My company has an ill-advised "try to do it with AI first" rule in place.
For senior devs, that means rejecting the first twenty AI-generated code suggestions as they're bollocks, and then having to fix it up by hand anyway. Takes 10x as long as it should do.
For junior devs, who don't know enough to reject the 'bollocks first suggestion', it means raising pull requests that take 10x as long for the seniors to review, since they have to untangle the original intention and then explain why all the code is terrible and why they need to go and fix it.
I'm in exactly the same boat as you, and it's driving me insane. I asked Cursor to generate a simple JavaScript function the other day, and it gave back a 200-line mess of garbage. I ended up just writing the function myself, and was able to do the same thing in 20 lines... and it's not like I was using some "clever" techniques that no future maintainer will ever understand, it was all just basic code.
AI is definitely generating 10x the amount of code... but not 10x the quality or productively.
Fuel economy is normally expressed as "distance per volume" (mpg) or "volume per distance" (l/100km). Normalise that, it has units of area. I like to visualise that as the cross-section of a string of fuel dragging a car along, being thicker when accelerating and thinner when running down hill.
Acre-feet is a weird unit for volume, but if you know how big your reservoir is in acres then at least it's an easy conversion to depth. And at least it's just a conversion factor away from a sane unit that can be used for pump sizing and such.
My experience exactly, from the last 3 companies ive worked for! And now there are no job advancement options internally because the openings have moved to India.
The company does have any money left after giving it all to AI grifters and letting their garbage code sink their business. Net negative on the jobs there
It's unbelievable how people overestimate the functions of LLMs. I know people who had to flunk their students because they had ChatGPT do their essays and ended spitting out wrong info.
Kernighan's Law states "Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it."
Since no thought went into writing the code, I can only believe that no thought is required to fix. Therefore my involvement is unnecessary. Have fun!
Professionally, I am unable to use AI due to my company not wanting to produce code that could be legally challenged.
In my spare time & projects I've tinkered with Github Co-pilot & Cursor.
GitHub co-pilot has been useful when I'm coding in a strongly typed and well documented language, however the benefit here is that it will suggest pretty much exactly what I wanted when producing suggestions for the rest of the line or the next 3-5 code lines. However, this is after I configured it to only offer suggestions for small chunks of code and the accept is a two key combo instead of overriding tab accept for standard autocomplete. Anything over that and it starts hallucinating like crazy. I've also found it useful for converting a script from one language to another (mostly bash to psh & bat, but a bit of python to JS.).
Cursor has been useful for creating a functional prototype for an idea on functional projects and it does an ok job at code tracing and explaining the design patterns used and where to inject behaviors.
That said, I stopped using cursor after my free trial ended, and the 1 year GitHub Co-pilot sub that was given to me runs up in November and I don't plan on renewing.
I produce far better code without the tools. The gains achieved with co-pilot were nice, but in the end I felt disconnected from my code and when problems arose, I spent more time remembering where to find it. And if I really wanted to understand the design patterns used in a code base there are tools out there to do that already with great visualization outputs.
As of late, I've mostly been using Jetbrains IDEs and their built-in tooling for that. VS also has tooling, but it's a bit dated on UX IMHO. I've also played with Emerge for data flow and data clustering visualization. There's also some in-house tools we use for memory allocation.
There's a bunch of other code to UML tooling and other diagraming tools out there for specialized cases.
25 Comments
SalamenceFury@lemmy.world · 74 pts · 330d
Oh no! The consequences of my own actions!
Pay me 10 times what you get paid and I'll fix it.
fullsquare@awful.systems · 34 pts · 330d
ah yes that's what 10x programmer must mean
addie@feddit.uk · 35 pts · 330d
My company has an ill-advised "try to do it with AI first" rule in place.
For senior devs, that means rejecting the first twenty AI-generated code suggestions as they're bollocks, and then having to fix it up by hand anyway. Takes 10x as long as it should do.
For junior devs, who don't know enough to reject the 'bollocks first suggestion', it means raising pull requests that take 10x as long for the seniors to review, since they have to untangle the original intention and then explain why all the code is terrible and why they need to go and fix it.
We have truly embraced 10x programming.
very_well_lost@lemmy.world · 10 pts · 330d
I'm in exactly the same boat as you, and it's driving me insane. I asked Cursor to generate a simple JavaScript function the other day, and it gave back a 200-line mess of garbage. I ended up just writing the function myself, and was able to do the same thing in 20 lines... and it's not like I was using some "clever" techniques that no future maintainer will ever understand, it was all just basic code.
AI is definitely generating 10x the amount of code... but not 10x the quality or productively.
bitjunkie@lemmy.world · 8 pts · 330d
Elegance is doing the same thing in fewer lines without sacrificing interpretability
JeeBaiChow@lemmy.world · 63 pts · 330d
Lol. Sure. What's your rate? Multiply that by 5 or so, and I'll consider.
