Hacker News Reader

Comment by xrd | Hacker News Reader

This stuff feels revolutionary, and I say that as someone that has watched the progress even the last few months and said to myself that it won't radically change the software industry. I think I was wrong.

Also, as they say, maintenance is where the true cost of software lies. Has to anyone asked ChatGPT to refactor existing horrible code? It's fascinating to think about what happens to refactoring existing code if the AI doesn't care if it gets laid off, and can refactor again (if the new code fails) within seconds rather than days.

And, what happens when rewrite from scratch takes seconds instead of months. Joel's law of never rewrite from scratch might completely fall away.

This feels really game changing. If anyone can come in and just pitch ideas to AI with zero barriers for cost analysis, it's going to open the doors for software to eat every tiny segment that has not already been devoured.

View on HN

Honestly I'm not that interested in the code part. It does well but it's not perfect, it's good for some stuff - just like with the copilot. For more concrete things like coding I like to have the control and not use a middleman.

What blows my mind is that I suddenly gain extra powers that I previously did not posses.

For example, let's say I'm developing an adventure game and I need to write content but writing is not my speciality. Fear not, I have access to Janet from the Good Place and it's called ChatGPT here on Earth.

So I'm writing a dialog for my character who is a gentlemen in England that came from old money.

I simply tell chatGPT: rewrite the sentence to sound more posh "I want to eat cookies but I am not hungry yet"

chatGPT: "I wish to partake in some biscuits, but I am not yet famished."

Okay, maybe its a bit overdone but I like it and can already imagine his moustache and the hat, I definitely couldn't have written that by myself and if "AI" think this is about right then it will probably sound right to many people because that "AI" is essentially a result of a study of other peoples work.

IMHO this AI stuff enables access to mastery and experience, not really replacing the thoughts of real people. It's the same with Dall-E or Stable Diffusion, you still need to be able to think ideas but the ability to paint like Van Gogh comes for free(or something like 0.01$).

Imagine applying it to the tech job search: unique resumes and cover letters tailor made for each job application.

Recruiters are going to use it too for personalized messages to every single candidate. I suspect future developments are going to make it increasingly difficult to filter out spam from your inbox

I think that most of the internet is going to get overrun by spam at this rate. The only way to avoid it will be to move into walled gardens with strict human-ness verification.

And it solves silly coding puzzles for you...

good idea!

> IMHO this AI stuff enables access to mastery and experience, not really replacing the thoughts of real people. It's the same with Dall-E or Stable Diffusion, you still need to be able to think ideas but the ability to paint like Van Gogh comes for free(or something like 0.01$).

No, it generates images of paintings ;)

Paintings are physical objects that exist in the real world. It doesn’t make those.

Lol you are faking responses though Doesn't sound genuine.

One of the most likely professions eaten by software, is software.

You are on point. I've been quietly mentioning this to software engineer friends, and poo pooed. I am more of the computer science-y than software engineer-y persuasion - and have seen this as a clear risk.

edit 1: originally I thought it will mainly be for new code bases (green field) - but with the ability for code to "git pull" and then "chatgpt refactor" may be a game changer.

edit 2: I think this is great - it can allow engineers to work on a higher level. I remember in CompSci training always thinking "why in the heck am I typing into a computer how to do its job?" and "why should engineers be dealing with stupid $hit like NaN and pointer overflows (C++) and heap issues an CRUD apps. It all seems like low brain stuff"

AI will eat software engineering the same way it'll eat art - from the bottom, and quickly at first then very very slowly. Engineers will be responsible for directing, verifying and correcting AIs at increasingly high levels over time. Hand written code will be limited to optimizations and working around weaknesses in the model much the same as inline assembly is still used with compiled code.

Fully agree with your assessment. Bottom up is vulnerable.

I can then see engineers running entire department of AI by themselves..... someone who is an architect level and understands the business and can both envision the services, but also fix black box systems creating content that no one understands under the covers.

This could be a very good time to create a startup that can make that command line and: 1. Document its functioning overall ("chatgpt doc") 2. Find bugs / logical flaws ("chatgpt debug") 3. Refactor the code for readability ("chatgpt refactor") 4. Improve its compute/memory performance ("chatgpt speed") 5. Improve the quality of comments ("chatgpt comment")

Or just "chatgpt fix" to do all the above ;)

Seems that the main limitation at this point would be the OpenAI Terms of Use. How much freedom is given to build commercial services on top of their API?

