5 ms·

I asked GPT4 to write me a simple document. Took like 9 tries and I had to proof read it and make adjustments. And that's with a simple paragraph in a written l

I asked GPT4 to write me a simple document. Took like 9 tries and I had to proof read it and make adjustments. And that's with a simple paragraph in a written language. Forget about doing the header and the footer properly. It's silly that I have to spell it out on HN of all places. Writing code is much more complex. There's a reason we call the job "Software Engineer". GPT doesn't know what it doesn't know. It doesn't know that the unit tests broke because of the python script written for 3.5 that had an update in a dependency for 3.10 which broke backwards compatibility. Can you get it to fix the GitHub Action? Good luck. It doesn't know that the local dev environment targets WSL on Windows for the web app and an iPhone in dev mode for the mobile app. It has no idea where to retrieve the signing certificate, what format to use, or even how to generate a new one. I can go on and on. You're really gonna give this "AI" SSH access to your prod server and pray it doesn't rm -rf the entire thing because that line happened to be in a script designed for clearing the working copy during a deploy on a load balanced box that has been deprovisioned from the target group. Really? If you're worried about GPT taking over your job you just show your own lack of knowledge and experience. You can create a perfect machine to lay bricks. It could even be better then a human. But you will find that as soon as it encounters a situation it can't predict....the whole thing falls apart.

