'. All that AI is replacing (generalizing? automating?) is execution. Prompt engineering is the 'what' of the 'how'.

"They should have sent a poet"- not NASA, but a screenplay written by some sort of poet.

If there's a singularity it will be in this: you speak of real programmers doing the work to make this AI, but it seems like a hell of a lot of this 'real programmer work' is not poetry. Someone put together the ideas that make code resonate with how people think (for good or ill) and then a lot of real programmers groveled over tedious Python code or whatever, to make it happen at scale.

That tedious groveling and scale is what took the effort, but it's that which can be replaced by machine learning.

Art is prompt engineering. 'Always has been'. We don't master penmanship in the age of the keyboard. People still buy very expensive pens and write in longhand for the beauty of it, but it ceased to be a requisite chore many years ago.

Value the correct side of the balance. If you can't beat the guy with poorer execution by having a better eye and better prompt-imagining, you're not the better artist, you just have better penmanship.

Same for programmers. Best put your effort into trying to understand… because that's going to be a seller's market. You're right that most people won't keep up with understanding, and that's going to directly set their value.

Good points. Understanding what and how it really works will be valueble.

Explaining the AI needs the knowledge of what can be done. For example, you cannot say AI "build systems that process the data". You need to explain AI what "kind of processing should be done". That by itself requires knowledge of what kind of processing exist and when to use which. Also, explaining to AI what should be done takes time and effort which should not be underestimated. Sometimes when I am exlaining someone which way data should be cleaned in many cases I think that it maybe faster to do it by myself and explainig to AI will most likly be a harder than to AI. So, the question about is it easier to implement by myself than explain what I need to AI is a valid and important question.

Also, with data important part even without AI is to verify how well system works. Does it correctly handles differents types of data and different use cases? Just having a system that processes data but you are not sure what output can be is in most cases useleless data system.

So, in any serious data systems understanding of what they can be done and how code and data systems are working are important and I think will become even more important than now.

Also, someone needs to creat AI and systems around them.

It will probably end up being more "test driven development"-like.

Write your tests/expectations and the tooling automatically prompts/writes the code to meet the tests.

That sounds pretty miserable, because test writing is one of the least fun aspects of my job

My experience with Copilot is the other way around. I spend time writing my types/data structures, my functions, then Copilot tends to easily spit several variations of the test by just me giving it the name of the test.

The LLM will write the tests too. Your job would be guiding it.

why not have a system to write the tests as well?

Won't the prompt itself be the test? TDD is literally just providing an expected output on a given input, before implementing the functionality. The prompt we give AIs follows the same pattern, "I have a data that looks like ___, visualize it in a graph that ____"

I agree with the loss of understanding being a bad thing, but I cannot mourn the loss of creativity.

Programming is a core part of our society and infrastructure, yet it's still one like a craft rather than engineering. Creativity should not be part of infrastructure efforts on a level beyond embellishments and experiments.

> Creativity should not be part of infrastructure efforts on a level beyond embellishments and experiments.

You could say the same about architecture, but only modern architects think that way, because they were brainwashed by fashion.

I do say the same about architecture, but with added nuance: the Sagrada Familia is creative in form, but not in function. It functions as a regular cathedral (or at least an art museum). The creativity is there, but it is subservient to the basic functional goals of structural integrity and fitness for humans to visit it.

Buckminster Fuller is creativity. Geodesic domes, tensegrity.

Failing to understand that is evidence you can be replaced by an AI that encompasses all of what's ALREADY known and understood.

The reality is, we're looking at an arms race between imagination and execution. And just implementing that which we already know is profoundly devalued in the age of machine-assisted intelligence.

People thought the same about OOP when it was new. The promise was that programming would turn more into factory work. You’d get a spec for a component, implement it, then move onto the next. How soul crushing!

> prompt engineering

Sounds like coding to me.

Not from my current experience.

I'm intending to try it out a lot more — perhaps turn my old Java/OpenGL desktop shareware games into browser based JavaScript freeware — but so far, it's like messaging a junior dev and having them write code for you: 98% of the code works fine, 2% is total rubbish that you spend a normal amount of time (i.e. ages) debugging.

> I still think you will have to be a good developer to understand if the solution you get is good, but is that even going to matter when you can just ask again and try again? I think there will be lots of code bases that no-one really understands, but that it won't matter much.

Really? If you have a million dollar bug because of one edge case and you are going to go to your AI and ask it to try again?

I think that's going to be a very likely scenario. And the AI will very likely be able to fix the issue.

I just don't want to work that way, as what I enjoy most about programming is understanding things.

Then you are in the wrong space of programming if you want to continue working. AI will improve over time and the scenario I described is going to be a thing of the past, but it won't go as fast as people expect because it is a question of trust and saving that hypotetical money by actually doing the work or at least understanding the output of AI.

I've experimented with both Copilot and GPT and I've gotten some nasty UB and even known security issues. Just looking at some of the GPT output lacking exceptions and errors checking that is clearly trained on tutorials. It will improve over time but currently it is not as good as people say it is.

I've had experience lately of getting GPT4 to output some code, I notice there's a bug, I point out the bug (not even 'it's at this line with this variable', but just that it exists) and, yes, it does fix successfully fix it. So, yes, I think that's likely.

Well that required you to understand the code. My reply is disbelief off codebases that no one understands in the near term when looking at the current output. People don't trust the output and do what you did, review it and catch the potential bug before going further with it.

The issue take with it is that we are able to trust the output of AI models as we trust the output of optimizing compilers. It isn't even close that we can just fire and forget the prompt like I can code up something and trust that it 99,999% the optimizer produces the desired binary that works as intended.

Reply on news.ycombinator.com