> Has to anyone asked ChatGPT to refactor existing horrible code?

Yes it works well with stuff like React class components to functional ones with hooks. It even tells you what to optimize and does it for you, if you ask it to.

Whoa. Do you have a good guide on how to do this?

you prompt with “rewrite this code, but use React Hooks”

I think it's too early to tell. It feels like early translators. It's indeed very impressive, but could be 5-10 years from actually having something useful and it's unclear what the practical impact will be.

But it's also true that I wouldn't know what career to recommend to someone now in their early 20s who has 30 years ahead. As I said too difficult to predict right now.

I don’t think it’s hard to predict at all.

The diffusion art models make it very clear that these technologies go from toys to replacing people’s jobs very very quickly.

Short of a (not impossible) breakdown in civilization, the current rate of change suggests that 5-10 years is probably unrealistic.

More like 2-4 years, it will be technically possible to be reasonably high level software engineering.

It’s already possible to chain high level “how would I do x?” Planning to “write code to do y…” code generation.

It’s very very obvious that technically it’s going to be possible to do a great deal of mechanical work like refactoring and adding trivial features automatically.

Companies do not need 50 engineers doing 100x productivity.

They need like, 5.

What remains to be seen is if governments and legal systems allow that scale of displacement of human labor to occur or not.

…buuut, I wouldn’t be holding my breath.

To my mind though this is only half the question. Let's assume that every company goes from 100 to 5 engineers - I could see that. But today, almost every company (aside from behemoths like FB, Google, etc) are engineer starved - they'd like to do so much more, but they don't have enough engineering manpower. One possibility is that, yes, engineers become much less in demand. But there's another possibility where the number of companies explodes because the amount of software that can be built is no longer bottlenecked on the number of engineers you can hire.

> The diffusion art models make it very clear that these technologies go from toys to replacing people’s jobs very very quickly.

Do you have any evidence that art jobs are being replaced now? I was talking with a designer friend of mine recently and he's not heard of Stable Diffusion, etc. replacing jobs yet. This is something that people keep bringing up but I haven't yet seen any evidence to back it up.

It will never replace anyone’s job, only tasks from it. A person who only ever did one task was always pretty underemployed.

Also, we’re getting this for free and don’t know what the real cost is compared to an employee.

There are plenty of intelligent animals and dextrous animals on the planet. The difference between them and humans is that humans can do both.

Nah, the differences are (at least) that we're future-oriented, motivated, social, have language, and live long enough to be grandparents. Take that, octopuses.

Also as far as the economy goes, we actually try to participate in it instead of sitting around getting eaten. Maybe we won't succeed in the future though…

The 'proko' channel on youtube has some discussions on this topic from people who know what they're talking about, rather just anecdotal evidence from a sample size of 1.

No, it has probably created more jobs at this point.

So, take-home code tests are dead?

Yes. I gave ChatGPT two favorite coding screens I'd used in the past. They're nothing fancy—ordinary, everyday coding tasks.

ChatGPT implemented the first program successfully, and explained how it worked. (It compulsively explains code, which is nice.)

ChatGPT's performance on the second program was more impressive. I explained what I wanted the program to do, and gave sample input and output data. ChatGPT wrote a clear, clever Python implementation with a single bug, and then explained that it would produce incorrect output. It showed what the incorrect output would be. The it explained how to fix the bug, and provided code for the fix.

In both cases, it used some clever, clean tricks I had never seen before, and explained how they worked.

It can absolutely pass coding screens better than half the people I saw submit resumes to one of my old jobs. Or you can ask it to answer essay questions about French literature, in French. It does that, too.

(I have managed to break it. Try giving it perfectly correct code and telling it find 2 bugs and explain them.)

> Has to anyone asked ChatGPT to refactor existing horrible code?

I'm not sure if "horrible" but I did experiment with getting it to refactor some code. I also gave it code with a bug and asked it to fix it (I described the bug), which it was able to do. I didn't test very complicated scenarios though. In the bug fix case, it was code to wrap, indent, and prefix a block of text.

Someone fed it legacy buggy code and it fixed the bugs.

Reply on news.ycombinator.com