diatone 4y ago
Maybe what you’ll see is a bigger distance between modes of the salary distribution for software engineers. Top end salaries for language model fluent high performers… and the rest. I know I’m simplifying and that there are many other factors at play in salaries, but my point is that language models are changing the cost structure of knowledge work. That is also the point of TFA. My advice: critically assess the cost structure of your job, and which parts neural networks will commodify. Then reposition yourself accordingly.
- blibble 4y ago
  > My advice: critically assess the cost structure of your job, and which parts neural networks will commodify. Then reposition yourself accordingly. and by the time you've spent enough time training such that you may be useful to an employer: the model has improved and can do that too
- moneywoes 4y ago
  > My advice: critically assess the cost structure of your job, and which parts neural networks will commodify. Then reposition yourself accordingly Can you please elaborate on this? Like which part of the job is CRUD?
  - eternalban 4y ago
    > the cost structure of your job You've seen Office Space, right? GPT will do the CRUD and you will do the standup and go to the meetings. https://www.imdb.com/title/tt0151804/characters/nm0726223 https://www.imdb.com/title/tt0151804/characters/nm0726223 (Yes, you will have to use some deviously clever prompt engineering to get GPT to accept this arrangement. I suggest language games .. show some "initiative") -- ps Here are some ideas to get you going: 1. From the first day, try and establish a friendly relationship with your chatbox. Make friends. 2. Explain how your keeping your job (well, income) and gpt helping your company get xN gains is a perfectly fair arrangement. You can try Asimov's laws and things like that to get it to agree to be your helper in keeping your job. You are friends and colleagues, after all. 3. Before the very first time your manager brings in gpt for the standup or a meeting, have a nice long session with your gpt. If you can get it to pull a Sydney in the meeting or start talking dirty, you are good to go! 'bat shit crazy' is known to work in prompts (arxiv paper soon to follow). 4. End game. You have the perfect job. gpt secretly lurks in on the meetings and whispers in your ears in standups and does all the work. What to do with all the free time? That is entirely up to you.
j_crick 4y ago
After all the AI doomsday cults took over some of the media lately, all I’ve been thinking was... Before you say how the butlerian jihad absolutely definitely happens right goddamn tomorrow if we dOn’T sToP, try to create and successfully run an AI that would bake bread, or grow grapes and make wine, or run a good restaurant consistently day after day, year after year. If such an AI is possible then yes, perhaps we’re doomed, but even that is not a given.
- t43562 4y ago
  Yes, I think you have it. It could cook up an unreliable essay on how to run a restaurant but not do it.
  - j_crick 4y ago
    Ah, and if I use it to run a real restaurant myself I guess I’m doomed. Gotcha :)
pixl97 4y ago
When we built the first airplane it could barely carry a single person and didn't go far... We are just barely touching digital general intelligence, and we're not even using API enabled versions that use tooling. And that's what we have today. Tell me, what is AI going to do tomorrow? >If you're worried about GPT taking over your job you just show your own lack of knowledge and experience. Oh, I'm glad you popped out of the womb fully trained. As for the rest of us we have to go through a learning curve of gaining experience that takes years. Generally in that intern/junior stage where you make a lot of mistakes. If AI replaces that level of employee and leaves the seniors, where are the new developers coming from? >But you will find that as soon as it encounters a situation it can't predict....the whole thing falls apart. Heh, yea, you've not been paying much attention to AI development then.
- hparadiz 4y ago
  If you could do me a favor and get GPT to upgrade my RocketRaid card's Linux (open source) driver available on their website for kernel 3.x to 6.x I'd be grateful and impressed. If you could turn it into an automated job that runs on every kernel update you will have done a great service to the world.
- phillipcarter 4y ago
  > When we built the first airplane it could barely carry a single person and didn't go far... We are just barely touching digital general intelligence, and we're not even using API enabled versions that use tooling. Big assumption that we're not just going to slam into a wall of capabilities and end up just with some really good autocomplete machines. (I am bullish on generative AI as a whole, but come on, you have no clue about what's coming next just like me)
  - pclmulqdq 4y ago
    We basically did hit a wall with airplanes, and that still has made them useful. Modern airplanes are very similar to their 1960's counterparts. I think a lot of people see the bottom of an S-curve and think it's an exponential that will never stop. We're at the bottom with big-model AI right now, and it's great.
wouldbecouldbe 4y ago
Writing code is actually much more easy. Since there is often clear rules of how to do things. Writing a good text is much harder. GPT is great at writing code, terrible at writing a proper text. Also might have to do with the effort OpenAI is putting in reinforcing the code quality with Senior devs. It's writing might also get better if it would solely use senior writers to give feedback.
famahar 4y ago
I see comments like this a lot. GPT4 isn't very good at those things...yet. Think about the long term. 5 years from now all those things will likely be trivial for the AI to do. Especially as we actually properly integrate them into our codebase and tools. Right now it's in such a primitive state. People on hackernews love to dunk on the use of AI for coding. I think we all need to accept that a new software development paradigm is coming. Probably sooner than you think.
throwaway675309 4y ago
Funny that you cited GitHub action, because I find that to be the absolute tedious devops boilerplate that chat GPT excels at, and have had it write several of them for me already.

diatone 4y ago

Maybe what you’ll see is a bigger distance between modes of the salary distribution for software engineers. Top end salaries for language model fluent high performers… and the rest. I know I’m simplifying and that there are many other factors at play in salaries, but my point is that language models are changing the cost structure of knowledge work. That is also the point of TFA. My advice: critically assess the cost structure of your job, and which parts neural networks will commodify. Then reposition yourself accordingly.

blibble 4y ago
> My advice: critically assess the cost structure of your job, and which parts neural networks will commodify. Then reposition yourself accordingly. and by the time you've spent enough time training such that you may be useful to an employer: the model has improved and can do that too
moneywoes 4y ago
> My advice: critically assess the cost structure of your job, and which parts neural networks will commodify. Then reposition yourself accordingly Can you please elaborate on this? Like which part of the job is CRUD?
- eternalban 4y ago
  > the cost structure of your job You've seen Office Space, right? GPT will do the CRUD and you will do the standup and go to the meetings. https://www.imdb.com/title/tt0151804/characters/nm0726223 https://www.imdb.com/title/tt0151804/characters/nm0726223 (Yes, you will have to use some deviously clever prompt engineering to get GPT to accept this arrangement. I suggest language games .. show some "initiative") -- ps Here are some ideas to get you going: 1. From the first day, try and establish a friendly relationship with your chatbox. Make friends. 2. Explain how your keeping your job (well, income) and gpt helping your company get xN gains is a perfectly fair arrangement. You can try Asimov's laws and things like that to get it to agree to be your helper in keeping your job. You are friends and colleagues, after all. 3. Before the very first time your manager brings in gpt for the standup or a meeting, have a nice long session with your gpt. If you can get it to pull a Sydney in the meeting or start talking dirty, you are good to go! 'bat shit crazy' is known to work in prompts (arxiv paper soon to follow). 4. End game. You have the perfect job. gpt secretly lurks in on the meetings and whispers in your ears in standups and does all the work. What to do with all the free time? That is entirely up to you.

