Using ChatGPT for technical advice #11895
Replies: 5 comments 1 reply
-
Completely agree. LLMs are a marvel and the'yre for instance awesome for higher level management talk and if you're lucky they can even produce proper working code, but I'm truly amazed by how seemingly large number of people do not realize what they cannot do (or rather what they're not suitable for, but still produce results for). Teaching you details of intrinsically hard and nuanced skills/crafts is definitely something they cannot. |
Beta Was this translation helpful? Give feedback.
-
Agree , ill not disengage my brain when considering the answers. I like to write code that is mine , on know on what it is built. So as a learning tool, I have been able to create functions to create and extend matplotlib plots in multiple steps that needed with just a little tweaking needed. |
Beta Was this translation helpful? Give feedback.
-
LLMs are just tools. It would be a shame if we just wrote them off. None of us will want to write in assembler if we have the choice. We have compilers now. I asked ChatGPT these questions:
I got reasonable answers that I can work on as a starting point. Of course you should know what you are asking for and can judge the usefulness of the answers. I can ask about specifics in quantum mechanics and get answers, but that will be useless since I know next to nothing about quantum mechanics. When it comes to LLMs, asking the right questions is the key. This is why prompt engineering is a hot topic right now. |
Beta Was this translation helpful? Give feedback.
-
Thanks @peterhinch yes I completely agree, and it's been a recent source of frustration to see examples of "chatgpt told me to do this so why doesn't it work?" (Both in MicroPython-related discussions but also other projects). The level of "confidently wrong" in the answers is very worrying. On a related point... I had a fascinating experience recently attempting to use "Cody" (from SourceGraph) as a tool to explore and understand the MicroPython codebase. Inspired by Stevey's blog posts:
The key thing is that SourceGraph/Cody has is the additional codebase-specific context ("embeddings") which gives it a huge advantage over the generic tools like ChatGPT, and it was definitely a fairly magical experience trying it out on the MicroPython codebase asking where/how/why sort of questions. (I've been otherwise fairly uninspired by ChatGPT/Bard, this felt much cleverer) Of course when you ask it questions that aren't specific to the codebase it still gives ChatGPT-style answers (and the answer I got about circular buffers was no better than your example). It's very hard for me to judge whether mostly-correct-with-insidious-errors is a useful tool or not. I am worried that by sowing misconceptions while helping someone explore a codebase it might be overall more harmful, but at the same time the ease at which it gave most of the information was quite astounding. To illustrate:
In the same way that ChatGPT/Bard's greatest strength seems to be as a summarising tool, at this stage I think it could be a great tool to help someone who can effectively fact check it write better documentation and comments. I only experimented briefly with it's "unit test generation" feature, and that showed some promise too. |
Beta Was this translation helpful? Give feedback.
-
@jimmo Thanks for those interesting links. I get it that it's early days - and that I'm the guy moaning that his chainsaw did a lousy job of getting the back off his laptop. Evidently it's a hugely powerful tool for grokking large codebases. Perhaps I've led a charmed life, but my work largely involved designing new systems rather than maintaining existing ones. Problems where I welcomed help were things like:
If you get those right the actual coding almost comes as a relief. However I do recall being in a team of developers pondering an early internet connection - and wondering at its application. I have no doubt that future iterations of AI will be able to help with these more rather nebulous areas. And some of those coding demos are impressive. |
Beta Was this translation helpful? Give feedback.
-
tl;dr Avoid. Apologies if this is OT but I thought a word of warning was in order. Some might find some humour in the outcome.
I asked "Under what conditions is a circular buffer thread safe?" and received the following answer:
As an answer this grovels in the no-mans-land between "poor" and "lousy". It is hopelessly generic and misses the point that, under very specific conditions, a circular buffer can be thread safe without locks and mutexes. I had hoped it might identify those conditions...
The first point is OK but the rest, including the buffer overflow stuff, is pure waffle. All IMO, of course.
Beta Was this translation helpful? Give feedback.
All reactions