On Thu, Apr 3, 2025, 3:51 PM Will Cooke via cctalk <[email protected]> wrote:
> > > > On 04/03/2025 5:37 PM EDT Chuck Guzis via cctalk <[email protected]> > wrote: > > > > > > On 4/3/25 14:21, John Foust via cctalk wrote: > > > At 10:22 PM 4/1/2025, Chuck Guzis via cctalk wrote: > > > > > In another task where I wanted to quickly generate a text-based > > > profile of a computer's specifications and components, it was very > > > helpful in writing a bunch of PowerShell. Did it hallucinate? > > > Perhaps once out of several dozen things I asked it, and in that > > > case, it was easy to see that the two clauses of the if-then were > > > the same. > > > > Interesting. I got no help at all from ChatGPT when working through > > some IBM 7090 code. > > > > --Chuck > > > My experience has been much like John's. In my company of 10,000 ish > employees, the entire company is "strongly encouraged" to use various "AI" > tools made available. The encouragement comes directly from the CEO (and > all his minions echo it) during his monthly video addresses. > So, all of us who write code have copilot licenses and they actually track > how much we use it (can you say layoff? I knew you could.) What I have > found is that the short suggestions of a half dozen lines or more are often > pretty good. It's the bad ones I have to fight the system to NOT take that > eat up the time. I usually could have just written those half dozen lines > faster. But sometimes it helps a lot. Apparently it tracks names and kind > find the correct variable or constant faster than I could. A couple of > coworkers have given a good description: good for boilerplate code. > That's what I've found too. Generate the boilerplate for me: not very creative but usuall 90+% right for code I'll be rototilling anyway. Small code: usually ok. Generate about 10-50 lines it's ok. But much more than that and the errors are too high to be worth it. Debugging time becomes much higher than the just write it from scratch time. Small snippets are also easier to catch errors on too. It's not terrible. On par with autocomplete emacs macros I've sometimes written but with less effort. But it's not great either, at least for the code I write. Warner As for chatgpt not doing well, I suspect that is largely due to it being a > general purpose "question answerer" rather than a tool specifically for > programming. I know of at least one person who did, however, get hired > based mostly on code written by chatgpt. But he didn't last long. > > Will >