Semi_Hemi_Demigod@lemmy.world · 43 pts · 330d
Bill them in kilowatt hours and acre feet of water
GissaMittJobb@lemmy.ml · 15 pts · 330d
Few units of measurements can claim to be as deranged as acre-feet. Truly horrifying
addie@feddit.uk · 10 pts · 330d
Fuel economy is normally expressed as "distance per volume" (mpg) or "volume per distance" (l/100km). Normalise that, it has units of area. I like to visualise that as the cross-section of a string of fuel dragging a car along, being thicker when accelerating and thinner when running down hill.
Acre-feet is a weird unit for volume, but if you know how big your reservoir is in acres then at least it's an easy conversion to depth. And at least it's just a conversion factor away from a sane unit that can be used for pump sizing and such.
Lorgres@lemmy.world · 3 pts · 330d
Another nice way to visualize the area of fuel economy is just the cross section of the fuel hose going to the engine.
puppinstuff@lemmy.ca · 52 pts · 330d
Nomecks@lemmy.ca · 17 pts · 330d
Like the messes cheapskates make by hiring a team of Indian programmers?
ShieldsUp@startrek.website · 7 pts · 330d
My experience exactly, from the last 3 companies ive worked for! And now there are no job advancement options internally because the openings have moved to India.
puppinstuff@lemmy.ca · 6 pts · 330d
Blackmist@feddit.uk · 35 pts · 330d
Real programmers can botch their own code.
FigMcLargeHuge@sh.itjust.works · 1 pts · 329d
If they are real programmers they also have the skills to debug and correct their own code.
psx_crab@lemmy.zip · 31 pts · 330d
And they say AI destroy job. It create more job apparently, because AI often generate trainwreck and real expert needed to rope in to fix the wreck.
_stranger_@lemmy.world · 12 pts · 330d
The company does have any money left after giving it all to AI grifters and letting their garbage code sink their business. Net negative on the jobs there
TanteRegenbogen@feddit.org · 29 pts · 330d
It's unbelievable how people overestimate the functions of LLMs. I know people who had to flunk their students because they had ChatGPT do their essays and ended spitting out wrong info.
AA5B@lemmy.world · 7 pts · 330d
I just now got finished updating an escalation ticket where someone had responded with an ai hallucination
groucho@lemmy.sdf.org · 29 pts · 330d
Kernighan's Law states "Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it."
Since no thought went into writing the code, I can only believe that no thought is required to fix. Therefore my involvement is unnecessary. Have fun!
xep@discuss.online · 21 pts · 330d
Nobody has any time for that, clean up your own mess.
bitjunkie@lemmy.world · 9 pts · 330d
HAHAHHAHAHAHA no
odelik@lemmy.today · 8 pts · 330d
Professionally, I am unable to use AI due to my company not wanting to produce code that could be legally challenged.
In my spare time & projects I've tinkered with Github Co-pilot & Cursor.
GitHub co-pilot has been useful when I'm coding in a strongly typed and well documented language, however the benefit here is that it will suggest pretty much exactly what I wanted when producing suggestions for the rest of the line or the next 3-5 code lines. However, this is after I configured it to only offer suggestions for small chunks of code and the accept is a two key combo instead of overriding tab accept for standard autocomplete. Anything over that and it starts hallucinating like crazy. I've also found it useful for converting a script from one language to another (mostly bash to psh & bat, but a bit of python to JS.).
Cursor has been useful for creating a functional prototype for an idea on functional projects and it does an ok job at code tracing and explaining the design patterns used and where to inject behaviors.
That said, I stopped using cursor after my free trial ended, and the 1 year GitHub Co-pilot sub that was given to me runs up in November and I don't plan on renewing.
I produce far better code without the tools. The gains achieved with co-pilot were nice, but in the end I felt disconnected from my code and when problems arose, I spent more time remembering where to find it. And if I really wanted to understand the design patterns used in a code base there are tools out there to do that already with great visualization outputs.
TonyOstrich@lemmy.world · 2 pts · 330d
What tools do you use for visualizing design patterns in programs!
odelik@lemmy.today · 1 pts · 326d
As of late, I've mostly been using Jetbrains IDEs and their built-in tooling for that. VS also has tooling, but it's a bit dated on UX IMHO. I've also played with Emerge for data flow and data clustering visualization. There's also some in-house tools we use for memory allocation.
There's a bunch of other code to UML tooling and other diagraming tools out there for specialized cases.
Pogogunner@sopuli.xyz · 6 pts · 329d
If it wasn't worth actually coding in the first place, it's not worth paying what it would cost to fix.