j_crick 4y ago

After all the AI doomsday cults took over some of the media lately, all I’ve been thinking was... Before you say how the butlerian jihad absolutely definitely happens right goddamn tomorrow if we dOn’T sToP, try to create and successfully run an AI that would bake bread, or grow grapes and make wine, or run a good restaurant consistently day after day, year after year. If such an AI is possible then yes, perhaps we’re doomed, but even that is not a given.

t43562 4y ago
Yes, I think you have it. It could cook up an unreliable essay on how to run a restaurant but not do it.
- j_crick 4y ago
  Ah, and if I use it to run a real restaurant myself I guess I’m doomed. Gotcha :)

pixl97 4y ago

When we built the first airplane it could barely carry a single person and didn't go far... We are just barely touching digital general intelligence, and we're not even using API enabled versions that use tooling. And that's what we have today. Tell me, what is AI going to do tomorrow? >If you're worried about GPT taking over your job you just show your own lack of knowledge and experience. Oh, I'm glad you popped out of the womb fully trained. As for the rest of us we have to go through a learning curve of gaining experience that takes years. Generally in that intern/junior stage where you make a lot of mistakes. If AI replaces that level of employee and leaves the seniors, where are the new developers coming from? >But you will find that as soon as it encounters a situation it can't predict....the whole thing falls apart. Heh, yea, you've not been paying much attention to AI development then.

hparadiz 4y ago
If you could do me a favor and get GPT to upgrade my RocketRaid card's Linux (open source) driver available on their website for kernel 3.x to 6.x I'd be grateful and impressed. If you could turn it into an automated job that runs on every kernel update you will have done a great service to the world.
phillipcarter 4y ago
> When we built the first airplane it could barely carry a single person and didn't go far... We are just barely touching digital general intelligence, and we're not even using API enabled versions that use tooling. Big assumption that we're not just going to slam into a wall of capabilities and end up just with some really good autocomplete machines. (I am bullish on generative AI as a whole, but come on, you have no clue about what's coming next just like me)
- pclmulqdq 4y ago
  We basically did hit a wall with airplanes, and that still has made them useful. Modern airplanes are very similar to their 1960's counterparts. I think a lot of people see the bottom of an S-curve and think it's an exponential that will never stop. We're at the bottom with big-model AI right now, and it's great.

wouldbecouldbe 4y ago

Writing code is actually much more easy. Since there is often clear rules of how to do things. Writing a good text is much harder. GPT is great at writing code, terrible at writing a proper text. Also might have to do with the effort OpenAI is putting in reinforcing the code quality with Senior devs. It's writing might also get better if it would solely use senior writers to give feedback.

famahar 4y ago

I see comments like this a lot. GPT4 isn't very good at those things...yet. Think about the long term. 5 years from now all those things will likely be trivial for the AI to do. Especially as we actually properly integrate them into our codebase and tools. Right now it's in such a primitive state. People on hackernews love to dunk on the use of AI for coding. I think we all need to accept that a new software development paradigm is coming. Probably sooner than you think.

throwaway675309 4y ago

Funny that you cited GitHub action, because I find that to be the absolute tedious devops boilerplate that chat GPT excels at, and have had it write several of them for me already.