Time goes in one direction for *humans*... computers like to do this thing where they set their clocks backwards/forwards in time & really screw with apps looking at time deltas (my work in embedded/firmware has traumatized me...)
Tbh AI completition is useful when you just don't want to type verbose stuff, but for all the stuff that actually requires logic is very stupid and tend to introduce strange code that you might not even notice. It happened to me and to my collegues. Don't use it for stuff that requires logic.
I wonder if society is prepared for the amount of undecipherable bugs and tech debt ML/LLM users might be introducing in the name of short term productivity
Im just hoping the ai bubble bursts before it does real damage. It's garbage incapable of doing any real work and people just trust that the asshats who built the misinformation generator are honest about what it's actually doing because they don't know any better. I do. Been following it for 15 yrs
This is my concern exactly - a billion subtle errors in the software plumbing of the world, lines that look good even on examination with errors you'd only notice if you'd written the line in the first place instead of just vetting output for an LLM.
It's not even short term productivity. We tried Devin at work and using it to complete tickets and create PRs requires you to do more context switching, you have to hand hold it anyways, and requires code reviewing skills of you, which usually aren't as good as the code writing ones. Bugs go in.
Overall our team seems to be coming to the conclusion that we'd do the thing faster and better without Devin. Some are still trying but many opt out of using it. Maybe in a year or something...
As a coder who thrives on tidying up and trying to resolve tech debt as I go, this gives me a small glimmer of hope that my career might still have some longevity, despite the looming threats posed by AI.
It is truly the mirror image of our modern society. Long term never matters, it's always short term gains. And once it caught up, those who lead us there are already gone with their golden parachutes.
Surely there's been some software that got shipped with LLM code in the last year or so?
I'd expect that with any important project the issues would likely be caught internally (prior to release). Essentially a transfer of work from the junior programmers to seniors/leads and QA.
More work for people that know what they do. I'm hoping every company is blindly buying into the hype for about 4-8 years. Long enough to let the parasite proliferate so you can cash in on all those bugs. The era of thinking meat is coming back. I'm pretty sure. Fingers crossed. :D
True. The biggest issue is that generated code looks correct on the first glance. These types of bugs are so hard to spot when debugging. I still remember struggling with camera projection code for many hours, a few years back, when i messed up .y with .x in some very long formula. One never forgets
Comments
It averaged 1 UB per 3 suggestions for me
Or it’s just dumb.
Can you imagine how boring that’d be?
If you are using this tech to build something you don't understand in a language you barely know - There WILL be bugs.
Suggestions will look correct but will have subtle errors.
I'd expect that with any important project the issues would likely be caught internally (prior to release). Essentially a transfer of work from the junior programmers to seniors/leads and QA.