> > The LLMs are good at finding patterns, yes. They are also spectacularly > good at confidently producing bullshit. >
I can't agree with this more. This sort of feeds into the reason I started this topic in the first place. It takes skill in coding, etc. to spot the slop. It would be really easy to accept the "working" production of an LLM as good code. Unfortunately, I don't think any of us have sufficient time to scrutinize something as simple as HALScope changes (which while increasing functionality also killed some functionality and introduced a few bugs) let alone a trajectory planner (not trying to pick on Luca, just recent examples). > > So, yes, they can be used to detect potential problems, but each and > every report must be manually scrutinized very carefully. Just > submitting what an LLM found is by definition the slop we want to > prevent. So, yes, you can use it privately, but only submit reports that > have been properly vetted. > Agreed. LLMs are pretty good at finding copy/paste errors, obvious logical flaws, etc. in _small_ sections of code, with most of the entire context entact. But it also depends on the particular day, time of day, tilt of the earth, whether or not it's windy, etc. One day it will spit out coherent easy to follow and vet bug reports, and the next day it will confidently spit out complete garbage. I do find it to be helpful for very specific niche questions I used to spend hours pouring over stack overflow posts to find answer to (think working through qt5 oddities). > > FWIW, you do not need a PhD to see the problems in the LinuxCNC code > base. Just the common sense of a reasonably versed programmer will > detect over 90% of the problems at first or second glance. > I echo Andy's sentiment. I am not a software engineer. I am a mechanical engineer with a love for CNC machines and writing code as a hobby. I've come light years from when I started, but I still have a ton to learn. > > Do not use LLMs for patch(sets). That is a large problem on the rights > front, as mentioned in the other thread. The rights issues will take > many years to be solved. You do not want to contaminate the code base > and risk costly cleanups. > I agree with this. LLMs should not be creating patches, etc. Unfortunately, I do wonder how we control some of this. Many do not subscribe to this email group, or even know it exists. Fewer still probably care about it. Many are legitimately trying to help by "vibe coding" a solution and "gifting" it to the community. It's hard to find the line between encouraging participation and ideas and discouraging it because the toolset they have is not adequate. Enforcing these policies is a whole other bag of worms. -Greg _______________________________________________ Emc-developers mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/emc-developers
