Learning/Coding with Claude 3/GPT4 is not like following turn-by-turn directions. It's like having a city guide next to you that knows everything about every street and the history of every building.
Very often, when I ask 'how do I do x' and had no idea how to do it in the first place.
I know what your trying to say, but if you ask how to do your problem, instead of something like how do I use a for loop to do it, you'll be given a solution.
Yeah, I've definitely seen it come up with API methods and the like, that I'd never have known to look for in docs. That's not quite what I have in mind here, though, but more how well LLMs tend to replicate the ancillary context that is available when reviewing documentation. My experience has been that they tend more narrowly task-focused than I do, although it occurs to me I don't think I've tried prompts around "suggest some solutions for the following..."
In general, I still find it frequently surprising just how much subtle differences in formulating a question can yield significant differences in response.
I’ve found them to be quite good but I often give an example of what this would look like in another language or I ask follow up questions on are there other ways and what are the tradeoffs etc. it’s all what you make it. Not that different from talking to another engineer.
The best part by far is I can ask to dive deeper or explain something that I didn’t understand very easily. If I was reading a book or in school I can’t stop the world and have it refocus on what I need to understand to continue easily.
I haven’t really had that many hallucinations in coding questions - sometimes it’s almost typos where the library and method do exist but typed slightly differently
You can modify the system prompt to describe who you are and your ability level, and to instruct the LLM to fill in gaps of knowledge you wouldn't know to ask about.
